Method and device for industrial process fault detection based on time convolution autoencoder

Abstract

The application discloses an industrial process fault detection method and device based on a time convolution autoencoder, and belongs to the field of industrial process detection. The method comprises the following steps: pre-processing training data and test data to obtain training time series data and test time series data; inputting the training time series data into an industrial process fault detection model to train a target model, wherein the industrial process fault detection model comprises at least two layers of TCN and CAE; inputting the training time series data into the target model to calculate a control limit; inputting the test time series data into the target model to calculate an SPE statistic; and judging whether the to-be-detected working condition data is abnormal according to the control limit and the SPE statistic. The application introduces TCN on the basis of CAE, designs a novel time convolution autoencoder, and creates an industrial process fault detection model. At least two stacked time convolution networks can gradually expand the convolution receptive field to change the time scale of time series feature extraction, so that time series modeling can be more accurately realized.

Description

Technical Field

[0001] This invention relates to a method and apparatus for industrial process fault detection based on a temporal convolutional autoencoder, belonging to the field of industrial process detection technology. Background Technology

[0002] To maintain the safe and reliable operation of industrial processes and the stability of product quality, effective fault detection technologies are crucial. With the continuous improvement of automation and informatization levels in industrial processes, a wealth of data information has accumulated during production, providing strong support for data-driven fault detection methods. Data-driven fault detection methods utilize data analysis and preprocessing techniques, leveraging prior knowledge from historical data to construct statistics from real-time data for process monitoring. As a primary method for data-driven fault detection, multivariate statistical methods have been widely applied in the industrial process field. Representative methods include principal component analysis, partial least squares, and independent component analysis, as well as corresponding derived nonlinear methods such as kernel principal component analysis and kernel independent component analysis. However, their feature extraction capabilities are limited, and selecting suitable kernel functions is difficult.

[0003] Deep learning, with its ability to more effectively handle complex nonlinear data and excellent adaptability, has attracted widespread attention in the field of fault detection. Commonly used deep learning methods include deep belief networks, convolutional neural networks, recurrent neural networks, and autoencoders. Among them, autoencoders can effectively reduce data dimensionality and extract features, but they cannot fully consider the local features of the data. To enhance the network's ability to extract local features, researchers have combined convolutional neural networks and autoencoders, using convolutional layers instead of fully connected layers to propose the convolutional autoencoder (CAE). CAE has strong nonlinear feature extraction capabilities and better generalization ability and computational efficiency when processing high-dimensional data. However, the improved CAE method does not fully consider the time-dependent nature of the data, which limits its application in industrial process fault detection that requires processing time-series data.

[0004] To effectively extract temporal features from sequence data, fault detection methods based on convolutional long short-term memory (LSTM) autoencoders use LSM networks to learn effective features of temporal data. Autoencoders based on convolutional gated recurrent units (GRUs) and attention combine convolutional neural networks and GRUs to extract temporal features from complex process data. Related researchers have proposed temporal convolutional networks (TCNs) based on one-dimensional convolutional neural networks. TCNs employ dilated causal convolutions to skip some convolutional units, allowing the convolutional operations to cover a larger sequence range, thereby effectively capturing long-term temporal information of the data.

[0005] The above fault detection methods have strong generalization ability and can reconstruct sampled data well. However, these methods do not fully consider the multi-scale characteristics of time series data, and can only extract features from a single scale. If features are extracted from a smaller time scale, only the short-term characteristics of the time series data can be considered, while its long-term characteristics are ignored; if features are extracted from a larger time scale, only the long-term characteristics of the time series data can be considered, while its short-term characteristics are ignored. Therefore, the features extracted by the above methods are relatively singular, and the feature extraction capability is poor, resulting in low accuracy in industrial process fault detection.

[0006] Therefore, how to improve the accuracy of fault detection in industrial processes has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the problems of limited feature scale extraction and low detection accuracy in current deep learning-based industrial process fault detection methods, this invention provides an industrial process fault detection method and apparatus based on a temporal convolutional autoencoder.

[0008] As one aspect of the present invention, an industrial process fault detection method based on a temporal convolutional autoencoder is provided, comprising:

[0009] The acquired normal operating condition data of the industrial process is used as training data, and the acquired test operating condition data of the industrial process is used as test data. The training data and the test data are preprocessed respectively to obtain training time series data and test time series data.

[0010] The training time series data is input into the industrial process fault detection model for training, and the target industrial process fault detection model is obtained based on the training results. The industrial process fault detection model is constructed based on a convolutional autoencoder and at least two layers of temporal convolutional networks.

[0011] The training time series data is input into the target industrial process fault detection model, and the control limit of the squared prediction error statistic of the training time series data is calculated.

[0012] The test time series data is input into the target industrial process fault detection model, and the squared prediction error statistic of the test time series data is calculated.

[0013] Based on the control limits of the squared prediction error statistic of the training time series data and the squared prediction error statistic of the test time series data, it is determined whether the test condition data is abnormal.

[0014] Furthermore, the industrial process fault detection model includes an encoder and a decoder; the output of the encoder is connected to the input of the decoder.

[0015] The encoder is used to extract temporal features of the training time series data based on at least two temporal convolutional network layers with different dilation coefficients, and to obtain the latent variables corresponding to the training time series data based on the extracted temporal features of the training time series data.

[0016] The decoder is used to restore the latent variables output by the encoder to temporal features based on at least two temporal convolutional network layers with different dilation coefficients, and to obtain reconstructed data with the same dimension as the training temporal data based on the restored temporal features.

[0017] Furthermore, the encoder includes a first temporal convolutional network layer, a second temporal convolutional network layer, a third temporal convolutional network layer, and a first one-dimensional convolutional network layer connected in sequence;

[0018] The decoder includes a fourth temporal convolutional network layer, a fifth temporal convolutional network layer, a sixth temporal convolutional network layer, and a second one-dimensional convolutional network layer connected in sequence.

[0019] The first one-dimensional convolutional network layer is connected to the fourth temporal convolutional network layer;

[0020] The dilation coefficients of the first temporal convolutional network layer and the sixth temporal convolutional network layer are 2. x The dilation coefficients of the second and fifth temporal convolutional network layers are 2. y The dilation coefficients of the third and sixth temporal convolutional network layers are 2. z x < y < z;

[0021] The first temporal convolutional network layer, the second temporal convolutional network layer, and the third temporal convolutional network layer are all used to extract the temporal features of the training temporal data at the corresponding time scale according to the corresponding dilation coefficient;

[0022] The first one-dimensional convolutional network layer is used to compress the feature dimension of the temporal features of the training time series data output by the third temporal convolutional network layer, and to obtain the latent variables corresponding to the training time series data;

[0023] The fourth, fifth, and sixth temporal convolutional network layers are all used to restore the latent variables output by the first one-dimensional convolutional network layer to temporal features according to the corresponding dilation coefficients at the corresponding time scales.

