JPH03134706A - Knowledge acquiring method for supporting operation of sewage-treatment plant - 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 control method for automatically acquiring state evaluation knowledge or operation knowledge of a sewage treatment plant, and further for properly operating the plant.

[Conventional technology]

In a sewage treatment plant, ■physical, chemical, and biological characteristics of the inflow water, ■state quantities of the plant, ■season and time are taken into account, ■evaluates the plant condition, and ■operates the plant based on this. are doing. Plant maintenance and management means performing this evaluation and operation (or control) appropriately. However, due to differences in plant-specific shapes and treatment methods, as well as differences in the quantity and quality of inflow water, the relationship between is also not completely elucidated. For these reasons, evaluation and operation are performed based on intuition and experience accumulated by operators based on their past performance.

As a method of automating this evaluation and operation, physical, chemical,
Proposed methods include constructing biological mathematical models and utilizing simulations to perform proper driving, as well as driving support and control methods based on knowledge engineering and fuzzy theory that simulate operator driving operations.

[Problem to be solved by the invention]

No satisfactory method for simulation has yet been developed.

In addition, through extensive data analysis and interviews with operators, we have learned that in knowledge engineering, rules such as "if... then..." are created, while in fuzzy control, membership is a characteristic function that quantifies ambiguity. Although functions are created and rules for inference are created, these tasks also require experience and a great deal of effort. Furthermore, software maintenance such as adding, deleting, and improving rules, membership functions, and inference rules requires a great deal of effort.

Furthermore, in sewage treatment plants, plant operations are manually controlled based on the above ■■■■ by focusing on past experience and operational records and taking into account the current situation, which eliminates the need for complicated and continuous monitoring. It was essential.

[Means to solve the problem]

In addition to creating the aforementioned rules (including inference rules) and membership functions, we also use neural networks (neural circuit models) to easily add, delete, and improve them, and to achieve operations equivalent to operators. Take advantage of your own learning abilities.

The operator refers to the past operation history, takes into account various factors such as the physical, chemical, and biological characteristics of the inflow water, the state quantities of the plant, and the season and time, and predicts the plant state. We evaluate and operate the plant based on this. Neural networks are applied to perform these tasks artificially.

First, the neural network is applied to the physical, chemical, and biological characteristics of the inflow water at multiple points in the past, and the state quantities of the plant.
Learn each pattern of ■season and time, ■plant condition evaluation results, and ■plant operating status. The specific configuration of the neural network that executes these functions includes an input layer configured based on a neuron element model, and at least one
It has a hierarchical structure consisting of an intermediate layer, an output layer, a comparison layer, and a teacher layer, and the specific learning method is a known error backpropagation method. Next, the unlearned current or future ■■■ was input to the input layer of the trained neural network, and the output layer was made to output (predict) ■■. The plant is operated based on this predicted value.

[Effect]

A neural network can be used to calculate the above ■■■■■ at multiple points in the past.
By having the robot learn the following, it will acquire the same sense as an operator. The actual situation is that the distribution of the weighting coefficients of the neural network changes.

Next, by inputting unlearned ■■■ into the trained neural network, ■■ is predicted. This is a simulation of the method by which an operator performs ■■ by considering various past results. In this way, by using a neural network, evaluation and operation equivalent to that of an operator can be realized.

〔Example〕

The present invention focuses on the physical, chemical, and biological characteristics of inflow water of sewage treatment plants, ■state quantities of the plant, ■season and time, ■evaluated quantities of plant conditions, and ■plant operating quantities. ■
By having a neural network learn the past history of , knowledge about ■■■ and ■, knowledge about ■■■ and ■,
In addition to automatically acquiring the
This is to provide guidance.

The execution process includes (1) a learning process using a learning neural circuit model (learning neural network), (2) a knowledge acquisition process,
It consists of (3) knowledge diagnosis process, (4) inference process, (5) associative prediction process of trained neural network (prediction neural network), and (6) plant operation control process.

