JPH0573522A - Neural network and its structuring method, and process control system using neural network - Google Patents

Abstract

PURPOSE:To obtain the optimized neural network and its learning and recollection system. CONSTITUTION:The neural network 2 is constituted by arranging neurons corresponding to variables of respective layers of the qualitative cause and effect network of an object process and coupling neurons in adjacent layers which have corresponding variables in cause and effect relation. Variable (a) and (b), and a variable (5) are in AND relation, so neurons (1) and (2) and a neuron (5) in a following layer are coupled by synapses (a) and (b) and neurons (3) and (4) and a neuron (6) are also coupled. The both (subnetwork), however, are not coupled since they do not have cause and effect relation. Consequently, the coupling by synapses is greatly decreased and the optimized neural network which is improved in both the processing ability and precision of learning and recollection is provided.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network, and more particularly, to a high-speed learning / recall suitable for real-time processing and a neural network faithfully reflecting a qualitative model (qualitative causal network) of a recall target. Regarding application.

[0002]

2. Description of the Related Art In recent years, a typical neural network introduced in many documents has a multi-layered structure of an input layer, an output layer and an intermediate layer (hiding layer), and each neuron between adjacent layers is a synapse. Are connected to each other by. The learning method is called the error backpropagation learning method,
The learning pattern (teacher signal) is given to the input of the neural network, and the ideal value of the output signal at that time (teacher signal)
This is a method of adjusting the degree of coupling (weight of synapse) so as to minimize the deviation with respect to.

That is, after the product of the input layer neuron output value (input pattern value of the teacher signal) and the weight of the synapse is calculated, the sum of them is calculated, and a nonlinear function (for example, a sigmoid function) for the sum is calculated, and the intermediate layer is calculated. Find the neuron output value. After calculating the product of the output value of this hidden layer neuron and the weight of the synapse, the sum of them is calculated,
Calculate the non-linear function (sigmoid function) for the sum,
Output layer neuron output value is obtained. Then, a known output pattern (teacher signal) having a pair of relationships with the input pattern is compared with the output value of the output layer neuron to calculate a synapse weight correction amount that minimizes the deviation, and this result is used. To a new weight.

After the above procedure is repeated to complete the learning, the synapse weights are fixed and the neural network is transplanted to an actual device. The operation of the neural network in the actual device is called recollection, and the recollection target, for example, the output information of the process can be obtained even for unknown input information that is not in the teacher signal. Recall is that after calculating the sum of products of input information and synapse weights, the output value of the hidden layer neuron is obtained by the nonlinear function operation, and similarly, the output layer neuron value, that is, recalled, is obtained by the output value of the hidden layer neuron and the synapse weight. Get results.

The above-mentioned prior art is detailed in, for example, "Pattern recognition and learning algorithm" (Sentence General Publishing). A concrete circuit configuration of the neural network is described in, for example, Japanese Patent Laid-Open No. 201764/2017.

The relatively effective application of such a neural network is, for example, stock price prediction and pattern recognition. These are to learn a certain event, that is, a change in stock price over time and a feature of image distribution, and perform prediction and recognition. However, when applying a neural network to process control and various diagnoses, it is necessary to deal with a wide variety of parameters and actuators. If these are directly applied to the input / output of the neural network, the convergence of learning and the speed of recall can be improved. In many cases, both sex cannot be put to practical use. As described above, the neural network currently has great difficulty in application to complicated problems.

[0007]

The problems of the above-mentioned prior art will be described by taking the pattern recognition of the numbers in FIG. 18 as an example. In the neural network shown in FIG. 9A, an image is divided into 64 as an input signal, and the corresponding input layer has 64 neurons. Twenty pieces are assigned to the middle-layer neurons. One output neuron is assigned to each number 0-9. According to this neural network, learning converges in a shorter time and recall accuracy is higher than that in FIG. However, since the number of neurons becomes large, the speed of recall becomes slow.

On the other hand, the neural network shown in FIG.
The same image is divided into 16 parts, 16 neurons in the input layer, 10 neurons in the intermediate layer, and 10 neurons in the output layer. According to this neural network, since the number of neurons is smaller than that in (a), the recall is faster, but the convergence of learning and the precision of recall are reduced. That is, (1) When the total number of neurons becomes large, the product-sum calculation becomes enormous, and the processability (processing speed) is greatly reduced.

