A sewage treatment process fault monitoring method of a high-efficiency extraction dynamic feature width learning network

Abstract

The application discloses a sewage treatment process fault monitoring method of a high-efficiency width learning network for extracting dynamic characteristics, and is used for solving the problem of inaccurate monitoring results caused by the dynamicity and nonlinearity of sewage treatment process data. The application comprises two stages of "off-line training" and "on-line monitoring". The "off-line training" comprises the following steps: firstly, integrating normal data and fault data collected into training data, performing dynamic expansion on the training data, and then performing normalization processing; and then, establishing an off-line training model by using a width learning network. The "on-line monitoring" comprises the following steps: performing dynamic expansion on newly collected data, and then performing normalization processing; and then, realizing on-line monitoring by using the model trained in the off-line state. The application extracts the features containing nonlinearity and dynamicity information in the data by using the high-efficiency width learning network for extracting dynamic characteristics, so that the rapid and accurate monitoring of the sewage treatment process can be better realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data-driven fault monitoring, in particular to a fault monitoring technology for data with nonlinearity and dynamics of a complex treatment process. The present application is based on the specific application of a high-efficiency width learning network method for extracting dynamic features in the field of sewage treatment process fault monitoring. BACKGROUND

[0002] With the continuous construction of modern cities, the amount of urban sewage and industrial sewage is also increasing. Direct discharge of a large amount of sewage not only destroys the natural ecological environment but also causes more water resources to be polluted in nature, leading to a more scarce water resource. Therefore, effective treatment of sewage can not only alleviate the shortage of water resources but also maintain a green ecological environment. However, the sewage treatment process is extremely susceptible to interference and influence of many factors such as season, weather, and flow, and therefore, sewage treatment process faults occur frequently, resulting in substandard effluent quality, increased process treatment energy consumption, and economic losses. In order to ensure qualified effluent quality and reduce secondary treatment energy consumption caused by faults, process monitoring of the sewage treatment process must be strengthened. The data-driven fault monitoring method extracts feature information from the historical data set for fault monitoring, and the effectiveness of the monitoring depends on the data itself. Therefore, the data-driven fault monitoring method is more suitable for fault monitoring of sewage treatment processes with complex mechanism models. However, it is worth noting that the sewage treatment process is a complex physical and biochemical reaction process, and the process variables are severely coupled, and the process data presents strong nonlinearity and dynamics. Therefore, for the data-driven fault monitoring method, enhancing the feature extraction of the nonlinearity and dynamic characteristics hidden in the process data can further improve the monitoring accuracy of the network.

[0003] At present, deep learning has been widely used in process monitoring due to its strong nonlinear explanation ability, such as Deep Stacked Auto-Encoder (DSAE), Deep Belief Network (DBN) and so on. Although deep neural networks have strong nonlinear processing ability, it is very difficult to set and adjust the complex hyperparameters in deep neural networks. Most deep neural networks use back propagation algorithm to adjust network parameters, which often causes long training time, gradient explosion and gradient disappearance. Moreover, when new training samples are added to the training set, deep neural networks need to be retrained to update the entire model. Obviously, the complexity of network structure makes it very time-consuming to retrain the deep neural network model, which cannot meet the real-time monitoring requirements of the wastewater treatment process. The emergence of Broad Learning System (BLS) provides a solution to the time-consuming problem of deep neural network training and updating. BLS network calculates the pseudo-inverse by ridge regression algorithm, which converts the weight calculation process into simple matrix operation, so that the network weight can be quickly obtained. Moreover, when the monitoring accuracy of the network does not meet the actual industrial requirements, BLS can expand horizontally to increase the enhanced nodes to improve the monitoring accuracy of the model, without the need to retrain the entire network, which greatly saves the time cost of retraining the network.

