Chemical process fault diagnosis method based on multi-scale spatial-temporal feature fusion

Abstract

The invention relates to the technical field of chemical process fault diagnosis, in particular to a chemical process fault diagnosis method based on multi-scale spatial-temporal feature fusion, which comprises the following steps: regularly acquiring fault data in a chemical production process, preprocessing the fault data and inputting the preprocessed fault data into an MTS-F model; the MTS-F network comprises a time sequence feature extraction module and a spatial feature extraction module which are parallel, and the time sequence feature extraction module comprises a long short-term memory network, a channel attention mechanism module, a one-dimensional convolution layer and a global maximum pooling layer which are connected in sequence; the spatial feature extraction module comprises a GS-CNN module, a maximum pooling layer, two groups of spatial separable convolution layers and maximum pooling layers, and a global maximum pooling layer which are connected in sequence; and after the outputs of the two parallel networks are spliced, a prediction result of fault classification is output through a discarding layer and a full connection layer. According to the method, the time sequence and space features are extracted simultaneously in a parallel mode, the fault diagnosis capability is improved, the model parameter quantity and the calculation quantity are reduced, and good generalization performance is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of chemical process fault diagnosis, and in particular relates to a chemical process fault diagnosis method integrating multi-scale spatiotemporal features. Background Art

[0002] With the rapid development of artificial intelligence, modern chemical industry has continuously improved its requirements for intelligent, large-scale and automated production processes and equipment, and the complexity and comprehensiveness of chemical systems have risen sharply. As a result, quality, safety, property and other issues have inevitably occurred. In recent years, the basic common technology research on dynamic system fault detection and abnormal diagnosis has attracted widespread attention in the chemical industry and academia. If faults can be discovered in time within the controllable range of the chemical process, and minor faults can be detected in the early stage, and fault identification and tracing can be performed, control measures such as system compensation can be used to reduce the occurrence of abnormal events. Therefore, reasonable fault detection and diagnosis (FDD) of complex dynamic systems is a key technical problem that needs to be solved urgently. At present, FDD modeling methods can be mainly divided into three categories: mechanism-driven, knowledge-driven and data-driven. However, with the widespread application of distributed control systems and the continuous development of data storage, transmission and processing technologies, facing the massive state data and monitoring variables in the production process and equipment operation, relying solely on traditional mechanism analysis methods can no longer fully extract the fault signs, causal logic and other relationships implied in the data. At the same time, faced with the characteristics of complex systems such as dynamic randomness, multi-source uncertainty, high coupling and strong interference, it is not only difficult to establish an accurate mathematical model, but also difficult to build a complete expert knowledge system.

[0003] In the era of big data, the widespread application of automation technology has accumulated a huge amount of data resources in chemical sites, making the data-driven FDD method more applicable. This method can directly and effectively perform statistical analysis and information extraction on massive, multi-source, high-dimensional data. By collecting monitoring data from different sources and different types and using various data mining techniques, implicit useful information can be extracted from them to characterize the normal mode and failure mode of system operation.

[0004] Among them, deep learning methods have gradually become a research hotspot in this field due to their powerful feature extraction capabilities. High-dimensional chemical data usually reflect both time-related patterns and spatial dependencies, and are considered as spatiotemporal sequences. At present, the research on spatiotemporal sequences based on deep learning is mainly carried out from two directions: spatial models and sequence models. In the research on spatial models, convolutional neural networks (CNNs) and improved methods have shown excellent performance; in the research on sequence models, long short-term memory (LSTM) networks and improved methods have achieved the purpose of effectively mining time series information in data. However, since it is difficult to fully capture the nonlinear, strong coupling, and strong dynamic characteristics of complex chemical processes by only extracting temporal features or spatial features, the most popular solution in the academic community is to combine sequence models and spatial models and construct fusion models in various ways. In the field of spatiotemporal information fusion research, the accurate expression, efficient extraction, and effective fusion of spatiotemporal information have important research value and significance for the development of modern chemical industry. Existing technologies mostly focus on deep feature extraction. Although this method can capture the complex patterns of data, it also leads to the problems of too many model parameters and slow reasoning speed. In addition, some methods extract temporal features and spatial features in series, which may lead to the loss of key information in the early stage of feature extraction, thus affecting the overall performance of the model. Furthermore, some methods have only been verified for their effectiveness on specific data sets and lack a systematic evaluation of their generalization capabilities, which makes their reliability and applicability in different application scenarios questionable. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a chemical process fault diagnosis method that integrates multi-scale spatiotemporal features to diagnose and identify faults in chemical process data sets.