[0024] The second one-dimensional convolutional network layer is used to obtain reconstructed data of the same dimension as the training temporal data based on the restored temporal features output by the sixth temporal convolutional network layer.

[0025] Furthermore, the encoder further includes a first feature fusion module, and the decoder further includes a second feature fusion module;

[0026] The first temporal convolutional network layer, the second temporal convolutional network layer, the third temporal convolutional network layer, the first feature fusion module, and the first one-dimensional convolutional network layer are connected in sequence;

[0027] The fourth temporal convolutional network layer, the fifth temporal convolutional network layer, the sixth temporal convolutional network layer, the second feature fusion module, and the second one-dimensional convolutional network layer are connected in sequence;

[0028] The first feature fusion module is used to weightedly fuse the temporal features of the training time series data extracted by the first temporal convolutional network layer, the second temporal convolutional network layer and the third temporal convolutional network layer respectively, to obtain the final temporal features of the training time series data;

[0029] The first one-dimensional convolutional network layer is used to compress the feature dimension of the final temporal features of the training temporal data output by the first feature fusion module, and to obtain the latent variables corresponding to the training temporal data.

[0030] The second feature fusion module is used to weightedly fuse the temporal features restored by the fourth temporal convolutional network layer, the fifth temporal convolutional network layer and the sixth temporal convolutional network layer respectively, to obtain the final temporal features of the restored training temporal data;

[0031] The second one-dimensional convolutional network layer is used to obtain reconstructed data of the same dimension as the training time series data based on the final time series features of the restored training time series data.

[0032] Furthermore, the first feature fusion module includes a first fusion unit, a first weight calculation unit, and a first weighted output unit; the second feature fusion module includes a second fusion unit, a second weight calculation unit, and a second weighted output unit.

[0033] The first fusion unit is used to obtain a first fusion feature based on the temporal features output by the first temporal convolutional network layer, the second temporal convolutional network layer and the third temporal convolutional network layer respectively;

[0034] The first weight calculation unit is used to calculate the weights corresponding to the first temporal convolutional network layer, the second temporal convolutional network layer, and the third temporal convolutional network layer respectively based on the first fusion feature through an efficient channel attention mechanism;

[0035] The first weighted output unit is used to perform weighted fusion of the temporal features output by the first temporal convolutional network layer, the second temporal convolutional network layer, and the third temporal convolutional network layer according to the weights corresponding to the first temporal convolutional network layer, the second temporal convolutional network layer, and the third temporal convolutional network layer, respectively, to obtain the final temporal features of the training temporal data;

[0036] The second fusion unit is used to obtain a second fusion feature based on the restored temporal features output by the fourth temporal convolutional network layer, the fifth temporal convolutional network layer, and the sixth temporal convolutional network layer, respectively.

[0037] The second weight calculation unit is used to calculate the weights corresponding to the fourth temporal convolutional network layer, the fifth temporal convolutional network layer, and the sixth temporal convolutional network layer respectively based on the second fusion feature through an efficient channel attention mechanism;

[0038] The second weighted output unit is used to perform weighted fusion of the temporal features output by the fourth, fifth, and sixth temporal convolutional network layers according to the weights corresponding to the fourth, fifth, and sixth temporal convolutional network layers, respectively, to obtain the final temporal features of the restored training temporal data.

[0039] Further, the step of inputting the training time-series data into the industrial process fault detection model for training, and obtaining the target industrial process fault detection model based on the training results, includes:

[0040] The training time series data is input into the industrial process fault detection model for forward calculation to obtain the reconstructed data corresponding to the training time series data.

[0041] The loss function value of the reconstructed data corresponding to the training time series data is calculated using the mean squared error function.

[0042] The industrial process fault detection model is updated using gradients based on the loss function value to obtain the target industrial process fault detection model.

[0043] Further, the step of performing gradient updates on the industrial process fault detection model based on the loss function value to obtain the target industrial process fault detection model includes:

[0044] Based on the loss function value, the Adam optimization algorithm is used to adjust the network weights of the industrial process fault detection model to obtain the target network weights.

[0045] Based on the loss function value, the network hyperparameters of the industrial process fault detection model are adjusted using a grid search method to obtain the target hyperparameters;

[0046] The target industrial process fault detection model is obtained based on the target network weights and the target hyperparameters.

[0047] Furthermore, the preprocessing includes data standardization and sliding window resampling.

[0048] Further, the step of determining whether the test condition data is abnormal based on the control limit of the squared prediction error statistic of the training time series data and the squared prediction error statistic of the test time series data includes:

[0049] When the squared prediction error statistic of the test time series data is greater than the control limit of the squared prediction error statistic of the training time series data, the test condition data is judged to be abnormal.

[0050] When the squared prediction error statistic of the test time series data is less than or equal to the control limit of the squared prediction error statistic of the training time series data, the test condition data is judged to be normal.

[0051] As another aspect of the present invention, an industrial process fault detection device based on a temporal convolutional autoencoder is provided, applied to the above-mentioned industrial process fault detection method based on a temporal convolutional autoencoder, comprising:

[0052] The preprocessing module is used to preprocess the acquired normal operating condition data of the industrial process as training data to obtain training time series data, and to preprocess the acquired test operating condition data of the industrial process to obtain test time series data.

[0053] The model training module is used to input the training time series data into the industrial process fault detection model for training, and obtain the target industrial process fault detection model based on the training results. The industrial process fault detection model is constructed based on a temporal convolutional network and a convolutional autoencoder.

[0054] The control limit calculation module is used to input the training time series data into the target industrial process fault detection model and calculate the control limit of the squared prediction error statistic of the training time series data.

[0055] The squared prediction error statistic calculation module is used to input the test time series data into the target industrial process fault detection model and calculate the squared prediction error statistic of the test time series data.

[0056] The judgment module is used to determine whether the test condition data is abnormal based on the control limit of the squared prediction error statistic of the training time series data and the squared prediction error statistic of the test time series data.

[0057] The beneficial effects of this invention are:

[0058] (1) The industrial process fault detection method based on temporal convolutional autoencoder provided by the present invention further introduces temporal convolutional network on the basis of convolutional autoencoder, designs a new type of temporal convolutional autoencoder (TCAE), and creates an industrial process fault detection model based on the new type of temporal convolutional autoencoder. After at least two stacked temporal convolutional networks, the convolutional receptive field can be gradually expanded to change the time scale of temporal feature extraction, and achieve temporal modeling more accurately.

