WO2017036363A1 - 光纤周界入侵信号的识别方法、装置及周界入侵报警系统 - Google Patents

光纤周界入侵信号的识别方法、装置及周界入侵报警系统 Download PDF

Info

Publication number
WO2017036363A1
WO2017036363A1 PCT/CN2016/096951 CN2016096951W WO2017036363A1 WO 2017036363 A1 WO2017036363 A1 WO 2017036363A1 CN 2016096951 W CN2016096951 W CN 2016096951W WO 2017036363 A1 WO2017036363 A1 WO 2017036363A1
Authority
WO
WIPO (PCT)
Prior art keywords
decision tree
signal
intrusion
classification
training data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2016/096951
Other languages
English (en)
French (fr)
Inventor
丛宇殊
姜婷
徐骏
杨捷
高柏松
徐惠康
仝义安
董坤
毛献辉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nuctech Co Ltd
Original Assignee
Nuctech Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nuctech Co Ltd filed Critical Nuctech Co Ltd
Priority to EA201792283A priority Critical patent/EA036070B1/ru
Priority to BR112017024488A priority patent/BR112017024488A2/pt
Priority to MX2017015387A priority patent/MX374173B/es
Priority to EP16840792.2A priority patent/EP3321859B1/en
Publication of WO2017036363A1 publication Critical patent/WO2017036363A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G08—SIGNALLING
    • G08B—SIGNALLING SYSTEMS, e.g. PERSONAL CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
    • G08B29/00—Checking or monitoring of signalling or alarm systems; Prevention or correction of operating errors, e.g. preventing unauthorised operation
    • G08B29/18—Prevention or correction of operating errors
    • G08B29/185—Signal analysis techniques for reducing or preventing false alarms or for enhancing the reliability of the system
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • G—PHYSICS
    • G08—SIGNALLING
    • G08B—SIGNALLING SYSTEMS, e.g. PERSONAL CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
    • G08B13/00—Burglar, theft or intruder alarms
    • G08B13/18—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength
    • G08B13/181—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using active radiation detection systems
    • G08B13/183—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using active radiation detection systems by interruption of a radiation beam or barrier
    • G08B13/186—Actuation by interference with heat, light, or radiation of shorter wavelength; Actuation by intruding sources of heat, light, or radiation of shorter wavelength using active radiation detection systems by interruption of a radiation beam or barrier using light guides, e.g. optical fibres
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Definitions

