JPH0465632A - Measuring instrument - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

Detailed Description of the Invention (Field of Industrial Application) The present invention relates to a measuring device for measuring physical values such as distance, temperature, humidity, velocity, acceleration, etc. Regarding measuring devices that can measure
More typically, the present invention relates to a distance measuring device used, for example, to accurately determine the amount of grinding when cutting an object.

(Background of the Invention and Prior Art) Currently, various sensors are used in various fields depending on the purpose, and for this purpose, various sensors are used in various fields.
Sensor elements, probes, or pressure detection information processing systems have been proposed.

Particularly in recent years, with the development of industrial robots and the increasing demand for ultra-precision machining, it has become extremely important to accurately detect the positional relationship between processing jigs and workpieces. Therefore, there is a need for a distance measuring device that can measure distance with high precision over a wide range, for example. Further, such a measuring device is desirably required to be stable and highly accurate, with as little influence as possible from variations in accuracy due to changes in the working environment.

However, it is not easy to develop a measuring device that can satisfy these requirements. For example, to explain the case where an eddy current sensor is used as the above-mentioned distance measuring device, generally,
It is difficult to require high accuracy from an eddy current sensor that can measure over a wide range, and conversely, it is difficult to require an eddy current sensor that can perform highly accurate measurements to simultaneously measure over a wide range. is normal.

For example, when comparing a sensor prepared to measure a distance of about 0 to 1 Omm and a sensor prepared to measure a distance of about 0 to 2 mm, it is assumed that the measured value obtained for both is a few μm. In the case where a numerical value of the order of magnitude is included, it is not necessarily easy or practical to create a sensor that requires high precision and reliability up to the numerical value of the order of several micrometers in the former sensor. On the other hand, in the latter sensor,
If the numerical value on the order of several micrometers is not reliable, there are many cases where the requirements for the sensor cannot be satisfied. As described above, the performance required of a sensor is generally set at a level of required accuracy in consideration of the range to be measured.

Therefore, considering the technical difficulty of achieving the contradictory requirements of measuring over a wide range and measuring with high precision and high sensitivity with a single sensor, we decided to use multiple sensors with different characteristics and performance. Systems that pseudo-satisfy the above-mentioned requirements by combining and properly using these have been proposed, and these are generally used for sensor integration.
system (abbreviated as rS I SJ).

Specifically, taking the distance sensor mentioned above as an example, the measurable distance range is wide, but the measured distance! A sensor element whose measurement accuracy cannot be said to be high, a sensor element whose measurable distance range is narrow but can obtain measurement information with high accuracy, and a sensor element whose measurement ability is intermediate between these two are combined into a single sensor element. By configuring a composite sensor (hereinafter referred to as a "composite sensor") and using the measurement information from one of the sensor elements as the output of this composite sensor depending on the usage situation, the user can feel as if it were a simple device. The aim is to ensure the same ease of handling as using one sensor.

By the way, since this rSI SJ uses multiple sensor elements, it is necessary to select a value that is close to the true measured value and has high accuracy from among the multiple measured values obtained from the multiple sensor elements. As a method to deal with this problem, the following proposals have been made.

For example, hypothesis testing using a sensor model and a Bayesian approach (Durrant-White, H.F., Sen.
sormodels and multisensors
in'tegration, J, Rob.

t, Res, 7-6 (Dec, 1988), pp, 9
7-113), method using confidence distance matrix (
Luo, R.C., Lin, M.

Dynamic multi-sensor data
fusion system for intllig
ent robots, IEEE J, R'obot.

Automat, 4-4-4 (198, 386-39
6), Performance and cost evaluation method (Zheng, Y
,F, ,Integration of multip
LE 5ENSOR 1NTO ROBOTIC SY
stem and performance
ation IEEE Trans, Robot
icsand Auto+at,,5-5(Oct,
1989), pp65g-669), cost evaluation method using Bayesian method (Richardson, J.M., Marsh
, K.A., ,Fusion of multitise
nsor data, Int, J, Robot,
Res...? -6(Dec.

1988), pp78-96), method using Kalman filter (Nakaa+ura, Y, Xu, Y,...
Geometrical fusions+method
Prof, IEEE Int, Conf, Robot
ics and Automat, (1989),
pp, 66g-673).

(Problems to be Solved by the Invention) However, the above-mentioned conventional rS I SJ'' () does not fully consider the problems associated with using a composite sensor, specifically, for example, the measurement sensitivity There is a problem that different sensors such as Another problem is that since it is a composite sensor formed by combining multiple sensor elements, it is difficult to avoid variations in the position of the sensor elements in a strict sense (, If used uniformly for high-precision measurements without taking into account such variations between products,
This poses a problem in that it is difficult to improve accuracy, and it is difficult to respond flexibly because replacing the sensor significantly changes the characteristics of the device.

The above will be explained in more detail with reference to FIGS. 10 to 14, using a case where a fan blade of a large turbo engine is repaired as an example of mechanical grinding.