[0059] (2) The industrial process fault detection method based on temporal convolutional autoencoder provided by this invention adds a TCN structure to the basic convolutional autoencoder network to construct the TCAE model, and designs a feature fusion module. By using skip connections, the feature fusion module is combined with the TCAE to establish an industrial process fault detection model based on feature fusion and temporal convolutional autoencoder (TCAE-FF). This model can make full use of temporal features at different scales, construct feature-rich temporal data through encoding and decoding, and obtain significant feature representations of the input data. This increases the degree of differentiation of reconstruction errors of normal data and fault data, enhances the model's ability to identify fault data, and improves the accuracy of fault detection. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart of an industrial process fault detection method based on a temporal convolutional autoencoder provided in an embodiment of the present invention;

[0062] Figure 2 This is a diagram of the temporal convolutional network structure provided in an embodiment of the present invention;

[0063] Figure 3 This is a flowchart of the control limit calculation steps provided in an embodiment of the present invention;

[0064] Figure 4 This is a flowchart of the fault diagnosis steps provided in an embodiment of the present invention;

[0065] Figure 5 This is a structural diagram of the TCAE provided in an embodiment of the present invention;

[0066] Figure 6This is a structural diagram of TCAE-FF provided in an embodiment of the present invention;

[0067] Figure 7 This is a structural diagram of the first feature fusion module provided in an embodiment of the present invention;

[0068] Figure 8 This is a structural diagram of the second feature fusion module provided in an embodiment of the present invention;

[0069] Figure 9 This is a flowchart of obtaining a target industrial process fault detection model provided in an embodiment of the present invention;

[0070] Figure 10 This is a flowchart of the gradient update steps provided in an embodiment of the present invention.

[0071] Figure 11 This is a schematic diagram of the TE industrial process provided in an embodiment of the present invention;

[0072] Figure 12 This is a schematic diagram of the TE industrial process fault detection results based on KPCA provided in an embodiment of the present invention;

[0073] Figure 13 This is a schematic diagram of the CAE-based TE industrial process fault detection results provided in an embodiment of the present invention;

[0074] Figure 14 This is a schematic diagram of the TE industrial process fault detection results based on LSTM-AE provided in an embodiment of the present invention;

[0075] Figure 15 This is a schematic diagram of the TE industrial process fault detection results based on TCAE provided in an embodiment of the present invention;

[0076] Figure 16 This is a schematic diagram of the TE industrial process fault detection results based on TCAE-FF provided in an embodiment of the present invention;

[0077] Figure 17 This is a structural diagram of an industrial process fault detection device based on a temporal convolutional autoencoder provided in an embodiment of the present invention. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0079] This invention provides an industrial process fault detection method based on a temporal convolutional autoencoder, such as... Figure 1 As shown, it includes:

[0080] S100. Use the acquired normal operating condition data of the industrial process as training data and the acquired test operating condition data of the industrial process as test data. Preprocess the training data and test data respectively to obtain training time series data and test time series data.

[0081] Specifically, the preprocessing in this embodiment mainly includes data standardization and sliding window resampling. For training data, the mean and standard deviation of the training data are first calculated, and then z-score (z-score) standardization is performed on the training data based on the mean and standard deviation to obtain standardized training data. Then, the standardized training data is resampled using a sliding window to obtain the corresponding training time-series data. Similarly, for test data, the mean and standard deviation of the test data are first calculated, and then z-score standardization is performed on the test data based on the mean and standard deviation. Then, the test data is resampled using a sliding window with the same width as the training data to obtain the test time-series data.

[0082] S200. Input the training time series data into the industrial process fault detection model for training, and obtain the target industrial process fault detection model based on the training results. The industrial process fault detection model is constructed based on a convolutional autoencoder and at least two layers of temporal convolutional networks.

[0083] A convolutional autoencoder (CAE) is an unsupervised deep neural network model whose basic structure consists of an encoder and a decoder. The encoder extracts features from the input data, and its output data dimension is generally smaller than the input dimension, achieving dimensionality reduction. The decoder, on the other hand, transforms the feature representation output by the encoder into the same dimension as the input data, thus reconstructing the input data.

[0084] Temporal convolutional networks are based on one-dimensional convolutional neural networks, incorporating dilated convolution and causal convolution structures to effectively extract temporal information. Their network structure is as follows: Figure 2 As shown, the temporal convolutional network employs a causal convolutional structure. By padding with zeros, it ensures that the output at each time step depends only on the current time step and previous historical data, thus avoiding the future information leakage problem that may arise from non-causal convolution and guaranteeing the rationality of temporal series prediction. To handle long-term sequence tasks, the temporal convolutional network also incorporates a dilated convolutional structure. By setting dilation coefficients, holes are added to the convolutional kernel, significantly expanding the receptive field of the convolutional operation. For a one-dimensional sequence input x∈R... n Given a convolution kernel f:{0,…,k-1}, the expression for calculating F by the dilated convolution of sequence elements s is:

[0085]

[0086] In the formula, k is the kernel size, d is the dilation coefficient, and x s-diThis represents the sequence corresponding to the elements of the convolution kernel. Dilated convolution expands the receptive field of the convolution by continuously increasing the dilation coefficient d, and in a typical temporal convolutional network structure, the dilation coefficient of each layer increases with d = 2. l Increase in the order of (l = 0, 1, 2, ...). For example Figure 2 As shown, the first convolutional layer has a receptive field size of 3, processing data from 3 time points at a time; the second convolutional layer has a receptive field size of 2, increasing to 7; and the third convolutional layer has a receptive field size of 15, increasing to 4, and extracting features from the input data at 15 time points.

[0087] This invention further introduces a temporal convolutional network (TCAE) on top of a convolutional autoencoder, designing a novel TCAE. Based on this novel TCAE, an industrial process fault detection model is created. By using at least two stacked TCAE networks, the receptive field of the convolution can be gradually expanded to change the time scale of temporal feature extraction, achieving more accurate temporal modeling and improving fault response capability and fault detection accuracy. Specifically, improving the accuracy of temporal modeling means improving the model's accuracy in reconstructing normal process data. When fault data is input, the reconstruction error will be significantly larger. Combining the reconstruction error to construct the SPE statistic can more clearly distinguish between normal and fault data. Generally, the more accurately the model reconstructs normal data, the smaller the corresponding SPE statistic, and the smaller the control limit. However, when fault data is input, the model's reconstruction error is larger, and the corresponding SPE statistic is larger, making it easier to exceed the control limit, thus enabling timely detection of fault data.

[0088] S300. Input the training time series data into the target industrial process fault detection model and calculate the control limit of the squared prediction error statistic of the training time series data.

[0089] Specifically, such as Figure 3 As shown, step S300 further includes:

[0090] S310. Calculate the squared prediction error (SPE) of the training time series data based on the reconstruction error of the target industrial process fault detection model, as shown in the following formula (2):

[0091]

[0092] Where L is the width of the sliding window, e t The input x of the autoencoder at time t t and reconstruction difference.

[0093] S320. Calculate the control limits of the squared prediction error statistic of the training time series data by kernel density estimation.

[0094] Specifically, as shown in formula (3):

[0095]

[0096] In the formula, n is the number of samples, x i Let K be the sample point, K(·) be the kernel function, and h be the bandwidth.