(2) Since the number of neurons decreases, the number of processing elements decreases, so that the convergence and the accuracy of recall during learning decrease.

As described above, in the conventional technique, there is a contradictory problem between the processability and the accuracy, which makes it difficult to apply the neural network to a real problem.

Various proposals have been made to solve the problems of the prior art. For example, JP-A-64-
The publication No. 82133 describes that a neuron that has been calmed down during the learning process is initialized and re-learned, thereby improving the accuracy by several percent. Further, JP-A-1-248268 discloses that
It has been proposed to improve the processing speed of the neural network propagation processing by using a plurality of parallel arithmetic circuits and pipeline processing.

[0012] However, all of these solutions are based on the efficiency or speedup of information processing, and do not give an essential solution to the above-mentioned problems of improvement in processability and accuracy. ..

Recently, studies have been conducted to consider the brain as an assembly of various subsystems, to elucidate the contents of the constituent subsystems, and to grasp the essence of the brain.
This seems to be useful when trying to optimize the network structure as much as possible when modeling the brain with a neural network. However, at present, no effective partitioning method for subsystems has been proposed.

An object of the present invention is to solve the problems of the above-mentioned conventional method and to provide an optimized neural network configuration and its construction method that enable improvement of processing speed and processing accuracy.

Another object of the present invention is to provide a learning method of a neural network which can converge early without deteriorating the processing accuracy.

Yet another object of the present invention is to provide a process control method and apparatus which is one of the preferred applications of the optimized neural network.

[0017]

The present invention was made by paying attention to the following properties of an object. For example, in a process (whether natural or not), there is a direct causal relationship between a large number of quantitative or qualitative process information that fluctuates according to its behavior. Some are not. In other words, it cannot be said that not all the events represented by information are organically connected in the process, but the causal relationship between the events is accumulated to determine the overall movement. The causal relationship in the present invention means that an event at an arbitrary step of the process directly contributes to an event at a subsequent step,
An event has a relationship between a cause (or part of the cause) and an effect (or part of the result).

The present invention has a basic structure of a sub-neural net (hereinafter referred to as a sub-net) whose unit is a direct and qualitative causal relationship between events which are treated as variables in the object. By stacking the nets,
It is characterized by eliminating qualitatively meaningless connections and constructing an optimized neural network.

The present invention is based on a qualitative causal network which is one of the description methods of a qualitative model, and defines a sub-net, that is, between directly connected neurons by a qualitative causal relationship between the variables, The neural network is constructed by sequentially expanding this to the neurons of each layer.

The neural network learning method of the present invention is characterized in that processing is first performed only between neurons whose connection relation is defined by a qualitative causal relation.

The neural network recall method of the present invention is characterized in that processing is performed only between neurons connected by a qualitative causal relationship.

The process control method of the present invention is characterized in that the qualitative prediction of the process behavior is performed by using a neural network into which a qualitative causal network is transplanted instead of the fuzzy theory, for example.

[0023]

According to the neural network of the present invention, the division unit of the neural network is clarified based on the a priori qualitative causal network of the input information and the recall target, and the sub-nets having a causal relationship are accumulated. Therefore, the configuration is optimized and the processability and accuracy can be improved at the same time.

Further, since synapse connections of uncorrelated neurons, that is, asphyxia synapses, are removed from the beginning, there is no redundant operation in the learning of the neural network and the learning can be converged in a short time.

Further, according to the process control method of the present invention, since the optimized neural network that behaves as if a qualitative causal network of the process is transplanted is used for the qualitative inference, the control system can be easily constructed and Highly accurate prediction and its real-time processing can be realized.

According to the optimized neural network and its learning function of the present invention, the conventional fuzzy reasoning can be replaced with high accuracy, and the membership is a bottleneck in applying the fuzzy theory to a complicated process. It eliminates the need for function tuning and inference rule definitions.