In large turbo engines, it is necessary to restore and repair the leading edges of fan blades that have worn out due to erosion or the like due to long-term use. This work is shown in Figure 10(a)~
As shown in (g), fan blade 100 before repair.
Cut the front edge 101 of the fan blade to a certain width (see Fig. 10(b)), apply a member 102 called a batch that is thicker than the fan blade to this cut part, and perform electron beam welding (see Fig. 10(+n)). ), the fan blade 103 is ground into a predetermined shape using a grinding machine, and then repaired and finished into a predetermined fan blade 103. In this case, since the individual shapes, etc. of the fan blades before repair are not the same in a strict sense, it is necessary to perform a grinding process (see FIG. 11) that matches each shape, etc. In this case, as already mentioned, there is generally a limit to accurately detecting the attitude of the blade set in the processing device, the curved trajectory of the surface, etc. using a normal single sensor. For this reason, conventionally, N
It was common to cut with a C processing machine, leaving a certain amount of finishing allowance, and then finish by hand.

However, such manual finishing is extremely time-consuming.

Therefore, the inventor of the present invention set out to develop the present invention in order to automate the process of this manual finishing work.

That is, a composite sensor 104 as shown in FIG. 12 is configured in which a plurality of sensors s, , S, . The composite sensor 104 is scanned along the surface of the fan blade 100 to be processed, and the surface shape is detected. Then, based on the results, for example, a control signal for a working robot is outputted, and the grinding is performed by controlling the drive.

At this time, when detecting the shape pressure of a fan blade using a composite sensor, measurement information from multiple sensors is input to a computer via an A/D converter as shown in Figure 13, and a predetermined value is determined from the multiple measurement information. By selecting, etc., information at each measurement time point is obtained.

However, if a method of selecting measurement information obtained from the plurality of sensors is adopted here, using a so-called if-close program (case classification based on a threshold value) shown in FIG. 14, The initial objectives were not realized.

That is, FIG. 14 shows three distance sensors S, , S2 . S3 (S is measurement range a =
0-20mm, similarly 82 is 520-10mm, S
An example is shown in which a composite sensor (not shown) is configured using a sensor (a is c=0 to 3 mm) and measurement information I is selected by a so-called if close program, and the measurement information I measured by the sensor is Each range a of sensors S, , S, , S,.

It is sequentially determined whether or not it is within the ranges b and c, and the selection of sensor information is switched according to the determination result. However, with this method, the error rate stemming from the measurement principle of these sensors, changes in the error rate due to environmental factors,
In addition, the error rate near the limits is not constant compared to near the center of the measurement range, and if the measurement limits are exceeded, the detected value may become saturated.Therefore, it is unclear whether highly accurate detection is possible at all times. Furthermore, if the accuracy of each sensor installation is not exactly the same, it may lead to the problem of so-called data skipping. For example, I Omm, 3ffIff+
Even if the sensor selection is switched at the level of
.

The measured values are not necessarily consistent, but rather 9.99 mm for Sl and 10.01 mm for S 5
It is conceivable that the distance may be different from 10.03 mm for s. In this case, it is unclear which measurement value is the most appropriate reference for performing the above switching, and the reliability of the selected measurement information becomes a problem.

Furthermore, the error contained in the sensor measurement information is influenced by environmental factors. That is, when using a semiconductor sensor, it is affected by temperature drift, and when using an optical sensor, it is affected by brightness.

In many cases, the degree of influence on measurement results is ambiguous.

These things ultimately lead to rS using the above-mentioned composite sensor.
In reality, this means that it is not always easy to take advantage of the advantages of ISJ.

Based on the various findings and studies described above, the present invention has been developed through further research and has developed an rS system using conventional composite sensors.
This was done with the aim of solving the problems of ISJ and providing a measuring device that can detect excellent measurement methods by suitably responding to various error factors when using a composite sensor. .

Another object of the present invention is that, in a system typically using a plurality of distance sensor elements, the assembly of each sensor element is not affected by variations in position, etc., and is highly sensitive over a wide range. An object of the present invention is to provide a measuring device that can measure distances with high accuracy.

Still another object of the present invention is to provide a distance measuring device that can be suitably used for cutting and the like.

(Means for Solving the Problem) The sensor information integrated measuring device of the present invention, which has been made to achieve the above object, is characterized by a composite sensor configured by combining a plurality of sensors, and a composite sensor configured by combining a plurality of sensors. Means for estimating the error range that may be included in each sensor measurement information, for example, by fuzzy inference, using each of the sensor measurement information measured by the sensor as input and using sensor characteristics given in advance for each sensor as auxiliary information; The multiple sensor measurement information measured by each sensor of the above composite sensor and the multiple error range information obtained by the above error range estimation means are input, and the measured value at each measurement point of the composite sensor is calculated from all these inputs. This neural network uses a composite sensor that has a combination of the plurality of sensors described above, and typically uses learning data constructed in advance using a backpropagation learning algorithm. It has a configuration in which it has as a calculation element.