  • the invention relates to a photoelectric signal processing and identification technology, in particular to a fiber perimeter intrusion signal identification method, an identification device and an optical fiber perimeter intrusion alarm system.
  • Fiber Bragg Grating technology is widely used in intelligent perimeter security systems due to its simple structure, dielectric insulation, high voltage resistance, corrosion resistance, electromagnetic interference and high sensitivity.
  • the main problems are the identification of intrusion signals and the shielding of external signals such as wind, rain, snow or vehicle traffic.
  • the identification of intrusion events mainly uses classical machine learning methods such as neural networks and support vector machines (SVM).
  • the present invention provides an optical fiber perimeter intrusion signal identification method, an identification device, and a fiber perimeter intrusion alarm system to meet the intrusion event identification and alarm requirements of the fiber perimeter intrusion system.
  • An aspect of the present invention provides a method for identifying a fiber perimeter intrusion signal, comprising: obtaining a real-time signal from a monitoring signal of a fiber perimeter monitoring system; and pre-processing the real-time signal to remove the real-time signal Noise; determining whether the real-time signal after the noise is removed is an intrusion signal; if the intruding signal is, intercepting the intrusion signal to obtain a signal segment from the intrusion signal, and separately extracting the correlation of the signal segment a feature quantity; and classifying the intrusion signal according to a classification rule determined by a pre-created decision tree model and a correlation feature quantity of the signal segment; wherein the decision tree model includes at least one decision tree.
  • the classification rule determined by each decision tree includes: a classification rule on each path from the root of the decision tree to its leaf node; wherein each classification rule includes: along each of Each attribute condition on the path A conjunction of the front piece of the classification rule formed and a posterior piece of the classification rule formed by the leaf nodes on the path.
  • the decision tree model includes a plurality of decision trees; classifying the intrusion signal according to a classification rule determined by a pre-created decision tree model and a correlation feature quantity of the signal segment: And classifying the intrusion signal according to a classification rule determined by each decision tree in the plurality of decision trees and a correlation feature quantity of the signal to obtain a corresponding plurality of classification results; and the plurality of classification results
  • the same classification result is divided into one group, at least one group of classification results is obtained, and the number of classification results in each group classification result is separately counted, and the classification result in the group with the largest number of classification results is used as the final result of the intrusion signal. Classification results.
  • the method further includes: creating the decision tree model, comprising: acquiring, in advance, a plurality of sets of known classified intrusion signals and corresponding classification results from the monitoring signals of the fiber perimeter monitoring system; Performing pre-processing on the plurality of sets of known classified intrusion signals to remove noise of the plurality of sets of known classified intrusion signals; respectively, intercepting the plurality of sets of known classified intruded signals after noise removal, to Obtaining a plurality of sets of intrusion signal segments in the plurality of sets of known classification intrusion signals, and respectively extracting correlation feature quantities of the plurality of sets of intrusion signal segments; and composing according to the plurality of sets of related feature quantities of the plurality of sets of intrusion signal segments
  • the first training data set creates the decision tree model; wherein a set of related feature quantities constitutes an input vector of the first training data set.
  • the decision tree model includes a plurality of decision trees; and the creating the decision tree model includes: performing random sampling with the returning of the first training data set to obtain a plurality of second trainings a data set; wherein each second training data set has the same number of input vectors as the first training data set; and according to the plurality of second training data sets, respectively, the corresponding multiple trees are established by using a decision tree algorithm Decision tree.
  • each split is randomly extracted from all relevant feature quantities of the input vector. Or all relevant feature quantities.
  • the creating the decision tree model includes: establishing a decision tree by using a decision tree algorithm according to the first training data set, and performing post pruning on the decision tree to obtain the decision tree model .
  • the creating the decision tree model further comprises: training the decision tree model a plurality of times according to the first training data set to obtain an optimal decision tree model.
  • the decision tree algorithm includes: using a Gini indicator as a CART algorithm for split attribute metrics.
  • the related feature quantity includes: a peak mean difference, a variance, a skewness, a kurtosis, a number of poles, and some or all of the frequency domain feature quantities obtained by wavelet packet decomposition.
  • Another aspect of the present invention provides an apparatus for identifying a fiber perimeter intrusion signal, comprising: a signal acquisition module for acquiring a real-time signal from a monitoring signal of a fiber perimeter monitoring system; and a signal pre-processing module for The real-time signal is pre-processed to remove the noise of the real-time signal, and determine whether the real-time signal after the noise is removed is an intrusion signal; and the feature quantity extraction module is configured to: when the real-time signal is the intrusion signal, Interpolating the intrusion signal to obtain a signal segment from the intrusion signal, and separately extracting related feature quantities of the signal segment; and a decision tree classification module, configured to determine a classification rule according to a pre-created decision tree model and The correlation feature quantity of the signal segment classifies the intrusion signal; wherein the decision tree model includes at least one decision tree.
  • the classification rule determined by each decision tree includes: a classification rule on each path from the root of the decision tree to its leaf node; wherein each classification rule includes: along each of Each of the attribute conditions on the strip path forms a conjunction of the predecessor of the classification rule and a posterior of the classification rule formed by the leaf nodes on the path.
  • the decision tree model includes a plurality of decision trees; classifying the intrusion signal according to a classification rule determined by a pre-created decision tree model and a correlation feature quantity of the signal segment: And classifying the intrusion signal according to a classification rule determined by each decision tree in the plurality of decision trees and a correlation feature quantity of the signal to obtain a corresponding plurality of classification results; and the plurality of classification results
  • the same classification result is divided into one group, at least one group of classification results is obtained, and the number of classification results in each group classification result is separately counted, and the classification result in the group with the largest number of classification results is used as the final result of the intrusion signal. Classification results.
  • the signal acquisition module is further configured to separately acquire, from the monitoring signals of the fiber perimeter monitoring system, a plurality of sets of known classification intrusion signals and corresponding classification results thereof;
  • the processing module is further configured to separately perform pre-processing on the plurality of sets of known classified intrusion signals to remove noise of the plurality of sets of known classified intrusion signals; and the feature quantity extracting module is further configured to intercept the noise respectively The plurality of sets of known classified intrusion signals to obtain a plurality of sets of intrusion signal segments from the plurality of sets of known classified intrusion signals, and extract relevant feature quantities of the plurality of sets of intrusion signal segments respectively;
  • the decision tree classification module is further configured to create the decision tree model according to the first training data set formed by the plurality of sets of related feature quantities of the plurality of sets of intrusion signal segments; wherein a set of related feature quantities constitutes the first training data An input vector of the set.
  • the decision tree model includes a plurality of decision trees; and the creating the decision tree model includes: performing random sampling with the returning of the first training data set to obtain a plurality of second trainings a data set; wherein each second training data set has the same number of input vectors as the first training data set; and according to the plurality of second training data sets, respectively, the corresponding multiple trees are established by using a decision tree algorithm Decision tree.
  • each split automatically re-extracts some or all of the relevant feature quantities from all relevant feature quantities of the input vector.
  • the creating the decision tree model includes: establishing a decision tree by using a decision tree algorithm according to the first training data set, and performing post pruning on the decision tree to obtain the decision tree model .
  • the creating the decision tree model further comprises: training the decision tree model a plurality of times according to the first training data set to obtain an optimal decision tree model.
  • the decision tree algorithm includes: using a Gini indicator as a CART algorithm for split attribute metrics.
  • the related feature quantity includes: a peak mean difference, a variance, a skewness, a kurtosis, a number of poles, and some or all of the frequency domain feature quantities obtained by wavelet packet decomposition.
  • a further aspect of the present invention provides an optical fiber perimeter intrusion alarm system, including: an optical path subsystem for providing a monitoring signal of a fiber perimeter; and any of the above-mentioned optical fiber perimeter intrusion signal identifying devices for The intrusion signal in the monitoring signal is identified and classified; and an alarm subsystem is configured to perform an alarm according to the classification of the intrusion signal by the identification device.
  • the invention discloses an optical fiber perimeter intrusion signal identification device and a recognition method thereof, and obtains a training data set of a model by extracting a correlation feature quantity of an intrusion signal of a known classification signal to establish a corresponding decision tree model; According to the decision tree model created, the classification rules of the fiber perimeter intrusion signal are determined to identify and classify the fiber perimeter intrusion signals.
  • the method is simple and easy to implement, and has high classification ability for intrusion signals and good classification effect.
  • FIG. 1 is a flow chart showing a method of identifying a fiber perimeter intrusion signal, according to an example embodiment.
  • FIG. 2 is a schematic diagram of a decision tree for identifying fiber perimeter intrusion signals, according to an example embodiment.
  • FIG. 3 is a flow diagram of a method of creating a decision tree model for identifying fiber perimeter intrusion signals, according to an example embodiment.
  • FIG. 4 is a structural diagram of an apparatus for identifying an optical fiber perimeter intrusion signal, according to an exemplary embodiment.
  • FIG. 5 is a block diagram of a fiber perimeter intrusion alarm system, according to an example embodiment.
  • Figure 6 is a block diagram of an optical path subsystem using a Michelson interferometer.
  • FIG. 7 is a schematic diagram of forming a combined decision tree model from an original training data set, according to an example embodiment.
  • FIG. 1 is a flow chart showing a method of identifying a fiber perimeter intrusion signal, according to an example embodiment. As shown in FIG. 1, the method 10 includes:
  • Step S110 obtaining a real-time signal y(t) from the monitoring signal of the fiber perimeter monitoring system.
  • a signal of a fiber perimeter intrusion alarm system is monitored and a real-time signal y(t) is obtained from the monitoring signal.
  • the real-time signal y(t) is obtained after sampling at a sampling frequency.
  • step S120 the real-time signal y(t) is preprocessed to remove the noise, and the denoised signal y'(t) is obtained.
  • the real-time signal y(t) may be preprocessed using a wavelet transform to remove noise such as background white noise therein.
  • step S130 it is judged whether the denoised signal y'(t) is an intrusion signal.
  • the intrusion signals in the fiber perimeter system are roughly classified into climbing, tapping, and environment (such as wind, rain, snow, and vehicle traffic, etc.).
  • a method of comparing the denoised signal y'(t) with a threshold may be used to determine whether it is an intrusion signal. When the denoised signal y'(t) is higher than the threshold, it is judged that the denoised signal y'(t) is an intrusion signal.
  • Step S140 when it is judged that the denoised signal y'(t) is an intrusion signal, the related feature quantity is extracted.
  • the correlation feature amount includes, for example, part or all of the peak mean difference (Diff), the variance (D), the skewness (Skew), the kurtosis (K), the number of extreme points (Num), and the frequency domain characteristic amount.
  • the peak mean difference (Diff), the variance (D), the skewness (Skew), the kurtosis (K), and the number of extreme points (Num) belong to the time domain feature quantity.
  • the difference is calculated as follows:
  • the frequency domain feature quantity includes the feature quantities [T 1 ', T 2 ', ..., T r '] obtained by wavelet packet decomposition, and the calculation process is as follows:
  • Step S150 classify the intrusion signal according to the classification rule determined by the previously created decision tree model and the extracted related feature quantity.
  • the above-mentioned pre-created decision tree model may include a model containing a single decision tree, or may include A model of multiple decision tree combinations. Below we first introduce how to classify intrusion signals based on the classification rules and related feature quantities determined by the decision tree model containing a single decision tree.
  • the classification rules are extracted from the decision tree determined by the decision tree model.
  • 2 is a schematic diagram of a single decision tree for identifying fiber perimeter intrusion signals, according to an example embodiment.
  • the single decision tree shown in Figure 2 is used to illustrate how to obtain the classification rules determined by the decision tree and to classify them.
  • each attribute is, for example, one of the above related feature quantities; wherein A 1 corresponds to three different attribute conditions a 1, 1 , a 1 , 2 and a 1 , 3 , A 2 correspond to two different attribute conditions a 2,1 , a 2,2 , A 3 correspond to two different attribute conditions a 3,1 , a 3,2 ; C 1 And C 2 represent two categories of intrusion signals, such as those described above, including climbing, tapping, and the environment (eg, wind, rain, snow, vehicle travel, etc.).
  • a classification rule is created from the root (A 1 in the figure) to each path of the leaf, respectively represented by the IF-THEN form, wherein each attribute condition along the path forms a combination of the rule predecessor (ie, the IF part).
  • the classification rules determined by the above obtained decision tree model are merely illustrative and not limiting.
  • the extracted related feature quantity is applied to the classification rule to classify the intrusion signal. For example, if the attribute conditions of the feature vectors A 1 and A 2 extracted from the intrusion signal satisfy the conditions a 1,1 and a 2,1 , respectively, the intrusion signal is classified into C 1 .
  • the pre-created decision tree model can also be a decision tree combination model T, which includes a plurality of decision trees T 1 , T 2 , . . . , T k .
  • the following focuses on how to combine the model T according to the decision tree to determine the classification results.