[0027]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to FIGS. FIG. 2 shows a basic configuration of an apparatus for constructing, learning and recalling a neural network of the present invention, which is usually realized as an expert system. This apparatus includes a central processing unit that performs calculation and control according to a program and a computer apparatus 10 that has a memory that stores programs and data, a man-machine apparatus 11 that inputs instructions from an operator and displays processing results, an input signal or learning. The input device 12 for inputting a teacher signal for input and the output device 13 for outputting a recall signal. The computer device 10 includes a defining unit 15 for defining a neural network, a learning / remembering unit 16 for executing learning or recollection, a storing unit 18 for the defined or learned neural network, and these devices or units directly or through an interface. An arithmetic control unit 17 for controlling via 19 is provided. In the above embodiment, the means 15, 16 and 17 are realized by a program, and the neural network is constructed by storing the neurons, the connection relations of the neurons and the weights of the synapses in a file. Of course, these may be configured by known hardware.

FIG. 3 shows a configuration in which the learning / remembering means 16 functions as a learning means. The learning means 16 is input with the learning pattern by the control means 17, and the memory 1
Based on the neural network stored in FIG. 8, the non-linear function calculation of the product sum value of the neuron output value and the synapse weight is sequentially performed for each layer, and the recollection calculation unit 161 that outputs the output value h (x) and the known teacher Pattern d
It comprises a synapse weight calculation correction means 162 for minimizing the error between (x) and h (x).

In the present embodiment having such a configuration, the definition of the neural network is based on the qualitative causal network shown in FIG. 1A, and the synapse connection state of the neural network shown in FIG. Establish.

Here, the qualitative causal network means that a model such as a process is output from its input information based on a qualitative causal relationship between a large number of information (represented by variables in the description) representing an event. It is a hierarchical description of the connection relationship to information. Such a qualitative causal network is
It is used as a model for qualitative inference using empirical rules and fuzzy rules for objects for which it is difficult to express quantitatively using mathematical formulas, but the causal relationship between input and output is known to some extent. For example, an actual example thereof is described in detail in, for example, Japanese Patent Application Laid-Open No. 1-243102 by the applicant of the present application.

The qualitative causal network in FIG. 1A is composed of the first to third layers, and the variable set representing the event (information) is described by circled numbers ~. The input of the causal network corresponds to the input variables of the first layer, the output of the network corresponds to the output variable of the third layer, and the variable of the second layer is the output variable of the first layer. In the figure, the # number at the upper left of each variable is the consistent number of the variable for each layer. The qualitative causal relationship between each variable is
D), the variable, and the logical product (AND), the variable,
And the variable indicate the combination (COMB).

The neural network of the present invention defines a neuron ~ in a three-layer structure of an input layer, an intermediate layer and an output layer as shown in FIG. Only weighted synapses a to f connect only the neurons having the connection relation between the variables of the causal net. As a result, the input neurons in the input layer are connected to the neurons in the intermediate layer at the synapses a and b to form one sub-net. The input neuron is connected to the neurons of the intermediate layer by synapses c and d to form another sub net. The neurons in the middle layer are synapses e,
At f, it is connected to the neurons in the output layer to form another sub-net.

As described above, a neural network is defined so that a plurality of sub-nets in the same layer are arranged in parallel independently of each other and are connected in series with sub-nets in other layers. It is as if the causal network was projected into a neural network.

Next, the processing procedure for defining (constructing) a neural network from the qualitative causal network shown in FIG. 1 will be described with reference to the definition processing flow of FIG.

Step 1 (S101): Definition of variables of qualitative causal network Here, all variable Nos of the qualitative causal network (given by consistent numbers from the input side), variable names, etc. are set from the man-machine device 11, and 5 (a), store in qualitative causal network storage file.

Step 2 (S102); Definition of Qualitative Causal Relationship Using the variable No defined in S101, the minimum unit of the qualitative causal relationship between a plurality of input variables and one output variable is shown in FIG.
Set as shown in (b).

Step 3 (S103): Definition of Qualitative Causal Network From the contents of the file in FIG. 5B, the variable No that appears only in the input variable No. is the first layer, and the variable No that appears only in the output variable No. is final. Other variables are automatically assigned to each layer sequentially from the qualitative causal relationship to the layers, and the variable No. and its order for each layer shown in FIG. 5C are stored in a file to define a causal network. Since the variable Nos are assigned in the order of the consistent numbers, automatic definition is possible even when the variables are related across layers.