The configuration and operation will be explained using the embodiment shown in FIG. First, the flow of the sewage treatment process will be explained below. In the initial settling tank 9, a portion of suspended solids in the inflowing sewage is removed by gravity settling. The sewage that has overflowed from the initial settling tank 9 and the return sludge from the return sludge pipe 16P1 flow into the aeration tank 15. Control valves 17A and 17B are connected to the aeration tank 15 from the blower 17.
Air is supplied through the system, and the sewage and returned sludge are mixed and stirred.

The returned sludge (activated sludge) absorbs oxygen from the supplied air, decomposes organic matter in the sewage, and is led to the final settling tank 16. In the final settling tank 16, the activated sludge settles due to gravity settling, and the supernatant water is discharged. The activated sludge settled in the final settling tank 16 is pulled out, and a portion is discharged as surplus sludge by the surplus sludge pipe 16P2 and pump 16C2. Most of the remaining activated sludge that was not discharged is transferred as return sludge from the return sludge pump 16C1 to the return sludge pipe 16.
It is returned to the aeration tank 15 through P1.

Next, the measuring device will be explained. First sedimentation tank 9゜Aeration tank 1
5. Each final settling tank 16 has a measuring device of 9M.

15M and 16M will be installed. Measurement items include inflow sewage volume, suspended solids concentration, chemical oxygen demand, pH5 nitrogen concentration,
Ammonia concentration, nitrate nitrogen concentration, nitrite nitrogen concentration,
Phosphorus concentration, dissolved oxygen concentration, sludge interface height (SVI: Sl
udge Volume Index, etc.), as well as image information of microorganisms and floating substances.

Next, the configuration of the predictive operation control device 80 will be explained.

The predictive operation control device 80 is a computer system, and a flowchart of the processing is shown in FIG. 1 to facilitate explanation of the present invention. In addition, the part described as network in the figure indicates a large number of wirings (or information communication routes) corresponding to a solid line 702 in FIG. 4, which will be described later.

The history pattern data file 71F includes: ■Inflow water characteristics measured by measuring device 9M; ■Measuring device 15M.

Plant status quantities measured by 16M, ■Season and time, ■Plant status evaluation quantities that are part of the measured quantities of measuring instruments 15M and 16M, ■Pump 16C2, return sludge pump 16C1,
Blower 17, control valve 17A.

Plant operation amounts such as 17B and time series data of the above items 1 to 2 are stored.

The learning neural network 71 automatically receives from the historical pattern data file 71F or by an instruction 71S from the communication means 46, ■ selected data string D1 of inflow water characteristics, and ■ selected data of plant state quantities. Column D2
．． 04 node and time selected data string D3, (1) selected data string D4 of plant condition evaluation amount, and (2) selected data string D5 of plant operation amount are output, respectively. In addition,
Each data string actually includes a plurality of items, but in this embodiment, five items are represented by representative symbols D1 to D5 for ease of explanation. In the learning neural network 71, data Di
, D2. D3. D4. Learning is executed using D5, and execution results 71S1 and 7182 are output to the knowledge acquisition process 73 and the prediction neural network 72.

In the knowledge acquisition step 73, the knowledge regarding ■■■ and ■ or ■■■ and ■ is converted into symbols or words based on 71S1 and 71S2. In the knowledge diagnosis step 74, the knowledge obtained in the knowledge acquisition step 73 is transferred to the knowledge candidate base 60B or the knowledge base 60 according to instructions from the communication means 46 (not shown).
Store in A.

The inference mechanism 61 receives the knowledge base 60A and the knowledge base 60C input in advance, performs inference, and outputs a signal 61S to the operation control process 75.

On the other hand, the prediction neural network 72 selects data necessary for prediction from the history pattern data file 71F and uses it for prediction. The prediction result is displayed on the communication means 46 as a prediction signal 72S, and is also output to the operation control process 75.

The operation control process 75 receives the prediction signal 72S and the signal 61S, outputs the signal 75S, and operates the pump 16c2, the return sludge pump 16C1, the blower 17, and the control valves 17A, 1.
Controls plant operation amounts such as 7B. At the same time, the control target value signal 75S is displayed on the communication means 46, and the actual control amount is corrected as required by the operator 101's selection. The correction value is output again.

Next, the present invention will be explained in detail using FIG.

FIG. 2 and subsequent figures will be used to explain the detailed configuration and operation.