[0006] In order to solve the above technical problems, the present invention provides a chemical process fault diagnosis method with multi-scale spatiotemporal feature fusion, comprising: at the chemical production site, time series data under different fault conditions in the chemical production process are regularly collected and recorded as fault data, and the fault data are pre-processed and input into a trained and tested MTS-F model;

[0007] The MTS-F network includes a parallel temporal feature extraction module and a spatial feature extraction module. The temporal feature extraction module includes a long short-term memory network, a channel attention mechanism module, a one-dimensional convolution layer and a global maximum pooling layer connected in sequence; the spatial feature extraction module is based on a CNNs network, including a GS-CNN module, a maximum pooling layer, a spatially separable convolution layer, a maximum pooling layer, a spatially separable convolution layer, a maximum pooling layer and a global maximum pooling layer connected in sequence; the output of the temporal feature extraction module is spliced ​​with the output of the spatial feature extraction module, and then the prediction result of the fault classification is output through a discard layer and a fully connected layer.

[0008] As an improvement of the chemical process fault diagnosis method of the present invention using multi-scale spatiotemporal feature fusion:

[0009] The preprocessing includes standardization and moving sliding window interception, and the standardization is:

[0010]

[0011] Where: Data is the fault sample data without preprocessing, μ is the mean of the normal sample data, s is the variance of the normal sample data, and Data* is the normalized data;

[0012] The moving window interception adopts a moving window with a width of L and a step size of S to slide and intercept along the time series direction to obtain a multivariate time series data set with an L×N structure, where N is the number of process variables;

[0013] The multivariate time series data is directly sent to the time series feature extraction module, and the multivariate time series data is converted into a single-channel 2D image and then sent to the spatial feature extraction module.

[0014] As a further improvement of the chemical process fault diagnosis method of the present invention using multi-scale spatiotemporal feature fusion:

[0015] The channel attention mechanism module includes compression and excitation stages, in which all feature values ​​of each feature dimension are aggregated into a single scalar.

[0016] As a further improvement of the chemical process fault diagnosis method of the present invention using multi-scale spatiotemporal feature fusion:

[0017] The GS-CNN module is the global feature extraction part of the LGS-CNN network, which includes first using a two-dimensional convolution operation. The two-dimensional convolution selects two strip convolution kernels, one with the same length as the input image height and the other with the same length as the input image width; then the feature map obtained by the two-dimensional convolution operation is calculated channel by channel through the vector outer product to obtain new features in three-dimensional form.

[0018] As a further improvement of the chemical process fault diagnosis method of the present invention using multi-scale spatiotemporal feature fusion:

[0019] The training and testing process of the MTS-F network includes:

[0020] 40 simulations are randomly selected from the training set of the TEP data set and preprocessed as the training set of the MTS-F network. 5 simulations are randomly selected from the test set of the TEP data set and preprocessed as the test set of the MTS-F network. The original category labels in the TEP data set are converted into one-hot encoding form as real labels. The training set is input into the MTS-F network to continuously optimize the network according to the cross entropy loss function to obtain weights closer to the labels. Then the test set is input into the trained model to generate predicted classification results, the differences between the predicted results and the real labels are compared, and the best results are selected.

[0021] As a further improvement of the chemical process fault diagnosis method of the present invention using multi-scale spatiotemporal feature fusion:

[0022] The moving sliding window interception adopts a width L of 60 and a step length S of 20.

[0023] The beneficial effects of the present invention are mainly reflected in:

[0024] 1. The present invention completes the chemical process fault identification and diagnosis tasks with a parameter amount far lower than that of the current mainstream network. While meeting the real-time performance required for fault identification, it significantly reduces the factory's requirements for hardware configuration; while reducing the complexity of the model, the accuracy is still improved, ensuring the reliability of the network.

[0025] 2. The method of the present invention ensures the recognition effect of small sample data, reduces the cost of data collection, and thus improves the deployment efficiency of production.

[0026] 3. The chemical process fault diagnosis network of the present invention adopts a parallel approach to simultaneously extract temporal and spatial features, thereby improving fault diagnosis capabilities and significantly reducing the amount of model parameters and calculations, while having good generalization performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0028] Figure 1 A process schematic diagram of a chemical process fault diagnosis method of the present invention using multi-scale spatiotemporal feature fusion;

[0029] Figure 2 It is a schematic diagram of the structure of the chemical process fault diagnosis network of the multi-scale spatiotemporal feature fusion of the present invention;

[0030] Figure 3 It is a structural diagram of the LSTM network;

[0031] Figure 4 It is a structural diagram of the SEnet module;

[0032] Figure 5 It is a schematic diagram of the structure of the CNNs network;

[0033] Figure 6 It is a structural diagram of the GS-CNN module;

[0034] Figure 7 Schematic diagram of the process of spatially separable convolution;

[0035] Figure 8 The accuracy graph of 5-fold cross validation for different networks (MCNN-LSTM, GF-CNN, MIC-GCN, Transformer, LGS-CNN, and MTS-F);

