A method for mode classification and monitoring of multimodal industrial processes

Abstract

This invention discloses a method for mode segmentation and monitoring of multimodal industrial processes, mainly comprising two stages: offline modeling and online monitoring. In the offline modeling stage, historical operating data of the multimodal industrial process is first collected and standardized. Based on the nonlinear characteristic distribution of the data, multiple local autoencoder models are established. Then, based on the reconstruction error of each local autoencoder model, the loss weights of each model are calculated to construct a multi-local autoencoder model. Through iterative training, the model with the minimum global error is obtained, thus achieving mode segmentation. In the online monitoring stage, real-time data is first standardized and then input into the multi-local autoencoder model to determine its corresponding mode, thereby calculating monitoring indicators for state monitoring. This invention dynamically calculates the weights of each model and data allocation based on the model reconstruction error, automatically determines the optimal number of local models, achieves clustering with minimum global error, and enhances the state monitoring capability of multimodal industrial processes.

Description

Technical Field

[0001] This invention belongs to the field of multimodal industrial process monitoring technology, and particularly relates to a method and system for multimodal industrial process mode classification and monitoring. Background Technology

[0002] Monitoring the condition of industrial processes is crucial for ensuring production efficiency, product quality, and equipment safety. It enables timely detection of anomalies, early fault diagnosis, and preventative maintenance, avoiding production stoppages and equipment damage, reducing operating costs, and improving overall productivity. In industries such as chemicals, energy, and food, due to the uncertainty of the production environment, the diversity of raw materials, and the complexity of production processes, industrial processes exhibit multiple continuous steady states, with the mathematical characteristics of process variables remaining unchanged over time in each steady mode, demonstrating multimodal characteristics. Traditional single nonlinear models can exhibit good performance under specific operating conditions, but they have inherent limitations in accurately characterizing global multimodal nonlinear processes.

[0003] Two commonly used methods for monitoring multimodal processes are single-model methods and multi-model methods. Single-model methods are simple and efficient in their modeling process, but they assume that different modes are statistically homogeneous or have similar central tendencies, ignoring the heterogeneity and uniqueness between data modes. Therefore, single-model methods cannot effectively monitor data with significant differences and cannot accurately characterize the subtle differences across all modes. Multi-model methods first divide historical data of industrial processes according to different operating modes using clustering algorithms, prior knowledge, and other methods. Sub-models are then built for each independent mode, forming a comprehensive model library covering all modes. Multi-model methods can comprehensively capture the characteristic information of each local mode and have recently been widely used; however, their model accuracy heavily depends on the accuracy of mode division. Currently, many clustering algorithms have been applied to mode division, such as K-means clustering, fuzzy C-means clustering, density peak clustering, and Gaussian mixture models. Some of these methods use distance as the basis for data classification without considering the relationships between variables, which can lead to feature loss during clustering. Others use probability-based clustering, which requires prior knowledge of the number of local models, making it difficult to determine the number of modalities in practice.

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

[0005] Purpose of the Invention: The purpose of this invention is to provide a method for mode segmentation and monitoring of multimodal industrial processes. This invention collects operational data from all operating conditions and constructs a corresponding autoencoder model for each operating condition. Based on the reconstruction error of different autoencoder models, the weights of each model are established, thereby automatically enhancing the relevant data of the mode to which the cluster model belongs and weakening the data of other modes. It can automatically select an appropriate number of local models without having to try different numbers of clusters, thus improving the objectivity and reliability of mode segmentation.

[0006] Technical Solution: To achieve the above objectives, this invention discloses a method for mode division and monitoring of multimodal industrial processes, comprising the following steps:

[0007] S1: Establish local autoencoder models of different modalities offline;

[0008] S1.1: Collect a sample set of relevant historical data and perform standardization processing;

[0009] S1.2: Set the number of iterations and the initial number of modes, and establish a local autoencoder model with the same number of modes;

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

[0011] S1.4: Add the loss weight w to the loss function to establish and train a multilocal 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 mode to obtain the mode division results, and establish local autoencoder models of different modes based on the mode division results;

[0014] S2: Real-time data is monitored online based on a local autoencoder model;

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

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

[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. If the value exceeds the threshold, it is an abnormal sample; otherwise, it is a normal sample.

[0018] Optionally, the standardization process in step S1.1 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.

[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 consists 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 features extracted from the hidden layer of the autoencoder.

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

[0025]

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

[0027] Optionally, in step S1.3, the reconstruction error is the difference between the input sample and the output sample, calculated using the following formula:

[0028]

[0029] Among them, e i,c For reconstruction error, x i For the input sample, For the output sample, the subscript c represents the c-th autoencoder model.

