JPH087684B2 - Real-time expert system using network-type knowledge representation - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Field of Industrial Application) The present invention is applicable to a wide range of fields such as process control in plants such as petroleum, chemicals, and steel, CIM, communications, logistics, financial transactions, and the like. Status monitoring, abnormality / fault diagnosis, driving support,
The present invention relates to a real-time expert system used for automatic control, dealing support, trader support, decision support, etc.

(Prior Art) As an application of an expert system, intelligent production equipment has been proposed centering on the manufacturing industry, and its practical application has been attempted from early on. Of these, the real-time expert system is receiving particular attention. A real-time expert system constantly monitors changes in a dynamic world where the state changes momentarily, makes inferences while capturing a large amount of fluctuating data in real time, and outputs necessary processing and advice depending on the situation. It is a thing.

Traditionally, as a real-time AI tool, Intelligent
Control's “PICON (Process Intelligent CONtrol sy
stem) "and" G2 "from Gensysm. "PICO
"N" is a dedicated tool specialized in the field of process control and is a pioneer in online real-time expert systems. "G2" follows the same flow as [PICON], but the modeling function of inference target, simulation function, man-machine interface function, etc. are enhanced.
These conventional AI tools employ frames and production rules as knowledge representation methods. Frames are used to describe static information about an object, and if-then format production rules are used to describe human empirical knowledge and object model expressions.

A conventional method will be described with reference to an example of a plant operation support expert system. The plant operation support expert system aims to provide various types of guidance and support to operators on site. Basically, the system development is connected to the actual plant and closely related to the field work. The development procedure goes through the steps of collecting knowledge, organizing knowledge, establishing knowledge base, establishing system, and maintaining knowledge base. At the stage of, the person in charge of development collects the operating know-how of the target plant and analyzes the process, and elucidates various events leading up to the abnormal situation of the plant and the countermeasures accompanying it. First, a plant engineer conducts brain storming, lists major abnormal items in the target plant, and analyzes causal relationships between various events leading to each abnormal situation. The hearing from the plant operator will confirm how the abnormal situation clarified here is recognized in the actual site and how the countermeasure is implemented. At the stage of, the complicated causal relationships between the events collected and analyzed at the stage of are organized. FIG. 2 shows an example of analysis created in this way. The organized knowledge is rewritten as if then rules to form a knowledge base. The constructed knowledge base is checked for operation by the simulator. The completed expert system will contact the plant control system for system testing, and then maintain the knowledge base while continuing practical operation.

(Problems to be solved by the invention) Used in the conventional expert system described above.
The knowledge expression method using if-then format production rules is easy to add and modify individual rules, and the rule expression itself is highly flexible. However, in reality, there is a large gap between the causal relationship analysis method by the field experts and the expression method by the production rules, so it is almost impossible for the field experts to understand the content as it is. It was possible. In addition, since it is difficult to understand the relationships between rules in the knowledge expressed in production rules, it is very difficult to verify and maintain the entire knowledge base when the knowledge base has thousands of rules. There was a problem that became.

According to the present invention, a field expert himself can build a knowledge base, maintainability is maintained even for a large-scale knowledge base, and the person who operates and manages the application can easily modify the knowledge base. Large-scale real-time expert system development using changeable knowledge representation method,
The purpose is to operate and maintain.

[Structure of the Invention] (Means for Solving Problems) The present invention captures knowledge as a causal relationship between various events, and expresses each event in a network form in which the causal relationships are connected as nodes. A knowledge base is constructed by converting the internal code that represents the causal relationship between the pointers between node objects that perform the same function as if-then rules inside the computer, and the execution of specific rules in these knowledge bases. It is a system in which the inference mechanism operates in real time with a function of performing bidirectional automatic inference in real time by designating a cycle and preferentially inferring and executing a rule group related to an emergency event by an interrupt processing routine.

Since the knowledge is organized in a network format, there is little gap between the problem analysis methods by field experts and the understanding is easy to obtain.Further maintenance work is done on the knowledge organized in a network format. It can be reflected in the knowledge base each time.

In addition, interrupt at any time during real-time inference,
The ability to modify or add knowledge bases makes it easy to create and maintain knowledge bases for large, complex real-time expert systems.

(Operation) A real-time expert system using the knowledge representation method in the network form of the present invention will be described.