[0036] Fig. 9 The t-SNE visualization results of the mainstream models for all fault types (where (a) is the visualization output of MCNN-LSTM, (b) is the visualization output of GF-CNN, (c) is the visualization output of MIC-GCN, (d) is the visualization output of Transformer, and (e) is the visualization output of LGS-CNN;

[0037] Fig.10 This is the t-SNE visualization result diagram of the MTS-F model of the present invention for all fault types. DETAILED DESCRIPTION

[0038] The present invention is further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto:

[0039] Example 1: A chemical process fault diagnosis method based on multi-scale spatiotemporal feature fusion, such as Figure 1 As shown, firstly, a chemical process fault diagnosis network (MTS-F network) with multi-scale spatiotemporal feature fusion of the present invention is constructed. The data for offline modeling is the data in the Tennessee Eastman (TE) process. The data in the TE process is collected, preprocessed, and a training set and a test set are constructed. The modeling process is completed through model training, and then the trained and tested MTS-F model is applied to the actual chemical production process for fault detection.

[0040] 1. Build MTS-F network

[0041] By constructing an MTS-F network for chemical process fault diagnosis, it includes two parallel temporal feature extraction modules and spatial feature extraction modules to fully learn the temporal and spatial information in the fault data. The MTS-F network structure is as follows: Figure 2 shown.

[0042] 1.1. Constructing the time series feature extraction module

[0043] In the time series feature extraction module, the long short-term memory (LSTM) network, the channel attention mechanism module (SEnet), and the one-dimensional convolutional layer (1D CNN) are connected and integrated in sequence to fully extract the time series characteristics.

[0044] 1.1.1 Long Short-Term Memory (LSTM) Network

[0045] LSTM network is a special type of recurrent neural network (RNN) that is specifically designed to solve long sequence dependency problems. For general feedforward neural networks, there is no feedback connection between neurons in each layer, that is, the current input does not pay attention to previous information. In time series, information between different moments is likely to affect each other. Reasonable use of moment t and previous inputs to process information at moment t can obviously learn features more reasonably and fully. There are three inputs to the LSTM network: the current moment input value x t , the value of the hidden state at the previous moment h t-1 And the cell state C at the previous moment t-1 . The cell state is the memory unit in the LSTM network, which is used to save previous information and calculate new information. The LSTM network controls the flow of cell states through three gates, namely the forget gate, input gate, and output gate. First, two gates are used to control the content of the cell state. One is the forget gate, which determines how much of the hidden state of the previous moment is retained to the current moment; the other is the input gate, which determines the current moment’s input x t How much is saved to the cell state C t Next, according to the current input x t and cell state C t , determines what information to output to the next layer or output layer. The LSTM network can generate a state X containing all the information of the previous time steps at each time step. The structure is as follows Figure 3 shown.

[0046] The calculation process of the LSTM network is shown in formula (1):

[0047]

[0048] Among them, f t 、i t , C t , o t 、h tare the values ​​of the forget gate, input gate, candidate memory unit, cell state (long-term memory output), output gate, hidden state (short-term memory output) at time t, respectively. f 、Wi、W C , W o are the weights of the forget gate, input gate, candidate memory unit, and output gate, respectively. f 、b i 、b C 、b o are the biases of the forget gate, input gate, candidate memory unit, and output gate, respectively. t-1 is the value of the hidden state of the LSTM network at time t-1, x t is the input of the LSTM network at time t.

[0049] 1.1.2. Introducing the channel attention mechanism module (SEnet) after the LSTM network

[0050] After extracting features through the LSTM network, although the input feature dimensions are enhanced, not all dimensions are equally important in the final fault diagnosis task. To address this problem, it is crucial to introduce a suitable attention mechanism method to assign different weights to different feature dimensions so that the model can focus on processing more informative and discriminative features. However, the attention mechanism method usually requires a large number of parameters. In order to reduce parameters, this paper adopts the channel attention mechanism module (SEnet) mainly used in computer vision tasks. SEnet achieves its goal through two key operations: Squeeze and Excitation. SEnet can dynamically adjust the relative weights of each channel feature without increasing too much computational complexity, strengthen the feature expression that is critical to the task, and suppress irrelevant or redundant information.

[0051] In the compression phase, a global average pooling operation is performed on the feature map. This operation aggregates the feature values ​​of all spatial locations of each channel into a single scalar, resulting in a vector with the same number of channels but a spatial size of 1×1. This process compresses the spatial dimensions and extracts the global features of each channel.

[0052] In the excitation phase, the compressed vector is processed through a neural network with two fully connected layers. The first layer usually uses the ReLU activation function for nonlinear transformation, and the second layer uses the Sigmoid activation function. These two layers of fully connected networks together form an adaptive feature recalibrator to learn the importance weights of each channel.