[0004] However, the BLS-based method and the deep neural network-based method rarely consider the dynamics of data. At present, the dynamic feature extraction for process monitoring has also been deeply studied, such as Dynamic Principal Component Analysis (DPCA), Dynamic Partial Least Squares (DPLS) and Dynamic Independent Component Analysis (DICA) methods. However, the above methods cannot effectively handle the nonlinearity and dynamics of the wastewater treatment process at the same time. If the nonlinearity or dynamics of the data is ignored, it will lead to insufficient feature extraction of the data and the decline of the monitoring accuracy of the model. Therefore, the wastewater treatment process monitoring model established should have the ability to extract the nonlinear and dynamic features of the data at the same time. SUMMARY

[0005] In order to solve the problem of inaccurate monitoring result caused by dynamicity and nonlinearity of sewage treatment process data, the sewage treatment process fault monitoring method of a high-efficiency dynamic broad learning system (DBLS) extracting dynamicity characteristics is provided.The technology based on the augmented time-delay extended matrix provides a new idea for solving the processing of dynamicity data, which can better replace the original data to complete data analysis and dynamicity feature extraction by deep time delay to the original time series and dynamic expansion to the variables of the original data.Meanwhile, the nonlinear activation function of the broad learning system can also process the nonlinearity of the data.Meanwhile, the network uses the ridge regression algorithm to calculate the pseudo-inverse matrix, which can quickly update the weight and improve the updating speed of the model.When the monitoring accuracy of the network does not meet the actual industrial demand, the BLS can expand horizontally to increase the enhanced nodes, so as to improve the monitoring accuracy of the model without retraining the network.In this study, the dynamic broad learning system based on matrix expansion is used to process the dynamicity and nonlinearity of the data.Firstly, the current sample variable is combined with the previous sample variables to construct an augmented matrix with time-delay process variables, and then the matrix is normalized and input into the broad learning system to establish an offline model to extract the nonlinearity and dynamicity characteristics of the data.Subsequently, the new data at the current time is collected and the augmented matrix is constructed and normalized, and the online monitoring is carried out by using the offline state trained model to judge the running state of the current system.

[0006] The technical scheme and implementation steps adopted by the present application are as follows:

[0007] A. Offline training stage:

[0008] 1) A batch of normal data fault and fault data x fault are combined to obtain the training data matrix and label matrix as follows:

[0009]

[0010]

[0011] Wherein represents the label of normal data, and Y fault represents the label of fault.

[0012] 2) The current sample variable is combined with the previous l sample variables to construct the augmented matrix X of the training data X train with time-delay process variables as follows:

[0013] where m is the number of sample variables, n is the number of samples, and l is the dynamic order of the matrix expansion (in this embodiment, l = 1), X e R (n -l)×(l+1)m .

[0014] 3) Normalize X using the following formula:

[0015]

[0016] wherein, represents the (l+1)th m variable value of the nth sample, mean(·) and std(·) are the mean and standard deviation of the (l+1)th m variable, respectively, and the normalized data set is obtained.

[0017] 4) Input the normalized data to the width learning system to generate the construction formula of the feature nodes of the ith window.

[0018]

[0019] wherein, W ei is the weight matrix of the feature nodes mapped to the ith window, β ei is the bias vector of the feature nodes mapped to the ith window, and W ei and β ei are random numbers uniformly distributed in the interval [0, 1] generated randomly during initial training.

[0020] 5) wherein the feature nodes of the ith feature window p is the number of feature nodes of the ith feature window, and the feature nodes of the ith window are normalized to the interval [-1, 1] using the following formula.

[0021]

[0022] wherein, the normalized feature nodes of the ith window T i = [t1, t2,..., t p ],

[0023] 6) In order to solve the problem caused by the randomness of the network weights and biases, the obtained feature nodes of the ith window T i are represented in a sparse manner. The sparse matrix W s is obtained by the following formula:

[0024] ​​

[0025] where λ1 is the L1 regularization parameter. by sparse matrix W s Mapping, get the final i-th window feature node

[0026] 7) Repeat steps 4-6 to complete the generation of feature nodes of all feature windows, and connect these nodes to represent Z I = [Z1, Z2,..., Z i ]. Then use the feature nodes to construct the j-th enhanced node:

[0027]

[0028] where is a random weight and bias uniformly distributed in the interval (0, 1), and ξ is the activation function tansig.