First, a method for storing data in the history pattern data file 71F will be explained. The history pattern data DI(0) to D5(0) with time=O are stored in the history pattern data file 71F. Repeat this until t=Q,
-1, -2, . . . data are stored sequentially. In this embodiment, a case of one hour will be explained as an example of the time interval, but
Implementation of the present invention is not restricted by this time setting.

Next, the learning neural network 71 in the learning process
The operation will be explained below. Learning neural network 71
Now, learning is executed by receiving data selected from the history pattern data file 71F. This data selection and learning method will be explained below using FIGS. 1 and 2.

As shown in FIG. 2, any time 1 for D1 to D5
=1□ as a reference, go back to the past and calculate t-work 1, t, -2,
First learn... Similarly, 1=12(tz≠t□
) based on t2-1. t2-2. ... and learns a total of nine pattern data. It is desirable that the q patterns be selected from past representative patterns. Learning is performed by dividing this group of data into input data and teacher data. As shown in FIG. 1, input N710 has an arbitrary time 1 (1=1□,tl,
), and Di(t-1) to D5(t-1) are input sequentially starting from time t. Here, D4(t-1), D5(t, -
Note that 1) is added. Below t=t-2
, t-3, . . . On the other hand, D4(t) and D5(t) are input to the teacher layer 750. Learning is performed at the input layer 710. This is performed using a learning neural network 71 consisting of an intermediate layer 720, an output layer 30, a comparison layer 740, and a teacher layer 750.

The signal processing method in the learning method will be explained below using FIG. The configuration and signal processing method shown in FIG. 3 are known, except for the input data and the setting method of the teacher M750 described above. That is, the configuration and signal processing method shown in FIG.
This is a known technique devised by Lhart et al., and details can be found in the literature (Parallel Distributed
Processing, HIT Press, Vol.
, 1° (1986)).

The configuration and operation of FIG. 3 will be outlined. In Figure 3, O is
This is a neuron element model 701 having a product-sum operation and a sigmoid transformation function, and a solid line 702 connecting O and O indicates that information is exchanged between the neuron element models 701. Here, each layer consists of a finite number of neuron element models, and all neuron element models in adjacent layers are connected. Although there may be a plurality of intermediate layers 720, this embodiment shows an example in which the number of intermediate layers is one for ease of explanation. In addition, in FIG. 3, an output layer 730, a comparison layer 7
40. There are two teacher N750s, D4(t) and D5(t), but a plurality of cases is illustrated as a general expression. The configuration shown in FIG. 3 is called a neural network (neural circuit model).

Next, basic operations of the neuron element model 701 will be explained with reference to FIG. Data D1 to D5 input to input layer 710
Collectively, each time series (assuming there are n in total),
As shown in FIG. 4, n variable values are written as Y□ to Yn. The operation (product-sum operation) of multiplying each of the input signal values Y1 to Yn by the weighting coefficient Wji and further adding these values is performed as (1)
Calculate by formula.

Here, Yi (1): Yi value of input layer (first layer), Wj
i (2←1): Weighting coefficient from the i-th variable of the input layer (first layer) to the j-th neuron element model of the hidden layer (second layer), Zj (2): Weight coefficient of the hidden layer (second layer) ) is the total input value to the j-th neuron element model.

In the neuron element model 701, the output value here is calculated using equation (2) (sigmoid transformation) depending on the magnitude of Zj(2).

Although the calculation content of equation (2) is nonlinear transformation as shown in FIG. 5, the same effect can be obtained even if linear transformation is applied. The calculated value Yj(2) is further sent to the output layer, and the same calculation is performed in the output layer.

Next, an overview of the calculation method using a neural network will be explained. The aforementioned variable value Y 1 (1) is input to the input layer 710 in FIG. 3, and this signal value is output to the neuron element model of the intermediate layer 720. In the intermediate M720 neuron element model, these output values yi(1) and the weighting coefficient W
The sum of products Zj(2) with ij(2←l) is calculated using equation (1), and the output value Yj(2) to the output layer 730 is calculated according to the magnitude.
is determined using equation (2). Similarly, intermediate fi 720
The output value Yj (2) is further calculated by the weighting coefficient Wij (3←2) between the intermediate layer (second calendar) 720 and the output layer (third layer) 730.
The sum of products Z j (3) is calculated using equation (3).