In the above configuration, the composite sensor may be configured using multiple sensors with the same measurement principle and measurement range, or may be configured using multiple sensors with different measurement principles, and may also be configured using multiple sensors with the same measurement principle and measurement range. It may be configured by using a combination of different materials.

Furthermore, in the above, "each measurement point" refers to each point in time at which measurement is required. Therefore, when measurements are performed continuously, it is necessary to integrate information using a neural network at each point in time. means. In addition, when the measuring device is a distance measuring device, it may be necessary to strictly consider the positional deviation when installing multiple distance sensors. It goes without saying that a shift register or the like may be used to shift the detection points of the measurement information to match each other, as necessary, using information such as the scanning speed during scanning and the amount of positional deviation of each sensor.

Furthermore, the sensor information selection type measuring device of the present invention is characterized by a composite sensor configured by assembling a plurality of distance sensors with different upper limits of measurable ranges on a common substrate; A means for estimating, for example, fuzzy inference, the error range included in each sensor measurement information using sensor measurement information as input and sensor characteristics given in advance for each sensor as auxiliary information; selecting means for selecting sensor measurement information corresponding to the smallest error range and outputting the sensor measurement information at each measurement time point of the composite sensor; The distance sensor with a higher upper limit of the measurable range is assembled to the substrate in a positional relationship that is not closer to the object to be measured than the distance sensor with a lower upper limit of the measurable range.

In the above, the means for estimating the error range that may be included in the sensor measurement information can typically be a means for performing fuzzy inference as illustrated in FIG. 9, but is particularly limited to this. (Instead of fuzzy inference, a program corresponding to this can also be constructed using multi-value inference means. For example, the error range that may be included in sensor measurement information can be calculated based on the sensor measurement information and auxiliary information. (characteristics related to resolution recorded in the sensor database as sensor specifications, errors that may be included, and auxiliary variables such as temperature that affect sensor measurement errors and their effects, etc.) Taking the input of one on the horizontal axis and the other input on the vertical axis, the extent of the error range corresponding to the size of the two-man power is determined in advance and written in a matrix, and when the two-man power is actually input, A means for estimating the error range can be constructed by selecting numerical values of the error range described in the matrix.To give a concrete example of the method for forming the above matrix, for example, a composite sensor is constructed from a plurality of distance sensors. When using the auxiliary sensor as a temperature sensor, create a matrix with an appropriate resolution by taking distance on the horizontal axis and temperature on the vertical axis, and write a statement in which the corresponding numerical value of the error range is stated in each square. Prepare a storage device and use it to record the actual measured distance and temperature.
It is possible to form a means for estimating the error range by multi-value inference using human power as input. The numerical values to be written in each square can also be obtained empirically by fuzzy inference techniques, verification work, etc. Note that the means for estimating the miscalculation range by multi-value inference is not limited to the case of the above inputs (it can also be applied to the case of 3 or more inputs, and the measurement item is the above-mentioned distance).

It is not limited to temperature.

As is known, the fuzzy inference explained in Figure 9 is different from a definite set in the normal mathematical sense, which deals only with a set of objects that can be clearly determined whether they belong to a definite set or not. We deal with fuzzy sets in areas where the criteria for determining whether or not they belong to is not clearly defined.For example, the fuzzy rule is rif x=A+
and y=B+ then z=c, J. Here, x and y are input variables, and in the example of FIG. 9, each measurement value signal (x+
...X, ...xn) and the rule data signal (y+
...y4...yn) correspond to this. On the other hand, z(
z, . . . z, . . . Z, .) are the output variables of each sensor, and in the example of FIG.
. . . xn) is a signal (z+ . . . Z, .

In simplified fuzzy inference with two people and one output,
Assuming that the input variables of each sensor are (x, y) and the output variable is (2), the i-th rule for one sensor is described as follows.

Rl: if x is At and y is B+
then z is C+(i=1.2,...
n) (1) C1 is a real number given empirically or experimentally, and the degree of conformity of the rule is given by the integration of equation (2).

μm=μe1(x)・μm,(y) (2
) Then, the inference error 2 is calculated using the defuzzification equation (3) below.

Reference numeral 3 in FIG. 9 indicates a fuzzy inference part, and the part in which the conditions for the above fuzzy rule to hold are written is the antecedent part,
The conclusion part is called the consequent part.

In fuzzy inference, a certain amount of human power is converted into a value between 0 and 1 for calculation, and this conversion is defined by a membership function, and this membership function is a proposition handled by fuzzy rules (AB in the above) Defined by

Then, referring to the membership function, the degree to which the input amount satisfies the fuzzy rule is calculated, and if there are multiple variables in the antecedent part, a min operation is performed to find the smallest value among them. Next, the results of the min operation for each fuzzy rule are synthesized. Furthermore, by comparing the consequent part of each rule, m
Create a new membership function by calculating a x and finding its maximum value. Compute the centroid of this membership function,
The center of gravity value becomes the inference result (output).

