JPH04100152A - Parallel fuzzy controller - Google Patents

Abstract

PURPOSE:To obtain the fuzzy controller which is fast at low cost by providing >=2 CPUs, dividing the definition area of a membership function into parts as many as the CPUs and assigning the divided areas to the CPUs, and performing parallel fuzzy reasoning. CONSTITUTION:The definition area is divided and assigned to, for example, four CPUs 1 - 4. The respective CPUs 1 - 4 are stored with output variables and conclusion fuzzy values corresponding to only the assigned parts of the definition area of the membership function. The CPU group consists of one main CPU and many sub-CPUs. The sub-CPUs perform the calculation of the output variable membership function which requires the majority of the processing time of the fuzzy reasoning and gravity center calculation in parallel. Consequently, reasoning speed improvement which is proportional to the number of the sub-CPUs is obtained and processes of the main CPU and sub-CPUs are performed on a pipeline basis simultaneously, so that the reasoning speed is improved more.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a fuzzy control device that executes fuzzy inference in parallel to perform high-speed control.

[Conventional technology]

In fuzzy control, control rules are given using fuzzy rules as shown in FIG. Here, x, y are input variables, 2
is an output variable. The output variables are fuzzy set variables. Also, At, A2. Bl, B2. C1, C2
is a fuzzy value, that is, a constant of a fuzzy set such as greater or lesser. A fuzzy set is represented by its membership function. After tp is the conditional part, TH[! After N is the conclusion part. Each consists of multiple conditional clauses and conclusion clauses. Rule 1 is if X is A1 and y is B1 then 2 is C1
It represents a control rule that tells you to do something.

When an input variable is given, each rule is applied to it to perform fuzzy inference and obtain the membership function of the output variable. Taking the rule of FIG. 5 as an example, FIG. 6 shows the inference process by MAX-MIN synthesis.

In the fuzzy value membership function, the horizontal axis represents the domain, and the vertical axis represents the degree of membership normalized to 1.

First, the degree of suitability of each rule to the input values xo and yO is determined. The value of the rule conditional clause membership function for the input value at this time is taken as the fitness degree, and when there are multiple conditional clauses, the minimum value thereof is taken. In this example, the degree of conformance is 0.5 for rule 1 and 0.3 for rule 2.

Next, the MI between the goodness of fit and the membership function of the conclusion clause
N is taken to obtain the membership function of the output variables for each rule.

Then, by calculating the mutual MAX of the membership functions of each rule, the final membership functions of output variables for all rules can be obtained.

Finally, the center of gravity position zO of this output membership function is determined and used as the manipulated variable to be output.

FIG. 7 shows a flowchart of the fuzzy inference based on the MAX-MIN synthesis described above.

First, clear all output fuzzy variables.

Next, we calculate the goodness of fit for each rule, and use that value to calculate the membership function of the output variable.

Finally, find the centroid of the membership function of the output variable.

Here, in the membership function calculation of the output variable, MAX-
MIN synthesis is performed.

First, a membership function is obtained by taking the conclusion fuzzy value and the MIN of the degree of fitness.

Then, take the MAX of the output membership function calculated using the previous rules and set it as a new output membership function.

As shown in FIG. 8, the membership function is expressed as an array of function values for a domain. In this figure, the domain is 25
6, the degree of membership of the membership function is divided into 256. When performing MAX-MIN synthesis, calculations must be performed on this entire array. A detailed flowchart of this part is shown in FIG. In this example, the array index is i and the output fuzzy variable is OUT[il,
The conclusion fuzzy value is A[il, and the goodness of fit is K.

As you can see from this example, in order to perform MAX-MIN composition, all array elements of the membership function must be
It is necessary to calculate MIN. As a result, this part of the processing occupies most of the total computational processing of fuzzy inference. In particular, if the domain of the membership function is divided into smaller parts, the size of the array will increase, and the calculation time will increase proportionally.

Conventional fuzzy control devices include those that perform the above-mentioned fuzzy inference using software and those that perform it using hardware. Software-based methods have excellent programmability and can perform complex processing, but in general, the inference speed is slow (
, it is not suitable for high-speed fuzzy control. Further, if a CPU dedicated to fuzzy control is used, faster control becomes possible, but the system becomes very expensive. Although the method using dedicated hardware is the fastest, there are limits to the number of rules and the complexity of the rules, making it difficult to perform complex fuzzy control. Additionally, these systems generally tend to be very expensive.

[Problem to be solved by the invention]

When performing high-speed fuzzy control, use the dedicated fuzzy control C
Using a PU or dedicated hardware results in a very expensive system. On the other hand, fuzzy control using software using an inexpensive general-purpose CPU has high programmability and can handle complex control, but cannot achieve sufficient speed.

On the other hand, JP-A-63-113735. Japanese Unexamined Patent Publication No. 63-113
For 736, a technique has been proposed in which a set of fuzzy rules is divided into several groups and processed in parallel, but this does not improve the processing speed.

The present invention uses a large number of inexpensive general-purpose CPUs to provide a fuzzy control device with high speed and high programmability at low cost through parallel processing.

[Means to solve the problem]

As shown above, when performing fuzzy inference using software, the part that requires the most calculation processing is the calculation of the membership function of the output variable and its center of gravity by performing MAX-MIN synthesis. The amount of calculation in this part is proportional to the width of the domain of the membership function. A feature of the present invention is that by dividing this domain into the number of CPUs, parallel calculation can be performed very efficiently.

This situation is shown in FIG. In this example, the domain is divided and allocated to four CPUs. Each CPU stores only the portion of the domain to which the membership function of the output variable and the conclusion fuzzy value is assigned.