Here, Yi(2): value of the middle layer (second layer), Wji(
3←2): Weighting coefficient from the i-th variable of the intermediate layer (second layer) to the j-th neuron element model of the output layer (third layer), Zj (3): Weight coefficient of the output layer (third layer) This is the total input value to the j-th neuron element model.

Furthermore, the output value Yj(3) to be output to the output layer 730 is calculated using equation (4) depending on the magnitude of Z j (3).

In this way, the calculated value Yj(3) of the output layer is obtained. Yj(3) is 04(t)*, D5(t) in this example
*It is.

Explain the learning procedure using neural networks. First, in the comparison layer 740, the signal 730S of the output N730 and the teacher signal layer 7
50 teacher signal 750S. For example, the magnitude of D4(t)*, which is the output signal 730S, and D4(t), which is the teacher signal 750S, is compared. The weighting coefficients Wji (3←2) and Wj are set so that the deviation between the two is small.
Correct the size of i(2←]). Using this corrected value, calculate the equations (1) to (4) again, and the output signal 730
The magnitude of S and the teacher signal 750S is compared again and the weighting coefficient value is corrected again. This process is repeated until the deviation becomes less than or equal to a predetermined value. At first, the weighting coefficients are given randomly by generating random numbers, so the deviation is large, but the output signal value gradually approaches the teacher signal value. At this time, since only the weighting coefficient value Wji is changed in equations (1) to (4), the learning results are reflected in the distribution of Wji values. Note that this Wji value is used in the prediction process.

The method of correcting the deviation in this manner is called the error backpropagation method, and uses a known technique devised by Rumelhart et al. For details, see the literature (Parallel Di
distributed processing, HIT
Press, Vol. 1. (1986)). Although this learning method itself is well known,
The present invention particularly provides a plurality of different times (t=t1゜1, .1□
2...tq) are repeatedly learned, and analog values are used as teacher signals.

As a result, pattern grasping ability comparable to the operator's past experience is accumulated and stored in each weighting coefficient Wji of the neural network. This operation was made to have the same effect as the operator's past experience.

A method of learning historical pattern data at a plurality of times 1 = 1 time + t2 + tl + . . . t will be described below. A plurality of historical pattern data groups are learned by arbitrarily changing the time t. This day t1ttz+t3t...t( may be selected by the operator or automatically, and both will be explained below.

The operator's selection is a typical pattern of variable values Y~Yn that he/she would like to reflect in the operation later, or a pattern for an abnormality that he/she would like to refer to later. As a result, the neural network outputs results according to what it has learned, so this selection is important. Leaving the selection to the operator relies on the operator's empirical and comprehensive ability to judge data. In this case, the patterns to be learned are patterns at different times in the past, and a plurality of patterns are repeatedly learned. As a result, the neural net acquires a pattern recognition ability comparable to the operator's past experience. The operator sets the time by man-machine conversation via the communication means 46. for example.

Learn each of 12 patterns from the typical pattern of January to the typical pattern of December. In particular, when the inflow water quality is typical, the treated water quality (suspended solids concentration, organic matter concentration, etc.) measured by the measuring device 16M of the I & final sedimentation tank 16 is good (
Learn the pattern when the standard is at its lowest temperature). In other words, it learns what operations were performed when the processing was good.

On the other hand, when performing automatically, statistical analysis of the data string is performed in advance. In other words, the system uses statistical analysis to find the most frequently occurring pattern and considers it as a representative example of normal conditions, and then learns this pattern.On the other hand, it learns patterns that occur less frequently, considering it as a typical example of abnormal conditions. . In the embodiment shown in FIG. 1, there is only one neural network, but it will be more effective if the neural network is divided into a neural network for normal times and a neural network for abnormal times (unsteady times).

Next, in the knowledge acquisition step 73, based on 71S1 and 71S2, the above ■■■ and ■, or ■■■ and ■, or ■ and ■
convert each knowledge about into symbols or words. As an index for comprehensively evaluating the relationship between variables set in the input ff 1710 and variables set in the output layer 730, a causality measure Iji defined by equation (5) is used.

