Multi-modal industrial process modal division and monitoring method

Abstract

The invention discloses a modal division and monitoring method for a multi-modal industrial process. The modal division and monitoring method mainly comprises an offline modeling stage and an online monitoring stage. In the off-line modeling stage, multi-modal industrial process historical operation data is collected for standardization processing, a plurality of local auto-encoder models are established based on data nonlinear feature distribution, and then the loss weight of each model is calculated according to the reconstruction error of the local auto-encoder models to construct a multi-local auto-encoder model. A model with the minimum global error is obtained through iterative training, and modal division is achieved; in the online monitoring stage, the real-time data is subjected to standardization processing firstly and then is input into the multi-local auto-encoder model, the mode to which the real-time data belongs is determined, and therefore monitoring indexes are calculated for state monitoring. According to the method, the weight and data distribution of each model are dynamically calculated based on the model reconstruction error, the number of the optimal local models is automatically determined, clustering of global error minimization is realized, and the state monitoring capability of the multi-modal industrial process is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of multimodal industrial process monitoring, and in particular relates to a multimodal industrial process modal division and monitoring method and system. Background Art

[0002] Condition monitoring of industrial processes is crucial to ensuring production efficiency, product quality and equipment safety. It can detect abnormal conditions in a timely manner, conduct early fault diagnosis and preventive maintenance, avoid production stagnation and equipment damage, reduce operating costs, and improve overall productivity. In the chemical, energy, food and other industries, due to the uncertainty of the production environment, the diversity of production raw materials and the complexity of production processes, industrial processes have multiple continuous stable states and the mathematical characteristics of process variables in each stable mode do not change with time, and have multimodal characteristics. Traditional single nonlinear models can perform well under specific operating conditions, but they have inherent limitations in accurately characterizing global multimodal nonlinear processes.

[0003] The two common methods for multimodal process monitoring are the single model method and the multi-model method. The single model method has a simple and efficient modeling process, but it assumes that different modes can be regarded as homogeneous or have similar central trends in statistics, ignoring the heterogeneity and uniqueness between data modes. Therefore, the single model method cannot effectively monitor data with significant differences and cannot accurately characterize the subtle differences of all modes. The multi-model method first divides the historical data of the industrial process according to different operating modes through clustering algorithms, prior knowledge and other methods, and constructs sub-models for each independent mode, thereby forming a comprehensive model library covering all modes. The multi-model method can fully capture the characteristic information of each local mode and has been widely used recently, but its model accuracy is heavily dependent on the accuracy of mode division. At present, many clustering algorithms have been applied to mode division, such as K-means clustering, fuzzy C-means clustering, density peak clustering, Gaussian mixture model, etc. Some of these methods use distance as the basis for data classification, without considering the relationship between variables, which may cause feature loss in the clustering process; the other part is based on probability clustering, which requires knowing the number of local models in advance, and it is difficult to determine the number of modes in actual operations.

[0004] Therefore, it is urgent to solve the above problems. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a multimodal industrial process modal division and monitoring method. The present invention collects operating data from all working conditions, and constructs a corresponding autoencoder model for each operating condition. The weights of each model are established according to the reconstruction errors of different autoencoder models, thereby automatically enhancing the relevant data of the modality to which the clustering model belongs and weakening other modal data. It can automatically select an appropriate number of local models without trying different numbers of clusters, thereby improving the objectivity and reliability of modal division.