[0029] 8) Repeat the formula in step 7 to generate all the required enhanced nodes and connect them to H J = [H1, H2,..., H j ].

[0030] 9) After obtaining the feature nodes and enhanced nodes, construct the width learning system, which can be represented by the following formula:

[0031] Y train = [Z I |H J ]W J

[0032] where Y train is the label matrix, W J = [Z I |H J ] + Y train is the output connection weight of the network.[·] + denotes the pseudo-inverse of a certain matrix, which can be obtained by the following formula:

[0033] [A J ] + = lim λ→0 (λ2I+A J [A J ] T ) -1 [A J ] T

[0034] where λ2 is the L2 regularization parameter, I is the identity matrix, and let A J = [Z I |H JThe output network weight W of the width learning system can be obtained j , so as to obtain the output matrix.

[0035] 10) When the accuracy of the model cannot meet the expected effect, increase the enhanced node to improve the nonlinear processing capacity, so as to improve the network performance. After increasing the enhanced node, the input layer is A J+1 , and the weight W is calculated according to the following formula: J+1 J+1 + J+1 , the calculation process is as follows:

[0036]

[0037] , A J+1 is an extended matrix after adding a new node, , W hJ+1 and β hJ+1 respectively represent the random weight and bias of the newly added node.

[0038]

[0039]

[0040]

[0041]

[0042]

[0043] 11) Perform step 10 until the network training accuracy reaches the desired requirement (the training accuracy expectation of this example is set to be greater than or equal to 95%), at this time, save the new training model parameters.

[0044] B. Online monitoring stage:

[0045] 12) Collect the sample data x new at the current time, and construct the augmented matrix

[0046] 13) According to the formula of step 3, normalize the extended to obtain the input data X new .

[0047] 14) Then, input X new to the trained width learning system, wherein the model parameters are obtained by offline modeling training, and the final output of the model is the label matrix ​​​If the label is 1, it indicates that a fault occurs here and an alarm is sent, otherwise it is normal.

[0048] Compared with the prior art, the present application uses a width learning system based on matrix expansion to expand the collected normal data and then performs offline training, and then uses the model parameters obtained by training to monitor online data. Unlike the known neural network-based analysis method, the method proposed by the present application can process the nonlinearity and dynamics of sewage data, extract features containing nonlinearity and dynamics, and the proposed features are more representative. The network structure is simple, the computational complexity is low, the update speed is fast, and it can be better applied to the monitoring of the sewage treatment process. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The BSM1 is defined for the overall view of the sewage treatment plant;

[0050] Figures 2-7 The monitoring chart for setting faults for Table 2;

[0051] Figure 8 The flow chart of the method of the present application. DETAILED DESCRIPTION

[0052] BSM1 (Benchmark Simulation Model 1, BSM1) is a benchmark simulation environment proposed by the International Water Association, which gives a basic structure of a sewage treatment plant, simulates the whole process from the input of sewage into the treatment plant to the final discharge of clean water, and can be used as a benchmark for a sewage treatment system. BSM1 is provided with five activated sludge reaction tanks (6000 m3) and a secondary sedimentation tank (4 m deep, 10 layers, 6000 m3), two of the five activated sludge reaction tanks are anoxic links and the rest are aeration links, see Figure 1 .

[0053] In real situations, sewage treatment plants will run continuously under different weather conditions, and sampling data will be significantly affected by weather interference, so BSM1 provides three test scenarios (sunny, rainy, and heavy rain). In this experiment, the closed-loop simulation was run for 14 days, i.e. two weeks, under three different environments, and sampling was performed every 15 minutes, a total of 1344 sample points were obtained.