Here, m is the number of neuron element models in the hidden layer.

This value is the sum of the products of the weighting coefficients of all routes from the input layer to the output layer. ■If jj takes a large value, there is a high possibility that there is a certain causal relationship between the i-th input and the j-th output. Compute ji for all combinations of the input layer and output layer, and convert the results into Japanese as knowledge candidates.

In this conversion method, if ji is a positive value and large, the conversion is performed as follows.

Knowledge candidate: "If the i-th input value is large, the j-th output value is large." ■If ji is a negative value and large, the following will occur.

Knowledge candidate = "If the i-th input value is large, the j-th output value is small." Since it is unclear whether the obtained knowledge candidate is valid knowledge or not, in the knowledge diagnosis step 74, the knowledge Acquisition process 7
After determining whether or not the knowledge candidate obtained in step 3 is knowledge based on an instruction from the communication means 46 (not shown), it is stored in the knowledge candidate base 60B or the knowledge base 60A. - If the knowledge candidate stored in the knowledge candidate base 60B occurs multiple times, the number of occurrences is counted and when it exceeds a predetermined number (for example, two times), the operator 101 is contacted again by the communication means 46 and knowledge diagnosis is performed. Repeat step 74.

The inference mechanism 61 executes inference upon receiving the knowledge base 60A and the knowledge base 60G, and sends a signal 61S to the operation control process 75.
Output. The knowledge base 60C stores knowledge acquired by the operator 101 in advance through a conventional interview, and uses the knowledge base 60C for supplementary purposes. Furthermore, knowledge base 60
This embodiment can be implemented without using C. The inference mechanism 61 executes inference (forward or backward) based on production rules or fuzzy rules. In the case of fuzzy inference, a membership function is required, but in the present invention, the membership function value for fuzzy inference can be set using the magnitude of the Iji value. Since the inference method is a well-known technique, a description thereof will be omitted. As a knowledge base, what kind of inflow conditions and plant state quantities,
■What kind of plant condition evaluation quantity will be obtained at the time of season and time, and as a result, ■How to operate the plant to improve the treatment.

Next, the prediction process by the prediction neural network 72 will be explained. The configuration of the prediction neural network 72 is shown in FIG. As shown in FIG. 1, the prediction neural network 72 uses the learning results of the learning neural network 71, that is, 1 weighting coefficient values Wji (3←2) and Wji (2←
1) as signals 71S1 and 71S2. In FIG. 1, the learning neural network 71 and the predictive neural network 72 are shown separately for explanation of the processing flow, but it goes without saying that the same neural network may be used.

Next, the operation of the prediction neural network 72 will be explained using FIG. 6 and FIG. 1. As shown in FIG. 6, the prediction neural network 72 has a configuration in which the comparison layer 740 and the teacher layer 750 are removed from the learning neural network 71. In the prediction neural network 72, corresponding to the above ■■■,
This is to predict the unknown above-mentioned ■■. For this purpose, first, as an input layer pattern in the input layer 710, variable values Y i (Di (0) to D3 (0)
) and variable values Yi (Di(i) to D5(i), C-1, -2,...
) is input to the input layer 710. As shown in FIG. 1, this variable value Yi is (1) the current known ■■■, and (2) the past (t=-1, -2, . . . ) ■ to ■.

Please note that all of these are actual values or known data. Based on these values, the above (1) −
Execute the calculation of equation (4) and calculate the unknown ■Plant status evaluation quantity (treated water quality) o4t(o) and ■Plant operation quantity (return/
Excess sludge amount, aeration air amount) OS book (O) and output layer 73
Output to 0. Regarding the above-mentioned (2), guidance is displayed on the communication means 46, and regarding (2), the control signal 72S is output to the operation control process 75.