The above fuzzy inference method is not limited to the case where measurement information from a sensor that detects distance is used as an input variable.
Measurable range, measurement accuracy, resolution, usable range due to usage environment, errors due to environmental influences such as temperature drift, dark current drift, and other characteristic information are listed, and the necessary information is selected by the fuzzy inference section. It can be used as another input variable to . Based on these, the range of error size that can be included in the measurement information can be output as an inference result.

In addition, the means for fuzzy inference includes not only the characteristics of the sensor used, but also auxiliary information used for fuzzy inference.
For example, environmental information that is thought to affect the measurement results depending on the measurement principle of the sensor is measured by an auxiliary sensor (indicated by reference numeral 12 in Fig. 9), and the measured information is used as auxiliary information for fuzzy inference. For example, when the composite sensor is an eddy current sensor, a temperature sensor or the like may be used as the auxiliary sensor.

Note that, as already stated, the measuring device of the present invention, whether it is the sensor information integration type device described above or the sensor information selection type device described below, sequentially performs measurement in response to the measurement. The fuzzy inference means is not limited to a means for estimating the error range. In other words, although it depends on the resolution of the measurement device and the number of inputs used to estimate the error range, if the combination of composite sensors is determined, the output (that is, the error) according to the input variables of the same combination of composite sensors For example, in the case of two-manpower, the degree of range) is determined in advance using fuzzy reasoning, and in the case of two-manpower, this is stored in advance in the computer's memory in the form of a two-dimensional matrix as described above, and the value of two-manpower is A means for estimating the error range of the apparatus can be configured by reading out a value indicating the degree of the corresponding miscalculation range from the storage unit by multi-value inference based on the above. In this case, the data to be stored in the storage unit may be obtained by using a method of fuzzy inference, but is not limited to this, and may be data obtained by modifying data obtained by fuzzy inference. , data obtained from other tests etc. may also be used.

The measuring device of the present invention can be configured not only as a sensor information integrated type measuring device using the neural network described above but also as a sensor information selective type distance measuring device having the above-described characteristic configuration. In other words, the smallest error range that can be included in the distance measurement value of each sensor obtained by the fuzzy inference method described above is the distance information that is closest to the true value (it is considered as the information that is probabilistically closest to the true value). For example, if the measurement information corresponding to the smallest inferred value indicating the extent of the miscalculation range is selected as the measurement information of the composite sensor at each point in time using a comparator, etc., the sensor information selection method can be used. This is because a measuring device can be configured.

However, regarding this sensor information integrated type distance measuring device, as will become clear in the examples described later, multiple distance sensors with different measurement ranges in the composite sensor can be measured in order to avoid so-called jumps in measured values. It is necessary that a distance sensor with a higher upper limit of the range is assembled to the substrate in a positional relationship that is not closer to the object to be measured than a distance sensor with a lower upper limit of the measurable range.

To explain the neural network in the sensor information integrated measuring device using the neural network of the present invention, as is well known, this is a computer program that artificially represents human nerve cells (neurons) using a mathematical model. By learning from a wide variety of input data using parallel distributed processing, the desired value can be output, and even unlearned patterns can be expected to produce output that matches the desired value to some extent. . For example, first input is given to a neural network that has applied backpropagation to a learning program, and the eight components of the neural network are examined. If the output is incorrect, the correct output is given to the neural network, and this allows the units in the neural network to be By repeating learning by changing the connection strength, it is possible to construct a neural network that can output desired values in response to input.

A specific example of the sensor information integrated measuring device using the neural network of the present invention will be explained by the following distance measuring device. That is, the measurement information from each distance sensor of the composite sensor made up of the plurality of distance sensors mentioned above and the inference value indicating the size of the error range obtained by the above-mentioned fuzzy inference method are sent to the neural network. It is given as all inputs and repeats learning to see if the output obtained by integrating these matches the desired value.

As a result, it is possible to construct specific learning data unique to a composite sensor in which a plurality of distance sensors have a specific combination. By using this neural network having pre-built learning data in conjunction with the corresponding composite sensor, a distance measuring device is constructed that can accurately measure the distance to the object to be measured.

Although the learning content of the neural network through the above learning is not particularly limited, its outline is as follows. That is, each unit as an intermediate layer formed corresponding to a neuron in the neural network receives the output from the previous unit as an input, converts it according to a certain rule, and outputs it. The input/output relationship is expressed by the following equation: a threshold function (when the sum of inputs exceeds a threshold, a value according to a certain rule is output) and a sigmoid function f (x
).