[0053] After being processed by the fully connected network, each channel corresponds to a weight value between 0 and 1. These weights reflect the model's judgment on the importance of the corresponding channel features: a weight close to 1 means that the channel has a greater contribution to the current task, while a weight close to 0 means that the channel is relatively unimportant or contains redundant information. Finally, the channel weight vector generated in the excitation phase is multiplied element-by-element with the original feature map. In this way, channel features with high importance are amplified, while channel features with low importance are suppressed. The final weighted feature map retains the original spatial structure, but performs adaptive attention adjustment in the channel dimension.

[0054] Considering that the input of the present invention is two-dimensional data without channels, SEnet is processed to a certain extent. In the compression stage, the step of summarizing the eigenvalues ​​of all spatial positions of each channel into a single scalar is changed to summarizing all eigenvalues ​​of each feature dimension into a single scalar. Therefore, in the present invention, SEnet does not learn the importance of each channel of the feature map, but focuses on the effective information on the feature dimension of the feature map. When used, SEnet is added after the LSTM network to enhance the network's attention to important features, thereby improving the performance and generalization ability of the model. The specific steps of SEnet in the present invention are as follows: the output of the LSTM network is first subjected to a global average pooling operation in the compression stage to obtain a vector of a single scalar in which the feature dimension remains unchanged but the time series eigenvalues ​​corresponding to each feature dimension are summarized. Then, after the full connection processing in the excitation stage, each feature dimension obtains a weight value between 0 and 1. Finally, the generated feature dimension weight vector is multiplied element-by-element with the output of the LSTM network.

[0055] SEnet can add a small amount of parameters while giving importance to the characteristic dimensions of the data, thereby improving the fault diagnosis performance. Figure 4 As shown in the figure, M represents the time series length of the output of the LSTM network, and F represents the feature dimension of the output of the LSTM network.

[0056] 1.1.3 Construction of time series feature extraction module

[0057] In the time series feature extraction module, the input is first sent to a layer of LSTM network to learn and represent the short-term characteristics and long-term dependencies in the data, and then sent to a layer of SEnet module to better capture the key features in the input data, enhance the network's feature representation ability and improve the model's performance in the task. Next, a layer of one-dimensional convolution with a convolution kernel length of 1 is introduced, which helps the model better understand the time dependency in the sequence data, thereby improving the classification performance. Finally, a global maximum pooling operation is added to extract key features and reduce the number of network parameters and calculations. The final output is processed in the next step.

[0058] 1.2. Constructing spatial feature extraction module

[0059] The baseline model of the spatial feature extraction module is CNNs (such as Figure 5 As shown in Figure 2, it includes 3 convolutional layers, 3 maximum pooling layers, a global maximum pooling layer and a fully connected layer (dense layer, Dense).

[0060] 1.2.1 GS-CNN module

[0061] The GS-CNN module is based on the LGS-CNN network. The LGS-CNN network was proposed by Wahaibi et al. in 2023 and consists of two parts: a local feature extraction part and a global feature extraction part. The present invention introduces the global feature extraction part of the LGS-CNN network (referred to as the GS-CNN module in the present invention). The structure of the GS-CNN module is as follows: Figure 6 shown.

[0062] The specific operation of the GS-CNN module is as follows: using a two-dimensional convolution operation, two strip convolution kernels are selected, one of which has the same length as the height of the input image and the other has the same length as the width of the input image. The two convolution kernels convolve the input image separately and generate two independent feature maps. Then, the obtained feature map is calculated channel by channel through the vector outer product to obtain new features in three-dimensional form.

[0063] In the spatial feature extraction module, the global feature extraction module GS-CNN in the LGS-CNN network is introduced to replace the first convolutional layer of the baseline model CNNs.

[0064] 1.2.2 Spatial Separable Convolution

[0065] In order to significantly reduce the number of network parameters and computational complexity, the present invention uses spatially separable convolution (SSC) to replace ordinary convolution. The specific operation of spatially separable convolution is as follows: the convolution kernel of size n×n is divided into two smaller convolution kernels of size n×1 and 1×n. First, the convolution kernel of size n×1 is used to extract features from the input, and then the convolution kernel of size 1×n is used to extract features based on the new features. The spatially separable convolution process is as follows: Figure 7 shown.

[0066] In the spatial feature extraction module, spatially separable convolution is introduced to replace the second and third layers of ordinary convolution of the baseline model CNNs. On the basis of greatly reducing network parameters and accelerating fault diagnosis, the network's ability to further extract features is enhanced, thereby improving the accuracy of fault diagnosis.

[0067] 1.2.3 Construction of spatial feature extraction module

[0068] The spatial feature extraction module of the present invention is based on the CNNs network, and firstly introduces the GS-CNN module to replace the first convolution layer of the CNNs network, so as to achieve the purpose of learning the global features of the data.