  • the decision tree model includes both a decision tree model with only a single decision tree and a plurality of decision trees participating in the voting. Combined model.
  • the method 20 includes:
  • Step S210 Acquire, in advance, a plurality of sets of known classified intrusion signals Y(t) from the monitoring signals of the fiber perimeter monitoring system, and corresponding classification results thereof.
  • the known classification includes, for example, the above-mentioned climbing, tapping, and environment (for example, wind, rain, snow, and vehicle transportation, etc.), but the present invention is not limited thereto.
  • step S220 the intrusion Y(t) of the plurality of sets of known classifications is respectively preprocessed to remove the noise, and the denoised plurality of sets of signals Y'(t) are respectively obtained.
  • a plurality of sets of known classified intrusion signals Y'(t) may be processed by wavelet transform to remove background white noise or the like therein.
  • the correlation feature amount includes, for example, part or all of peak mean value difference (Diff), variance (D), skewness (Skew), kurtosis (K), number of extreme points (Num), and frequency domain feature quantities.
  • Diff peak mean value difference
  • D variance
  • Skew skewness
  • K kurtosis
  • Num number of extreme points
  • the peak mean difference (Diff), the variance (D), the skewness (Skew), the kurtosis (K), and the number of extreme points (Num) belong to the time domain feature quantity.
  • the eigenvectors constitute an input vector, and multiple sets of input vectors form the training data set D of the decision tree model.
  • a decision tree model is created based on the overall training data set D.
  • the CART algorithm using the Gini indicator as the split attribute measurement method will be described as an example.
  • Gini indicator is defined as the following formula:
  • p i is the probability that category C i appears in training data set D.
  • the Gini index of the given D is as follows:
  • the Gini indicator is calculated for different subsets of the attribute, and the subset that produces the smallest Gini indicator is selected as the split subset of the attribute. Repeat the above process until the split stops.
  • the decision tree model containing a single decision tree is created.
  • Pruning mainly includes pre-pruning and post-pruning.
  • the pre-pruning is to stop the tree by prematurely, such as by deciding that the tree is not pruned or divided into a subset of the training tuple at a given node, and once stopped, the node becomes a leaf node.
  • the pruning is then to cut the subtree from the fully grown tree.
  • the subtree of the given node is clipped by deleting the branch of the node and replacing it with the leaf node.
  • the leaf node uses the most frequent class in the subtree being replaced. mark.
  • the principle of post pruning includes the principle of minimum description length and the principle of minimum expected error rate. According to the statistical metric, the unreliable branch is subtracted, and the correct classification ability of the tree independent of the test data is improved.
  • the CART algorithm usually uses a post pruning strategy. For example, pruning is performed using the CCP (cost-complexity) method.
  • CCP cost-complexity
  • a series of trees ⁇ T 0 , T 1 , ..., T k ⁇ are obtained from the bottom up by trimming the original decision tree, where T 0 is the original tree without any clipping, T k is a tree with only one node, and T i+1 is obtained by replacing one or more subtrees of T i .
  • the best tree is selected as the final pruned decision tree based on the true error rate.
  • the above is the decision tree obtained as a training. It is also possible to perform multiple trainings based on the training data set and select an optimal decision tree to determine the decision tree model.
  • the decision tree algorithm used may further include: an ID3 algorithm that uses information gain as a split attribute metric or a C4.5 algorithm that uses information gain as a split attribute metric.
  • FIG. 7 is a schematic diagram of forming a combined decision tree model from an original training data set, according to an example embodiment.
  • the training data set D is subjected to random sampling with a return, and one training data set D 1 , D 2 , ..., D l is obtained .
  • Each training data set contains the same number of input vectors as in the original training data set D.
  • a corresponding decision tree model is established for each of the training data sets D 1 , D 2 , . . . , D l .
  • Each decision tree model uses the CART algorithm as an example, but the invention is not limited thereto.
  • the number of attributes extracted earlier be M.
  • the initial data is at the root node and will be split later.
  • F and F ⁇ M For each splitting process, F attributes are randomly extracted from all M attributes as a split attribute set.
  • the node to extract the F attributes choose the best way to split.
  • the Gini indicator is used as a measurement method. During the establishment of the decision tree, the number of F remains the same.
  • Gini indicator is defined as the following formula:
  • p i is the probability that category C i appears in training data set D.
  • the Gini indicators are as follows:
  • the Gini index is calculated for different subsets of the attributes, and the subset that produces the smallest Gini indicator is selected as the split subset of the attribute.
  • Each of the training data sets D 1 , D 2 , ..., D l is subjected to such a process to obtain a plurality of decision trees T 1 , T 2 , ..., T k .
  • a plurality of decision tree combinations T also called a random forest
  • T 1 , T 2 , ..., T k are obtained .
  • the invention provides a method for identifying a fiber perimeter intrusion signal by using an intrusion signal for a known classification signal
  • the relevant feature quantity is extracted, and the training data set is obtained to establish the corresponding decision tree model.
  • the classification rules of the fiber perimeter intrusion signal are determined to identify and classify the fiber perimeter intrusion signal.
  • the method is simple, easy to implement, high in classification ability and good in classification effect.
  • the identification device 30 includes a signal acquisition module 310, a signal pre-processing module 320, a feature quantity extraction module 330, and a decision tree classification module 340.
  • the signal acquisition module 310 is configured to obtain a real-time signal y(t) from the monitoring signal of the fiber perimeter monitoring system. Typically, the real-time signal y(t) is obtained after sampling at a sampling frequency.
  • the signal preprocessing module 320 is configured to preprocess the real-time signal y(t) to remove the noise, obtain the denoised signal y'(t), and determine whether the denoised signal y'(t) is an intrusion. signal.
  • the real-time signal y(t) may be preprocessed using a wavelet transform to remove noise such as background white noise therein.
  • the intrusion signals in the fiber perimeter system are roughly classified into climbing, tapping, and environment (such as wind, rain, snow, and vehicle traffic, etc.).
  • a method of comparing the denoised signal y'(t) with a threshold may be used to determine whether it is an intrusion signal. When the denoised signal y'(t) is higher than the threshold, it is judged that the denoised signal y'(t) is an intrusion signal.
  • the feature quantity extraction module 330 is configured to extract the correlation feature quantity when the denoised signal y'(t) is determined to be an intrusion signal.
  • the correlation feature amount includes, for example, part or all of the peak mean difference (Diff), the variance (D), the skewness (Skew), the kurtosis (K), the number of extreme points (Num), and the frequency domain feature amount.
  • the peak mean difference (Diff), the variance (D), the skewness (Skew), the kurtosis (K), and the number of extreme points (Num) belong to the time domain feature quantity.
  • the decision tree classification module 340 is configured to classify the intrusion signal according to the classification rule determined by the pre-created decision tree model and the extracted related feature quantity.
  • the above-mentioned pre-created decision tree model may include a model containing a single decision tree or a model containing a plurality of decision tree combinations. Below we first introduce how to classify intrusion signals based on the classification rules and related feature quantities determined by the decision tree model containing a single decision tree.
  • the classification rules are extracted from the decision tree determined by the decision tree model.
  • 2 is a schematic diagram of a single decision tree for identifying fiber perimeter intrusion signals, according to an example embodiment.
  • the single decision tree shown in Figure 2 is used to illustrate how to determine the classification rules based on the decision tree and to classify them.
  • each attribute is, for example, one of the above related feature quantities; wherein A 1 corresponds to three different attribute conditions a 1,1 , a 1,2 and a 1,3 ,A 2 correspond to two different attribute conditions a 2,1 , a 2,2 , A 3 correspond to two different attribute conditions a 3,1 , a 3,2 ; C 1 and C 2 denotes two categories of intrusion signals, such as as described above, including climbing, tapping, and the environment (such as wind, rain, snow, and vehicle travel, etc.).
  • a classification rule is created from the root (A 1 in the figure) to each path of the leaf, respectively represented by the IF-THEN form, wherein each attribute condition along the path forms a combination of the rule predecessor (ie, the IF part).
  • the classification rules determined by the above obtained decision tree model are merely illustrative and not limiting.
  • the extracted related feature quantity is applied to the classification rule to classify the intrusion signal. For example, if the attribute conditions of the feature vectors A 1 and A 2 extracted from the intrusion signal satisfy the conditions a 1,1 and a 2,1 , respectively, the intrusion signal is classified into C 1 .
  • the pre-created decision tree model can also be a decision tree combination model T, which includes a plurality of decision trees T 1 , T 2 , . . . , T k .
  • the following focuses on how to combine the model T according to the decision tree to determine the classification results.
  • the fiber perimeter intrusion signal identification device 30 is further configured to create the above decision tree model before identifying the intrusion signal, the decision tree model includes a decision tree model having only a single decision tree, and corresponding participation in voting A combined model of multiple decision trees.
  • the signal acquisition module 310 is further configured to separately acquire multiple sets of known points from the fiber perimeter monitoring signal.
  • the known classification includes, for example, the above-mentioned climbing, tapping, and environment (for example, wind, rain, snow, and vehicle transportation, etc.), but the present invention is not limited thereto.
  • the signal pre-processing module 320 is further configured to pre-process the intrusion Y(t) of the plurality of sets of known classifications respectively to remove the noise, and respectively obtain the de-noised multiple sets of signals Y'(t).
  • the correlation feature amount includes, for example, part or all of peak mean value difference (Diff), variance (D), skewness (Skew), kurtosis (K), number of extreme points (Num), and frequency domain feature quantities.
  • Diff peak mean value difference
  • D variance
  • Skew skewness
  • K kurtosis
  • Num number of extreme points
  • the peak mean difference (Diff), the variance (D), the skewness (Skew), the kurtosis (K), and the number of extreme points (Num) belong to the time domain feature quantity.
  • the eigenvectors constitute an input vector, and multiple sets of input vectors form the training data set D of the decision tree model.
  • a decision tree model is created based on the overall training data set D.
  • the CART algorithm using the Gini indicator as the split attribute measurement method will be described as an example.
  • Gini indicator is defined as the following formula:
  • p i is the probability that category C i appears in training data set D.
  • the Gini index of the given D is as follows:
  • the Gini indicator is calculated for different subsets of the attribute, and the subset that produces the smallest Gini indicator is selected as the split subset of the attribute. Repeat the above process until the split stops. After that, the established decision tree is pruned by post-pruning method to form a training decision tree model. Repeated training for multiple times to get the best decision tree model.
  • a decision tree model that contains multiple decision trees.
  • random sampling with the put back is performed on the training data set D, and one training data set D 1 , D 2 , . . . , D l can be obtained.
  • the training data set contains the same number of input vectors as in the original training data set D.
  • a corresponding decision tree model is established for each of the training data sets D 1 , D 2 , . . . , D l .
  • Each decision tree model is exemplified by the above CART algorithm, but the invention is not limited thereto.
  • the total number of attributes extracted earlier is M, and the number of attributes specified is F, which satisfies F ⁇ M. During the establishment of the decision tree, the number of F remains the same. In each splitting process, F attributes are randomly extracted from the M attributes as a split attribute set. Use the Gini indicator as a measure to choose the best split mode until the split stops.
  • Each of the training data sets D 1 , D 2 , ..., D l is subjected to such a process to obtain a plurality of decision trees T 1 , T 2 , ..., T k .
  • a plurality of decision tree combinations T also called a random forest
  • T 1 , T 2 , ..., T k are obtained .
  • the device for identifying an intrinsic fiber intrusion signal obtains a training data set by extracting a correlation feature quantity of an intrusion signal of a known classification signal to establish a corresponding decision tree model; and according to the created decision tree model The classification rules of the fiber perimeter intrusion signals are determined to identify and classify the fiber perimeter intrusion signals.
  • the identification method adopted by the device is simple, easy to implement, high in classification ability and good in classification effect.
  • FIG. 5 is a block diagram of a fiber perimeter intrusion alarm system, according to an example embodiment.
  • the fiber perimeter intrusion alarm system 40 includes: an optical path subsystem 410, and an optical fiber perimeter intrusion signal.
  • the device 30 and the alarm subsystem 420 are included in the fiber perimeter intrusion alarm system 40.
  • the optical path subsystem 410 is used to provide a monitoring signal for the perimeter of the fiber.
  • the optical path system 410 will be described below by taking a Michelson interferometer as an example, but the invention is not limited thereto.
  • Figure 6 is a block diagram of an optical path subsystem using a Michelson interferometer. As shown in FIG. 6, the optical path subsystem 410 includes a laser 4110, a coupler 4120, a reference arm 4130, a sensitive arm 4140, a reflective end face 4150, and a detector 4160.
  • the narrow-band laser emitted by the laser 4110 is reflected by the reflective end face 4150 and interferes with the light reflected by the reference arm 4130 on the detector 4160. If there is an intrusion event, the detection signal is phase-changed by the modulation of the disturbance signal on the sensitive arm 4140, causing a change in the interference fringes, thereby causing the intensity of the light detected by the detector 4160 to change.
  • the detector 4160 is, for example, a photodetector that converts the detected optical signal into an electrical signal output to provide a monitoring signal for the perimeter of the fiber.
  • the optical fiber perimeter intrusion signal identification device 30 identifies the intrusion signal and classifies the intrusion signal based on the monitoring signal. How the identification device 30 performs intrusion signal identification and classification has been described in detail above and will not be described herein.
  • the alarm subsystem 420 Based on the classification information of the intrusion signal output by the identification device 30, the alarm subsystem 420 performs an alarm accordingly.
  • the alarm subsystem 420 can include an alarm output control module (not shown) for outputting a control alarm peripheral for alarming.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Computer Security & Cryptography (AREA)
  • Image Analysis (AREA)
  • Radar Systems Or Details Thereof (AREA)
  • Complex Calculations (AREA)
  • Preliminary Treatment Of Fibers (AREA)
  • Length Measuring Devices By Optical Means (AREA)