U nit+ = Σ w ,, o , 100,
=f (Unite)
(4) f (x) = 1 / (1+ exp
(-X)) However, Unit: Input of unit i (C1, 2, -
-1) o4: Output of unit j in the previous layer (j = 1.2
,...J) W, , : coupling coefficient θ of unit j and unit i, : threshold value of unit i.

Then, the integrated output value of the neural network is
If it does not match the externally given desired value, learning is performed by changing the coupling coefficient wl between each unit according to a predetermined rule. Here, changing the coupling coefficient changes the output of unit i at input cover turn p to 0.1
．． If the desired output is pl, then the error function E is E
:Σ (D□−Op,) ” /2 (
5), and learning is performed so that E becomes the minimum value. The amount of change in the coupling coefficient Δw+ is given by: n is a learning coefficient.

Note that the above equation (6) means that the coupling coefficient W is modified in the direction in which the error function when the pattern p is input decreases, that is, in the negative direction.

Note that it is also preferable to prepare a learning database in which specific learning data is separated from the neural network, and to load learning data corresponding to a specific composite sensor into the neural network as necessary. In this way, the neural network program does not need to be changed (it is convenient because it can support various composite sensors and learning databases).

Furthermore, the sensor information integrated measuring device having the above configuration has the following features. In other words, when a large number of composite sensors are manufactured in which the combination of sensors to be used is determined, in a strict sense, it is normal for these multiple composite sensors to have variations in the positions of the sensors. There is a difference in the amount of offset between. In this case, it is of course possible to create learning data for each individual composite sensor, but in a preferred embodiment, when creating learning data using one of the multiple composite sensors,
If you create learning data by incorporating various possible offset amounts into the program and simulating them, even if the actual products have different amounts of offset, as long as the combination of sensors is the same, any complex It can also be used with sensors. Therefore, there is an advantage that the composite sensor can be quickly replaced. Specifically, for example, in a composite sensor composed of a combination of the above-mentioned plurality of distance sensors, if the plurality of distance sensors are expected to have an offset of about 300μ at maximum due to manufacturing technology problems, the sensor By incorporating offsets into the program at appropriate intervals within a range of ±300 μm while taking into account the resolution of For composite sensors having the same combination of sensors, single learning data created including this simulation can be used in common.

Therefore, if the composite sensor has a common combination configuration, there is no need to change the learning data of the neural network even if the composite sensor is replaced (and parts replacement etc. can be quickly handled).

The measuring device of the present invention typically comprises a composite sensor using a plurality of distance sensors with different upper limit values of the measurable range, and further uses a temperature sensor etc. as an auxiliary sensor as necessary. I can give an example of what happened.

(Example of implementation) The present invention will be described below based on embodiments shown in the drawings.

Embodiment 1 FIG. 1 shows an embodiment applied to the sensor information integrated type distance measuring device according to the present invention, in which a composite sensor indicated by the symbol l is used as a main sensor for distance measuring devices with different measurement ranges. S, -S, , are fixedly provided on a - substrate (not shown). The measurement information detected by each of these distance sensors is input to a computer (not shown, λ central processing circuit (hereinafter referred to as "CPU") 50).

The distance sensors S and -Si are mounted on a single substrate, for example.
They are assembled to have different measurable ranges of O~2IIIm, O~5 mm-, and O~20 mm to form a composite sensor. This composite sensor is installed on a moving body having a predetermined driving means so as to be able to move toward and away from the object to be measured, and to scan along the surface of the object to be measured.

CP that receives A/D conversion of the signal from the above composite sensor
The U50 includes a fuzzy inference unit 3 that performs fuzzy inference that is modularized by a built-in program, and a neural network 4, and the plurality of measurement information described above is input to each of these modules.

In addition, in the measuring device of this example, in addition to the distance sensor that is the main sensor, a temperature sensor S1 is installed on the base of the composite sensor.
An auxiliary sensor 12 consisting of .degree. to S, .degree. is provided, and the outputs of these sensors are also A/D converted and input to the fuzzy inference section 3 of the CPLI 50.

2 is a sensor characteristic recording unit that records sensor characteristics;
In this example, a floppy disk (FD) on which the characteristics of each sensor constituting a composite sensor are magnetically recorded is set in an input section (disk drive) installed in a computer, and the contents are input to the fuzzy inference section 3. However, it is not limited to this configuration. For example, the characteristics of the sensors that may be used can be recorded in advance in the memory (RAMROM) of the computer, and the characteristics of the composite sensor to be used can be loaded into the computer memory (RAMROM). Initialize the type of sensor installed in the device in the recording unit (composite sensor database) that records the characteristics of the composite sensor, and when a composite sensor is selected, the necessary sensor characteristics are automatically It may also be loaded into the inference section.

In this example, the above-mentioned eddy current type distance sensor is used as the main sensor mounted on the composite sensor, and the recorded contents regarding one of the sensor characteristics of the sensor provided on the board are given as an example. It is shown as follows.