[0069] Secondly, spatially separable convolution with a kernel size of 3×3 is used to replace the ordinary convolutions in the second and third layers of CNNs, which greatly reduces the number of model parameters and computational complexity while ensuring network performance, thereby improving the efficiency and performance of the model.

[0070] Finally, through the global maximum pooling operation, more global feature information is extracted to reduce the complexity of the model. The final output is sent to the network for the next step of processing.

[0071] 1.3 MTS-F Network

[0072] The MTS-F network of the present invention includes a parallel temporal feature extraction module and a spatial feature extraction module, such as Figure 2 As shown in the figure, the time series feature extraction module includes a long short-term memory (LSTM) network, a channel attention mechanism module (SEnet), a one-dimensional convolutional layer (1D CNN) and a global maximum pooling layer connected in sequence. Through the introduction of the long short-term memory (LSTM) network, the channel attention mechanism module (SEnet) and the one-dimensional convolutional layer (1D CNN), the long-term and short-term time characteristics contained in the time series data are fully utilized, and the network feature extraction capability is enhanced.

[0073] The spatial feature extraction module of the present invention is based on a CNNs network, and comprises a GS-CNN module, a maximum pooling layer, a spatially separable convolution layer, a maximum pooling layer, a spatially separable convolution layer, a maximum pooling layer and a global maximum pooling layer which are connected in sequence; the global features of the data are learned and extracted by introducing the GS-CNN module, and ordinary convolution is replaced by spatially separable convolution, so that the model performance is improved, the model parameters are greatly reduced, and the model detection speed is improved.

[0074] Then, the output of the temporal feature extraction module and the output of the spatial feature extraction module are spliced ​​in the last dimension, and the output information of the two modules is integrated to form a richer feature representation (including both temporal characteristics and spatial correlation), which improves the generalization ability of the model to obtain high-level features that fully represent temporal information and spatial information, and then the fault classification results are output through the dropout layer and the fully connected layer: fault category. The dropout layer is used to improve the generalization ability of the network and reduce the overfitting phenomenon of the model, so as to achieve accurate identification of different faults.

[0075] 2. Model training

[0076] 2.1、Dataset establishment

[0077] The present invention uses the TEP dataset generated by the large-sample benchmark Tennessee Eastman process (TEP) provided by Rieth et al. (2017) of Harvard University. The TEP dataset contains 52 process variables, including 41 measured variables and 11 manipulated variables. These variables come from multiple simulations of a chemical process consisting of a reactor, a condenser, a compressor, a separator, and a stripper. Table 1 shows the details of these variables.

[0078] Table 1 Variables involved in the TEP process and their descriptions

[0079]

[0080]

[0081]

[0082] The training set and test set of the TEP dataset each include 500 simulations, and each simulation includes 20 different fault types. The training set in the TEP dataset is obtained under 25h simulation, with one sample collected every 3 minutes, and the fault sample is introduced after 1h of simulation. The simulation time of the test set is 48h, and the fault sample is introduced after 8h.

[0083] The present invention randomly selects 40 simulations from the training set of the TEP data set for the training of the MTS-F network of the present invention, and randomly selects 5 simulations from the test set of the TEP data set for the testing of the present invention. The 40 simulations used for training contain a total of 16,000 groups of normal operating data and 384,000 groups of fault data, and the present invention uses 384,000 groups of fault data as the training set. The 5 simulations used for testing contain a total of 16,000 groups of normal operating data and 80,000 groups of fault data, and the present invention uses 80,000 groups of fault data as the test set. In the actual chemical industry, due to the difficulty of data collection, high cost or scarcity in a specific field, the data set size is often relatively small. Therefore, the present invention also sets a small sample performance test experiment to verify the generalization performance of the network proposed by the present invention. The present invention randomly selects 3 simulations from the training set of the TEP data set for small sample data training, and randomly selects 1 simulation from the test set of the TEP data set for small sample data testing. Therefore, the small sample training set contains 28,800 sets of fault data, and the small sample test set contains 16,000 sets of fault data.

[0084] Before training, all data of the training set and test set, small sample training set and test set of the present invention are preprocessed, including standardization and moving window interception. The calculation process of standardization is shown in formula (2):

[0085]

[0086] Where: Data is the fault sample data without preprocessing, μ is the mean of the normal sample data, s is the variance of the normal sample data, and Data* is the normalized data.

[0087] Then, a sliding window with a width of L and a step length of S is used to perform sliding interception along the time series direction. In the training set and test set, small sample training set and test set of the present invention, the data of each fault type corresponding to each simulation is processed as follows: A multivariate time series data set of L×N structure is constructed, wherein N is the number of process variables (N=52 in the present invention), T is the number of fault data of each fault type in each simulation, T=480 in the training set, and T=800 in the test set.