[0054] The experiment selected two of the most common failures in wastewater treatment: sludge bulking and toxic shock. Sludge bulking failure is caused by the complexity of wastewater and insufficient dissolved oxygen in the biological system, leading to lighter, larger, and deteriorated activated sludge settling properties. Toxic shock failure is caused by the large-scale discharge of toxic wastewater, resulting in reduced microbial activity in the activated sludge and substandard wastewater quality. On the BSM1 simulation platform, the occurrence of different degrees of sludge bulking and toxic shock failures was simulated by adjusting the values ​​of the maximum heterotrophic bacteria growth rate μ_h and the heterotrophic bacteria decay rate b_h (where the normal value of μ_h is 4 and the normal value of b_h is 0.3). The variables used for monitoring are shown in Table 1, and Table 2 shows the failure settings. The monitoring performance was quantified using the fault alarm rate (FAR), miss alarm rate (MAR), and accuracy (ACC), as shown in the following formulas:

[0055]

[0056]

[0057]

[0058] Table 1. Variables used for monitoring

[0059]

[0060]

[0061] Table 2 Fault Settings

[0062]

[0063] The application process of this invention on the aforementioned BSM1 simulation platform is described in detail below:

[0064] Offline modeling:

[0065] Operational data under normal and fault conditions were collected using the BSM1 simulation platform. Each batch of sampled data contained 1344 samples with 13 variables. The normal data and fault data were denoted as x. normal ∈R m*n and x fault ∈R m*n , where m and n represent the number of variables (13) and the number of samples (1344), respectively.

[0066] Step 1: Transfer a batch of normal data x normal and fault data x fault By combining the data, we obtain the training data matrix. and label matrix Ytrain where y represents the label of normal data, y

[0067] represents the label of failure.

[0068]

[0069] where y normal represents the label of normal data, y fault represents the label of failure.

[0070] Step 2: Construct training data of the augmented matrix X train of the process variable with time delay as follows. Where l is the dynamic order (in this example, l = 1).

[0071]

[0072] Step 3: Preprocess the expanded new data, the steps are as follows:

[0073] Step 3.1: Calculate the mean and variance of the expanded data set X train , the formula is as follows:

[0074]

[0075] where x i,j represents the measurement value of the jth variable at the ith sampling time in the expanded training data X train .

[0076] Step 3.2: Standardize the data, the formula is as follows:

[0077]

[0078] Step 3.3: Reconstruct the standardized data into an N x J matrix x(t) ∈ R N×J .

[0079] Step 4: Train the training data matrix and label matrix after expansion and normalization using the width learning system to obtain the connection weight W.

[0080] Step 4.1: According to the input matrix, construct the kth feature node Z k , the formula is as follows:

[0081]

[0082] where W ek , β ek are the weight and bias, respectively, which are uniformly distributed random numbers in the interval [0, 1]. The feature node of the kth window (in this example, k is 10), p is the number of feature nodes of each feature window (in this example, p is 10).

[0083] Step 4.1.1: Normalization of the feature node of each window is as follows

[0084]

[0085] Step 4.1.2: The normalized feature node T of the ith window k = [t1, t2,..., t p ] is obtained, and the feature node T of the ith window is represented by a sparse matrix. The final feature node of the kth window is obtained from the sparse matrix W k , and the sparse matrix W s is obtained by the following formula:

[0086]

[0087]

[0088] Where λ1 is the L1 regularization parameter (in this example, 0.001).

[0089] Step 4.2: Repeat step 4.1 to complete the generation of all feature nodes, and connect all feature nodes to represent Z K , as follows:

[0090] Z K = [Z1,..., Z K ]

[0091] Step 4.3: Generate enhanced nodes H from feature nodes Z, and the mth enhanced node is represented as:

[0092] H m = ξ(Z k W hm + β hm )

[0093] Where W hm and β hm are randomly generated weight and bias matrices, and ξ is a mapping function.