Abstract

一种光纤周界入侵信号的识别方法（10）、识别装置（30）及光纤周界入侵报警系统（40）。该识别方法（10）包括：从一光纤周界监测系统的监测信号中获取实时信号；对该实时信号进行预处理，以去除噪声；判断去除噪声后的该实时信号是否为入侵信号；如果为入侵信号，则截取该入侵信号，以从该入侵信号中获得信号片段，并分别提取该信号片段的相关特征量；以及根据预先创建的决策树模型所确定的分类规则及该信号片段的相关特征量，对该入侵信号进行分类。该方法简单、易于实现，对入侵信号的分类能力高并且分类效果较好。

Description

光纤周界入侵信号的识别方法、装置及周界入侵报警系统 技术领域
本发明涉及光电信号处理及辨识技术，尤其涉及一种光纤周界入侵信号的识别方法、识别装置及光纤周界入侵报警系统。
背景技术
随着社会的发展，为了防止非法入侵和破坏，越来越多的重要设施都需要进行周界监测和安防。而光纤光栅技术由于其结构简单、介质绝缘、耐高压、耐腐蚀、不受电磁干扰及具有较高灵敏度等特性，广泛地被应用于智能周界安防系统中。
对于采用光纤光栅技术的周界安防系统，其主要面临的问题在于对入侵信号的识别，及对外界环境如风、雨、雪或车辆来往等干扰信号的屏蔽。目前对于入侵事件的识别主要采用神经网络、支持向量机(SVM)等经典机器学习方法。
如何提供一种简单易实现、分类能力高、且分类效果好的入侵信号识别方法成为业界的一个研究方向。
发明内容
有鉴于此，本发明提供了一种光纤周界入侵信号的识别方法、识别装置及光纤周界入侵报警系统，以满足光纤周界入侵系统对入侵事件的识别及报警要求。
本发明的额外方面和优点将部分地在下面的描述中阐述，并且部分地将从描述中变得显然，或者可以通过本发明的实践而习得。
本发明一方面提供了一种光纤周界入侵信号的识别方法，包括：从一光纤周界监测系统的监测信号中获取实时信号；对所述实时信号进行预处理，以去除所述实时信号的噪声；判断去除噪声后的所述实时信号是否为入侵信号；如果为所述入侵信号，则截取所述入侵信号，以从所述入侵信号中获得信号片段，并分别提取所述信号片段的相关特征量；以及根据预先创建的决策树模型所确定的分类规则及所述信号片段的相关特征量对所述入侵信号进行分类；其中所述决策树模型包含至少一棵决策树。
于一实施例中，其中每棵决策树所确定的分类规则包括：从该决策树的根到其叶结点的每条路径上的分类规则；其中每个分类规则包括：以沿着其每条路径上的每个属性条件 形成的该分类规则的前件的一个合取项以及以该路径上的叶结点形成的该分类规则的后件。
于另一实施例中，其中所述决策树模型含有多棵决策树；根据预先创建的决策树模型所确定的分类规则及所述信号片段的相关特征量对所述入侵信号进行分类包括：依次根据所述多棵决策树中的每棵决策树所确定的分类规则及所述信号的相关特征量对所述入侵信号进行分类，以获得相应的多个分类结果；将所述多个分类结果中相同的分类结果分为一组，获得至少一组分类结果，分别统计各组分类结果中的分类结果的数量，以其中分类结果数量最多的一组中的分类结果作为所述入侵信号最终的分类结果。
于再一实施例中，该方法还包括创建所述决策树模型，包括：预先从所述光纤周界监测系统的监测信号中分别获取多组已知分类的入侵信号及其对应的分类结果；分别对所述多组已知分类的入侵信号进行预处理，以去除所述多组已知分类的入侵信号的噪声；分别截取去除噪声后的所述多组已知分类的入侵信号，以从所述多组已知分类的入侵信号中获得多组入侵信号片段，并分别提取所述多组入侵信号片段的相关特征量；以及根据所述多组入侵信号片段的多组相关特征量所组成的第一训练数据集，创建所述决策树模型；其中一组相关特征量组成所述第一训练数据集的一个输入向量。
于再一实施例中，其中所述决策树模型含有多棵决策树；创建所述决策树模型包括：对所述第一训练数据集进行有放回的随机抽样，以获得多个第二训练数据集；其中每个第二训练数据集与所述第一训练数据集所包含的输入向量数目相同；根据所述多个第二训练数据集，采用决策树算法分别建立相应的所述多棵决策树。
于再一实施例中，其中在建立每棵所述决策树时，针对每个所述第二训练数据集的输入向量，每次分裂都重新从该输入向量的所有相关特征量中随机抽取部分或全部的相关特征量。
于再一实施例中，其中创建所述决策树模型包括：根据所述第一训练数据集，采用决策树算法建立决策树，对所述决策树进行后剪枝，以获得所述决策树模型。
于再一实施例中，其中创建所述决策树模型还包括：根据所述第一训练数据集，多次对所述决策树模型进行训练，以获得最佳的决策树模型。
于再一实施例中，其中所述决策树算法包括：采用Gini指标作为分裂属性度量的CART算法。
于再一实施例中，其中所述相关特征量包括：峰均值差、方差、偏度、峭度、极点数目及采用小波包分解获得的频域特征量中的部分或全部。
本发明另一方面提供了一种光纤周界入侵信号的识别装置，包括：信号获取模块，用于从一光纤周界监测系统的监测信号中获取实时信号；信号预处理模块，用于对所述实时信号进行预处理，以去除所述实时信号的噪声，及判断去除噪声后的所述实时信号是否为入侵信号；特征量提取模块，用于当所述实时信号为所述入侵信号时，截取所述入侵信号，以从所述入侵信号中获得信号片段，并分别提取所述信号片段的相关特征量；以及决策树分类模块，用于根据预先创建的决策树模型所确定的分类规则及所述信号片段的相关特征量对所述入侵信号进行分类；其中所述决策树模型包含至少一棵决策树。
于一实施例中，其中每棵决策树所确定的分类规则包括：从该决策树的根到其叶结点的每条路径上的分类规则；其中每个分类规则包括：以沿着其每条路径上的每个属性条件形成的该分类规则的前件的一个合取项以及以该路径上的叶结点形成的该分类规则的后件。
于另一实施例中，其中所述决策树模型含有多棵决策树；根据预先创建的决策树模型所确定的分类规则及所述信号片段的相关特征量对所述入侵信号进行分类包括：依次根据所述多棵决策树中的每棵决策树所确定的分类规则及所述信号的相关特征量对所述入侵信号进行分类，以获得相应的多个分类结果；将所述多个分类结果中相同的分类结果分为一组，获得至少一组分类结果，分别统计各组分类结果中的分类结果的数量，以其中分类结果数量最多的一组中的分类结果作为所述入侵信号最终的分类结果。
于再一实施例中，其中，所述信号获取模块还用于预先从所述光纤周界监测系统的监测信号中分别获取多组已知分类的入侵信号及其对应的分类结果；所述预处理模块还用于分别对所述多组已知分类的入侵信号进行预处理，以去除所述多组已知分类的入侵信号的噪声；所述特征量提取模块还用于分别截取去除噪声后的所述多组已知分类的入侵信号，以从所述多组已知分类的入侵信号中获得多组入侵信号片段，并分别提取所述多组入侵信号片段的相关特征量；以及所述决策树分类模块还用于根据所述多组入侵信号片段的多组相关特征量所组成的第一训练数据集，创建所述决策树模型；其中一组相关特征量组成所述第一训练数据集的一个输入向量。
于再一实施例中，其中所述决策树模型含有多棵决策树；创建所述决策树模型包括：对所述第一训练数据集进行有放回的随机抽样，以获得多个第二训练数据集；其中每个第二训练数据集与所述第一训练数据集所包含的输入向量数目相同；根据所述多个第二训练数据集，采用决策树算法分别建立相应的所述多棵决策树。
于再一实施例中，其中在建立每棵所述决策树时，针对每个所述第二训练数据集的输 入向量，每次分裂都重新从该输入向量的所有相关特征量中随机抽取部分或全部的相关特征量。
于再一实施例中，其中创建所述决策树模型包括：根据所述第一训练数据集，采用决策树算法建立决策树，对所述决策树进行后剪枝，以获得所述决策树模型。
于再一实施例中，其中创建所述决策树模型还包括：根据所述第一训练数据集，多次对所述决策树模型进行训练，以获得最佳的决策树模型。
于再一实施例中，其中所述决策树算法包括：采用Gini指标作为分裂属性度量的CART算法。
于再一实施例中，其中所述相关特征量包括：峰均值差、方差、偏度、峭度、极点数目及采用小波包分解获得的频域特征量中的部分或全部。
本发明再一方面提供了一种光纤周界入侵报警系统，包括：光路子系统，用于提供光纤周界的监测信号；上述任一种光纤周界入侵信号的识别装置，用于对所述监测信号中的入侵信号进行识别及分类；以及报警子系统，用于根据所述识别装置对所述入侵信号的分类，相应地进行报警。
本发明提供的光纤周界入侵信号的识别装置及其采用的识别方法，通过对已知分类信号的入侵信号的相关特征量的提取，获得模型的训练数据集，以建立相应的决策树模型；根据所创建的决策树模型，确定光纤周界入侵信号的分类规则，以对光纤周界入侵信号进行识别与分类。该方法简单、易于实现，对入侵信号的分类能力高并且分类效果较好。
附图说明
通过参照附图详细描述其示例实施方式，本发明的上述和其它特征及优点将变得更加明显。
图1为根据一示例实施例示出的光纤周界入侵信号的识别方法的流程图。
图2为根据一示例实施例示出的用于识别光纤周界入侵信号的决策树的示意图。
图3为根据一示例实施例示出的创建用于识别光纤周界入侵信号的决策树模型的方法的流程图。
图4为根据一示例实施例示出的光纤周界入侵信号的识别装置的结构图。
图5为根据一示例实施例示出的光纤周界入侵报警系统的结构图。
图6为采用迈克尔逊干涉仪的光路子系统的结构图。
图7为根据一示例实施例示出的通过原始训练数据集形成组合决策树模型的示意图。
具体实施方式
现在将参考附图更全面地描述示例实施方式。然而，示例实施方式能够以多种形式实施，且不应被理解为限于在此阐述的实施方式；相反，提供这些实施方式使得本发明将全面和完整，并将示例实施方式的构思全面地传达给本领域的技术人员。在图中相同的附图标记表示相同或类似的结构，因而将省略对它们的重复描述。
所描述的特征、结构或特性可以以任何合适的方式结合在一个或更多实施方式中。在下面的描述中，提供许多具体细节从而给出对本发明的实施方式的充分理解。然而，本领域技术人员应意识到，没有所述特定细节中的一个或更多，或者采用其它的方法、组元等，也可以实践本发明的技术方案。在其它情况下，不详细示出或描述公知结构或者操作以避免模糊本发明。
图1为根据一示例实施例示出的光纤周界入侵信号的识别方法的流程图。如图1所示，该方法10包括：
步骤S110，从光纤周界监测系统的监测信号中获取实时信号y(t)。
例如，对一光纤周界入侵报警系统的信号进行监测，并从该监测信号中获取实时信号y(t)。通常，该实时信号y(t)经以一采样频率采样后获得。
步骤S120，对实时信号y(t)进行预处理，以去除其噪声，获得去噪后的信号y’(t)。
例如，可以采用小波变换对实时信号y(t)进行预处理，以去除其中的背景白噪声等噪声。
步骤S130，判断去噪后的信号y’(t)是否为入侵信号。