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

[0031]

[0032] Where 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 verification region of the local autoencoder model m The distance is given by the following formula:

[0033]

[0034] in, Let be the center point of the sample validation region for the local autoencoder model m;

[0035] The influence function formula for Fair estimation is as follows:

[0036]

[0037] Where l is an adjustment constant.

[0038] Optionally, after adding the loss weight w to the loss function in step S1.4, the resulting objective function of the multilocal autoencoder is as follows:

[0039]

[0040] Where, when a sample point belongs to the local autoencoder model s but not to the local autoencoder model m, we have and c∈{1,2,…,M;c≠m,s}, where the weight w of the sample point in the local autoencoder model m is... i,m ≈0, meaning samples that do not belong to the local autoencoder model m will not affect the training of the local autoencoder model m; conversely, when a sample belongs to the local autoencoder model m, there is... and Furthermore, if the sample is close to the center point of the sample in 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 difference in weight between the sample points and the sample points in the non-local autoencoder model m will be even greater.

[0041] Optionally, in step S1.6, when classifying data with larger loss weights in each local autoencoder into the same mode, the weights of all local autoencoder models of the sample are compared, and the sample points are classified into the autoencoder model with the largest weight, that is, classified into the mode corresponding to the autoencoder model, thereby obtaining the mode classification result, and establishing autoencoder models of different modes based on the mode classification result.

[0042] Optionally, the SPE statistics information based on the encoder input x and decoder output y in step S1.6 is constructed as follows:

[0043]

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

[0045] Optionally, the threshold in step S1.6 The formula is derived from kernel density estimation and is as follows:

[0046]

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

[0048] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention collects operational data from all operating conditions and simultaneously constructs a corresponding autoencoder model for each operating condition. It establishes the weights of each model based on the reconstruction errors of different autoencoder models, thereby automatically enhancing the relevant data of the mode to which the clustering model belongs and weakening the data of other modes. It can automatically select an appropriate number of local models without needing to attempt different numbers of clusters, thus improving the objectivity and reliability of mode partitioning. This invention constructs multiple local autoencoder models and dynamically calculates the weights and data allocation of each model based on the model reconstruction error, achieving clustering with minimized global error and enhancing the state monitoring capability of multimodal industrial processes. This invention can automatically determine the optimal number of local models, reducing human intervention, and improve the accuracy of mode partitioning through a dynamic weight adjustment mechanism. Especially when the number of modes is unknown, it can still accurately perform mode partitioning, demonstrating strong adaptive capabilities. Attached Figure Description

[0049] Figure 1 This is a flowchart of the present invention;

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

[0051] Figure 3 This is a scatter plot of the modal division results of the heater using the K-means method in this invention.

[0052] Figure 4 This is a scatter plot of the modal classification results of the heater in this invention.

[0053] Figure 5 These are the monitoring results of the heater using the K-means and AE methods in this invention;

[0054] Figure 6 The results of heater monitoring using the SDAE method in this invention;

[0055] Figure 7This is a graph showing the monitoring results of the heater in this invention. Detailed Implementation

[0056] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0057] This invention collects operational data from all operating ranges and constructs a corresponding autoencoder model for each operating condition. It establishes weights for each model based on the reconstruction errors of different autoencoder models, thereby automatically enhancing the relevant data of the clustering model's modality and weakening data from other modalities. For example... Figure 1 As shown, the present invention provides a method for mode division and monitoring of a multimodal industrial process, comprising the following steps:

[0058] S1: Establish local autoencoder models of different modalities offline;

[0059] S1.1: Collect and run relevant historical data sample sets and perform standardization processing; the standardization processing 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;

[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 with the same number of modes.

[0061] The local autoencoder model consists of an encoder and a decoder. The encoding and decoding processes are 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 features extracted from the hidden layer of the autoencoder.

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

[0066]

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

[0068] S1.3: Calculate the reconstruction error of each local autoencoder model and 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 the calculation formula is as follows:

[0070]

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

[0072] The loss weight w of the local autoencoder model m is calculated using the following formula:

[0073]

[0074] Where 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 verification region of the local autoencoder model m The distance is given by the following formula:

[0075]

[0076] in, Let be the center point of the sample validation region for the local autoencoder model m;

[0077] The influence function formula for Fair estimation is as follows:

[0078]

[0079] Where l is an adjustment constant;

[0080] S1.4: Add the loss weight w to the loss function to establish and train a multilocal autoencoder model;

[0081] After adding the loss weight w to the loss function, the resulting objective function of the multilocal autoencoder is as follows:

[0082]