Next, the operation control process 75 will be explained below. In the operation control process 75, the OS*(O) signal 72S and the inference mechanism 61
Upon receiving the result signal 61S, first, the consistency between the two is checked. If the signal 72S does not contradict 618, the signal 75S is sent to the pump 16C2, the return sludge pump 16C1, the blower 17, and the control valve 17A as the target value of the plant operation amount.
, 17B, etc. On the other hand, if there is a contradiction, the means of communication 46
The operator 101 is notified of the change and corrections are made. Note that although the control frequency is every hour in this embodiment, this time unit can be set arbitrarily. Of course, the smaller the time interval, the better the prediction accuracy. If the set time interval (one hour in this embodiment) is too long to predict the amount of chlorine to be injected over a short period of time (for example, one minute), the prediction is made by mathematical interpolation. At the same time, the injection amount target value signal 75S of the target value (prediction result) is displayed on the communication means 46, and the actual operation is corrected according to selection by the operator 101 as necessary.

〔Effect of the invention〕

The effects of the present invention will be described below. Conventional control methods for sewage treatment plants involve analyzing large amounts of data, acquiring knowledge through interviews with operators, and creating inference rules.
Additions, corrections, deletions, improvements, etc. required a lot of effort. In knowledge engineering methods, the knowledge acquired was dependent on the subjective judgment of a particular operator. However, in the present invention,
Knowledge is automatically acquired from past performance data, and driving guidance and control can be performed using neural network predictions. Therefore, if the present invention is applied.

Operators can easily perform ``operations based on past experience and precedent'' with less effort. Furthermore, since learning can be performed at any time, it is possible to quickly follow changes in the situation, learn and control.

In particular, the present invention focuses on ■physical, chemical, and biological characteristics of inflow water, ■plant state quantities, ■season and time, ■plant status evaluation quantities, and ■plant operation quantities, and analyzes past ■■■
Since it is possible to predict (1) plant status table quantity and (2) plant operation amount based on the following, it is possible to perform operations equivalent to those of an operator.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram showing the configuration of an embodiment of the present invention, FIG. 2 is an explanatory diagram of a learning pattern, FIG. 3 is a configuration diagram of a neural network, FIG. 4 is an explanatory diagram showing a neuron element model, and FIG. The figure is a characteristic diagram showing signal conversion in the neuron element model, and FIG. 6 is an explanatory diagram explaining the prediction process. 9... First settling tank, 15... Aeration tank, 16... Final settling tank, 17... Blower, 16C1... Return sludge pump, 16C2... Pump, 9M, 15M, 16
M... Measuring instrument, 46... Communication means, 101... Operator, 71F... History pattern data file,
71... Neural network for learning, 72... Neural network for prediction, 73... Knowledge acquisition process, 74...
・Knowledge diagnosis process, 75... Operation control process, 61...
Reasoning mechanism, 710...input layer, 720...middle layer,
730... Output layer, 740... Comparison layer, 750...
・Teacher layer, 701...2nd factor Figure 6 input layer

Claims (1)

[Claims] 1. For (1) physical, chemical, and physical characteristics of inflow water, (2) state quantities of the plant, and (3) season and time, which change with time, (4) ) Evaluation quantity of plant condition, (5)
It outputs the plant operation amount, and includes a neural circuit model with a hierarchical structure consisting of an input layer, at least one intermediate layer, and an output layer,
), (2), and (3), and the neural circuit model learns the above (4) and (5) corresponding to the input as a teacher pattern, thereby obtaining the above (1), (2), and (3).
(4) A knowledge acquisition method for sewage treatment plant operation support, characterized by expressing knowledge describing the relationship between (5) in symbols or words. 2. Regarding (1) physical, chemical, and biological characteristics of inflow water, (2) plant state quantities, (3) season and time, and (4) plant state evaluation quantities that change over time, (5)
It outputs and controls the plant operation amount, and is equipped with a neural circuit model with a hierarchical structure consisting of an input layer, at least one intermediate layer, and an output layer, and has a neural network model that outputs and controls plant operation amounts at multiple points in the past. )(3), and the neural circuit model learns the above-mentioned (4) and (5) corresponding to the input as a teacher pattern, and the learned neural circuit model is used to input (1) which is expected at present or in the future. (2)(3)
A knowledge acquisition method for supporting operation of a sewage treatment plant, the method comprising: inputting the following information, outputting (4) and (5) corresponding to the input, and controlling the plant operating amount (5).