通常，光纤周界系统中的入侵信号大致被分为攀爬、敲击及环境(例如风、雨、雪及车辆来往等)等几类。在一些实施例中，例如可以采用将去噪后的信号y’(t)与一阈值进行比较的方法来判断其是否为入侵信号。当去噪后的信号y’(t)高于该阈值时，判断该去噪后的信号y’(t)为入侵信号。
步骤S140，当判断去噪后的信号y’(t)为入侵信号时，对其进行相关特征量的提取。
首先，截取该去噪后的信号y’(t)，以获得离散的信号片段{x(ti)},i＝1,...,N。
之后，提取信号片段{x(ti)},i＝1,...,N的相关特征量。相关特征量例如包括峰均值差(Diff)、方差(D)、偏度(Skew)、峭度(K)、极值点数目(Num)及频域 特征量中的部分或全部。其中峰均值差(Diff)、方差(D)、偏度(Skew)、峭度(K)、极值点数目(Num)属于时域特征量。下面分别定义信号片段{x(ti)},i＝1,...,N的上述特征量：
峰值均值差(Diff)定义为信号片段{x(ti)},i＝1,...,N中最大值max(x(t))和平均值
Figure PCTCN2016096951-appb-000001
的差值，其计算公式如下：
Figure PCTCN2016096951-appb-000002
方差(D)的计算公式如下：
Figure PCTCN2016096951-appb-000003
偏度(Skew)的计算公式如下：
Figure PCTCN2016096951-appb-000004
峭度(K)的计算公式如下：
Figure PCTCN2016096951-appb-000005
极值点数目(Num)用于统计信号片段{x(ti)},i＝1,...,N中极值点的数目。
频域特征量包括采用小波包分解获得的特征量[T1',T2',......,Tr']，其计算过程如下：
将信号片段{x(ti)},i＝1,...,N进行j层小波包分解，设采样频率为2f，则可形成2j个等宽的频带，每个频带宽度为
Figure PCTCN2016096951-appb-000006
分解后，得到j层小波包系数
Figure PCTCN2016096951-appb-000007
其中k＝1,2,...,2j-1，m为位置指标。选择其中r个对能量最为敏感的频带，求出其能量并归一化处理，其计算公式如下：
Figure PCTCN2016096951-appb-000008
步骤S150，根据预先创建的决策树模型所确定的分类规则与所提取的相关特征量，对入侵信号进行分类。
上述预先创建的决策树模型既可以包括含有单棵决策树的模型，也可以包括含有 多棵决策树组合的模型。下面我们首先介绍如何根据包含有单棵决策树的决策树模型所确定的分类规则及相关特征量，对入侵信号进行分类。
当预先创建的决策树模型建立后，相应的分类规则也已经确立。具体地，从该决策树模型确定的决策树中抽取分类规则。图2为根据一示例实施例示出的用于识别光纤周界入侵信号的单棵决策树的示意图。以图2所示的单棵决策树来说明如何获得决策树所确定的分类规则，及进行分类。
如图2所示，其中A1、A2及A3表示一棵决策树的三个属性，每个属性例如为上述的相关特征量之一；其中A1对应三个不同属性条件a1,1、a1,2及a1,3，A2对应两个不同属性条件a2,1、a2,2，A3对应两个不同属性条件a3,1、a3,2；C1和C2表示入侵信号的两个类别，例如如上所述，包括攀爬、敲击及环境(例如风、雨、雪及车辆来往等)等。从根(如图中A1)到叶的每一条路径创建一个分类规则，分别以IF-THEN形式表示，其中沿着路径上的每个属性条件形成规则前件(即IF部分)的一个合取项，包含分类(Cn)的叶结点，形成规则后件(即THEN部分)。例如，如果(IF)被分类信号对A1属性值满足条件a1,2，则(THEN)分类为C1；如果被分类信号对A1属性值满足条件a1,1且对A2属性值满足条件a2,2，则分类为C2。以此类推，不再一一赘述。需要说明的是，上述获得决策树模型所确定的分类规则仅为示例说明，而非限制本发明。
获得分类规则后，将所提取的相关特征量套用于该分类规则，以对该入侵信号进行分类。例如，从该入侵信号所提取的特征向量A1、A2的属性条件分别满足条件为a1,1和a2,1，则将该入侵信号分类为C1。
如上所述，预先创建的决策树模型也可以为一决策树组合模型T，该决策树组合模型T包括多棵决策树T1,T2,...,Tk。下面着重说明如何根据决策树组合模型T，确定分类结果。采用简单投票法，对输入的入侵信号进行分类。对每个输入的入侵信号，决策树组合模型T中的每棵决策树T1,T2,...,Tk都会如上述单棵决策树的方法得到一个分类结果，即相当于对某个类型Ci,i＝1,2,...,n投了一票。统计各个类型Ci,i＝1,2,...,n的票数，最终选择获得票数最多的分类Cj,1≤j≤n，以此作为对该入侵信号的最终分类结果。或者，对每棵决策树T1,T2,...,Tk的分类结果进行分组，从而获得至少一组分类结果，分别统计各组分类结果中所包含的分类结果的数量，以其中分类结果数量最多的一组中的分类结果作为最终的分类结果。
下面将具体介绍在进行上述识别方法之前，如何创建上述的决策树模型，该决策树模型既包括仅具有单棵决策树的决策树模型，也包括参与投票相应的多棵决策树的 组合模型。
图3为根据一示例实施例示出的创建用于识别光纤周界入侵信号的决策树模型的方法的流程图。如图3所示，该方法20包括：
步骤S210，预先从光纤周界监测系统的监测信号中分别获取多组已知分类的入侵信号Y(t)，以及其对应的分类结果。
为了建立用于入侵信号分类的决策树模型，需要预先从光纤周界监测信号中分别获取多组已知分类的入侵信号Y(t)。所述已知的分类例如包括上述的攀爬、敲击及环境(例如风、雨、雪及车辆来往等)等，但本发明不以此为限。
步骤S220，分别对多组已知分类的入侵Y(t)进行预处理，以去除其噪声，分别获得去噪后的多组信号Y’(t)。
例如，可以采用小波变换对多组已知分类的入侵信号Y’(t)进行处理，以去除其中的背景白噪声等。
步骤S230，分别截取去除噪声后的多组信号Y’(t)，以从中获得多组离散的入侵信号片段{X(ti)},i＝1,...,N，并分别提取多组离散的入侵信号片段{X(ti)},i＝1,...,N的相关特征量。
具体地，相关特征量例如包括峰均值差(Diff)、方差(D)、偏度(Skew)、峭度(K)、极值点数目(Num)及频域特征量中的部分或全部。其中峰均值差(Diff)、方差(D)、偏度(Skew)、峭度(K)、极值点数目(Num)属于时域特征量。上述相关特征向量的定义与计算如前所述，在此不再赘述。
步骤S240，根据多组入侵信号片段{X(ti)},i＝1,...,N的相关特征量，建立训练数据集，创建决策树模型。
具体地，设每组入侵信号片段{X(ti)},i＝1,...,N的相关特征量包括{A1,A2,......,AM}，这些特征向量组成了一个输入向量，多组输入向量组成了决策树模型的训练数据集D。
根据总体训练数据集D，创建决策树模型。首先介绍包含有单棵决策树的决策树模型的创建。下面以采用Gini指标作为分裂属性度量方法的CART算法为例进行说明。
首先，Gini指标定义为如下公式：
Figure PCTCN2016096951-appb-000009
其中，pi是类别Ci在训练数据集D中出现的概率。
对于训练数据集D中某个属性A，如果将训练数据集D划分为两个部分D1和D2，则给定的D的Gini指标如下：
Figure PCTCN2016096951-appb-000010
对该属性的不同子集进行上述Gini指标的计算，选择产生最小Gini指标的子集作为该属性的分裂子集。重复以上过程，直到分裂停止。此时包含有单棵决策树的决策树模型创建完毕。
在建立决策树后，为防止由训练数据集样本带噪声或者训练数据集样本不充分等原因引起的过拟合现象，需要对建立的决策树进行剪枝。剪枝主要包括预剪枝和后剪枝。
其中，预剪枝是通过提前停止树的构造，如通过决定在给定的节点不再分裂或划分训练元组的子集，而对树剪枝，一旦停止，该节点即成为叶结点。而后剪枝是由完全生长的树剪去子树，通过删除节点的分支，并用叶结点替换它而剪掉给定节点的子树，叶结点使用被替换的子树中最频繁的类标记。后剪枝的原则包括最小描述长度原则和最小期望错误率原则。根据统计度量，减去不可靠的分支，提高树独立于测试数据的正确分类能力。
CART算法通常采用后剪枝策略。例如采用CCP(代价-复杂度)方法进行剪枝。首先，自下而上地通过对原始决策树的修剪得到一系列的树{T0,T1,......,Tk}，其中T0为未经任何修剪的原始树，Tk为只有一个结点的树，Ti+1是Ti的一个或多个子树被替换所得到的。之后，根据真实误差率来选择最优秀的树作为最后被剪枝的决策树。
上面是作为一次训练得到的决策树。还可以根据训练数据集进行多次训练，选出最优的决策树，从而确定决策树模型。
此外，所采用的决策树算法还可以包括：采用信息增益作为分裂属性度量的ID3算法或采用信息增益作为分裂属性度量的C4.5算法。
接下来介绍包含有多棵决策树的决策树模型的创建。图7为根据一示例实施例示出的通过原始训练数据集形成组合决策树模型的示意图。
在上述确定总体训练数据集D后，如图7所示，对训练数据集D进行有放回的随机抽样，可获得l个训练数据集D1,D2,...,Dl，每个训练数据集中都包含与原训练 数据集D中的相同数目的输入向量。对这l个训练数据集D1,D2,...,Dl分别建立相应的决策树模型。每棵决策树模型都以采用CART算法为例，但本发明不以此为限。
设前面提取的属性数目为M个。每棵决策树建立时，初始数据都在根结点的位置，之后会对其进行分裂。指定一个属性数为F，且F≤M。对每次分裂过程中，在所有的M个属性中随机抽取F个属性作为分裂属性集。对结点以抽取的F个属性，选择最好的方式进行分裂。分裂过程中，以Gini指标作为度量方法。在决策树的建立过程中，F的数目保持不变。
下面以采用Gini指标作为分裂属性度量方法的CART算法为例进行说明。这里，假设F＝1。
首先，Gini指标定义为如下公式：
Figure PCTCN2016096951-appb-000011
其中，pi是类别Ci在训练数据集D中出现的概率。
对于训练数据集D中的抽取的某个属性Aj(j＝1,…,M)的一个子集A，如果将训练数据集划分为两个部分D1和D2，则给定的D的Gini指标如下：
Figure PCTCN2016096951-appb-000012
分别对该属性的不同子集进行上述Gini指标的计算，选择产生最小Gini指标的子集作为该属性的分裂子集。
每次分裂都重新从所有的M个属性中随机抽取F个属性(F的数目保持不变)。
重复以上过程，直到所有结点都是叶子结点，分裂停止。
此时单棵决策树创建完毕，不需要进行剪枝。