Step 4 (S104); Neuron Definition Here, based on the variable definition of FIG. 5A, the same number of neurons No (assigned with a consistent number) and their names are shown in FIG.
It is automatically generated as shown in (a). Variables and neurons are not always 1: 1 but 1: n (n = 1, 2,
...), but this will be described in another embodiment.

Step 5 (S105); Correspondence definition of neuron and variable S104 for all variables No set in S101
The neuron No. automatically generated in step 6 is automatically associated as shown in FIG.

Step 6 (S106): Definition of Synapse Connection State As the final step, the connection state of synapses between the neurons is defined by the flow shown in FIG. 7, and stored in the synapse connection state file of FIG. In step S201, all the synapse connection state files are initialized to the non-connection state. In S202, the following process is repeated from the second layer to the final layer (maximum layer), and in S203, the following process is repeatedly executed from the minimum neuron No. to the maximum neuron No. of the layer. S204 is a determination unit for the presence or absence of synapse binding,
It is determined whether or not the neuron is connected to the previous neuron based on the qualitative causal relationship file (step S102). In the example of FIG. 1, the second layer neuron No = (j
= 1) and the neuron No = (i = 1) of the first layer and the neuron No = (i = 2) have a causal relationship.
Of the second layer (i, j) file of i = 1, j = 1 and i
= 2, j = 1, and 1 is stored (S205).

The synapse connection state file stores the connection state between the first and second layers of the neural network on the second layer (i, j) file and the connection state between the second and third layers. On the 3rd layer (j, k) file, the final layer is prepared such as ... The number of neurons in each layer can be obtained from the file of the qualitative causal network defined in S103 and the correspondence between the variable No and the neuron No defined in S105. By the way, in the case of the qualitative causal network and the neural network shown in FIG. 1, the second layer (i, j)
In the file, n = 4 and m = 2, and the third layer (j, k)
In the file, m = 2 and l = 1. In the calculation of the degree of excitement of neurons in learning / remembering, which will be described later, this file is referred to and the calculation is performed only between connected neurons.

As described above, according to the present invention, the optimized neural network is automatically defined based on the variables of the qualitative causal network and the causal relationships between the variables, and does not have the wasteful connection that would die by learning as in the conventional case. Is configured. It should be noted that the network of FIG.
Since each file of S106 to S106 can be displayed on the man-machine device 11 on the screen, the operator can confirm the definition status.

Next, the learning method of the neural network according to the present invention will be described. Learning is based on the well-known error back-propagation learning method, and its outline will be described.

As shown in FIG. 3, the learning pattern set X =
_Given {x ₁ , x ₂ , ... X _n } and teacher information d (x) for x, the defined neural network performs learning as follows.

(1) Intermediate layer neuron output value calculation: After calculating the input layer neuron output value, that is, the product of the learning pattern value and the weight of the synapse, the sum of them is calculated, and the nonlinear function for the sum is calculated, and the intermediate layer is calculated. Find the neuron output value.

[0046]

[Equation 1]

Here, W _ij is a weighting coefficient of the synapse connecting the input neuron i and the intermediate neuron j, and is a real value that takes positive, zero, or negative values. x _i (i = 0 ... n) is the output value of the input layer neuron. y _j is the output value of the hidden layer neuron,
σ is a sigmoid function for obtaining the nonlinearity of the neuron output, and is represented by [Equation 2].

[0048]

[Equation 2]

(2) Output layer neuron output value calculation: After calculating the product of the output value y _i of the intermediate layer neuron obtained in (1) above and the synaptic weights connecting the neurons of the intermediate layer and the output layer, Is calculated, a sigmoid function for the sum is calculated, and the output layer neuron output value is obtained.

[0050]

[Equation 3]

Here, z _k is the output value of the output layer neuron k, and v _jk is a weight value of the synapse connecting the intermediate layer neuron j and the output layer neuron k, and is a real value that takes positive, zero, or negative values.

(3) Calculation of synapse weight correction amount: teacher pattern d (X) for a known input X (xεX) and output z _{k of the} neural network obtained in (2) above.
The synapse weight correction amounts δ _2k (X) and δ _1j (X) that minimize the error from (= h _k (X)) are calculated.

[0053]

[Equation 4]

Where α (X) is the degree of importance of the proximity of the output vector h (X) to the input vector X,
It is a known item.