[0006] Technical solution: To achieve the above objectives, the present invention discloses a multi-modal industrial process mode division and monitoring method, comprising the following steps:

[0007] S1: Offline establishment of local autoencoder models of different modalities;

[0008] S1.1: Collect and standardize the sample set of historical data related to operation;

[0009] S1.2: Set the number of iterations and the number of initial modalities, and build a local autoencoder model equal to the number of modalities;

[0010] S1.3: Calculate the reconstruction error of each local autoencoder model to obtain the loss weight w of each local autoencoder model;

[0011] S1.4: Add the loss weight w to the loss function, build and train the multi-local autoencoder model;

[0012] S1.5: Repeat S1.3-S1.4 until the loss weight w converges or the maximum number of iterations is reached;

[0013] S1.6: Divide the data with larger loss weights in each local autoencoder into the same modality to obtain the modality division result, and establish local autoencoder models of different modalities according to the modality division result;

[0014] S2: Online monitoring of real-time data based on the local autoencoder model;

[0015] S2.1: Perform the same standardization process on real-time data;

[0016] S2.2: Put the real-time data into each modal model to calculate its reconstruction error and determine the mode to which it belongs;

[0017] S2.3: Calculate the SPE value based on the modal autoencoder model to which the real-time sample belongs and compare it with the control threshold. Samples exceeding the threshold are abnormal samples, otherwise they are normal samples.

[0018] Optionally, the standardization process in step S1.1 is Z-Score standardization, and the standardized data will have a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0019] Optionally, in step S1.2, the number of initial modes is greater than the number of known operating conditions.

[0020] Optionally, the local autoencoder model in step S1.2 is composed of an encoder and a decoder, and the encoding and decoding processes are as follows:

[0021] H=g(W1X+b1)

[0022]

[0023] Where X is the input sample; is the reconstructed sample of the autoencoder; W1 and b1 are the encoder weights and biases; W2 and b2 are the decoder weights and biases; g is the RELU activation function; H is the variable feature extracted by the hidden layer of the autoencoder.

[0024] The autoencoder training process adjusts the parameters to minimize the reconstruction error to obtain the optimal representation of the input features; the squared reconstruction error is used as the loss function of the autoencoder, which is defined as follows:

[0025]

[0026] in, is the input sample; is the output sample; ||·||2 represents the L2 norm.

[0027] Optionally, the reconstruction error in step S1.3 is the difference between the input sample and the output sample, and the calculation formula is as follows:

[0028]

[0029] Among them, e i,c is the reconstruction error, x i is the input sample, is the output sample, and the subscript c represents the cth autoencoder model.

[0030] Optionally, the loss weight w of the local autoencoder model m in step S1.3 is calculated as follows:

[0031]

[0032] Among them, F(a) is the influence function of the Fair estimate of the independent variable a, M is the number of local autoencoder models, and d i,m For x iThe center point of the data validation area with the local autoencoder model m The distance is as follows:

[0033]

[0034] in, is the center point of the sample verification area of ​​the local autoencoder model m;

[0035] Among them, the influence function formula of Fair estimation is as follows:

[0036]

[0037] Among them, l is a tuning constant.

[0038] Optionally, after the loss weight w is added to the loss function in step S1.4, the objective function of the multi-local autoencoder formed is as follows:

[0039]

[0040] Among them, when a sample point belongs to the local autoencoder model s, but not to the local autoencoder model m, there is and c∈{1,2,…,M;c≠m,s}, the weight w of the sample point in the local autoencoder model m i,m ≈0, the samples that do not belong to the local autoencoder model m will not affect the training of the local autoencoder model m; on the contrary, when a sample belongs to the local autoencoder model m, and And if the sample is closer to the sample center point of the local autoencoder model m, the weight of the sample point in the local autoencoder model m will be greater than zero, and the difference with the sample point weight of the non-local autoencoder model m will be greater.

[0041] Optionally, in step S1.6, when the data with larger loss weights in each local autoencoder are divided into the same modality, the weights of all local autoencoder models of the sample are compared, and the sample point is classified as the autoencoder model with the largest weight, that is, classified as the modality corresponding to the autoencoder model, thereby obtaining the modal division result, and establishing autoencoder models of different modalities according to the modal division result.

[0042] Optionally, the construction of the SPE statistical information according to the encoder input x and the decoder output y in step S1.6 is as follows:

[0043]

[0044] Where (xy) is the model residual vector; To calculate the L2 norm of the residual vector.