对所有训练数据集D1,D2,...,Dl分别进行这样的过程，得到多棵决策树T1,T2,...,Tk。这样就获得了一个由T1,T2,...,Tk构成的多棵决策树组合T(也称为一个随机森林)。
以未被抽到的原始训练集D中的数据作为检测输入，根据每组输入向量对应的分类结果Ci,i＝1,2,...,n，以及其通过决策树模型最终分类结果，还可以计算整个决策树模型的分类准确率。
本发明提供的光纤周界入侵信号的识别方法，通过对已知分类信号的入侵信号的 相关特征量的提取，获得训练数据集，以建立相应的决策树模型；并根据所创建的决策树模型，确定光纤周界入侵信号的分类规则，以对光纤周界入侵信号进行识别与分类。该方法简单、易于实现，分类能力高并且分类效果较好。
图4为根据一示例实施例示出的光纤周界入侵信号的识别装置的结构图。如图4所示，该识别装置30包括：信号获取模块310、信号预处理模块320、特征量提取模块330以及决策树分类模块340。
其中，信号获取模块310用于从光纤周界监测系统的监测信号中获取实时信号y(t)。通常，该实时信号y(t)经以一采样频率采样后获得。
信号预处理模块320用于对实时信号y(t)进行预处理，以去除其噪声，获得去噪后的信号y’(t)，并判断去噪后的信号y’(t)是否为入侵信号。
例如，可以采用小波变换对实时信号y(t)进行预处理，以去除其中的背景白噪声等噪声。
通常，光纤周界系统中的入侵信号大致被分为攀爬、敲击及环境(例如风、雨、雪及车辆来往等)等几类。在一些实施例中，例如可以采用将去噪后的信号y’(t)与一阈值进行比较的方法来判断其是否为入侵信号。当去噪后的信号y’(t)高于该阈值时，判断该去噪后的信号y’(t)为入侵信号。
特征量提取模块330用于当判断去噪后的信号y’(t)为入侵信号时，对其进行相关特征量的提取。
首先，截取该去噪后的信号y’(t)，以获得信号片段{x(ti)},i＝1,...,N。
之后，提取信号片段{x(ti)},i＝1,...,N的相关特征量。相关特征量例如包括峰均值差(Diff)、方差(D)、偏度(Skew)、峭度(K)、极值点数目(Num)及频域特征量中的部分或全部。其中峰均值差(Diff)、方差(D)、偏度(Skew)、峭度(K)、极值点数目(Num)属于时域特征量。上述相关特征向量的定义与计算如前所述，在此不再赘述。
决策树分类模块340用于根据预先创建的决策树模型所确定的分类规则与所提取的相关特征量，对入侵信号进行分类。
上述预先创建的决策树模型既可以包括含有单棵决策树的模型，也可以包括含有多棵决策树组合的模型。下面我们首先介绍如何根据包含有单棵决策树的决策树模型所确定的分类规则及相关特征量，对入侵信号进行分类。
当预先创建的决策树模型建立后，相应的分类规则也已经确立。具体地，从该决策树模型确定的决策树中抽取分类规则。图2为根据一示例实施例示出的用于识别光纤周界入侵信号的单棵决策树的示意图。以图2所示的单棵决策树来说明如何根据决策树确定分类规则，及进行分类。
如图2所示，其中A1、A2及A3表示决策树的三个属性，每个属性例如为上述的相关特征量之一；其中A1对应三个不同属性条件a1,1、a1,2及a1,3，A2对应两个不同属性条件a2,1、a2,2，A3对应两个不同属性条件a3,1、a3,2；C1和C2表示入侵信号的两个类别，例如如上所述，包括攀爬、敲击及环境(例如风、雨、雪及车辆来往等)等。从根(如图中A1)到叶的每一条路径创建一个分类规则，分别以IF-THEN形式表示，其中沿着路径上的每个属性条件形成规则前件(即IF部分)的一个合取项，包含分类(Cn)的叶结点，形成规则后件(即THEN部分)。例如，如果(IF)A1的属性条件满足条件a1,2，则(THEN)分类为C1；如果A1的属性条件满足条件a1,1且A2的属性条件满足条件a2,2，则分类为C2。以此类推，不再一一赘述。需要说明的是，上述获得决策树模型所确定的分类规则仅为示例说明，而非限制本发明。
获得分类规则后，将所提取的相关特征量套用于该分类规则，以对该入侵信号进行分类。例如，从入侵信号中所提取的特征向量A1、A2的属性条件分别满足条件为a1,1和a2,1，则将该入侵信号分类为C1。
如上所述，预先创建的决策树模型也可以为一决策树组合模型T，该决策树组合模型T包括多棵决策树T1,T2,...,Tk。下面着重说明如何根据决策树组合模型T，确定分类结果。采用简单投票法，对输入的入侵信号进行分类。对每个输入的入侵信号，决策树组合模型T中的每棵决策树T1,T2,...,Tk都会如上述单棵决策树的方法得到一个分类结果，即相当于对某个类型Ci,i＝1,2,...,n投了一票。统计各个类型Ci,i＝1,2,...,n的票数，最终选择获得票数最多的分类Cj,1≤j≤n，以此作为对该入侵信号的最终分类结果。或者，对每棵决策树T1,T2,...,Tk的分类结果进行分组，从而获得至少一组分类结果，分别统计各组分类结果中所包含的分类结果的数量，以其中分类结果数量最多的一组中的分类结果作为最终的分类结果。
此外，光纤周界入侵信号的识别装置30还用于在识别入侵信号之前，创建上述的决策树模型，该决策树模型既包括仅具有单棵决策树的决策树模型，也包括参与投票相应的多棵决策树的组合模型。
其中信号获取模块310还用于预先从光纤周界监测信号中分别获取多组已知分 类的入侵信号Y(t)。
为了建立用于入侵信号分类的决策树模型，需要预先从光纤周界监测系统的监测信号中分别获取多组已知分类的入侵信号Y(t)，以及其对应的分类结果。所述已知的分类例如包括上述的攀爬、敲击及环境(例如风、雨、雪及车辆来往等)等，但本发明不以此为限。
信号预处理模块320还用于分别对多组已知分类的入侵Y(t)进行预处理，以去除其噪声，分别获得去噪后的多组信号Y’(t)。
特征量提取模块330还用于分别截取去除噪声后的多组信号Y’(t)，以从中获得多组离散的入侵信号片段{X(ti)},i＝1,...,N，并分别提取多组离散的入侵信号片段{X(ti)},i＝1,...,N的相关特征量。
具体地，相关特征量例如包括峰均值差(Diff)、方差(D)、偏度(Skew)、峭度(K)、极值点数目(Num)及频域特征量中的部分或全部。其中峰均值差(Diff)、方差(D)、偏度(Skew)、峭度(K)、极值点数目(Num)属于时域特征量。上述相关特征向量的定义与计算如前所述，在此不再赘述。
决策树分类模块340还用于根据多组入侵信号片段{X(ti)},i＝1,...,N的相关特征量，建立训练数据集，创建决策树模型。
具体地，设每组入侵信号片段{X(ti)},i＝1,...,N的相关特征量包括{A1,A2,......,AM}，这些特征向量组成了一个输入向量，多组输入向量组成了决策树模型的训练数据集D。
根据总体的训练数据集D，创建决策树模型。首先介绍包含有单棵决策树的决策树模型的创建。下面以采用Gini指标作为分裂属性度量方法的CART算法为例进行说明。
首先，Gini指标定义为如下公式：
Figure PCTCN2016096951-appb-000013
其中，pi是类别Ci在训练数据集D中出现的概率。
对于训练数据集D中的某个属性A，如果将训练数据集D划分为两个部分D1和D2，则给定的D的Gini指标如下：
Figure PCTCN2016096951-appb-000014
对该属性的不同子集进行上述Gini指标的计算，选择产生最小Gini指标的子集作为该属性的分裂子集。重复以上过程，直到分裂停止。之后，用后剪枝方法对已建立的决策树进行剪枝，形成一次训练的决策树模型。多次反复训练，以获得最佳决策树模型。
接下来介绍包含有多棵决策树的决策树模型的创建。继续参考图7，在上述确定总体训练数据集D后，对训练数据集D进行有放回的随机抽样，可获得l个训练数据集D1,D2,...,Dl，每个训练数据集中都包含与原训练数据集D中的相同数目的输入向量。对这l个训练数据集D1,D2,...,Dl分别建立相应的决策树模型。每棵决策树模型都以采用上述的CART算法为例，但本发明不以此为限。
前面提取的属性总数目为M个，指定一个属性数为F，满足F≤M。在决策树的建立过程中，F的数目保持不变。在每次分裂过程中，都从M个属性中随机抽取F个属性作为分裂属性集。以Gini指标作为度量方法，选择最佳分裂方式，直到分裂停止为止。
多棵决策树模型中的单棵决策树的具体创建如方法部分所述，在此不再赘述。创建完毕每棵决策树后，不需要进行剪枝。
对所有训练数据集D1,D2,...,Dl分别进行这样的过程，得到多棵决策树T1,T2,...,Tk。这样就获得了一个由T1,T2,...,Tk构成的多棵决策树组合T(也称为一个随机森林)。
以未被抽到的原始训练集D中的数据作为检测输入，根据每组输入向量对应的类别Ci,i＝1,2,...,n，以及其通过决策树模型最终分类结果，还可以计算整个决策树模型的分类准确率。
本发明提供的光纤周界入侵信号的识别装置，通过对已知分类信号的入侵信号的相关特征量的提取，获得训练数据集，以建立相应的决策树模型；并根据所创建的决策树模型，确定光纤周界入侵信号的分类规则，以对光纤周界入侵信号进行识别与分类。该装置所采用的识别方法简单、易于实现，分类能力高并且分类效果较好。
下面将介绍一种包括本发明提供的光纤周界入侵信号的识别装置的光纤周界入侵报警系统。图5为根据一示例实施例示出的光纤周界入侵报警系统的结构图。如图5所示，该光纤周界入侵报警系统40包括：光路子系统410、光纤周界入侵信号的识 别装置30以及报警子系统420。
其中光路子系统410用于提供光纤周界的监测信号。下面以采用迈克尔逊干涉仪为例，介绍光路系统410，但本发明不以此为限。图6为采用迈克尔逊干涉仪的光路子系统的结构图。如图6所示，光路子系统410包括：激光器4110、耦合器4120、参考臂4130、敏感臂4140、反射端面4150及探测器4160。
其中，激光器4110发出的窄带激光，经过反射端面4150的反射，与参考臂4130反射的光在探测器4160上干涉。如果有入侵事件，则探测信号在敏感臂4140上受扰动信号的调制而发生相位改变，引起干涉条纹变化，从而使探测器4160探测的光强发生变化。探测器4160例如为光电探测器，将探测到的光信号转换为电信号输出，以提供光纤周界的监测信号。
光纤周界入侵信号的识别装置30根据该监测信号，识别入侵信号并进行入侵信号的分类。关于该识别装置30如何进行入侵信号识别与分类已在上文中详细描述，在此不再赘述。
根据识别装置30输出的入侵信号的分类信息，报警子系统420相应地进行报警。报警子系统420例如可以包括报警输出控制模块(图中未示出)，用于输出控制报警外设以进行报警。
以上具体地示出和描述了本发明的示例性实施方式。应该理解，本发明不限于所公开的实施方式，相反，本发明意图涵盖包含在所附权利要求范围内的各种修改和等效置换。