[0055]

[Equation 5]

(4) Weight correction: The new weight v _jK and w _ij of the current synapse are calculated by using the result of the above (3).
Modify to _jK and w _ij .

[0057]

[Equation 6]

[0058]

[Equation 7]

The above is the error backpropagation learning method. The above process, particularly the number of times the product of the neuron output and the weighting coefficient shown in [Equation 1] and [Equation 3] is calculated, is enormous in the conventional neural network shown in FIG. become.

On the other hand, as shown in FIG. 9B, in the neural network of the present invention, the connection relation between the input layer and the intermediate layer or the adjacent layer such as the intermediate layer and the output layer is a finite subset of neurons ( Xa) and (Yb) and (X-
Xa) and (Y-Yb). Therefore, the calculation of [Equation 1] is also

[0061]

[Equation 8]

[0062]

[Equation 9]

The number of product operations is divided into [Equation 1], that is, the number of product operations is reduced to 1/2 as compared with that before the division in FIG.
Of course, the number of product operations is reduced to 1 / n if it is divided into n, so that the learning can be performed at high speed.

Next, it will be explained that the optimized neural network of the present invention is configured not only to reduce the connection relation but also to retain the property of the qualitative causal network (qualitative causal relation).

FIG. 10A is an example of a fuzzy inference model called MAX-MIN composition method described by a qualitative causal network. The reasoning method is a qualitative relation of AND in the conditional part,

[0066]

[Equation 10]

[0067] and, among the adaptability ty ₂ fitness ty ₁ and variables x ₂ variables x _1, and adaptability ty of the variable y the smallest.

On the other hand, the reasoning of the goodness of fit in the conclusion part is a qualitative relation of the combination (COMB),

[0069]

[Equation 11]

Then, the maximum one of ty ₁ and ty ₂ is set to the variable y.
Let ty be the goodness of fit.

This fuzzy inference model is excellent in modeling a process that cannot be grasped quantitatively, but it takes a long time to tune the membership function, which is a relational expression between the process amount and the goodness of fit. Makes it difficult to apply to complex processes.

On the other hand, as shown in FIG. 10B, in the learning of the neural network of the present invention, if the product-sum function for the process quantities x ₁ and x ₂ is S and the fitness of the output value is ty,

[0073]

[Equation 12]

It becomes Therefore, the above AND and COM
The relationship such as B is reflected in the synapse connection status (weighting coefficients w ₁ , w ₂ , ...) In the learning process, and as a result, the weighting coefficient at the time of convergence is adjusted to include these qualitative relationships. This means that the conversion form of the process quantity and the process control quantity is embodied as a neural network in which the weighting coefficient is adjusted to the optimum value, which is substantially equivalent to the transplantation of the qualitative causal network. ..

After the learning is completed, the weights of the synapses are fixed, and the neural network becomes a recollecting means of the real device. As in the embodiment of FIG. 3, the recollection device takes in input data via the input device 12 adapted to the input layer of the neural network and performs the same forward-propagation operation as in steps (1) and (2) of the learning described above. .. At this time, the product operation is performed by referring to the synapse connection state file. As a result, the recall output obtained in the output layer is the diagnostic result for the cause input or the future prediction for the current input,
The data is appropriately converted and output via the output device 13. Alternatively, it is displayed on the man-machine device 11.

According to the recall of the present embodiment, since the connection of the neural nets is optimized in terms of both processing speed and precision, both the recall processability and precision are improved. Furthermore, qualitative reasoning similar to fuzzy reasoning can be easily realized without using a membership function.

Next, a process control device according to another embodiment of the present invention will be described with reference to FIGS. FIG. 11 shows an outline of a tunnel ventilation process which is an example of an actual process. Tunnel ventilation control seeks and controls the operation amount M for dust collectors and jet fans by predicting the pollution amount VI of soot and smoke from process amounts such as traffic volume, and tries to maintain the pollution amount within a standard range. Is.

FIG. 12 shows the process control device 30 of this embodiment.
This is an example of applying the method to a tunnel ventilation process. The inference device 31 uses the input / output interface 32 and the neural network 3 optimized by the causal network.
3 and is connected to the process 40 via the process input device 34 and the process output device 35.