The knowledge expression in the system of the present invention captures knowledge as a causal relationship between various events, and expresses each event on a screen in a network form in which they are connected by causal relationships as nodes. Most operations can be performed by operating the popup menu with guidance. Users don't need to know a language for AI such as Lisp
You can build a knowledge base just by learning basic mouse operations.

The real-time expert system of the present invention uses a network form as a knowledge base and a real-time inference function.
Fast forward / backward dynamic inference is performed at high speed. When an event that requires an urgent response occurs, the interrupt handling routine prioritizes and executes the rule group related to the event, and has a function that emphasizes real-time property.

In order to perform the maintenance work of the knowledge base easily and efficiently, it is designed so that the knowledge base can be modified and changed while executing the real-time inference. Knowledge such as real-time rule trace function that back racks related nodes that conclude specific rules and displays only related parts, time series display function of input online data, time series display function of success or failure of events It also has a base verification and maintenance support function.

FIG. 3 is an explanatory diagram showing an example of an operating environment of a network using the method of the present invention. The data needed for inference is
It is captured in real time by the network communication module via LAN and provided to the real time inference engine. As a result, at the time of inference, various data such as the database of the production management host computer and the process computer can be used.

The module structure of the knowledge network is shown in FIG. The knowledge base constructed by the knowledge base construction module is scheduled by the real time inference module and the inference is executed. The network communication module performs data communication with other devices connected on the LAN, collects real-time data necessary for inference, and transmits a message for displaying the inference result on other devices. The data simulator is used when performing a simulated inference while using a value desired to be generated by the simulator instead of the actual real-time data when checking the operation of the knowledge base.

The knowledge base construction procedure used in this system is
It can be roughly divided into three steps: object hierarchy, object visualization, and rule network construction. The user proceeds while using the corresponding submodule in the knowledge base construction module.

The "class editor" whose screen is shown in FIG. Information about the object of the knowledge base to be constructed is expressed as a frame, and its attribute is defined by positioning it in the overall hierarchical structure including the object. This target type is called a class. The entire hierarchical structure is expressed as a tree structure. The attributes of a class can be freely defined and have values by the user and are inherited by all child classes that have a parent-child relationship with the class.

Next, an "image editor" as shown in FIG. 6 is used to physically limit the scope of the target system of the knowledge base, and at the same time, to define the data source of the real-time data on which the inference is based. The object of the knowledge base to be constructed is represented as a schematic diagram using icons to define the physical model of the object. Schematic diagrams can be created by placing various icons and connecting them to each other. The example shown in FIG. 6 shows a general view of "NO.1-crude oil atmospheric distillation unit" which is the object of the knowledge base, and the sensor of the data source is defined by the sensor icon.

The information associated with each icon is defined as various attributes in the frame representation. Each attribute value can be set freely by the user. The inference engine executes the inference by referring to the attribute value of the icon object that is a real-time data source such as a sensor. Although non-standard icons are provided as an icon library for each application field, users can create their own.

Lastly, a chain of causal relationships of various states of the target system is expressed on the network using a knowledge editor as shown in the screen example of FIG. In the knowledge editor, first, each event corresponding to the condition part and the conclusion part of the rule is defined as a node (rectangular box). Here, the following six types of nodes were prepared to express the behavior of the target system.

Definition statement node Defines the status of the target system and intermediate events as variables.

Theoretical node Uses variables and intermediate event names defined in the definition statement node to form the conditional part of the rule as one node.

Logical symbol node Represents a logical symbol for forming the logical condition of a rule.

Message sentence node Displays the message of the inference result.

Action statement node Sends data to the outside and calls a user-defined function.

Comment sentence node Add a comment in the network.

When creating these node sentences, the words and phrases to be described are displayed in the pop-up menu, and the next word or phrase is selected from the candidate list displayed in the menu, or the event name, variable name, etc. are input from the keyboard. A rule is generated by connecting the nodes with arrows. The node at the base of the arrow represents the condition part of the rule, and the node at the tip of the arrow represents the conclusion part. It is also possible to connect rules in a chain, such as using a node in the conclusion part of a rule as a condition part of another rule. This network representation shows a list of network representations and their corresponding if-then rules in FIG. In the present invention, without outputting such a list or table of if-then rules to a file or a form, the causal relationship between pointers between the node objects having the same function as the if-then rule in the system is expressed. Convert to internal code.