Measuring Range Lou (μm
): OMeasuring Range H
igh (um): 2f)00Measu
ring Error Late (%): 0
．． 5Using Temp Range Lou (
"C): OLlsing Temp Ra
nge High ('C): 10100
Effective Late (%):
In a sensor with characteristics like 0.0,
For example, from the information on the measurement value measured at the time point and the change in Measuring Error Late determined in relation to Temp Effective Late due to the temperature compared by the temperature sensor which is the auxiliary sensor, it is possible to determine the information included in the measurement value. The error range obtained is fuzzy inferred.

These inferred values and measured values are obtained for each sensor used in the composite sensor, and all these inferred values and measured values are input to the neural network, which is trained in advance. The neural network will provide an integrated and relevant output of the measuring device.

For example, the fuzzy inference section 3 receives measurement information from the composite sensor 1 and rule data from the auxiliary sensor 12 and the sensor characteristic recording section 2 as auxiliary information.
The explanation in the figure can be applied, but in order to perform fuzzy inference on the size of the error range that may be included in the measured value of the eddy current distance sensor adopted in this example, the membership function is set to a trapezoidal shape. The membership function for the measurement target of the i-th sensor is expressed as follows.

mem [il [0] =fdouble) (r
[0] [il) mem [month [1] = (double
e) (r[0] [i]+ (1*p[il))me
m[il [2]・(double) (r[1] [
1l-(1*p(il>)mem[il [3J □(
double) fr11] [il) However, r[0][il Lower limit value of measurement range of sensor i r[1
] [il: Upper limit value of the measurement range of sensor i p[il
: Measurement accuracy of sensor i As a result, the distance measurement values measured by the distance sensors S, ~Sn, the usage specifications of the sensor recorded in the sensor characteristic recording section (sensor database), and the auxiliary sensor 12
The output of the fuzzy inference unit 3 is obtained by inputting the temperature measurement values measured by the temperature sensors S to So°, and indicating the extent of the error range that may be included in the distance measurement values of the respective distance sensors S1 to So.

Reference numeral 4 in FIG. 1 indicates a neural network module, and in this example, the content of pre-constructed learning data can be loaded into this module. That is, the above-mentioned loading can be performed by setting an FD (learning database) 5 on which the content of the learning data is magnetically recorded into an input unit provided in the computer. As with the sensor characteristics recording section above, multiple pieces of learning data can be recorded in advance in the computer's memory (RAM, ROM), and the corresponding learning data can be automatically selected based on the mark, etc. of the composite sensor being used. Needless to say, it is also possible to load the file using the following method.

As the human power of the neural network, the measurement signals of each sensor S1 to SL1 from the composite sensor 1 and the signal (z+ to zn) indicating the inference error value for each measurement signal obtained by the fuzzy inference section 3 are given. , the output of the neural network 4 is finally obtained as the integrated output of the composite sensor.

Specifically, the neural network is set to have a four-layer structure, for example, two intermediate layers, one input layer and one output layer each, and backpropagation is used as the learning algorithm, and learning is performed according to the program shown in Figure 2. It can be performed. Note that the inputs of the neural network are the measurement results and inference errors for each sensor as described above, and each value is unified by the following equation.

Ml 1= Ml /M□8 (7)E
, ”=E-,,/E, ('8) Note that M,,E, is the measurement result and inference error of sensor i, and M@&X (Em i. is the measurement result of all sensors. The maximum value and the minimum value of the estimation error.

According to the measuring device with the above configuration, when a composite sensor created using a large number of multiple sensors is used, a neural network with learning data constructed by learning outputs measured values at each measurement point with high accuracy. Therefore, mounting a plurality of sensors has an extremely excellent advantage in that there is no need to take into account variations in position.

In addition, the characteristics of the individual sensors installed in the composite sensor used,
Even if there are differences in the characteristics of each composite sensor, it is possible to replace the entire system by using a corresponding sensor characteristic recording section and by using the corresponding learning data in the neural network. It is possible to quickly change the use of the composite sensor and use it as a measuring device for various different measurement targets without requiring changes to the usage program. This has the extremely advantageous effect of being able to deal with the limitations of one sensor with a system that basically has a common basic configuration.

The present invention is not limited to the embodiments described above, and various modified embodiments are possible. These are briefly listed below.

Embodiment 2 FIG. 6 shows a sensor information integrated measuring device having a configuration in which the auxiliary sensor in the embodiment of FIG. 1 is omitted, and its basic configuration is the same as that in FIG. 1. In other words, the inference error obtained by fuzzy inference and the measurement information detected by each sensor can be integrated as the output of the measurement device.

In this case, if each sensor is substantially unaffected by the usage environment, or if the measurement accuracy is affected by temperature, etc., but the sensor is used in a state where the usage environment hardly changes. In this case, an auxiliary sensor is not required, and therefore, the measuring device having the configuration shown in FIG. 6 can be used to construct a device having approximately the same accuracy as that shown in FIG. 1.