Claims (21)

  1. 一种光纤周界入侵信号的识别方法，其特征在于，包括：
    从一光纤周界监测系统的监测信号中获取实时信号；
    对所述实时信号进行预处理，以去除所述实时信号的噪声；
    判断去除噪声后的所述实时信号是否为入侵信号；
    如果为所述入侵信号，则截取所述入侵信号，以从所述入侵信号中获得信号片段，并分别提取所述信号片段的相关特征量；以及
    根据预先创建的决策树模型所确定的分类规则及所述信号片段的相关特征量对所述入侵信号进行分类；
    其中所述决策树模型包含至少一棵决策树。
  2. 根据权利要求1所述的识别方法，其中每棵决策树所确定的分类规则包括：从该决策树的根到其叶结点的每条路径上的分类规则；其中每个分类规则包括：以沿着其每条路径上的每个属性条件形成的该分类规则的前件的一个合取项以及以该路径上的叶结点形成的该分类规则的后件。
  3. 根据权利要求1或2所述的识别方法，其中所述决策树模型含有多棵决策树；根据预先创建的决策树模型所确定的分类规则及所述信号片段的相关特征量对所述入侵信号进行分类包括：
    依次根据所述多棵决策树中的每棵决策树所确定的分类规则及所述信号的相关特征量对所述入侵信号进行分类，以获得相应的多个分类结果；
    将所述多个分类结果中相同的分类结果分为一组，获得至少一组分类结果，分别统计各组分类结果中的分类结果的数量，以其中分类结果数量最多的一组中的分类结果作为所述入侵信号最终的分类结果。
  4. 根据权利要求1所述的识别方法，还包括创建所述决策树模型，包括：
    预先从所述光纤周界监测系统的监测信号中分别获取多组已知分类的入侵信号及其对应的分类结果；
    分别对所述多组已知分类的入侵信号进行预处理，以去除所述多组已知分类的入侵信号的噪声；
    分别截取去除噪声后的所述多组已知分类的入侵信号，以从所述多组已知分类的入侵信号中获得多组入侵信号片段，并分别提取所述多组入侵信号片段的相关特 征量；以及
    根据所述多组入侵信号片段的多组相关特征量所组成的第一训练数据集，创建所述决策树模型；
    其中一组相关特征量组成所述第一训练数据集的一个输入向量。
  5. 根据权利要求4所述的识别方法，其中所述决策树模型含有多棵决策树；创建所述决策树模型包括：
    对所述第一训练数据集进行有放回的随机抽样，以获得多个第二训练数据集；其中每个第二训练数据集与所述第一训练数据集所包含的输入向量数目相同；
    根据所述多个第二训练数据集，采用决策树算法分别建立相应的所述多棵决策树。
  6. 根据权利要求5所述的识别方法，其中在建立每棵所述决策树时，针对每个所述第二训练数据集的输入向量，每次分裂都重新从该输入向量的所有相关特征量中随机抽取部分或全部的相关特征量。
  7. 根据权利要求4所述的识别方法，其中创建所述决策树模型包括：根据所述第一训练数据集，采用决策树算法建立决策树，对所述决策树进行后剪枝，以获得所述决策树模型。
  8. 根据权利要求7所述的识别方法，其中创建所述决策树模型还包括：根据所述第一训练数据集，多次对所述决策树模型进行训练，以获得最佳的决策树模型。
  9. 根据权利要求5或7所述的识别方法，其中所述决策树算法包括：采用Gini指标作为分裂属性度量的CART算法。
  10. 根据权利要求1或4所述的识别方法，其中所述相关特征量包括：峰均值差、方差、偏度、峭度、极点数目及采用小波包分解获得的频域特征量中的部分或全部。
  11. 一种光纤周界入侵信号的识别装置，其特征在于，包括：
    信号获取模块，用于从一光纤周界监测系统的监测信号中获取实时信号；
    信号预处理模块，用于对所述实时信号进行预处理，以去除所述实时信号的噪声，及判断去除噪声后的所述实时信号是否为入侵信号；
    特征量提取模块，用于当所述实时信号为所述入侵信号时，截取所述入侵信号，以从所述入侵信号中获得信号片段，并分别提取所述信号片段的相关特征量；以及
    决策树分类模块，用于根据预先创建的决策树模型所确定的分类规则及所述信号片段的相关特征量对所述入侵信号进行分类；
    其中所述决策树模型包含至少一棵决策树。
  12. 根据权利要求11所述的识别装置，其中每棵决策树所确定的分类规则包括：从该决策树的根到其叶结点的每条路径上的分类规则；其中每个分类规则包括：以沿着其每条路径上的每个属性条件形成的该分类规则的前件的一个合取项以及以该路径上的叶结点形成的该分类规则的后件。
  13. 根据权利要求11或12所述的识别装置，其中所述决策树模型含有多棵决策树；根据预先创建的决策树模型所确定的分类规则及所述信号片段的相关特征量对所述入侵信号进行分类包括：
    依次根据所述多棵决策树中的每棵决策树所确定的分类规则及所述信号的相关特征量对所述入侵信号进行分类，以获得相应的多个分类结果；
    将所述多个分类结果中相同的分类结果分为一组，获得至少一组分类结果，分别统计各组分类结果中的分类结果的数量，以其中分类结果数量最多的一组中的分类结果作为所述入侵信号最终的分类结果。
  14. 根据权利要求11所述的识别装置，其中，
    所述信号获取模块还用于预先从所述光纤周界监测系统的监测信号中分别获取多组已知分类的入侵信号及其对应的分类结果；
    所述预处理模块还用于分别对所述多组已知分类的入侵信号进行预处理，以去除所述多组已知分类的入侵信号的噪声；
    所述特征量提取模块还用于分别截取去除噪声后的所述多组已知分类的入侵信号，以从所述多组已知分类的入侵信号中获得多组入侵信号片段，并分别提取所述多组入侵信号片段的相关特征量；以及
    所述决策树分类模块还用于根据所述多组入侵信号片段的多组相关特征量所组成的第一训练数据集，创建所述决策树模型；
    其中一组相关特征量组成所述第一训练数据集的一个输入向量。
  15. 根据权利要求14所述的识别装置，其中所述决策树模型含有多棵决策树；创建所述决策树模型包括：
    对所述第一训练数据集进行有放回的随机抽样，以获得多个第二训练数据集；其中每个第二训练数据集与所述第一训练数据集所包含的输入向量数目相同；
    根据所述多个第二训练数据集，采用决策树算法分别建立相应的所述多棵决策树。
  16. 根据权利要求15所述的识别装置，其中在建立每棵所述决策树时，针对每个所述第二训练数据集的输入向量，每次分裂都重新从该输入向量的所有相关特征量中随机抽取部分或全部的相关特征量。
  17. 根据权利要求14所述的识别装置，其中创建所述决策树模型包括：根据所述第一训练数据集，采用决策树算法建立决策树，对所述决策树进行后剪枝，以获得所述决策树模型。
  18. 根据权利要求17所述的识别装置，其中创建所述决策树模型还包括：根据所述第一训练数据集，多次对所述决策树模型进行训练，以获得最佳的决策树模型。
  19. 根据权利要求15或17所述的识别装置，其中所述决策树算法包括：采用Gini指标作为分裂属性度量的CART算法。
  20. 根据权利要求11或14所述的识别装置，其中所述相关特征量包括：峰均值差、方差、偏度、峭度、极点数目及采用小波包分解获得的频域特征量中的部分或全部。
  21. 一种光纤周界入侵报警系统，其特征在于，包括：
    光路子系统，用于提供光纤周界的监测信号；
    根据权利要求11-20任一项所述的光纤周界入侵信号的识别装置，用于对所述监测信号中的入侵信号进行识别及分类；以及
    报警子系统，用于根据所述识别装置对所述入侵信号的分类，相应地进行报警。
PCT/CN2016/096951 2015-09-02 2016-08-26 光纤周界入侵信号的识别方法、装置及周界入侵报警系统 Ceased WO2017036363A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
EA201792283A EA036070B1 (ru) 2015-09-02 2016-08-26 Способ распознавания сигнала о вторжении в пределы оптоволоконного периметра, устройство распознавания и система сигнализации о вторжении в пределы оптоволоконного периметра
BR112017024488A BR112017024488A2 (pt) 2015-09-02 2016-08-26 método de identificação de sinal de intrusão de perímetro de fibra óptica, dispositivo de identificação e sistema de alarme de intrusão do perímetro de fibra óptica
MX2017015387A MX374173B (es) 2015-09-02 2016-08-26 Método de identificación de señal de intrusión de perímetro de fibra óptica, dispositivo de identificación y sistema de alarma de intrusión de perímetro de fibra óptica.
EP16840792.2A EP3321859B1 (en) 2015-09-02 2016-08-26 Optical fiber perimeter intrusion signal identification method and device, and perimeter intrusion alarm system

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201510556766.2 2015-09-02
CN201510556766.2A CN105184319B (zh) 2015-09-02 2015-09-02 光纤周界入侵信号的识别方法、装置及周界入侵报警系统

Publications (1)

Publication Number Publication Date
WO2017036363A1 true WO2017036363A1 (zh) 2017-03-09

Family

ID=54906383

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/096951 Ceased WO2017036363A1 (zh) 2015-09-02 2016-08-26 光纤周界入侵信号的识别方法、装置及周界入侵报警系统

Country Status (6)

Country Link
EP (1) EP3321859B1 (zh)
CN (1) CN105184319B (zh)
BR (1) BR112017024488A2 (zh)
EA (1) EA036070B1 (zh)
MX (1) MX374173B (zh)
WO (1) WO2017036363A1 (zh)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107331097A (zh) * 2017-08-01 2017-11-07 中科融通物联科技无锡有限公司 基于目标位置信息融合的周界防入侵装置与方法
CN110502979A (zh) * 2019-07-11 2019-11-26 哈尔滨工业大学 一种基于决策树的激光雷达波形信号分类方法
CN110852187A (zh) * 2019-10-22 2020-02-28 华侨大学 一种周界入侵事件的识别方法与系统
CN111079705A (zh) * 2019-12-31 2020-04-28 北京理工大学 一种振动信号分类方法
CN112309063A (zh) * 2020-10-30 2021-02-02 魏运 混合光纤入侵信号特征谱提取方法及装置
CN113409305A (zh) * 2021-07-15 2021-09-17 中国南方电网有限责任公司 一种光缆资源数字化管理方法
CN115311534A (zh) * 2022-08-26 2022-11-08 中国铁道科学研究院集团有限公司 基于激光雷达的铁路周界入侵辨识方法、装置及存储介质
CN115479219A (zh) * 2022-09-20 2022-12-16 无锡科晟光子科技有限公司 智能管道状态监测方法、监测装置及智能管道系统
CN116453278A (zh) * 2023-03-30 2023-07-18 上海波汇科技有限公司 一种深度学习智能检测与光纤振动传感结合的入侵监测方法
CN116977705A (zh) * 2023-05-29 2023-10-31 齐鲁工业大学(山东省科学院) 基于深度学习的光纤入侵检测方法及系统
CN116318787B (zh) * 2022-12-08 2023-11-07 中国科学院信息工程研究所 一种基于可解释人工智能的实时入侵检测方法及系统
CN118051830A (zh) * 2024-04-16 2024-05-17 齐鲁工业大学(山东省科学院) 一种周界安防入侵事件识别方法