[0083] Where, when a sample point belongs to the local model s but not to the local autoencoder model m, we have and The weight w of the sample point in the local autoencoder model m i,m ≈0, meaning samples that do not belong to the local autoencoder model m will not affect the training of the local autoencoder model m; conversely, when a sample belongs to the local autoencoder model m, there is... and Furthermore, if the sample is close to the center point of the local autoencoder model m, then the weight of this sample point in the local autoencoder model m will be greater than zero, and the difference in weight between it and the sample points in the non-local autoencoder model m will be even greater. In summary, if a sample is similar to most samples in the local autoencoder model m and the training error is small, then the data point is close to the region of the local autoencoder model m and its weight increases; if the sample is far from most samples in the local autoencoder model m and the training error is large, then the data point must belong to another local autoencoder model, and therefore its weight decreases.

[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 mode to obtain the mode division results, and establish autoencoder models of different modes based on the mode division results;

[0086] Specifically, when classifying data with larger loss weights w in each local autoencoder into the same mode, the weights of all local autoencoder models of the sample are compared, and the sample points are classified into the autoencoder model with the largest weight, that is, classified into the mode corresponding to that autoencoder model, thereby obtaining the mode classification result, and establishing autoencoder models of different modes based on the mode classification result.

[0087] S2: Real-time data is monitored online based on a local autoencoder model;

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

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

[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. If it exceeds the threshold, it is an abnormal sample; otherwise, it is a normal sample.

[0091] SPE (Specific Performance Parameter) is a commonly used process monitoring statistical indicator that reflects the degree of deviation between process variables and the model. The SPE statistics based on encoder input x and decoder output y are constructed as follows:

[0092]

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

[0094] threshold The formula is derived from kernel density estimation and is as follows:

[0095]

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

[0097] Based on actual data, the present invention provides the following optional specific implementation schemes:

[0098] The effectiveness of the method proposed in this invention was verified using actual operating data from the air heater equipment in a thermal power plant. The air heater is an important device in the flue gas waste heat system, typically installed between the blower outlet and the air preheater inlet. Its location, for example... Figure 2 As shown. When the outdoor air temperature is low, the performance of the fan will decrease, and condensation on the equipment pipe walls will cause dust accumulation, reducing heat transfer efficiency. The air heater utilizes the waste heat from the boiler exhaust to heat the air entering the boiler, increasing the temperature of the cold air entering the system, thereby optimizing the boiler's thermal cycle efficiency. The application of the air heater not only improves the performance of the blower but also significantly reduces condensation on the equipment pipe walls, reduces dust accumulation, lowers system resistance, and improves overall heat cycle efficiency. The air heater equipment lacks valve position measurement points for controlling the internal circulating water flow, and this valve is manually controlled by operators, making it impossible to confirm its precise position. The operating characteristics of the air heater are related to the condensate flow rate; the outlet air temperature of the air heater increases with the increase of the condensate flow rate. Therefore, different valve openings will result in different operating characteristics of the air heater, resulting in multiple different modes.

[0099] The inlet temperature, outlet temperature, outlet pressure, inlet air volume, inlet temperature, outlet pressure, and outlet air volume of the blower A heater, as well as the temperature and pressure of the left and right hot secondary air headers of the boiler, are used as parameter variables for the heater model. Partial operational data from April 2022 was collected from the DCS system, totaling 6000 samples, as the training set. Although it is known that the condensate flow valve changed twice during this period, and the number of modes in the training set should be 3, the number of modes was preset to 4 to verify that the method can automatically determine the number of modes. The hyperparameters of this method are set as c=1, a=8, k=0.2, and the autoencoder network structure is 12-8-4-2-4-8-12. The K-means method was also applied to modally partition the training set, with the number of modalities set to 3. For example... Figure 3 and Figure 4As shown, the modality division results of the training set are displayed through scatter plots of the outlet temperatures of the blower A and blower B, which are parameter variables. The method proposed in this invention can accurately divide the three modes.

[0100] The effectiveness of the proposed method for process monitoring was verified using a test set. The test set contained 600 samples, of which samples 111 to 586 were samples of air heater leakage failures. When leakage occurs, the outlet temperature of the air heater in the blower is lower than the normal temperature, thus affecting the operation of the entire flue gas waste heat recovery system. The proposed method was also compared with a local model (a method combining K-means and AE) and a single SDAE model that does not differentiate between modes. Figure 5 , Figure 6 and Figure 7 The monitoring results of the three methods are shown, with the dashed line representing the control limit at a confidence level of 99%. It can be seen that the SPE value of the method in this invention exceeds the control limit first, outperforming the other two methods. This is because a single model models three modes of data simultaneously, ignoring mode information, making it insensitive to small-amplitude faults and unable to diagnose faults in their early stages, exhibiting a certain lag. While the method combining K-means and AE establishes AE models on different modes, the distance-based clustering method leads to inaccurate mode segmentation, resulting in decreased performance of local models. The method in this invention considers the interrelationships between variables and segments the data from a model perspective, making the mode segmentation results more interpretable and enabling faster detection of fault data.