The following shows an example of the internal code that represents the causal relationship between pointers between node objects. The following objects, object pointer # <FI1-HIGH>#<FEED-LINE-IS-ABNORMAL>#<SEND“~”>#<NOT>#<NORMAL-OPERATION> internal code # <internal code: if Expressed with # <> then # <> else # <>>. Here, “# <>” indicates an object.

Moreover, in order to maintain the real-time property of inference when the number of rules increases, the types of rules are classified hierarchically. It is divided into rules that constantly monitor the target state and rules that perform deeper analysis such as cause search. The former is called the primary rule and the latter is called the secondary rule or the tertiary rule depending on the depth of the hierarchy. By limiting the rules that should always be evaluated, the load of inference processing is reduced.

In the primary rule, the timing at which the inference engine executes the evaluation is specified as the cycle time (scan interval). Information about each rule, such as the rule group attributed to the rule and the certainty factor, is expressed as a frame. Table 1 shows a list of attributes described in this frame. The user can freely set these attribute values.

An inference mechanism for real-time processing will be described. The inference engine uses real-time data that reflects the state of the target system to recognize the state change of the target system and infers the cause of the state change. The main work of the inference engine is as follows.

△ Decide the rule to be evaluated at a certain time.

△ Evaluate the condition part of the rule and execute the conclusion part according to the result.

△ Output messages and operations to the outside.

△ When a event interrupt occurs, the specified rule group is the target of evaluation.

△ If the value of the sensor variable referenced by the rule is invalid, fetch new valid data from the database of the network server. At the time of simulation, the value is determined by the simulation evaluation formula.

△ Calculate the certainty factor of each event and propagate between rules.

The inference method is a bidirectional inference that accompanies periodic scanning of rules specified in advance. When there is a description that determines the value of another event in the conclusion part, all the subordinate rule groups that refer to that event in the condition part are subject to evaluation. In the evaluation of the conditional part of the rule, if the value of the event is unknown, the event is used as a hypothesis and backward inference is automatically performed to verify the hypothesis. If a reference to external data is required during backward inference, the latest value on the database of the network server at that time is referred to.

An example of a knowledge network is shown in FIG. In this figure, (Rule 1) “If A and B then C” (Rule 2) “If C or D then E” (Rule 3) “If E then G” (Rule 4) “If H and G then send” Four rules, "to OT-1" for Tokyo branch and "to OT-1" for Tokai branch, are generated.The inference engine periodically evaluates rule 3 specified as the primary rule. At that time, first, in order to determine the success or failure of the condition part “E” of the rule 3, the rule 2 is changed to the condition part “C” of the rule 2.
Rule 1 and the inference are repeated backward to determine
When the success or failure of “C” is determined in rule 1, inference is repeated positively with rules 2, 3, and 4 this time.

In this way, it is not necessary to specify forward / backward, just by specifying the rule to be the target of periodic inference and its cycle, the inference engine automatically performs bidirectional inference and outputs the execution result. To do. In other words, the user does not have to worry about the evaluation order of rules and the inference method when constructing the knowledge base.

One of the major features of this tool in real-time inference is the activation of inference by external interrupt. In periodic inference using primary rules, if the number of target primary rules increases or the number of related rules and rule groups in evaluation / execution increases, the time required for one inference cycle increases and There is a risk that it will not be possible to compensate for the execution of the rules that are executed at regular intervals.
In such a case, in order to maintain real time,
It is possible to activate a specified rule group using an external interrupt.

Figure 9 shows an outline of inference activation by external interrupt.
An external interrupt is specified by writing an interrupt condition expression of a real-time data source with an image editor and writing a group of rules to be processed preferentially at the time of interrupt. When the network server fetches real-time data, it evaluates the interrupt conditional expression, and if the condition is satisfied, it makes an interrupt request to the inference engine. The inference engine that received the request suspends the periodic inference up to that point and starts inference of the specified rule group. When the inference of the rule group is finished, the original periodic inference returns.

The real-time expert system of this example can proceed with inference while communicating data with a device connected to a LAN. The real-time data referred to during inference can be collected from an external computer via a local data station (LDS) by a network server, as shown in FIG.

The network server exists as one process of the system of this example, and exchanges real-time data and messages with communication devices on the LAN such as LDS and message terminals, and interrupts the inference engine.

The LDS is one of the communication devices connected to the LAN, and it collects data from an external computer on a regular basis and sends the data to a network server via a communication network.
Conversely, the data sent from the network server based on the inference result can be sent to an external computer. If you set the LDS communication interface appropriately,
Data can be collected by connecting to various external computers such as host computers, process computers, and plant control systems. It is also possible to display the inference result in a place other than the console through the data communication network.