[0088] In addition, the multivariate time series data with dimensions of L×N is directly fed into the temporal feature extraction module; the multivariate time series data with dimensions of L×N is converted into a single-channel 2D image with dimensions of L×N and then fed into the spatial feature extraction module. In fact, each process variable represents a pixel value in the converted image.

[0089] The six data sets obtained by sliding and intercepting along the time series direction with moving windows of different widths L and step sizes S are shown in Table 2, which details the widths L and step sizes S of the moving windows of data sets 1-6 and the corresponding number of training samples and test samples. Through experimental verification, the present invention uses data 1 with a width of 60 and a step size of 20 to slide and intercept along the time series direction as the training set and test set. Data set 6 is a data set simulating a small sample situation, with only 66 training samples for each category.

[0090] Table 2 Six datasets with different structures

[0091]

[0092] 2.2 Training and testing process

[0093] The training samples in Dataset 1 of Table 2 are input into the MTS-F network for training. In the temporal feature extraction module, the temporal features are first extracted through the LSTM network, and then the SEnet module and the one-dimensional convolution layer are used to further extract and fuse the information, and finally the global maximum pooling layer generates the output. In the spatial feature extraction module, the global features are first fused through the GS-CNN module, and then two sets of spatially separable convolutions and maximum pooling layers are used to further extract features, and finally the global maximum pooling layer generates the output. The outputs of the two modules are spliced ​​and sent to the fully connected layer for classification to generate the predicted output. The original category labels in the dataset are converted into one-hot encoding form as the true labels, and the network is continuously optimized according to the cross entropy loss function to train the model to obtain weights closer to the labels.

[0094] The training epoch is set to 150, the batch size is set to 128, the optimization algorithm adopts the adaptive momentum estimation (Adam) algorithm, the preset learning rate is set to 0.001, and an early stopping mechanism is added to prevent overfitting.

[0095] The testing process is to put the test samples in Dataset 1 in Table 2 into the trained model, generate predicted classification results, compare the differences between the predicted results and the true labels, and use precision, F1 score, and error rate as evaluation indicators. If the test results meet the requirements, the model is saved for online diagnosis. If not, the model is redesigned and verified to obtain an MTS-F model that can be used online.

[0096] 3. Use the trained MTS-F model online

[0097] In order to realize the online diagnosis of chemical faults, at the chemical production site, through sensors and computers and other technical means, the values ​​of variables such as material parameters (such as pressure, corresponding to the variable XMEAS (7) in Table 1), equipment parameters (such as power usage, corresponding to the variable XMV (5) in Table 1), and state parameters (such as switch state, corresponding to the variable XMEAS (17) in Table 1) in the chemical production process are collected regularly to obtain the time series data of each variable under different fault conditions, that is, fault data. The collected real-time data is standardized and preprocessed by a moving window, where the width L of the moving window is 60 and the step length S is 20. The preprocessed data is input into the MTS-F model that can be used online in step 2, so as to obtain the fault classification diagnosis results of real-time data. The collected data is saved for model maintenance and updating.

[0098] experiment:

[0099] 1. Evaluation indicators

[0100] The experimental evaluation indicators include the trainable parameters (Trainable params) from the perspective of the model and the precision (Precision, P), F1 score (F1 score, F), and error rate (error rate, E) from the perspective of data. The F1score formula is shown in formula (3):

[0101]

[0102] Among them, P is the precision rate, which refers to the ratio of the number of samples that are actually positive to the number of samples that are predicted to be positive among all the samples predicted by the model; R is the recall rate, which refers to the proportion of samples that are successfully predicted as positive among all the samples that are actually positive.

[0103] The error rate formula is shown in formula (4):

[0104] E=1-A (4)

[0105] Where A is the accuracy, which is the ratio of the number of samples correctly predicted by the model to the total number of samples, and is used to measure the degree of overall classification accuracy.

[0106] 2. Mobile window comparison experiment

[0107] In order to verify the influence of different width and step size settings of the moving window on the model of the present invention, the performance of the MTS-F model is compared on the data sets 1-5 shown in Table 2. The comparative test results are shown in Table 3.

[0108] Table 3 Comparison of moving window performance with different settings

[0109]

[0110] As shown in Table 3, on Dataset 1 after sliding interception along the time series direction using a moving window with a width of 60 and a step size of 20, the MTS-F model of the present invention has the best performance, with an accuracy of 98.2% and an F1 score of 98.16%. Although the number of parameters of the model on Dataset 1 is higher than that on Datasets 3, 4, and 5, it performs better in terms of the accuracy of the fault diagnosis task.

[0111] 3. Comparative experiment

[0112] The performance of the MTS-F model of the present invention is compared with that of five other mainstream fault diagnosis models (MCNN-LSTM, GF-CNN, MIC-GCN, Transformer, and LGS-CNN), and the experiment uses the data set 1 shown in Table 2. The comparative test results are shown in Table 4. In order to reduce the randomness of the experimental results, we repeated each experiment 8 times and averaged the obtained results.