Embodiment 3.4 Figures 7 and 8 show examples of sensor information selection type measurement devices. The cases in which the sensor is omitted are shown in each figure.

The feature of these examples is that, based on the inference error obtained by the fuzzy inference section 3, a selection section 6 is provided that selects the signal with the smallest error among the 11 constant values of the composite sensor I as the output of the measuring device. be.

As already mentioned, the fuzzy inference unit 3 that estimates the error range can be replaced with a means for estimating by multivalued inference, for example, using a storage unit in which the results of previously inferred results equivalent to fuzzy inference are written in a matrix. It is.

A measuring device with such a configuration omits a neural network, so it is systemically simple, and it does not require the creation of learning data, so it is effective in providing an inexpensive and highly versatile measuring device. be.

In order to confirm the effectiveness of the measuring device with the above configuration,
Using the simple measuring device shown in Figs. 6 and 8 constructed using two eddy current type distance sensors as an example, the following distance measurement test was conducted and the results are shown in Figs. 3 to 5. Ta.

Combined sensor: Eddy current sensor 1, 2...Sensor 1: Measuring range 0-2mm measurement accuracy
1μm Sensor 2: Measuring range 0~10mm Measurement accuracy 10μm The learning data of the neural network uses back propagation as the learning algorithm, and the following (1
) to (3), and according to the program shown in Figure 2, assemble the composite sensor to the tip of the working arm of a 5-axis industrial robot, and while operating the robot, follow the flowchart shown in Figure 2. Ta. That is, while changing the distance between the composite sensor and the object to be measured, the measurement information of each sensor was input into the computer via the A/D board, and learning was performed.

(1) Measurement range (2) Resolution of the A/D board (3) Installation of the two sensors is due to the location As a problem with composite sensors, as mentioned above, when multiple sensors are installed, variations in position (offset) may occur. Since this offset amount is an unknown amount, in order to confirm its influence, we used a test model 1 of a composite sensor with an offset of practically 0 μm. Test model 2, which is attached to the base in the state (this offset is considered "positive"),
and test model 3 with a "negative" offset in which sensor 1 was installed farther from the object than sensor 2 were tested.

Further, the sensor information selection type measuring device shown in FIG. 8 was also tested using three similar composite sensors of test models 1 to 3.

Furthermore, for comparison, instead of the fuzzy inference unit 3 and the neural network 4, the test model of the three composite sensors described above was used for a computer having a program based on the so-called if-close (binary theory) shown in FIG. The test was conducted.

The above test results are shown in Figures 3 to 5. Figure 3 shows the measurement results of test model 1, Figure 4 shows the measurement results of test model 2, and Figure 5 shows the measurement results of test model 3. It shows. In each of these figures, figure (a) is sensor 1
.
8 is a diagram showing the difference between the output value of the measuring device of FIG. 1 using a neural network and the output value of the sensor information selection type measuring device of FIG. 8. FIG.

Of these figures, as is clear from Fig. 3, in test model 1 where there is no offset in the assembly of sensors 1 and 2, there are 8 categories: if close (ifC) and neural network (N, N,). Based on the results of inference using fuzzy inference means (Fig. 8), there is no difference in the output values of the measured values of the device (F, 1.) and (N, N,). Therefore, if multiple sensors can be mechanically assembled to the target with extremely high precision and no offset, it can be seen that there is no problem in creating an SIS using if close.However, in reality, such It is extremely difficult to create a composite sensor that is strictly free of offset.

From the results in Figure 4, the output of 1f close and (N, N,
), whereas there is no difference between the output values (F, 1.) and (N, N,) of the measured values of the device in FIG. 8. Therefore, in a distance measuring device in which a composite sensor is configured with an offset in which the sensor 1, which has a small upper limit of the measurable range, is close to the measurement target,
It can be seen that whether it is an information selection type distance measuring device or an information integrated type distance measuring device, it is possible to perform excellent measurements that cannot be expected with an IF close type device.

From the results shown in Figure 5, there is a difference between the outputs of the if close (ifC) and the neural network (N,,N,), and the
The output value (F, 1.) of the measured value of the device and (N, N,
) It can be seen that there is also a difference in the output. In other words, in the created composite sensor, when there is an offset in which sensor 2, which has a large upper limit of the measurable range, is close to the measurement target, excellent measurement is only possible with the measuring device having the neural network shown in Fig. 6. I understand.

(Effects of the Invention) As described above, according to the measuring device of the present invention, the most appropriate information at each measurement point is selected or integrated from among the measurement information obtained by each sensor of the composite sensor. It can be obtained as information about the device, and as a whole, it simultaneously satisfies the contradictory requirements of excellent handling, allowing the composite sensor to be handled as a single sensor, measurement over a wide range, and high-precision measurement. It has the effect of being able to

In addition, the configuration of the measuring device of the present invention makes it possible to quickly and easily respond to composite sensors with different characteristics depending on the product by sharing the basic system outline, and when changing the usage of composite sensors. This has the advantage that it can be easily adapted without requiring any changes to the basic program.