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105184319B (zh) * 2015-09-02 2019-07-30 同方威视技术股份有限公司 光纤周界入侵信号的识别方法、装置及周界入侵报警系统
CN105654645B (zh) * 2016-01-22 2018-04-10 北京北邮国安技术股份有限公司 一种光纤安防信号处理方法及系统
CN105551163B (zh) * 2016-02-29 2018-05-08 刘海 一种光纤智能防盗装置
CN106226845B (zh) * 2016-08-11 2018-08-28 国家电网公司 一种从opgw振动信号中识别冰雹分布区域的方法
CN108961641B (zh) * 2018-07-24 2021-03-19 民航成都电子技术有限责任公司 一种基于分类树的降低电容围界报警系统虚警的方法
CN109489800A (zh) * 2018-12-14 2019-03-19 广东世港信息科技有限公司 一种分布式光缆振动安全预警系统中的扰动事件识别方法
CN110481600B (zh) * 2019-08-07 2024-06-11 深圳市信海通科技有限公司 无人自主综合报警系统
CN114268365B (zh) * 2021-12-02 2023-07-11 国网甘肃省电力公司酒泉供电公司 一种基于可视化技术的通信光缆智能预警方法及系统
CN114781437A (zh) * 2022-03-29 2022-07-22 中科朗劢技术有限公司 入侵信号的分类方法、系统及可读存储介质
US12517274B2 (en) 2022-08-18 2026-01-06 Network Integrity Systems, Inc. Intrusion detection algorithm with wind rejection heuristic
CN116131935A (zh) * 2022-11-09 2023-05-16 南京大学 基于数据重构和lstm入侵节点检测的方法、装置及存储介质
CN115798131B (zh) * 2023-02-13 2023-04-28 成都陆迪盛华科技有限公司 一种基于分布式光纤的多维度特征入侵检测方法
CN118553056A (zh) * 2024-05-17 2024-08-27 上海彩佩电子科技有限公司 一种基于目标识别的光纤周界安防报警方法及系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060156404A1 (en) * 2002-07-30 2006-07-13 Day Christopher W Intrusion detection system
CN103236127A (zh) * 2013-05-06 2013-08-07 无锡成电光纤传感科技有限公司 一种光纤围栏入侵监测系统及其模式识别方法
CN103345808A (zh) * 2013-06-26 2013-10-09 武汉理工光科股份有限公司 光纤光栅周界入侵模式识别方法及系统
CN104240455A (zh) * 2014-08-07 2014-12-24 北京航天控制仪器研究所 一种分布式光纤管线安全预警系统中的扰动事件识别方法
CN104751580A (zh) * 2015-04-08 2015-07-01 武汉理工光科股份有限公司 分布式光纤传感信号模式识别方法及系统
CN105184319A (zh) * 2015-09-02 2015-12-23 同方威视技术股份有限公司 光纤周界入侵信号的识别方法、装置及周界入侵报警系统

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10235605B2 (en) * 2013-04-10 2019-03-19 Microsoft Technology Licensing, Llc Image labeling using geodesic features

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060156404A1 (en) * 2002-07-30 2006-07-13 Day Christopher W Intrusion detection system
CN103236127A (zh) * 2013-05-06 2013-08-07 无锡成电光纤传感科技有限公司 一种光纤围栏入侵监测系统及其模式识别方法
CN103345808A (zh) * 2013-06-26 2013-10-09 武汉理工光科股份有限公司 光纤光栅周界入侵模式识别方法及系统
CN104240455A (zh) * 2014-08-07 2014-12-24 北京航天控制仪器研究所 一种分布式光纤管线安全预警系统中的扰动事件识别方法
CN104751580A (zh) * 2015-04-08 2015-07-01 武汉理工光科股份有限公司 分布式光纤传感信号模式识别方法及系统
CN105184319A (zh) * 2015-09-02 2015-12-23 同方威视技术股份有限公司 光纤周界入侵信号的识别方法、装置及周界入侵报警系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3321859A4 *

Cited By (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107331097A (zh) * 2017-08-01 2017-11-07 中科融通物联科技无锡有限公司 基于目标位置信息融合的周界防入侵装置与方法
CN110502979A (zh) * 2019-07-11 2019-11-26 哈尔滨工业大学 一种基于决策树的激光雷达波形信号分类方法
CN110502979B (zh) * 2019-07-11 2023-04-14 哈尔滨工业大学 一种基于决策树的激光雷达波形信号分类方法
CN110852187B (zh) * 2019-10-22 2023-04-07 华侨大学 一种周界入侵事件的识别方法与系统
CN110852187A (zh) * 2019-10-22 2020-02-28 华侨大学 一种周界入侵事件的识别方法与系统
CN111079705A (zh) * 2019-12-31 2020-04-28 北京理工大学 一种振动信号分类方法
CN111079705B (zh) * 2019-12-31 2023-07-25 北京理工大学 一种振动信号分类方法
CN112309063A (zh) * 2020-10-30 2021-02-02 魏运 混合光纤入侵信号特征谱提取方法及装置
CN112309063B (zh) * 2020-10-30 2022-09-09 魏运 混合光纤入侵信号特征谱提取方法及装置
CN113409305A (zh) * 2021-07-15 2021-09-17 中国南方电网有限责任公司 一种光缆资源数字化管理方法
CN115311534A (zh) * 2022-08-26 2022-11-08 中国铁道科学研究院集团有限公司 基于激光雷达的铁路周界入侵辨识方法、装置及存储介质
CN115479219A (zh) * 2022-09-20 2022-12-16 无锡科晟光子科技有限公司 智能管道状态监测方法、监测装置及智能管道系统
CN115479219B (zh) * 2022-09-20 2024-03-01 无锡科晟光子科技有限公司 智能管道状态监测方法、监测装置及智能管道系统
CN116318787B (zh) * 2022-12-08 2023-11-07 中国科学院信息工程研究所 一种基于可解释人工智能的实时入侵检测方法及系统
CN116453278A (zh) * 2023-03-30 2023-07-18 上海波汇科技有限公司 一种深度学习智能检测与光纤振动传感结合的入侵监测方法
CN116977705A (zh) * 2023-05-29 2023-10-31 齐鲁工业大学(山东省科学院) 基于深度学习的光纤入侵检测方法及系统
CN118051830A (zh) * 2024-04-16 2024-05-17 齐鲁工业大学(山东省科学院) 一种周界安防入侵事件识别方法

Also Published As

Publication number Publication date
EA036070B1 (ru) 2020-09-22
CN105184319B (zh) 2019-07-30
MX2017015387A (es) 2018-03-09
BR112017024488A2 (pt) 2018-07-24
EP3321859B1 (en) 2021-01-27
EP3321859A1 (en) 2018-05-16
MX374173B (es) 2025-03-05
CN105184319A (zh) 2015-12-23
EP3321859A4 (en) 2019-03-27
EA201792283A1 (ru) 2018-06-29

Similar Documents

Publication Publication Date Title
WO2017036363A1 (zh) 光纤周界入侵信号的识别方法、装置及周界入侵报警系统
US12366504B2 (en) Rolling bearing combined fault diagnosis method based on multi-domain feature construction
CN113780094B (zh) 基于混合高斯模型和隐马尔可夫模型的周界入侵识别方法
CN117436005B (zh) 一种环境空气自动监测过程中异常数据处理方法
Bulakh et al. Time series classification based on fractal properties
CN105930723A (zh) 一种基于特征选择的入侵检测方法
CN113608566B (zh) 一种纺织车间的环境监控调整方法及系统
CN110444011B (zh) 交通流高峰识别方法、装置、电子设备及存储介质
CN110874471B (zh) 保护隐私安全的神经网络模型的训练方法和装置
CN106650644B (zh) 驾驶员危险行为识别方法及系统
CN116838955B (zh) 一种两阶段的油气管道沿线干扰识别方法
CN108801437B (zh) 基于扰动信号特征提取的分布式光纤振动传感定位方法及装置
CN116305588B (zh) 一种风洞试验数据异常检测方法、电子设备及存储介质
CN117889945A (zh) 一种公路桥梁施工振动测试方法
Reif et al. Anomaly detection by combining decision trees and parametric densities
CN120234685A (zh) 基于深度学习的纺纱机故障检测方法与系统
CN111814689B (zh) 火灾识别网络模型的训练方法、火灾识别方法及相关设备
CN115719428B (zh) 基于分类模型的人脸图像聚类方法、装置、设备及介质
CN113569879A (zh) 异常识别模型的训练方法、异常账号识别方法及相关装置
CN119832673B (zh) 一种基于物联网的校园安防系统及其服务终端
CN107783942B (zh) 一种异常行为检测方法及装置
CN119853796A (zh) 一种宽固网用户光链路质量达标分析与监控方法
Ivanichenko et al. Exposing deviations in information processes using multifractal analysis
CN113919237B (zh) 一种风机设备在线工况分割及故障诊断的方法
CN117059124A (zh) 一种针对带噪场景的伪造音频检测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 16840792

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 201792283

Country of ref document: EA

WWE Wipo information: entry into national phase

Ref document number: MX/A/2017/015387

Country of ref document: MX

WWE Wipo information: entry into national phase

Ref document number: 139650140003010326

Country of ref document: IR

NENP Non-entry into the national phase

Ref country code: DE

REG Reference to national code

Ref country code: BR

Ref legal event code: B01A

Ref document number: 112017024488

Country of ref document: BR

ENP Entry into the national phase

Ref document number: 112017024488

Country of ref document: BR

Kind code of ref document: A2

Effective date: 20171114