[0113] Table 4 Performance comparison of mainstream fault diagnosis models

[0114]

[0115]

[0116] In order to verify the generalization performance of different networks on small sample data sets, the MTS-F model is compared with mainstream fault diagnosis models (MCNN-LSTM, GF-CNN, MIC-GCN, Transformer and LGS-CNN) on Dataset 6. The comparative test results are shown in Table 5.

[0117] Table 5 Comparison of generalization performance of mainstream fault diagnosis models

[0118]

[0119] By analyzing Table 4, it can be seen that the MTS-F model of the present invention has the best performance, with an accuracy rate and F1 score of 98.2%. For each type of fault, MTS-F has achieved good results. Especially for the more difficult to detect faults 3, 9, 15, and 18, MTS-F has achieved a higher accuracy rate compared to other models. Therefore, it can be proved that the MTS-F network proposed in the present invention has a higher detection accuracy. At the same time, it can be seen from the analysis of Table 4 that MTS-F requires fewer parameters than other networks. In addition, by analyzing the indicators in Table 5, it can be seen that compared with other models, the MTS-F network proposed in the present invention still maintains the best effect in the case of small samples, proving that the MTS-F network has good generalization performance.

[0120] 4. Ablation experiment

[0121] In order to evaluate the effects of different modules of the MTS-F model, an ablation experiment was conducted. The ablation experiment used data set 1. The ablation experiment comparison results of the present invention are shown in Table 6, where:

[0122] CNNs is the benchmark model. G-CNNs is a network that introduces the GS-CNN module to replace the first convolution layer of CNNs on the benchmark model CNNs. S-CNNs is a network that introduces spatially separable convolution to replace the second and third layers of ordinary convolution of CNNs on the G-CNNs model. The L-CNNs network includes two parallel branches, one branch is S-CNNs, and the other branch is the LSTM layer. The outputs of the two branches are spliced ​​and then output through the fully connected layer. LA-CNNs is a network that introduces SEnet after the LSTM layer of the L-CNNs network. MTS-F is the chemical fault diagnosis network of the present invention, which introduces the Dropout operation on the basis of LA-CNNs.

[0123] Comparing G-CNNs and CNNs, we can see that after introducing the GS-CNN module to replace the first convolutional layer of CNNs, the F1 score index increased by 2.97%, the error rate decreased by 2.89%, and the feature extraction ability of the model was improved, indicating that chemical data is rich in global spatial correlation. G-CNNs can learn global features, while CNNs focus more on learning local features.

[0124] Comparing S-CNNs and G-CNNs, we can see that after replacing the second and third layers of ordinary convolution in CNNs with spatially separable convolutions, although the F1 score index dropped by 1.34%, the number of model parameters was reduced by 83.69%. Therefore, the introduction of spatially separable convolutions is desirable.

[0125] By comparing L-CNNs and G-CNNs, we can see that after introducing the LSTM layer of the time series feature extraction module, the F1 score index increased by 3.47%, the error rate decreased by 3.31%, and the feature extraction ability of the model was improved, indicating that the introduction of the time series feature extraction module is necessary for chemical data.

[0126] By comparing LA-CNNs and L-CNNs, we can see that after introducing SEnet in the time series feature extraction module, the number of model parameters increased by 3.88%, the F1 score index increased by 0.8%, and the error rate decreased by 0.84%. This shows that the introduction of SEnet can improve the feature extraction ability of the model while increasing a small amount of parameters.

[0127] Comparing MTS-F and LA-CNNs, we can see that after the introduction of the Dropout layer, the F1 score index increased by 0.3%, indicating that the introduction of the Dropout layer can improve the feature extraction capability of the model without increasing the number of parameters.

[0128] Comparing MTS-F and CNNs, we can see that the proposed model has 70.31% fewer parameters and a 6.2% increase in F1score. This shows that the improved network in this paper achieves higher detection accuracy with fewer parameters.

[0129] Table 6 MTS-F ablation experiment results

[0130]

[0131]

[0132] 5. Generalization performance experiment supplement

[0133] In order to more comprehensively evaluate the generalization performance of the network of the present invention, the k-fold cross-validation method (k=5) was used in the experiment. Specifically, the training set of data set 1 was divided into 5 subsets of equal size. In each iteration, 4 subsets were used for model training, and the remaining 1 subset was used for validation. This process was repeated 5 times, so that each subset was used as a validation set. Figure 8 The experimental results show that the network of the present invention performs well in the 5-fold cross validation, with an average accuracy of 98.97% and a standard deviation of 0.27%, which indicates that the performance of MTS-F on different data sets is relatively consistent, further verifying that the network has good generalization performance.