[0094] Step 4.4: Repeat step 4.3, and connect all enhanced nodes to H M , as follows:

[0095] H M = [H1,..., H M ]

[0096] Step 4.5: Z is obtained by using the results of step 4.2 and step 4.4 K and H M , a wide width learning system is constructed. The wide width model can be represented as:

[0097] Y

[0098] = [φ(x train W e1 + β e1 ), …, φ(x train W ek + β ek ) | ξ(Z K W h1 + β h1 ), …, ξ(Z K W hM + β hM )]W

[0099] = [Z1, …, Z K |H1, …, H M ]W

[0100] = [Z K |H M ]W

[0101] wherein W = [Z K |H m ] + Y is the connection weight of the network, [·] + represents the pseudo-inverse of a matrix, which can be obtained by the following formula:

[0102]

[0103] wherein A K+m = [Z K |H m ], let K+M = J, λ2 is the L2 regularization parameter (2 is taken in the present example -30 ), I is the unit matrix, and Y is the label output matrix obtained by training.

[0104] Step 6: When the model precision is not high, the network structure of the wide width learning system is expanded horizontally by increasing the enhanced nodes to improve the nonlinear feature processing capability of the data and improve the model precision.

[0105] Step 6.1: After increasing the enhanced nodes, the input layer is A J+1 , which can be replaced by the following formula:

[0106] A J+1 = [A J+1 |ξ(Z k W hm + β hm )]

[0107] Step 6.2: New Input Layer A of the Network J+1 The pseudo-inverse calculation method is as follows:

[0108]

[0109] D = [A J ] + ξ(Z k W hm +β hm )

[0110]

[0111] C=ξ(Z k W hm +β hm )-A J D

[0112] Step 6.3: Add new connection weights W to the network after enhancing the nodes. J+1 for:

[0113]

[0114] Step 7: Execute step 6 until the network training accuracy reaches the desired level. At this point, save the new training model parameters.

[0115] Online monitoring:

[0116] Step 8: Collect sample data containing the current fault. Using the dynamic order l during offline modeling, refer to step 2 for... Matrix expansion yields x test .

[0117] Step 9: For x test Refer to step 3 to perform data normalization to obtain X. test .

[0118] Step 10: Expand the X test The data X is input into the BLS system. Within the BLS system, the network parameters obtained through offline training are used to process the test data X. test The network is mapped to feature nodes and augmentation nodes, and then the network weight matrix W obtained from offline training is used to test the online test data X. test Reconstruction yields reconstructed data labels y test If y test A value of 1 indicates a fault and triggers an alarm; otherwise, it indicates normal operation.

[0119] The above steps are the specific application of the method of the present application in the field of BSM1 simulation platform fault monitoring. In order to verify the effectiveness of the method, 6 kinds of fault data were set for online monitoring under different weather conditions, and one batch of normal data and 4 batches of fault data were used for training each time, and 6 batches of test data were used for online monitoring, and the monitoring result graph is shown in Figures 2 to 7 If the vertical coordinate value in the monitoring graph is 1, it means that the model determines that the current sewage treatment process is a fault; if the vertical coordinate value in the graph is 0, it means that the model determines that the sewage treatment process is running normally. Table 3 is the monitoring index result of the width learning network method for extracting dynamic characteristics. FAR and MAR are the false alarm rate and the missed alarm rate of fault monitoring, and ACC is the accuracy of fault detection. The results in Table 3 show that the width learning network for extracting dynamic characteristics has achieved high ACC under 6 kinds of faults, with an average ACC of 92.23%, and the average values of FAR and MAR are 7.48% and 8.86% respectively.