[0097] S400. Input the test timing data into the target industrial process fault detection model and calculate the squared prediction error statistic of the test timing data.

[0098] Similarly, after inputting the test time series data into the target industrial process fault detection model, refer to the above formula (2) to calculate the squared prediction error statistic of the test time series data based on the reconstruction error of the target industrial process fault detection model.

[0099] S500. Based on the control limits of the squared prediction error statistic of the training time series data and the squared prediction error statistic of the test time series data, determine whether the test condition data is abnormal.

[0100] Specifically, such as Figure 4 As shown, step S500 further includes:

[0101] S510. When the squared prediction error statistic of the test time series data is greater than the control limit of the squared prediction error statistic of the training time series data, the test condition data is judged to be abnormal.

[0102] S520. When the squared prediction error statistic of the test time series data is less than or equal to the control limit of the squared prediction error statistic of the training time series data, the test condition data is judged to be normal.

[0103] In summary, this invention introduces a novel temporal convolutional autoencoder (TCAE) by further incorporating a temporal convolutional network on top of a convolutional autoencoder. Based on this novel TCAE, an industrial process fault detection model is created and trained for fault detection of industrial process-related data. By using at least two stacked TCAE networks, the receptive field of the convolution can be gradually expanded to change the time scale for extracting temporal features from industrial process-related data, enriching the time scale for extracting temporal features, and more accurately realizing temporal modeling. This improves the feature extraction capability of the industrial process fault detection model and further enhances the accuracy of industrial process fault detection.

[0104] Furthermore, such as Figure 5As shown, the industrial process fault detection model includes an encoder and a decoder; the output of the encoder is connected to the input of the decoder.

[0105] The encoder is used to extract the temporal features of the training time series data based on at least two temporal convolutional network layers with different dilation coefficients, and to obtain the latent variables corresponding to the training time series data based on the extracted temporal features of the training time series data.

[0106] The decoder is used to restore the latent variables output by the encoder to temporal features based on at least two temporal convolutional network layers with different dilation coefficients, and to obtain reconstructed data with the same dimension as the training temporal data based on the restored temporal features.

[0107] In addition to enriching the temporal scales for extracting and reconstructing the temporal features of training time-series data at different time scales by setting at least two temporal convolutional network layers with different dilation coefficients in the encoder and decoder respectively, and achieving more accurate temporal modeling, this invention also introduces residual connections into the temporal convolutional network to form a residual block structure. Specifically, a complete residual block structure includes two sets of dilated causal convolutions, weight normalization, activation functions, and random deactivation steps. The temporal convolutional network with residual connections can retain a large receptive field with fewer network layers, fully exploring the deep features of time-series data. At the same time, residual connections can effectively suppress problems such as overfitting and gradient explosion during the increase of network depth.

[0108] Furthermore, such as Figure 5 As shown, the encoder includes a first temporal convolutional network layer TCN1, a second temporal convolutional network layer TCN2, a third temporal convolutional network layer TCN3, and a first one-dimensional convolutional network layer Cov1 connected in sequence.

[0109] The decoder consists of a fourth temporal convolutional network layer TCN4, a fifth temporal convolutional network layer TCN5, a sixth temporal convolutional network layer TCN6, and a second one-dimensional convolutional network layer Cov2, which are connected in sequence.

[0110] The first one-dimensional convolutional network layer Cov1 is connected to the fourth temporal convolutional network layer TCN4;

[0111] The dilation coefficients of the first-time convolutional network layer TCN1 and the sixth-time convolutional network layer TCN4 are 2. x The dilation coefficients of the second temporal convolutional network layer TCN2 and the fifth temporal convolutional network layer TCN5 are 2. y The dilation coefficients of the third temporal convolutional network layer TCN3 and the fourth temporal convolutional network layer TCN6 are 2. z x < y < z;

[0112] The first temporal convolutional network layer TCN1, the second temporal convolutional network layer TCN2, and the third temporal convolutional network layer TCN3 are all used to extract the temporal features of the training temporal data according to the corresponding dilation coefficients at the corresponding time scales.

[0113] The first one-dimensional convolutional network layer Cov1 is used to compress the feature dimension of the temporal features of the training time series data output by the third temporal convolutional network layer, and to obtain the latent variables corresponding to the training time series data.

[0114] The fourth temporal convolutional network layer TCN4, the fifth temporal convolutional network layer TCN5, and the sixth temporal convolutional network layer TCN6 are all used to restore the latent variables output by the first one-dimensional convolutional network layer Cov1 to temporal features according to the corresponding dilation coefficients at the corresponding time scales.

[0115] The second one-dimensional convolutional network layer, Cov2, is used to obtain reconstructed data of the same dimension as the training temporal data based on the restored temporal features output by the sixth temporal convolutional network layer, TCN6.

[0116] In this embodiment of the invention, each TCN layer is a complete residual block structure, and its corresponding dilation coefficient d varies exponentially by 2. Specifically, in this embodiment, let x = 0, y = 1, z = 2, that is, the dilation coefficients of the first temporal convolutional network layer TCN1, the second temporal convolutional network layer TCN2, and the third temporal convolutional network layer TCN3 in the encoder are set to 1, 2, and 4 respectively, while the dilation coefficients of the fourth temporal convolutional network layer TCN4, the fifth temporal convolutional network layer TCN5, and the sixth temporal convolutional network layer TCN6 in the decoder are set to 4, 2, and 1 respectively. For high-dimensional temporal data, the encoder continuously expands the receptive field through the three-layer TCN structure, extracts the temporal features of the training temporal data at a larger time scale, and then further compresses the feature dimension through the first one-dimensional convolutional layer Cov1 to obtain the corresponding latent variables. The decoder takes the latent variables as input, and continuously shrinks the receptive field through the three-layer TCN structure, gradually reducing the time scale to restore the corresponding temporal features. Finally, the second one-dimensional convolutional layer Cov2 is used to transform it into reconstructed data with the same dimension as the training temporal data.

[0117] The temporal convolutional autoencoder provided in this embodiment of the invention maintains a typical encoder and decoder structure, replaces the fully connected layers of the autoencoder with TCN residual blocks, and adds a one-dimensional convolutional network to change the dimension of the data. The addition of TCN enables the model to flexibly change the receptive field of the convolution while maintaining low computational cost, capturing the temporal information of the input data at different time intervals, thereby improving the accuracy of the model in reconstructing the data.

[0118] Furthermore, such as Figure 6As shown, the encoder also includes a first feature fusion module 10, and the decoder also includes a second feature fusion module 20;

[0119] The first temporal convolutional network layer TCN1, the second temporal convolutional network layer TCN2, the third temporal convolutional network layer TCN3, the first feature fusion module 10, and the first one-dimensional convolutional network layer Cov1 are connected in sequence.