[Brief explanation of drawings]

Fig. 1 is a diagram showing an outline of the configuration of a sensor information integrated measuring device according to the present invention - an example, and Fig. 2 is a diagram showing a neural network learning program in the measuring device.
Figure 3 (a), (b). (c) to Fig. 5 (a), (b), and (c) are diagrams showing the results of testing the characteristics of the measuring device according to the present invention;
FIGS. 6 to 8 are diagrams showing the general configuration of other embodiments of the measuring device of the present invention. FIG. 9 is a diagram for explaining the configuration of a fuzzy inference means, and FIGS. 10(a) to (g) are diagrams for explaining the machining of a fan blade, which is an example of the grinding process to which the present invention is applied. , FIG. 11 is a diagram explaining the cutting state, and FIG. 12 (a
), (b) is a diagram for explaining the configuration of the composite sensor, FIG. 13 is a diagram showing the relationship with an industrial robot when the measuring device of the present invention is used, and FIG. 14 is a diagram showing if close as control. It is a figure which showed the flowchart when adopted. (4 others) Fig. Test results (Offset time (sec) (b) N-N!・C data time (sec) (CIN -N F・] data time (y) Fig.・Soto 0) (Al sensor element 1 2 data time (sec) (b+N -N. Ic data (c) N-N F.! data chart test result (offset +) (b) N-N 1 -C data (CAN -N. F-1 data time (seconds) Fig. Integrated detection value Fig. Fig. (a) Large 7-inch, 'L Shishin 4 Tow 1 Ladle 77 Shift -h' BL

Claims (1)

[Claims] 1. A composite sensor configured by combining a plurality of sensors, and sensor characteristics that are given in advance for each sensor, each of which receives sensor measurement information measured by the plurality of sensors. as auxiliary information, a means for estimating the error range that can be included in each sensor measurement information, and a plurality of sensor measurement information measured by each of the above-mentioned sensors of the composite sensor, and a plurality of sensor measurement information obtained by the means for estimating the above-mentioned error range. a neural network that receives error range information as input and calculates the measured value at each measurement point of the composite sensor as an integrated value of all these inputs; What is claimed is: 1. A measuring device that integrates sensor information, characterized in that it has learning data constructed in advance using a learning algorithm as a calculation element. 2. The measuring device according to claim 1, wherein the means for estimating the error range that may be included in the sensor measurement information is means for performing fuzzy inference. 3. The measuring device according to claim 1, wherein the means for estimating the error range that may be included in the sensor measurement information is a means for performing multi-value inference. 4. In any one of claims 1 to 3, the means for estimating the error range includes an input section that receives auxiliary information used for estimation from the outside, and detects information about the environment in which the sensor group is used and inputs the information to the input section. A sensor information integrated measuring device characterized by having an auxiliary sensor that outputs. 5. The sensor information integrated measuring device according to claim 4, wherein the composite sensor is formed by assembling a plurality of distance sensors having different upper limits of measurable ranges on a common substrate. 6. In claim 4, the composite sensor comprises a plurality of distance sensors having different upper limits of measurable ranges assembled on a common substrate, and the auxiliary sensor is a sensor that detects temperature. A sensor information selection type measuring device characterized by: 7. The sensor information integrated measuring device according to any one of claims 1 to 6, wherein the algorithm for constructing learning data for the neural network is backpropagation. 8. A composite sensor configured by assembling multiple distance sensors with different upper limits of measurable ranges on a common board, and sensor measurement information measured by each of the above sensors are input, and the sensor measurement information given in advance for each sensor is input. A means for estimating the error range that can be included in each sensor measurement information using the sensor characteristics as auxiliary information, and a method for selecting the sensor measurement information corresponding to the smallest error range among the error ranges obtained by this estimating means. and selecting means for selecting an output at each measurement time point of the composite sensor, wherein each distance sensor in the composite sensor is configured such that a distance sensor with a higher upper limit of a measurable range is selected as an output than a distance sensor with a lower upper limit of a measurable range. A sensor information selection type distance measuring device, characterized in that it is assembled to a board in a positional relationship that is not close to a target. 9. The sensor information selection type distance measuring device according to claim 8, wherein the means for estimating the error range that may be included in the sensor measurement information is means for performing fuzzy inference. 10. The sensor information selection type distance measuring device according to claim 8, wherein the means for estimating the error range that may be included in the sensor measurement information is a means for performing multi-value inference. 11. In any one of claims 8 to 10, the means for estimating the error range includes an input section that receives auxiliary information used for estimation from the outside, and an input section that detects information about the environment in which the composite sensor is used. A sensor information selection type distance measuring device characterized by having an auxiliary sensor that outputs. 12. The sensor information selection type distance measuring device according to claim 11, wherein the auxiliary sensor is a sensor that detects temperature.