[0120] Table 3 Monitoring index results of DBLS under 6 kinds of faults (%)

[0121]

Claims

1. A method for fault monitoring in wastewater treatment processes based on a wide-range learning network that efficiently extracts dynamic features, characterized by two stages: "offline training" and "online monitoring," with the specific steps as follows: A. Offline training phase: 1) Transfer a batch of normal data and fault data x fault By combining the data, we obtain the training data matrix and the label matrix as follows: in Labels representing normal data, Y fault Labels indicating malfunctions; 2) Combine the current sample variable with the previous l sample variables to construct training data X. train The augmented matrix X with time-delayed process variables is shown below: Where m is the number of sample variables, n is the number of samples, and l is the dynamic order of the matrix expansion, l = 1, X ∈ R. (n-l)×(l+1)m ; 3) Normalize X using the following formula: in, Let represent the (l+1)m-th variable value of the nl-th sample, and let mean(·) and std(·) be the mean and standard deviation of the (l+1)m-th variable, respectively. This will ultimately yield the normalized dataset. It is a dataset Elements in; 4) Normalize the data The input is mapped into the width learning system to generate feature nodes for the i-th window. The construction formula is as follows: Among them, W ei It is the weight matrix mapped to the feature nodes of the i-th window, β ei W is the bias vector mapped to the feature node of the i-th window during initial training. ei and β ei All are randomly generated random numbers that are uniformly distributed within the interval [0,1]. 5) The feature nodes of the i-th feature window p is the number of feature nodes in the i-th feature window. The following formula is used to normalize the feature nodes of the i-th window to the interval [-1, 1]. Among them, the feature node T of the i-th window after normalization i =[t1,t2,...,t p ], 6) To address the issues arising from the randomness in generating network weights and biases, the feature nodes T of the obtained i-th window are... i Perform sparsification representation; where the sparse matrix W s Obtained from the following formula: Where λ1 is the L1 regularization parameter, Through the sparse matrix W s Mapping yields the feature nodes of the final i-th window. Z i =[z1,z2,...,z p ], 7) Repeat steps 4-6 to generate all feature nodes for the feature windows, and connect these nodes to represent Z. I =[Z1,Z2,...,Z i Then, the j-th augmentation node is constructed using the feature nodes: in, ξ is a randomly generated weight and bias uniformly distributed in the interval (0,1), and ξ is the activation function tansig; 8) Repeat the formula in step 7 to generate all the required enhancement nodes and connect them as H. J =[H1,H2,...,H j ]; 9) After obtaining the feature nodes and augmentation nodes, construct the width learning system, which can be represented by the following formula: Y train =[Z I |H J ]W J Among them, Y train For the label matrix, W J =[Z I |H J ] + Y train These are the network's output connection weights; [·] + The pseudo-inverse of a matrix can be obtained using the following formula: [A J ] + =lim λ→0 (λ2I+A J [A J ] T ) -1 [A J ] T Where λ² is the L² regularization parameter, I is the identity matrix, and let A J =[Z I |H J The output network weights W of the width learning system can then be obtained. J Thus, the output matrix is ​​obtained; 10) When the model's accuracy fails to meet expectations, add enhancement nodes to improve nonlinear processing capabilities, thereby improving network performance; after adding enhancement nodes, the input layer becomes A. J+1 Since the network does not need to be retrained, it only needs to calculate A. J+1 The pseudo-inverse [A] J+1 ] + and weight W J+1 The calculation process is as follows: Among them, A J+1 It is the expanded matrix after adding new nodes. This indicates a newly added node. and These represent the random weights and biases of the newly added nodes, respectively. 11) Execute step 10 until the network training accuracy reaches the expected requirement, which is set to be greater than or equal to 95%. At this point, save the new training model parameters. B. Online monitoring phase: 12) Collect sample data x at the current moment. new Referring to step 2, construct an augmented matrix with time-delay process variables for the new collected data. 13) Apply the expanded formula from step 3. Normalization is performed to obtain the input data X new ; 14) Subsequently, X new The input is fed into a pre-trained width learning system, where the model parameters are obtained through offline modeling and training, and the final output of the model is a label matrix. If the label is 1, it indicates that a fault has occurred and an alarm has been issued; otherwise, it is normal.

2. The wastewater treatment process fault monitoring method based on a wide learning network for efficiently extracting dynamic features, as described in claim 1, is characterized in that: The width learning network for dynamic features, after incremental training during offline training, has 10 feature windows, 10 nodes per window, a total of 100 feature nodes, and 350 augmentation nodes.

Images