Functions for maintaining the knowledge base in the real-time expert system of the present invention will be described. When constructing an expert system, it is an important point in operation to maintain the knowledge base that has been constructed by repeating verification, modification, and changes. In this example, in order to efficiently perform this maintenance work, the knowledge base is designed to be interrupted during real-time inference. Upon completion of the correction by the network editor, it is immediately converted into a rule expression inside the system, and further inference can be continued.

It also has a data simulation function, real-time rule trace function, and time-series data display function for verification and maintenance support of the knowledge base.

The data simulation function simulates the value for each data source defined in the image editor. The model formula for that is freely defined by the user. This feature allows testing of the entire knowledge base in a static data environment focusing on specific data sources and rules.

With the real-time rule trace function, as shown in Figure 10, when the user specifies a specific rule or message from the trace display or message of the inference rule displayed on the console, all the conclusions are made. The relevant node of is extracted and displayed only from the relevant part from the entire network. History information is stored for each event, and the history information is used to display whether the related rule is established or not established. By using this function, it is possible to check the causal relationship between rules.

In the time-series data display function, by selecting the icon of the data source or the central node of the rule with the mouse during inference, the time-series change of the data value of the data source is displayed as a trend graph. It is also possible to display a time series change of success or failure of the event in a trend graph.
By using this function, the consistency of the entire knowledge base can be checked dynamically.

(Field of application) The knowledge representation method of the present invention is used in a wide range of fields such as process control in plants such as petroleum, chemicals, and steel, CIM, communication, logistics, financial transactions, condition monitoring, abnormality / fault diagnosis, and operation. Used for support, automatic control, dealing support, trader support, decision support, etc.

[Effect of the Invention] Compared with production rules, the knowledge representation in the network form of the present invention makes it easier for a human to visually understand the overall structure and the causal relationship between rules, even in a complex and large-scale knowledge base. It is also an expression close to intuition, so you can organize and acquire knowledge while building a knowledge base. By making it easier to understand the whole, there is a merit that not only the construction of the knowledge base but also the maintenance work such as verification, modification / addition / deletion becomes easy. As a result, plant operators who are not experts in the expert system can fully understand the contents of the knowledge base, and the number of rules can be increased to several thousand while smoothly reflecting it in the opinions and knowledge base of the field experts. It has the effect of being able to develop, operate, and maintain a system with a large-scale knowledge base.

[Brief description of drawings]

FIG. 1 is an explanatory view showing an example of a screen in which the causal relationship between states and events of the target system is expressed in a network format by the knowledge expressing method of the present invention, and FIG. 2 is an example of analysis of causal relationship between events. Explanatory diagram, FIG. 3 is an explanatory diagram showing an example of an operating environment of a network using the method of the present invention, FIG. 4 is an explanatory diagram showing a module configuration of a knowledge network, and FIG. 5 is a screen example of a class editor. FIG. 6, FIG. 6 is a diagram showing an example of a screen of the image editor, FIG. 7 is an explanatory diagram of the relationship between the network representation and the list in the if-thne format, and FIG. 8 is a diagram showing an example of the knowledge network. FIG. 10 is a schematic diagram of inference activation by external interrupt, and FIG. 10 is a diagram showing an example of a screen when operating the real-time rule trace function.

Front page continuation (72) Inventor Takashi Nishiwaki 63-2 Higashikawashima-cho, Hodogaya-ku, Yokohama-shi, Kanagawa (56) References JP-A-63-127331 (JP, A) JP-A-63-20529 (JP, A) No. Proc. Of the 11th Knowledge and Intelligent Systems Symposium (1990-3-13) 1-6 “Instrumentation” Vol. 32, No. 5 (1989-5) P. 21-26 "Meiden Shiki" No. 211 (1990-4) P. 42-46

Claims (1)

[Claims] 1. Means for expressing the knowledge of an expert system in a network form in which events are connected as causal relations as nodes, and the causal relation expressed in the network form as an internal code representing the causal relation between pointers between node objects. A means for converting, a means for performing automatic bidirectional inference using the internal code in real time, a means for preferentially executing a rule group related to an emergency event by an interrupt processing routine, and an interrupt in the middle of the reasoning to interrupt the network format. A real-time expert system characterized by means for changing or adding knowledge.