[0120] The fourth temporal convolutional network layer TCN4, the fifth temporal convolutional network layer TCN5, the sixth temporal convolutional network layer TCN6, the second feature fusion module 20, and the second one-dimensional convolutional network layer Cov2 are connected in sequence.

[0121] The first feature fusion module 10 is used to weightedly fuse the temporal features of the training time series data extracted by the first temporal convolutional network layer TCN1, the second temporal convolutional network layer TCN2, and the third temporal convolutional network layer TCN3 respectively, to obtain the final temporal features of the training time series data.

[0122] The first one-dimensional convolutional network layer Cov1 is used to compress the feature dimension of the final temporal features of the training temporal data output by the first feature fusion module 10, and to obtain the latent variables corresponding to the training temporal data.

[0123] The second feature fusion module 20 is used to weightedly fuse the temporal features restored by the fourth temporal convolutional network layer TCN4, the fifth temporal convolutional network layer TCN5, and the sixth temporal convolutional network layer TCN6 respectively, to obtain the final temporal features of the restored training temporal data.

[0124] The second one-dimensional convolutional network layer, Cov2, is used to obtain reconstructed data of the same dimension as the training time series data based on the final time series features of the restored training time series data.

[0125] To further enhance the ability of the industrial process fault detection model to extract temporal features from industrial process data, this embodiment of the invention combines TCAE with a feature fusion module to construct an industrial process fault detection model based on TCAE and feature fusion, namely, an industrial process fault detection model based on TCAE-FF (temporal convolutional autoencoder-feature fusion, TCAE-FF).

[0126] A first feature fusion module 10 is set between the last TCN layer and the one-dimensional convolutional network layer of the encoder, i.e., between the third temporal convolutional network layer TCN3 and the first one-dimensional convolutional network layer Cov1. A second feature fusion module 20 is set between the last TCN layer and the one-dimensional convolutional network layer of the decoder, i.e., between the sixth temporal convolutional network layer TCN6 and the second one-dimensional convolutional network layer Cov2. Figure 6 As shown.

[0127] For the encoder, the training temporal features of the outputs of the first temporal convolutional network layer TCN1, the second temporal convolutional network layer TCN2, and the third temporal convolutional network layer TCN3 are copied through a skip connection structure and used as inputs to the first fusion module 10 for weighted fusion. For the decoder, the restored training temporal features of the outputs of the fourth temporal convolutional network layer TCN4, the fifth temporal convolutional network layer TCN5, and the sixth temporal convolutional network layer TCN6 are copied through a skip connection structure and used as inputs to the second fusion module 20 for weighted fusion, thereby obtaining the final temporal features of the restored training temporal data.

[0128] In summary, this invention incorporates a TCN structure into a basic convolutional autoencoder network to construct a TCAE model. It also designs a feature fusion module, combining the feature fusion module with the TCAE through skip connections. This establishes an industrial process fault detection model based on feature fusion and temporal convolutional autoencoder (TCAE-FF). This model can fully utilize temporal features at different scales, constructing feature-rich temporal data through encoding and decoding, while simultaneously obtaining salient feature representations of the input data. This increases the differentiation of reconstruction errors between normal and fault data, enhancing the model's ability to identify fault data and improving fault detection accuracy.

[0129] Furthermore, the first feature fusion module 10 includes a first fusion unit 11, a first weight calculation unit 12, and a first weighted output unit 13; the second feature fusion module 20 includes a second fusion unit 21, a second weight calculation unit 22, and a second weighted output unit 23.

[0130] like Figure 7 As shown, the first fusion unit 11 is used to obtain the first fusion feature based on the temporal features output by the first temporal convolutional network layer TCN1, the second temporal convolutional network layer TCN2, and the third temporal convolutional network layer TCN3, respectively.

[0131] The first weight calculation unit 12 is used to calculate the weights corresponding to the first temporal convolutional network layer TCN1, the second temporal convolutional network layer TCN2, and the third temporal convolutional network layer TCN3 respectively based on the first fusion feature through an efficient channel attention mechanism.

[0132] The first weighted output unit 13 is used to perform weighted fusion of the temporal features output by the first temporal convolutional network layer TCN1, the second temporal convolutional network layer TCN2, and the third temporal convolutional network layer TCN3 according to the weights corresponding to the first temporal convolutional network layer TCN1, the second temporal convolutional network layer TCN2, and the third temporal convolutional network layer TCN3, respectively, to obtain the final temporal features of the training temporal data.

[0133] like Figure 8 As shown, the second fusion unit 21 is used to obtain the second fused feature based on the restored temporal features output by the fourth temporal convolutional network layer TCN4, the fifth temporal convolutional network layer TCN5, and the sixth temporal convolutional network layer TCN6, respectively.

[0134] The second weight calculation unit 22 is used to calculate the weights corresponding to the fourth temporal convolutional network layer TCN4, the fifth temporal convolutional network layer TCN5, and the sixth temporal convolutional network layer 6 respectively based on the second fusion feature through an efficient channel attention mechanism.

[0135] The second weighted output unit 23 is used to perform weighted fusion of the temporal features output by the fourth temporal convolutional network layer TCN4, the fifth temporal convolutional network layer TCN5, and the sixth temporal convolutional network layer TCN6 according to the weights corresponding to the fourth temporal convolutional network layer TCN4, the fifth temporal convolutional network layer TCN5, and the sixth temporal convolutional network layer TCN6, respectively, to obtain the final temporal features of the restored training temporal data.

[0136] The working principles of the first feature fusion module 10 and the second feature fusion module 20 are roughly the same, as detailed below:

[0137] First, the output features of the corresponding TCN layer are copied through a skip connection structure and fused into feature U by adding corresponding elements. Then, an efficient channel attention mechanism is used to calculate the corresponding weights, and global average pooling is performed on the fused feature U to obtain compressed feature S. Next, one-dimensional convolution is used to perform local cross-channel interaction and softmax is used to generate the weights of the corresponding branches. Finally, the obtained weights and the copied output features of each layer are weighted and fused by adding elements to obtain the final output feature.

[0138] From the overall structural analysis, it can be divided into three parts: separation, fusion, and selection. In the separation stage, time-series features U1, U2, and U3 at different scales are obtained through skip connections. In the fusion stage, the fusion feature U is calculated for the input features, as shown in equation (4):

[0139]

[0140] in This indicates the sum of corresponding elements. Then, using an ECA structure, global average pooling is applied to U to compress and represent the information of each channel from a temporal perspective:

[0141]

[0142] Where i represents the corresponding channel position, C is the feature dimension, L represents the temporal length of the fused feature U, and t represents the corresponding temporal position. One-dimensional convolution is performed on the compressed representation S to achieve channel interaction, generating weight matrices Z1, Z2, and Z3 for the corresponding branches. In the selection phase, unlike the ECA module which uses sigmoid to calculate weights, this feature fusion module uses softmax to transform the obtained weight matrices Z1, Z2, and Z3 into weights for each corresponding branch:

[0143]

[0144] Where K j This represents the attention weight for branch j. After obtaining each weight, the temporal features extracted from the corresponding branch are weighted, and finally, the weighted features are fused to obtain the final temporal feature V.