Displacement ΔTBt of the number of large trucks at time t (current) measured by the process quantity measuring machine 31
(Vehicle / 5 minutes), displacement of vehicle speed ΔTSt (km / h), displacement of pollution amount ΔVIt (%), displacement of traffic amount ΔTRt (vehicle /
5 minutes), the natural wind speed WNt (m / s), and the displacement ΔMt (m ³ / s) of the mechanical ventilation volume are periodically (5 minutes in this example) taken in by the process input device 34 (the displacement is the same as the previous value). Calculated by the difference). Each process amount is adapted by the interface 32 and input to the neural network 33.
In the neural network 33, the pollution amount displacement ΔVIt + 1 at time t + 1 is predicted by the recall operation of [Equation 1] to [Equation 3], and the process output device 35 controls so that the contamination amount is maintained within the standard. The operation amount of the device (variable amount from the current control amount) is obtained and output to the control device 42. The recall operation is performed according to the synapse connection.

Next, the qualitative causal network of the tunnel ventilation process and the neural network 3 based on it.
The configuration method of No. 3 will be described. FIG. 13 shows a qualitative causal network of the tunnel ventilation process. Each displacement (from t-1) of the number of large vehicles, vehicle speed, and pollution amount at the present time t has a causal relationship with the displacement of pollution amount in the tunnel. Also,
The displacements of traffic volume and mechanical ventilation and natural wind speed have a causal relationship with the displacement of total ventilation. Further, the displacement of the pollution amount in the tunnel and the displacement of the total ventilation amount have a causal relationship with the predicted displacement of the pollution amount after a certain time (t + 1). In this way, information that is a factor of the process and its causal relationship are obtained as a priori information from the causal network.

FIG. 14 shows an example of variable definition, causal relationship definition and qualitative causal network definition in S101 to S103 of FIG. 4 from this qualitative causal network. The figure (a) is 1-9 corresponding to each said process amount.
The variable No. and its name are stored, and the upper and lower limit values of each variable are also stored as appropriate. FIG. 2B defines the causal relationship of the minimum unit, and the variables 1, 2, and 1 of the first layer are defined between the first layer and the second layer.
3 and sub net of variable 7 of the second layer, variable 4 of the first layer 4,
It is divided into sub nets 5 and 6 and a variable 8 of the second layer. In the same figure (c), each variable is arranged in order of stratification and consistent number, and the structure of the qualitative causal network is defined.

FIG. 15 defines the relationship between neurons and variables and the synapse connection state in the definition flows S104 to S106. In the tunnel ventilation process, when the process amount is normalized by the interface 32 and input to the input layer neuron, each process amount is evaluated by three types, that is, μ + (increase), μο (maintain status quo), μ
Normalized by- (decrease), each is used as neuron input. The normalization is performed using, for example, a membership function (which is empirically known in advance) that represents the goodness of fit μ (value of 0 to 1) for each evaluation (+, 0, −) process amount. , Ordinary linear normalization may be used.

As a result, as shown in FIG. 7A, the displacement ΔTBt of the number of large vehicles is N-TBt, Z-T.
The evaluation values of Bt and P-TBt are input to neurons Nos. 1 to 3. In the present embodiment, as for the process amount of the second layer, neurons corresponding to three kinds of evaluation values are arranged respectively as in the input layer, but for the output layer, seven neurons corresponding to seven kinds of evaluation values are arranged. The neurons are arranged. That is, the predicted pollution amount displacement ΔVIt + 1 is μnb (significantly decreased), μnm (decreased), μns (slightly decreased), μz0 (currently maintained), μps (slightly increased), μpm (increased), μpb (significantly increased). From the viewpoint, each of the evaluations is performed, and the matching value μ is output to the neurons of Nos. 25 to 31. FIG. 6B shows a causal network variable No and neuron N.
The relationship between the variables and the neurons is 1: n (n = 1, 2,
...) can be defined. FIG. 7C shows a synapse connection state file. In the connection state of the second layer, the first layer neurons No. 1 to 9 are the second layer neurons No. 19 to 21.
(The second layer has consistent numbers 1 to 3), and the first layer neurons No. 10 to 18 have second layer neurons 22 to 24 (the second layer have consistent numbers 4 to 6). They are combined to form sub nets.