[0045] Optionally, the threshold in step S1.6 It is calculated by kernel density estimation, and the formula is as follows:

[0046]

[0047] Among them, z α is the upper quantile corresponding to the (1-α)% confidence region; λ k is the kth eigenvalue of the sample covariance matrix;

[0048] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the present invention collects operating data from all working conditions, and constructs a corresponding autoencoder model for each operating condition at the same time, and establishes the weight of each model according to the reconstruction error of different autoencoder models, so as to automatically enhance the relevant data of the mode to which the clustering model belongs, and weaken other modal data, and can automatically select an appropriate number of local models without trying different numbers of clusters, thereby improving the objectivity and reliability of modal division; the present invention constructs multiple local autoencoder models, dynamically calculates the weight of each model and data allocation based on the model reconstruction error, realizes clustering with global error minimization, and enhances the state monitoring capability of multimodal industrial processes; the present invention can automatically determine the optimal number of local models, reduce human intervention, and improve the modal division accuracy through a dynamic weight adjustment mechanism; especially when the number of modes is unknown, it can still accurately perform modal division, and has strong adaptive capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the present invention;

[0050] Figure 2 This is a flow chart of the flue gas waste heat system in the present invention;

[0051] Figure 3 It is a scatter plot of the results of the heater mode division using the K-means method in the present invention;

[0052] Figure 4 It is a scatter diagram of the modal division results of the air heater of the present invention;

[0053] Figure 5 The monitoring results of the air heater using the K-means and AE methods in the present invention;

[0054] Figure 6 This is the monitoring result of the air heater using the SDAE method in the present invention;

[0055] Figure 7This is a monitoring result diagram of the air heater of the present invention. DETAILED DESCRIPTION

[0056] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0057] The present invention collects operating data from all operating ranges, and constructs corresponding autoencoder models for each operating condition. The weights of each model are established according to the reconstruction errors of different autoencoder models, thereby automatically enhancing the relevant data of the modality to which the clustering model belongs and weakening the data of other modalities. Figure 1 As shown, the present invention provides a multi-modal industrial process mode division and monitoring method, comprising the following steps:

[0058] S1: Offline establishment of local autoencoder models of different modalities;

[0059] S1.1: Collect a sample set of historical data related to operation and perform standardization processing; the standardization processing is to perform Z-Score standardization, and the standardized data will have a standard normal distribution with a mean of 0 and a standard deviation of 1;

[0060] S1.2: Set the number of iterations and the initial number of modes. The initial number of modes is greater than the number of known working conditions, and establish a local autoencoder model equal to the number of modes.

[0061] The local autoencoder model consists of an encoder and a decoder. The encoding and decoding process is as follows:

[0062] H=g(W1X+b1)

[0063]

[0064] Where X is the input sample; is the reconstructed sample of the autoencoder; W1 and b1 are the encoder weights and biases; W2 and b2 are the decoder weights and biases; g is the RELU activation function; H is the variable feature extracted by the hidden layer of the autoencoder.

[0065] The autoencoder training process adjusts the parameters to minimize the reconstruction error to obtain the optimal representation of the input features; the squared reconstruction error is used as the loss function of the autoencoder, which is defined as follows:

[0066]

[0067] in, is the input sample; is the output sample; ||·||2 represents the L2 norm;

[0068] S1.3: Calculate the reconstruction error of each local autoencoder model to obtain the loss weight w of each local autoencoder model;

[0069] The reconstruction error is the difference between the input sample and the output sample, and is calculated as follows:

[0070]

[0071] Among them, e i,c is the reconstruction error, x i is the input sample, is the output sample, the subscript c represents the cth autoencoder model;

[0072] The loss weight w of the local autoencoder model m is calculated as follows:

[0073]