Claims

1. A method for mode classification and monitoring of a multimodal industrial process, characterized in that, Includes the following steps: S1: Establish local autoencoder models of different modalities offline; S1.1: Collect a sample set of relevant historical data and perform standardization processing; S1.2: Set the number of iterations and the initial number of modes, and establish a local autoencoder model with the same number of modes; S1.3: Calculate the reconstruction error of each local autoencoder model to obtain the loss weight of each local autoencoder model. The loss weights of the local autoencoder model m in step S1.3 The calculation formula is as follows: ， in, For the independent variable The influence function of the Fair estimate, where M is the number of local autoencoder models. for The center point of the data verification region of the local autoencoder model m The distance is given by the following formula: ， in, Let be the center point of the sample validation region for the local autoencoder model m; The influence function formula for Fair estimation is as follows: ， in, It is an adjustment constant; S1.4: Weight the loss Add it to the loss function to build and train a multi-local autoencoder model; S1.5: Repeat S1.3-S1.4 until the weights are lost. Convergence or reaching the maximum number of iterations; S1.6: Divide the data with larger loss weights in each local autoencoder into the same mode to obtain the mode division results, and establish local autoencoder models of different modes based on the mode division results; S2: Real-time data is monitored online based on a local autoencoder model; S2.1: Perform the same standardization process on real-time data; S2.2: Input the real-time data into each modal model to calculate its reconstruction error and determine its mode; 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. If the value exceeds the threshold, it is an abnormal sample; otherwise, it is a normal sample.

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

3. The method for mode division and monitoring of a multimodal industrial process 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. The method for mode division and monitoring of a multimodal industrial process according to claim 3, characterized in that, The local autoencoder model in step S1.2 consists of an encoder and a decoder, and the encoding and decoding processes are as follows: ， ， in, It is the input sample; These are reconstructed samples from an autoencoder; , For encoder weights and biases; , For decoder weights and biases; It is the ReLU activation function; These are the variable features extracted from the hidden layer of the autoencoder. The autoencoder training process adjusts parameters to minimize the reconstruction error, thereby obtaining the optimal representation of the input features. The squared reconstruction error is used as the loss function of the autoencoder, defined as follows: ， in, For input samples; For output samples; This represents the L2 norm.

5. The method for mode division and monitoring of a multimodal industrial process according to claim 1, characterized in that, In step S1.3, the reconstruction error is the difference between the input sample and the output sample, and the calculation formula is as follows: ， in, For reconstruction error, For the input sample, For the output sample, the subscript c represents the c-th autoencoder model.

6. The method for mode division and monitoring of a multimodal industrial process according to claim 1, characterized in that, In step S1.4, the loss weight is... After adding the loss function, the resulting objective function of the multilocal autoencoder is as follows: ， Where, when a sample point belongs to the local autoencoder model s but not to the local autoencoder model m, we have and The weight of the sample point in the local autoencoder model m Samples that do not belong to the local autoencoder model m will not affect the training of the local autoencoder model m; conversely, when a sample belongs to the local autoencoder model m, there is and If the sample is close to the center point of the sample in 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 difference between the weights of the sample points in the non-local autoencoder model m will be even greater.

7. The method for mode division and monitoring of a multimodal industrial process according to claim 1, characterized in that, In step S1.6, when classifying data with larger loss weights in each local autoencoder into the same mode, the weights of all local autoencoder models of the sample are compared, and the sample points are classified into the autoencoder model with the largest weight, that is, classified into the mode corresponding to the autoencoder model, thereby obtaining the mode classification result, and establishing autoencoder models of different modes based on the mode classification result.

8. The method for mode division and monitoring of a multimodal industrial process according to claim 1, characterized in that, The SPE statistics information based on the encoder input x and decoder output y in step S1.6 is constructed as follows: ， in, The model residual vector; labeled To calculate the L2 norm of the residual vector.

9. The method for mode division and monitoring of a multimodal industrial process according to claim 8, characterized in that, The threshold in step S1.6 The formula is derived from kernel density estimation and is as follows: ， in, for The upper quantile corresponding to the confidence region; , The first of the sample covariance matrix One eigenvalue; .

Images