[0134] 5. Visualization

[0135] In order to further investigate the superiority of the MTS-F network, the outputs of the proposed model and five mainstream fault diagnosis models are visualized using the t-distributed stochastic neighbor embedding (t-SNE) method. Fig. 9 The t-SNE visualization results of the output layers of five mainstream models MCNN-LSTM, GF-CNN, MIC-GCN, Transformer, and LGS-CNN for dataset 1 are shown in the figure. The fault samples shown are from the test set. The output is converted into a two-dimensional vector through t-SNE and represented in two-dimensional coordinates. Each point represents a sample, and samples with different faults are represented by different colors to distinguish them.

[0136] from Fig. 9From (a), we can see that for the output data of this network model, except for a few categories with slightly obvious differences, there is no obvious separation boundary between different types of fault data. From (b), (c), (d), and (e), we can see that the same type of fault data is clustered together, the fault category has taken shape, the interval between different categories is further expanded, and the boundary interval between different types of fault data is more obvious. However, it is still difficult to identify at least three types of faults. Fig.10 The t-SNE visualization results of the output of the MTS-F model proposed in this invention after the test set samples in data set 1 are shown. The results show that the same fault type is more closely clustered, the intervals between categories are very clear, and the model classification effect is significant. Only a small number of samples of fault types 3, 9, and 15 in the black box have some classification errors. From the t-SNE visualization results, we can draw the following conclusions: The MTS-F network is an effective means to fully extract spatiotemporal fusion information. Fig.10 It can be seen that the improved network proposed in this paper reduces the occurrence of misclassification and has stronger fault diagnosis capabilities.

[0137] Finally, it should be noted that the above examples are only some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by a person skilled in the art should be considered as the protection scope of the present invention.

Claims

1. A chemical process fault diagnosis method based on multi-scale spatiotemporal feature fusion, characterized by: The process includes: at the chemical production site, time series data under different fault conditions in the chemical production process are collected regularly and recorded as fault data, and the fault data are input into the trained and tested MTS-F model after preprocessing; The MTS-F network includes a parallel temporal feature extraction module and a spatial feature extraction module, wherein the temporal feature extraction module includes a long short-term memory network, a channel attention mechanism module, a one-dimensional convolution layer and a global maximum pooling layer connected in sequence; the spatial feature extraction module is based on a CNNs network, including a GS-CNN module, a maximum pooling layer, a spatially separable convolution layer, a maximum pooling layer, a spatially separable convolution layer, a maximum pooling layer and a global maximum pooling layer connected in sequence; The output of the temporal feature extraction module is concatenated with the output of the spatial feature extraction module, and then passes through the discard layer and the fully connected layer to output the prediction result of the fault classification.

2. The chemical process fault diagnosis method of multi-scale spatiotemporal feature fusion according to claim 1 is characterized by: The preprocessing includes standardization and moving sliding window interception, and the standardization is: Where: Data is the fault sample data without preprocessing, μ is the mean of the normal sample data, s is the variance of the normal sample data, and Data* is the normalized data; The moving window interception adopts a moving window with a width of L and a step size of S to slide and intercept along the time series direction to obtain a multivariate time series data set with an L×N structure, where N is the number of process variables; The multivariate time series data is directly sent to the time series feature extraction module, and the multivariate time series data is converted into a single-channel 2D image and then sent to the spatial feature extraction module.

3. The chemical process fault diagnosis method based on multi-scale spatiotemporal feature fusion according to claim 2 is characterized by: The channel attention mechanism module includes compression and excitation stages, in which all feature values ​​of each feature dimension are aggregated into a single scalar.

4. The chemical process fault diagnosis method of multi-scale spatiotemporal feature fusion according to claim 3 is characterized by: The GS-CNN module is the global feature extraction part of the LGS-CNN network, which includes first using a two-dimensional convolution operation. The two-dimensional convolution selects two strip convolution kernels, one with the same length as the input image height and the other with the same length as the input image width; then the feature map obtained by the two-dimensional convolution operation is calculated channel by channel through the vector outer product to obtain new features in three-dimensional form.

5. The chemical process fault diagnosis method of multi-scale spatiotemporal feature fusion according to claim 4 is characterized by: The training and testing process of the MTS-F network includes: 40 simulations are randomly selected from the training set of the TEP data set and preprocessed as the training set of the MTS-F network. 5 simulations are randomly selected from the test set of the TEP data set and preprocessed as the test set of the MTS-F network. The original category labels in the TEP data set are converted into one-hot encoding form as real labels. The training set is input into the MTS-F network to continuously optimize the network according to the cross entropy loss function to obtain weights closer to the labels. Then the test set is input into the trained model to generate predicted classification results, the differences between the predicted results and the real labels are compared, and the best results are selected.

6. The chemical process fault diagnosis method of multi-scale spatiotemporal feature fusion according to claim 5 is characterized by: The moving sliding window interception adopts a width L of 60 and a step length S of 20.

Images