[0074] Among them, F(a) is the influence function of the Fair estimate of the independent variable a, M is the number of local autoencoder models, and d i,m For x i The center point of the data validation area with the local autoencoder model m The distance is as follows:

[0075]

[0076] in, is the center point of the sample verification area of ​​the local autoencoder model m;

[0077] Among them, the influence function formula of Fair estimation is as follows:

[0078]

[0079] Among them, l is a regulation constant;

[0080] S1.4: Add the loss weight w to the loss function, build and train the multi-local autoencoder model;

[0081] After adding the loss weight w to the loss function, the resulting multi-local autoencoder objective function is as follows:

[0082]

[0083] Among them, when a sample point belongs to the local model s but not to the local autoencoder model m, there is and The weight w of this sample point in the local autoencoder model m i,m ≈0, the samples that do not belong to the local autoencoder model m will not affect the training of the local autoencoder model m; on the contrary, when a sample belongs to the local autoencoder model m, and And the sample is closer to the center point of the sample of the local autoencoder model m, then the weight of the sample point in the local autoencoder model m will be greater than zero, and the weight difference with the sample point of the non-local autoencoder model m will be greater; In summary, if the sample is similar to most of the samples in the local autoencoder model m and the training error is small, then the data point is closer to the local autoencoder model m area and the weight becomes larger; if the sample is far away from most of the samples in the local autoencoder model m and the training error is large, then the data point must belong to another local autoencoder model, so the weight becomes smaller;

[0084] S1.5: Repeat S1.3-S1.4 until the loss weight w converges or the maximum number of iterations is reached;

[0085] S1.6: Divide the data with larger loss weights in each local autoencoder into the same modality to obtain the modality division result, and establish autoencoder models of different modalities according to the modality division result;

[0086] Specifically, when the data with larger loss weight w in each local autoencoder are divided into the same modality, the weights of all local autoencoder models of the sample are compared, and the sample point is classified as the autoencoder model with the largest weight, that is, classified as the modality corresponding to the autoencoder model, thereby obtaining the modality division result, and establishing autoencoder models of different modalities according to the modality division result;

[0087] S2: Online monitoring of real-time data based on the local autoencoder model;

[0088] S2.1: Perform the same standardization process on real-time data;

[0089] S2.2: Put the real-time data into each modal model to calculate its reconstruction error and determine the mode to which it belongs;

[0090] S2.3: Calculate the SPE value based on the modal autoencoder model to which the real-time sample belongs and compare it with the control threshold. Samples exceeding the threshold are abnormal samples, and samples exceeding the threshold are normal samples.

[0091] SPE is a commonly used process monitoring statistical indicator that reflects the degree of deviation between process variables and models. The construction of SPE statistical information based on encoder input x and decoder output y is as follows:

[0092]

[0093] Where (xy) is the model residual vector; To calculate the L2 norm of the residual vector;

[0094] Threshold It is calculated by kernel density estimation, and the formula is as follows:

[0095]

[0096] Among them, z α is the upper quantile corresponding to the (1-α)% confidence region; λ k is the kth eigenvalue of the sample covariance matrix;

[0097] Combined with actual data, the present invention provides optional specific implementation schemes as follows:

[0098] The effectiveness of the method proposed in the present invention is verified by the actual operation data of the air heater equipment in the thermal power plant. The air heater is an important equipment in the flue gas waste heat system, usually installed between the outlet of the blower and the inlet of the air preheater. Figure 2 As shown. When the outdoor wind temperature is low, the fan performance will be reduced, and condensation on the equipment pipe wall will cause dust and other substances to accumulate, which will reduce the heat transfer efficiency. The air heater uses the waste heat of the flue gas discharged from the boiler to heat the air entering the boiler, increase the temperature of the cold air entering the system, and thus optimize the thermal cycle efficiency of the boiler. The application of the air heater not only improves the performance of the blower, but also significantly reduces the condensation on the equipment pipe wall, reduces dust accumulation, reduces system resistance, and improves the overall thermal cycle utilization efficiency. The air heater equipment does not have a valve position measurement point for controlling the internal circulating water flow, and the valve is manually controlled by the operator, and its precise valve position cannot be confirmed. The operating characteristics of the air heater are related to the condensate flow rate. The outlet air temperature of the air heater will increase with the increase of the condensate flow rate. Therefore, different valve openings will cause the air heater to have different operating characteristics, resulting in multiple different modes of the air heater.