[0145]

[0146] In the formula, This indicates element-wise multiplication. This indicates the summation of corresponding elements. This feature fusion module weights temporal features at different scales along the channel direction and obtains effective fused features, which can focus on important temporal information and enhance the richness of the overall features.

[0147] Furthermore, such as Figure 9 As shown, the training time-series data is input into the industrial process fault detection model for training, and the target industrial process fault detection model is obtained based on the training results, including:

[0148] S210. Input the training time series data into the industrial process fault detection model for forward calculation to obtain the reconstructed data corresponding to the training time series data;

[0149] Specifically, the processed training time series data is used as the input of the industrial process fault detection model. Reasonable initial parameter values ​​are selected, and the network model is trained using the time-based backpropagation algorithm. The output values ​​of each TCN layer, one-dimensional convolutional layer and feature fusion module are calculated. The calculation formulas are shown in (1), (4) and (7).

[0150] S220. Calculate the loss function value of the reconstructed data corresponding to the training time series data using the mean square error function;

[0151] Specifically, in this embodiment, the mean squared error (MSE) is used as the loss function to calculate its loss function value, and the calculation method is shown in formula (8):

[0152]

[0153] Where n is the number of samples, x i For sample data, For the corresponding reconstructed data, MSE reflects the training state of the network. The smaller the MSE value, the better the dataset reconstruction effect.

[0154] S230. Based on the loss function value, perform gradient updates on the industrial process fault detection model to obtain the target industrial process fault detection model.

[0155] like Figure 10 As shown, the gradient update process specifically includes:

[0156] S231. Based on the loss function value, the Adam optimization algorithm is used to adjust the network weights of the industrial process fault detection model to obtain the target network weights.

[0157] Based on the corresponding error term, the Adam optimization algorithm is used to update the network weights. The Adam optimization algorithm is a first-order optimization algorithm that can replace the traditional stochastic gradient descent algorithm. It has higher computational efficiency and better convergence performance in the same training period, while requiring less computational space. The target network weights here specifically refer to the network weights that minimize the loss function value through training.

[0158] S232. Based on the loss function value, the network hyperparameters of the industrial process fault detection model are adjusted using a grid search method to obtain the target hyperparameters;

[0159] Grid search (GS) was used to fine-tune the hyperparameters of the network. First, candidate confidence values ​​for different hyperparameters were determined based on prior knowledge, generating a search grid. Then, GS was used to fine-tune the hyperparameters of the industrial process fault detection model, selecting the optimal combination of target hyperparameters. The target hyperparameters are those that minimize the loss function value after training. Finally, based on the target hyperparameter combination, an appropriate number of training epochs were selected to complete the model training, and the network parameters were saved.

[0160] S233. Obtain the target industrial process fault detection model based on the target network weights and target hyperparameters.

[0161] Another embodiment of the present invention provides a time-convolutional autoencoder-based industrial process fault detection method for Tennessee Eastman (TE). The TE process consists of a reactor, condenser, circulating compressor, separator, and stripping tower, etc. See [link to relevant documentation]. Figure 11 The method includes:

[0162] S1: Collect SE industrial process data to form a historical training sample database:

[0163] This embodiment uses simulation experimental data, which was collected from the TE process simulation system. The TE process includes 12 operational variables, 22 continuous measurements, and 19 component variables. The collected data includes 22 subsets, one of which is data collected under normal operating conditions, and the remaining 21 are fault data samples. Each dataset is sampled every 3 minutes, for a total of two days, meaning each dataset contains 960 samples. To generate the fault dataset, a corresponding fault is introduced after 8 hours of normal operation; therefore, each fault dataset contains 160 normal samples and 800 fault samples sequentially.

[0164] S2: Preprocess the sample data in the historical training sample database:

[0165] By analyzing and preprocessing the sample data (including handling missing values, outliers, and constant values), and combining process mechanism information with relevant expert experience, a candidate input variable set consisting of 11 operational variables and 22 measurement variables was finally determined. The corresponding variable descriptions are shown in Table 1. Simultaneously, 18 commonly used fault types were selected from the 21 fault subsets for detection, excluding faults 3, 9, and 15. All fault types are shown in Table 2.

[0166] Table 1 Candidate Input Variables for the TE Process

[0167]

[0168]

[0169] Table 2 TE Process Fault Description

[0170]

[0171] Before fault detection, the collected data undergoes standardization and sliding window resampling. First, the normal operating condition data from the sample data is used as training data, and its mean and standard deviation are calculated, followed by z-score standardization. The fault data from the sample data is used as test data, and its mean and standard deviation are combined with those of the training data for z-score standardization. Finally, the standardized training and test data are resampled using a sliding window of the same width to obtain time-series data for training and testing the temporal convolutional autoencoder.

[0172] S3: Building and training industrial process fault detection models for Tennessee-Eastman:

[0173] The processed training time-series data is used as input to the industrial process fault detection model. Reasonable initial parameter values ​​are selected, and the network model is trained using a time-based backpropagation algorithm. The specific training process mainly includes three steps:

[0174] (1) Forward computation: Calculate the output values ​​of each TCN layer, one-dimensional convolutional layer and feature fusion module. The calculation formula is the same as the above formula (1), formula (4) and formula (7).

[0175] (2) Backward calculation: Calculate its loss function value. The loss function of the LSTM model is MSE, and the calculation method is shown in formula (8). MSE reflects the training state of the network. The smaller the MSE value, the better the data set reconstruction effect.

[0176] (3) Gradient update: Based on the corresponding error terms, the Adam optimization algorithm is used to update the network weights. The Adam optimization algorithm is a first-order optimization algorithm that can replace the traditional stochastic gradient descent algorithm. It has higher computational efficiency and better convergence performance in the same training period, while requiring less computational space.

[0177] The network hyperparameters were tuned using a grid search (GS) method. First, candidate confidence values ​​for different hyperparameters were determined based on prior knowledge, generating a search grid. Then, the GS method was used to tune the TCAE-FF network hyperparameters, selecting the optimal hyperparameter combination. Finally, based on the optimal hyperparameter combination, an appropriate number of training epochs were selected to complete the model training, and the network parameters were saved.

[0178] S4: Calculate the statistic and corresponding control limits for the normal sample:

[0179] Import the training time series data into the trained optimal model, and use the model reconstruction error to calculate the SPE statistic, specifically refer to formula (2);

[0180] After obtaining the SPE statistic, the corresponding control limits are calculated using kernel density estimation, as detailed in formula (3).