As shown in FIG. 2, the CPU group consists of one main CPU and many sub-CPUs. Each CPU has a local memory, and the main CPU and sub 020 can communicate with each other through message communication or shared memory.

FIG. 3 shows the calculation processing flow in the main CPU and sub CPU. The main CPU calculates the degree of suitability for each rule and broadcasts the value to the sub CPUs. Sub C
The PU receives the fitness degree from the main CPU, and the receiver calculates the membership function of the output variable with respect to the membership function domain it has. This part of the calculation is the same as the MAX-MIN synthesis flowchart shown in FIG.

The above process is repeated for the number of rules.

When the sub CPU finishes calculating the output variable membership function, it returns the moment and area of the partial area to the main CPU. The main CPU receives partial moments and partial areas from all sub-CPUs and calculates the center of gravity of the output variable. This process is repeated for the number of output variables. Here, sub CP
The partial moment Mi and partial area St of the output variable 2 calculated by U i are calculated as follows: Mi = Σ μ1 (z), where Di is the divided domain and μ1 (z) is the output variable membership function of that region. ) z Di Si = Σ μ1cz) Di is given by. Then, the output variable z(7) center of gravity 20 is
It is given by Σ Si.

The calculation processing performed by the main CPU consists only of determining the conformity of the rules and calculating the center of gravity from the divided moments, which is much smaller than the calculation processing performed by the sub CPU. Therefore, each sub CPU does not wait for processing by the main CPU and the overall processing speed does not decrease.

[Effect]

As shown in the flowchart of FIG. 3, the calculation of the output variable membership function and the calculation of the center of gravity, which require most of the processing time in fuzzy inference, can be executed in parallel by the sub CPU. As a result, the inference speed can be improved in proportion to the number of sub CPU0s. In addition, the main CPU and sub CPU
The processing of U is a vibe line processing in which processing is performed overlapping each other, and the inference speed is further improved.

Also, from the main CPU, only the conformance of the rule is sent to the sub-C
Passed to the PU, only the partial moments and areas of the divided output variable membership functions are returned.

Therefore, there is very little communication between CPUs, and communication costs do not become a bottleneck to processing speed.

When fuzzy control is processed in parallel by a plurality of CPUs, it is possible to consider a method of dividing by rule instead of dividing by the domain of the membership function as shown in the present invention.

However, since it is necessary to synthesize the output variable membership functions obtained by each CPU, the processing becomes complicated, and
This has the disadvantage that the amount of communication between CPUs becomes large.

The present invention eliminates these drawbacks and enables efficient parallel processing.

〔Example〕

FIG. 4 shows a configuration diagram of a parallel fuzzy control device consisting of one main CPU and four sub-CPUs as an embodiment of the present invention. The main CPU has an external local memory and an A/C for inputting and outputting observed values and manipulated variables from external control targets.
A D converter and a D/A converter are connected by a bus. Each sub-CPtJ has a small amount of local memory inside, in which data of divided membership functions and processing programs are stored. Each CPU is linearly connected by a communication link, and data from the main CPU can be passed through the sub-CPUs one by one and broadcast to all sub-CPUs. Similarly, data from the sub CPU is transmitted to the main CPU. each CP
The flowchart of the processing at U is the same as that shown in FIG.

This configuration has the advantage that the sub-CPLI does not require additional circuit elements such as external memory, and therefore has a very simple structure. Furthermore, simply by adding a new sub-CPU and connecting it via a communication link, the inference speed can be easily improved.

〔Effect of the invention〕

As described above, very efficient parallel processing can be realized by dividing the domain of the membership function into each CPU and assigning a receiver, and calculating the output membership function and its center of gravity in parallel. Furthermore, since a general-purpose CPU can be used, there is an effect that a high-speed fuzzy control device can be realized at low cost.

Furthermore, since the fuzzy control algorithm is implemented in software, it has the advantage of being able to adequately handle complex fuzzy control.

Furthermore, since the membership function is divided and assigned to each CPU, the amount of required data is reduced in proportion to the number of divisions. As a result, it has the effect of requiring less local memory to store it.

In particular, if you use a CPU with internal RAM, you can allocate all the necessary data and programs there.
Access to memory becomes faster, allowing faster processing. A further advantage is that the construction of the fuzzy controller is very simple, since no external memory is required.

[Brief explanation of the drawing]

Figure 1 shows the calculation of output variable membership functions using multiple CPs.
A diagram showing an example of division by U, Figure 2 is a configuration diagram of parallel fuzzy control, Figure 3 is a flowchart of parallel fuzzy control,
Fig. 4 is a diagram showing an embodiment of a fuzzy control device using the present invention, Fig. 5 is an example of a fuzzy rule, and Fig. 6 is a diagram showing an example of a fuzzy control device using the present invention.
-MIN Diagram showing the process of finding membership functions of output variables through synthesis, Figure 7 is a flowchart of fuzzy inference, Figure 8 is a diagram showing the conventional fuzzy value data structure, Figure 9 is a diagram showing membership functions calculated. This is a flowchart. Patent applicant Anyo Electric Manufacturing Co., Ltd.

Claims (1)

[Claims] It has a fuzzy rule in IF-THBN format, stores the fuzzy value membership function of the rule as a table of function values for its domain, and stores the MAX of the membership function of input variables and fuzzy values. - A fuzzy inference control device that calculates the membership function of an output variable from MIN synthesis, is equipped with two or more CPUs, divides the domain of the fuzzy membership function into the number of CPUs, and divides the domain into each CPU. A fuzzy control device that allocates and performs fuzzy inference in parallel.