[0099] The inlet temperature of the heater of fan A, the outlet temperature of the heater of fan A, the outlet pressure of fan A, the inlet air volume of fan A, the inlet temperature of the heater of fan B, the outlet temperature of the heater of fan B, the outlet pressure of fan B, the inlet air volume of fan B, the temperature of the hot secondary air main pipe on the left side of the boiler, the pressure of the hot secondary air main pipe on the left side of the boiler, the temperature of the hot secondary air main pipe on the right side of the boiler, and the pressure of the hot secondary air main pipe on the right side of the boiler are used as parameter variables of the heater model. Some operating data in April 2022 are collected from the DCS system, and a total of 6000 samples are used as training sets. Although it is known that the condensate flow valve has changed twice during this period, the number of modes in the training set should be 3, but in order to verify that the method can automatically determine the number of modes, the number of modes is preset to 4. The hyperparameters of this method are set to c=1, a=8, k=0.2, and the autoencoder network structure is 12-8-4-2-4-8-12. At the same time, the K-means method is also used to divide the training set into modes, and the number of modes is set to 3. Figure 3 and Figure 4As shown, the modal division results of the training set are displayed through a scatter plot of the parameter variables, the outlet temperature of the heater of blower A and the outlet temperature of the heater of blower B. The method proposed in the present invention can accurately divide the three modes.

[0100] The test set is used to verify the effectiveness of the proposed method for process monitoring. The test set contains 600 samples, of which 111 to 586 samples are heater leakage fault samples. When leakage occurs, the outlet temperature of the blower heater is lower than the normal temperature, thereby affecting the operation of the entire flue gas waste heat utilization system. At the same time, the proposed method is compared with the local model (the method combining K-means and AE) and the single SDAE model without mode distinction. Figure 5 , Figure 6 and Figure 7 The monitoring results of the three methods are shown respectively, where the dotted lines represent the control limits with a confidence level of 99%. It can be seen that the SPE value of the method of the present invention exceeds the control limit first, which is better than the other two methods. This is because a single model simultaneously models the three mode data, ignores the mode information, is insensitive to small-amplitude faults, cannot diagnose the fault in time at the beginning of the fault, and has a certain hysteresis. Although the method combining K-means and AE establishes AE models on different modes, the distance-based clustering method makes the mode division inaccurate, which reduces the performance of the local model. The method of the present invention takes into account the relationship between variables and divides the data from the perspective of the model, so that the mode division results are more interpretable and can detect fault data more quickly.

Claims

1. A multi-modal industrial process modal division and monitoring method, characterized in that: The steps include: S1: Offline establishment of local autoencoder models of different modalities; S1.1: Collect and standardize the sample set of historical data related to operation; S1.2: Set the number of iterations and the number of initial modalities, and build a local autoencoder model equal to the number of modalities; S1.3: Calculate the reconstruction error of each local autoencoder model to obtain the loss weight w of each local autoencoder model; S1.4: Add the loss weight w to the loss function, build and train the multi-local autoencoder model; S1.5: Repeat S1.3-S1.4 until the loss weight w converges or the maximum number of iterations is reached; S1.6: Divide the data with larger loss weights in each local autoencoder into the same modality to obtain the modality division result, and establish local autoencoder models of different modalities according to the modality division result; S2: Online monitoring of real-time data based on the local autoencoder model; S2.1: Perform the same standardization process on real-time data; S2.2: Put the real-time data into each modal model to calculate its reconstruction error and determine the mode to which it belongs; S2.3: Calculate the SPE value based on the modal autoencoder model to which the real-time sample belongs and compare it with the control threshold. Samples exceeding the threshold are abnormal samples, otherwise they are normal samples.