FIG. 16 shows a neural network constructed based on the above definition. In this example, since the space between the input layer and the middle layer is divided into two sub nets,
Synapses are connected only in the solid line. As a result, the number of 108 (18 × 6) connecting lines including the dotted line is 5 in the past.
It is halved to 4 (9 × 3 + 9 × 3), which can significantly improve the processability in the subsequent learning / remembering.

In the neural network initially defined in this way, the synaptic weight is determined by learning the teacher signal pattern by the measured data of the tunnel process, and after the qualitative causal network of the tunnel ventilation process is substantially transplanted, It is incorporated in the inference device 31 of the process control device. In addition, 7 from the output neuron
The two adaptive values are inversely converted from the fuzzy amount in the interface 32 to obtain the predicted pollution amount displacement. This inverse conversion is performed by the defuzzifying means 36, for example, by calculating the center of gravity when each matching value cuts on the membership function of each evaluation type.

In FIG. 17, the horizontal axis represents time and the vertical axis represents the transmittance TI (reciprocal of pollution degree VI) in the tunnel, and illustrates the operational relationship between TI and the ventilation control device. The thick solid line shows the measured value of TI, and the double line shows the predicted value of TI by the inference device 31. The target range of TI is 70% for the upper limit and 4 for the lower limit.
It is set to 0%.

TI actual measurement value at 16:50 at present (=
49%) is input and the TI is predicted to be 42% 5 minutes later (16:55), and the lower limit of TI may be interrupted soon after the current operating state is maintained. Therefore, the process output device 35 determines the power and the number of operating dust collectors and jet fans so as to be in accordance with the deviation between the present value and the predicted value, or to be about the intermediate value of the above target range (dotted line in the figure), and the control devices 42 in real time (16:50)
To control. As a result, the measured value at 16:55 is improved to 57%.

By optimizing the neural network, the effect of shortening the recall time increases as the process becomes more complicated and the number of input signals increases as in the present embodiment, and real-time predictive control is enabled. As a result, conventionally, it took 1-2 years in a complicated process for tuning a fuzzy control membership function, but it takes 1-2 days in learning the neural network of the present invention, and the same accuracy can be obtained. Be done.

[0089]

According to the present invention, based on the qualitative causal network which is the a priori information of the object, the minimum unit of the causal relationship between the variables is the sub-neural network, and the neural network is constructed by stacking the sub-neural networks. Therefore, the coupling of synapses is significantly reduced, and there is an effect that a neural network with improved learning / remembering processability and accuracy can be provided. Furthermore, since the connection status (weight of synapses) of the neural network is the one that substantially transfers the causal relationship of the qualitative causal network, qualitative inference equivalent to fuzzy inference can be realized, and membership like fuzzy Since it does not take a long time to tune functions, it has an effect that it can be easily applied to diagnosis and control of complicated processes.

[Brief description of drawings]

FIG. 1 is an explanatory diagram showing a basic concept for generating a neural network of the present invention from a qualitative causal network.

FIG. 2 is an embodiment of an expert system having means for constructing, learning and recalling the neural network of the present invention.

FIG. 3 shows an embodiment of the learning means shown in FIG.

FIG. 4 is a basic flow chart of a definition process of a qualitative causal network and a neural network.

5 is a definition file of the qualitative causal network in FIG.

6 is a definition file of the neural network shown in FIG.

FIG. 7 is a detailed flow for defining the synapse connection state of FIG.

FIG. 8 is a definition file of the synapse coupling state of FIG.

FIG. 9 is an explanatory diagram comparing a conventional neural network connection with a neural network connection of the present invention.

FIG. 10 is an explanatory diagram comparing the concepts of fuzzy inference and neural net inference.

FIG. 11 is an explanatory view of a tunnel ventilation process.

FIG. 12 shows an embodiment of the process control device of the present invention.

FIG. 13: Qualitative causal network of tunnel ventilation process.

FIG. 14: Definition file of qualitative causal network of tunnel ventilation process.

FIG. 15: Neural network definition file for tunnel ventilation process.

FIG. 16 is a configuration diagram of a neural network of a tunnel ventilation process.

FIG. 17 is an operation explanatory diagram of the control device of FIG. 12.