[0181] S5: Calculate the statistics for the fault samples:

[0182] Load the saved optimal parameters of the model, run the model in test mode, and input the time series data of the fault samples for testing. Then, calculate the SPE statistic based on the reconstruction error of the model to the corresponding fault data, and obtain 18 sets of data, corresponding to the 18 selected faults. The number of statistics in each set is consistent with the number of test data, that is, each fault sample corresponds to one SPE statistic.

[0183] S6: Perform fault detection:

[0184] The control limit calculated using the normal sample statistic is used as the judgment standard. The statistic corresponding to the test data is compared with it. If the SPE is greater than the control limit, the corresponding test sample is judged to be faulty data; otherwise, it is normal data.

[0185] To analyze the fault detection performance of the method in this application from multiple perspectives, this embodiment is compared with fault detection methods based on kernelized principal component analysis (KPCA), CAE, LSTM-AE, and TCAE. Furthermore, this application uses the following two parameters as evaluation indicators for fault detection for each method, and their calculation formulas are as follows:

[0186] The fault detection rate (FDR) is calculated as follows:

[0187]

[0188] The false alarm rate (FAR) is calculated as follows:

[0189]

[0190] Where N1 and N2 are the number of faulty samples and the number of normal samples, respectively, N 1,f and N 2,f These represent the number of samples detected as faulty from the faulty samples and the number of samples from the normal samples, respectively.

[0191] The KPCA algorithm uses a Gaussian kernel with 16 principal components; the CAE algorithms all use a four-layer CNN network with a 3*3 kernel size, 16, 32, 32, and 16 output channels, and ReLU and sigmoid activation functions; LSTM-AE uses a two-layer LSTM network with a hidden layer dimension of 32 and sigmoid activation function; TCAE uses a six-layer TCN with a kernel size of 5, 12 output channels, and dilation coefficients of 1, 2, 4, 4, 2, 1; TCAE-FF uses a six-layer TCN with a kernel size of 5, 16 output channels, and dilation coefficients of 1, 2, 4, 4, 2, 1.

[0192] The simulation environment was a deep learning server: a 12th Gen Intel(R) Core(TM) i7-12700 CPU processor, an NVIDIA GeForce RTX 3090 GPU, and 128GB of RAM. The initial learning rate was 0.001, the time step was 12, the maximum number of iterations was 500, and the Adam algorithm was used as the model gradient optimization algorithm. The fault detection results of different algorithms are shown in Table 3, with the bold numbers representing the highest detection rate for that type of fault.

[0193] Table 3. Detection results of the TE process using five methods.

[0194]

[0195] Comparing the detection results of the five methods above, overall, under the premise that the average false alarm rates are not significantly different, the fault detection methods based on TCAE and TCAE-FF provided by this invention both show good performance, with TCAE-FF performing even better. The average alarm rate of TCAE-FF is 25.09% higher than that of the traditional KPCA method, indicating that adding TCN to extract the temporal features of the input and fusing the extracted temporal features through the feature fusion module can effectively improve the fault detection accuracy. Meanwhile, the TCAE-FF method improves the accuracy by 14.59%, 10.08%, and 4.59% compared to similar deep learning methods CAE, LSTM-AE, and TCAE, respectively, further demonstrating the effectiveness of the feature fusion module. It can more effectively extract temporal features of different scales of the input data, enhance the model's ability to distinguish between normal and faulty samples, and improve the fault detection rate. Observing the detection results of various types of faults, the TCAE-FF method shows significant improvement over the other four methods in the detection of faults 5, 10, 16, 19, and 20. The SPE statistic curve for fault 10 is shown in the figure below. Figures 12 to 16 As shown.

[0196] Another embodiment of the present invention provides an industrial process fault detection device 30 based on a temporal convolutional autoencoder, which is applied to the above-mentioned industrial process fault detection method based on a temporal convolutional autoencoder, such as... Figure 17 As shown, it includes:

[0197] The preprocessing module 31 is used to preprocess the acquired normal operating condition data of the industrial process as training data to obtain training time series data, and to preprocess the acquired test operating condition data of the industrial process to obtain test time series data.

[0198] The model training module 32 is used to input the training time series data into the industrial process fault detection model for training, and obtain the target industrial process fault detection model based on the training results. The industrial process fault detection model is constructed based on a temporal convolutional network and a convolutional autoencoder.

[0199] The control limit calculation module 33 is used to input the training time series data into the target industrial process fault detection model and calculate the control limit of the squared prediction error statistic of the training time series data.

[0200] The squared prediction error statistic calculation module 34 is used to input the test time series data into the target industrial process fault detection model and calculate the squared prediction error statistic of the test time series data.

[0201] The judgment module 35 is used to determine whether the test condition data is abnormal based on the control limit of the squared prediction error statistic of the training time series data and the squared prediction error statistic of the test time series data.

[0202] This invention provides an industrial process fault detection device 30 based on a temporal convolutional autoencoder. By further introducing a temporal convolutional network on the basis of a convolutional autoencoder, a novel temporal convolutional autoencoder (TCAE) is designed. Based on this novel TCAE, an industrial process fault detection model is created and trained for fault detection of industrial process-related data. By using at least two stacked temporal convolutional networks, the receptive field of the convolution can be gradually expanded to change the time scale of temporal feature extraction of industrial process-related data, enrich the time scale of temporal feature extraction, and more accurately realize temporal modeling, thereby improving the feature extraction capability of the industrial process fault detection model and further improving the accuracy of industrial process fault detection.

[0203] The specific working principle of the industrial process fault detection device 30 based on a time convolutional autoencoder provided in this embodiment of the invention can be referred to the above description of the principle of an industrial process fault detection method based on a time convolutional autoencoder, and will not be repeated here.