2. A multi-modal industrial process modal division and monitoring method according to claim 1, characterized in that: The standardization process in step S1.1 is Z-Score standardization, and the standardized data will have a standard normal distribution with a mean of 0 and a standard deviation of 1.

3. A multi-modal industrial process modal division and monitoring method according to claim 1, characterized in that: In step S1.2, the number of initial modes is greater than the number of known operating conditions.

4. A multi-modal industrial process modal division and monitoring method according to claim 3, characterized in that: The local autoencoder model in step S1.2 is composed of an encoder and a decoder, and the encoding and decoding process is as follows: H=g(W1X+b1) Where X is the input sample; is the reconstructed sample of the autoencoder; W1 and b1 are the encoder weights and biases; W2 and b2 are the decoder weights and biases; g is the RELU activation function; H is the variable feature extracted by the hidden layer of the autoencoder. The autoencoder training process adjusts the parameters to minimize the reconstruction error to obtain the optimal representation of the input features; the squared reconstruction error is used as the loss function of the autoencoder, which is defined as follows: in, is the input sample; is the output sample; ||·||2 represents the L2 norm.

5. A multi-modal industrial process modal division and monitoring method according to claim 1, characterized in that: The reconstruction error in step S1.3 is the difference between the input sample and the output sample, and the calculation formula is as follows: Among them, e i,c is the reconstruction error, x i is the input sample, is the output sample, and the subscript c represents the cth autoencoder model.

6. A multi-modal industrial process modal division and monitoring method according to claim 5, characterized in that: The loss weight w of the local autoencoder model m in step S1.3 is calculated as follows: Among them, F(a) is the influence function of the Fair estimate of the independent variable a, M is the number of local autoencoder models, and d i,m For x i The center point of the data validation area with the local autoencoder model m The distance is as follows: in, is the center point of the sample verification area of ​​the local autoencoder model m; Among them, the influence function formula of Fair estimation is as follows: Where l is a tuning constant.

7. A multi-modal industrial process modal division and monitoring method according to claim 1, characterized in that: After the loss weight w is added to the loss function in step S1.4, the objective function of the multi-local autoencoder is as follows: Among them, when a sample point belongs to the local autoencoder model s, but not to the local autoencoder model m, there is and The weight w of this sample point in the local autoencoder model m i,m ≈0, the samples that do not belong to the local autoencoder model m will not affect the training of the local autoencoder model m; on the contrary, when a sample belongs to the local autoencoder model m, and And if the sample is closer to the sample center point of the local autoencoder model m, the weight of the sample point in the local autoencoder model m will be greater than zero, and the difference with the sample point weight of the non-local autoencoder model m will be greater.

8. A multi-modal industrial process modal division and monitoring method according to claim 1, characterized in that: Specifically, in step S1.6, when the data with larger loss weights in each local autoencoder are divided into the same modality, the weights of all local autoencoder models of the sample are compared, and the sample point is classified into the autoencoder model with the largest weight, that is, classified into the modality corresponding to the autoencoder model, thereby obtaining the modality division result, and establishing autoencoder models of different modalities according to the modality division result.

9. A multi-modal industrial process modal division and monitoring method according to claim 1, characterized in that: The construction of the SPE statistics according to the encoder input x and the decoder output y in step S1.6 is as follows: Where (xy) is the model residual vector; To calculate the L2 norm of the residual vector.

10. A multi-modal industrial process modal division and monitoring method according to claim 9, characterized in that: The threshold value in step S1.6 It is calculated by kernel density estimation, and the formula is as follows: Among them, z α is the upper quantile corresponding to the (1-α)% confidence region; λ k is the kth eigenvalue of the sample covariance matrix;

Images