FIG. 18 is an explanatory diagram of a conventional neural network.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Qualitative causal network, 2 ... Neural network, 10 ... Computer system, 15 ... Neural network definition means, 16 ... Learning / remembering means, 17 ... Arithmetic / control means, 18 ... Neural network storage memory, 30 ...
Process control device, 31 ... Inference device, 34 ... Process input device, 35 ... Process output device, 41 ... Measuring device, 42
… Control equipment.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Masakazu Yahiro, 5-2-1, Omika-cho, Hitachi-shi, Ibaraki Hitachi Ltd. Omika Plant, Hitachi, Ltd. (72) Noboru Abe 5-chome, Omika-cho, Hitachi-shi, Ibaraki No. 1 stock company Hitachi Ltd. Omika factory

Claims (11)

[Claims] 1. A neural network comprising a plurality of layers including an input layer and an output layer, wherein the neurons are connected in accordance with the causal relationship of the recall target. 2. In a neural network composed of a plurality of layers, each of which has a plurality of neurons in adjacent front and back layers, information input to a front layer neuron and information output from a rear layer neuron A neural network that connects these neurons with synapses only when there is a causal relationship between them. 3. A neural network consisting of two or more layers, an anterior layer and a posterior layer, which are adjacent to each other, and a posterior layer neuron to which a predetermined input signal is input, and an output signal which is causally related to the input signal. A neural network that is combined with a sub-network formed by connecting synapses between layer neurons and all anterior neurons to which an input signal causally related to the output signals of all the posterior neurons are connected. .. 4. The neural network according to claim 3, wherein the sub-networks are connected in parallel between the same layers and in series with other layers. 5. A qualitative causal network that qualitatively expresses the process by a plurality of information (variables) indicating each event of the process and a causal relationship between the information (variables) is stored, and a neuron corresponding to the information is stored. And a neural network is defined so as to connect the synapses in correspondence with the causal relationship. 6. A neural network which stores a qualitative causal network and defines neuron arrangement and synapse connection in each layer including an input layer and an output layer in correspondence with information (variables) of the qualitative causal network and a causal relationship. A pair of the input information pattern and the output information pattern of the qualitative causal network is stored as a teacher signal, the input information pattern is given to each neuron of the input layer to execute the neural network, and the result of the execution is A learning method for a neural network, which corrects synaptic weights so that a deviation between an output signal obtained in the output layer and the output information pattern becomes small. 7. A neural network recalling method in which a signal indicating a recalling condition is input to a neuron in an input layer for recalling, and a signal indicating a recalling result is output from a neuron in an output layer, wherein the recalling refers to a synapse connection state. A method of recalling a neural network, which is characterized in that it is carried out for a connection. 8. A method of predicting a process state and controlling the process, comprising inputting a process state quantity to a neural network,
A process control method characterized in that a predictive value of a state quantity of a process is obtained by recollecting a neuron output only for a neural network in which neurons are connected, and a control quantity is determined based on the predictive value. 9. A causal relation definition for defining a qualitative causal network to be recalled by a variable and an input / output relation between variables, which is a processing device having an input / output means and a storage means for arithmetically controlling the construction of a neural network. Means for determining the arrangement of neurons for each layer from the defined variables and the input / output relationship, and for connecting the neurons having a relationship by synapse with reference to the input / output relationship between the variables. An apparatus for constructing a neural network, comprising: 10. An input means for inputting a learning pattern and a teacher pattern, which are a pair of data, a neural network for sequentially calculating output values of neurons by the inputted learning pattern, and outputting a calculation result, the calculation result and the teacher. In a learning device of a neural network having weight correction means for calculating and correcting synapse weights of the neural network so that the deviation of the pattern becomes small, referring to a pre-stored connection state of the synapse A learning device for a neural network, comprising: a synapse connection state determining means for executing the calculation of the neural network. 11. A process control device for operating a control effector while predicting a process state, wherein a process information input device and neurons are connected by a qualitative causal relationship of the process, and the process state is based on the process information. A neural network that infers a predicted value of the quantity, a synapse connection state determining unit that executes a calculation of the neural network when there is a connection by referring to a connection state of the synapse of the neural network that is stored in advance, and is predicted. A process control device comprising: a process output device that obtains and outputs an operation amount of the control effector from the state quantity.