[0204] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0205] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for industrial process fault detection based on time convolution autoencoder, characterized in that, include: The acquired normal operating condition data of the industrial process is used as training data, and the acquired test operating condition data of the industrial process is used as test data. The training data and the test data are preprocessed respectively to obtain training time series data and test time series data. The training time series data is input into the industrial process fault detection model for training, and the target industrial process fault detection model is obtained based on the training results. The industrial process fault detection model is constructed based on a convolutional autoencoder and at least two layers of temporal convolutional network. The training time series data and the test time series data are input into the target industrial process fault detection model, and the control limit of the squared prediction error statistic of the training time series data and the squared prediction error statistic of the test time series data are calculated. Based on the calculated control limits and squared prediction error statistics, determine whether the test condition data is abnormal; The industrial process fault detection model includes an encoder and a decoder; The encoder includes a first temporal convolutional network layer, a second temporal convolutional network layer, a third temporal convolutional network layer, a first feature fusion module, and a first one-dimensional convolutional network layer; the first temporal convolutional network layer, the second temporal convolutional network layer, the third temporal convolutional network layer, the first feature fusion module, and the first one-dimensional convolutional network layer are connected in sequence; The decoder includes a fourth temporal convolutional network layer, a fifth temporal convolutional network layer, a sixth temporal convolutional network layer, a second feature fusion module, and a second one-dimensional convolutional network layer; the fourth temporal convolutional network layer, the fifth temporal convolutional network layer, the sixth temporal convolutional network layer, the second feature fusion module, and the second one-dimensional convolutional network layer are connected sequentially. The first one-dimensional convolutional network layer is connected to the fourth temporal convolutional network layer; The dilation coefficients of the first and sixth temporal convolution network layers are 2 x The dilation coefficients of the second and fifth temporal convolution network layers are 2 y The dilation coefficients of the third and sixth temporal convolution network layers are 2 z x < y < z; The first, second, and third temporal convolutional network layers are all used to extract temporal features of the training temporal data at the corresponding time scale according to the corresponding dilation coefficients. The first one-dimensional convolutional network layer is used to compress the feature dimension of the temporal features output by the third temporal convolutional network layer to obtain the latent variables corresponding to the training temporal data; The fourth, fifth, and sixth temporal convolutional network layers are all used to restore the latent variables output by the first one-dimensional convolutional network layer to temporal features according to the corresponding dilation coefficients at the corresponding time scales. The second one-dimensional convolutional network layer is used to obtain reconstructed data of the same dimension as the training temporal data based on the temporal features output by the sixth temporal convolutional network layer.

2. The method of claim 1, wherein, The output of the encoder is connected to the input of the decoder; The encoder is used to extract temporal features of the training time series data based on at least two temporal convolutional network layers with different dilation coefficients, and to obtain the latent variables corresponding to the training time series data based on the extracted temporal features of the training time series data. The decoder is used to restore the latent variables output by the encoder to temporal features based on at least two temporal convolutional network layers with different dilation coefficients, and to obtain reconstructed data with the same dimension as the training temporal data based on the restored temporal features.

3. The method according to claim 2, characterized in that, The first feature fusion module is used to weightedly fuse the temporal features of the training time series data extracted by the first temporal convolutional network layer, the second temporal convolutional network layer and the third temporal convolutional network layer respectively, to obtain the final temporal features of the training time series data; The first one-dimensional convolutional network layer is used to compress the feature dimension of the final temporal features of the training temporal data output by the first feature fusion module, and to obtain the latent variables corresponding to the training temporal data. The second feature fusion module is used to weightedly fuse the temporal features restored by the fourth temporal convolutional network layer, the fifth temporal convolutional network layer and the sixth temporal convolutional network layer respectively, to obtain the final temporal features of the restored training temporal data; The second one-dimensional convolutional network layer is used to obtain reconstructed data of the same dimension as the training time series data based on the final time series features of the restored training time series data.

4. The method of claim 3, wherein, The first feature fusion module includes a first fusion unit, a first weight calculation unit, and a first weighted output unit; the second feature fusion module includes a second fusion unit, a second weight calculation unit, and a second weighted output unit. The first fusion unit is used to obtain a first fusion feature based on the temporal features output by the first temporal convolutional network layer, the second temporal convolutional network layer and the third temporal convolutional network layer respectively; The first weight calculation unit is used to calculate the weights corresponding to the first temporal convolutional network layer, the second temporal convolutional network layer, and the third temporal convolutional network layer respectively based on the first fusion feature through an efficient channel attention mechanism; The first weighted output unit is used to perform weighted fusion of the temporal features output by the first temporal convolutional network layer, the second temporal convolutional network layer, and the third temporal convolutional network layer according to the weights corresponding to the first temporal convolutional network layer, the second temporal convolutional network layer, and the third temporal convolutional network layer, respectively, to obtain the final temporal features of the training temporal data; The second fusion unit is used to obtain a second fusion feature based on the restored temporal features output by the fourth temporal convolutional network layer, the fifth temporal convolutional network layer, and the sixth temporal convolutional network layer, respectively. The second weight calculation unit is used to calculate the weights corresponding to the fourth temporal convolutional network layer, the fifth temporal convolutional network layer, and the sixth temporal convolutional network layer respectively based on the second fusion feature through an efficient channel attention mechanism; The second weighted output unit is used to perform weighted fusion of the temporal features output by the fourth, fifth, and sixth temporal convolutional network layers according to the weights corresponding to the fourth, fifth, and sixth temporal convolutional network layers, respectively, to obtain the final temporal features of the restored training temporal data.

5. The method according to any one of claims 1 to 4, characterized in that, The step of inputting the training time-series data into the industrial process fault detection model for training, and obtaining the target industrial process fault detection model based on the training results, includes: The training time series data is input into the industrial process fault detection model for forward calculation to obtain the reconstructed data corresponding to the training time series data. The loss function value of the reconstructed data corresponding to the training time series data is calculated using the mean squared error function. The industrial process fault detection model is updated using gradients based on the loss function value to obtain the target industrial process fault detection model.

6. The method according to claim 5, characterized in that, The step of updating the industrial process fault detection model based on the loss function value to obtain the target industrial process fault detection model includes: Based on the loss function value, the Adam optimization algorithm is used to adjust the network weights of the industrial process fault detection model to obtain the target network weights. Based on the loss function value, the network hyperparameters of the industrial process fault detection model are adjusted using a grid search method to obtain the target hyperparameters; The target industrial process fault detection model is obtained based on the target network weights and the target hyperparameters.

7. The method of claim 1, wherein, The preprocessing includes data standardization and sliding window resampling.

8. The method of claim 1, wherein, The step of determining whether the test condition data is abnormal based on the control limit of the squared prediction error statistic of the training time series data and the squared prediction error statistic of the test time series data includes: When the squared prediction error statistic of the test time series data is greater than the control limit of the squared prediction error statistic of the training time series data, the test condition data is judged to be abnormal. When the squared prediction error statistic of the test time series data is less than or equal to the control limit of the squared prediction error statistic of the training time series data, the test condition data is judged to be normal.

9. An apparatus for industrial process fault detection based on time convolution autoencoder, applied to the method for industrial process fault detection based on time convolution autoencoder according to any one of claims 1 to 8, characterized in that, include: The preprocessing module is used to preprocess the acquired normal operating condition data of the industrial process as training data to obtain training time series data, and to preprocess the acquired test operating condition data of the industrial process to obtain test time series data. The model training module is used to input the training time series data into the industrial process fault detection model for training, and obtain the target industrial process fault detection model based on the training results. The industrial process fault detection model is constructed based on a temporal convolutional network and a convolutional autoencoder. The control limit calculation module is used to input the training time series data into the target industrial process fault detection model and calculate the control limit of the squared prediction error statistic of the training time series data. The squared prediction error statistic calculation module is used to input the test time series data into the target industrial process fault detection model and calculate the squared prediction error statistic of the test time series data. The judgment module is used to determine whether the test condition data is abnormal based on the control limit of the squared prediction error statistic of the training time series data and the squared prediction error statistic of the test time series data.

Images