Industrial process fault detection method and device based on time convolution auto-encoder

Abstract

The invention discloses an industrial process fault detection method and device based on a time convolution auto-encoder, and belongs to the field of industrial process detection.The method comprises the steps that training data and test data are preprocessed, and training time sequence data and test time sequence data are obtained; inputting the training time sequence data into an industrial process fault detection model for training to obtain a target model, wherein the industrial process fault detection model at least comprises two layers of TCN and CAE; inputting the training time sequence data into a target model to calculate a control limit; inputting the test time sequence data into the target model to calculate an SPE statistical magnitude; and judging whether the to-be-detected working condition data is abnormal or not according to the control limit and the SPE statistical magnitude. According to the method, TCN is introduced on the basis of CAE, a novel time convolution auto-encoder is designed, an industrial process fault detection model is created, a convolution receptive field can be gradually expanded through at least two stacked time convolution networks to change the time scale of time sequence feature extraction, and time sequence modeling is more accurately achieved.

Description

Technical Field

[0001] The invention relates to an industrial process fault detection method and device based on a time convolution autoencoder, and belongs to the technical field of industrial process detection. Background Art

[0002] In order to maintain the safe and reliable operation of industrial processes and the stability of product quality, it is particularly important to implement effective fault detection technology. With the continuous improvement of the level of automation and informatization of industrial processes, rich data information has been accumulated in the production process, providing favorable support for data-driven fault detection methods. Data-driven fault detection methods use data analysis and preprocessing techniques, use prior knowledge of historical data, and construct statistics in real-time data to achieve process monitoring. As the main method of data-driven fault detection, multivariate statistical methods have been widely used in the field of industrial processes. Its 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, its feature extraction ability is limited and it is difficult to select a suitable kernel function.

[0003] Deep learning can process complex nonlinear data more effectively and has excellent adaptive capabilities, which has attracted widespread attention in the field of fault detection. Common 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 characteristics of the data. In order to enhance the network's ability to extract local features, researchers combined convolutional neural networks with autoencoders, used convolutional layers instead of fully connected layers, and proposed convolutional autoencoders (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 dependence of data on the time dimension, and has certain limitations in industrial process fault detection applications that require processing time series data.

[0004] In order to effectively extract the temporal features of sequence data, the fault detection method based on convolutional long short-term autoencoder uses convolutional long short-term memory network to learn the effective features of temporal data, and the autoencoder based on convolutional gated recurrent unit and attention combines convolutional neural network and gated recurrent unit to realize the temporal feature extraction of complex process data. Relevant scholars proposed temporal convolutional network (TCN) based on one-dimensional convolutional neural network. TCN uses dilated causal convolution to skip some convolution units, so that the convolution operation covers a larger sequence range, thereby effectively capturing the long-term temporal information of the data.

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

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

[0007] In order to solve the problems that the existing industrial process fault detection methods based on deep learning have a single feature extraction scale and low fault detection accuracy, the present invention provides an industrial process fault detection method and device based on a time convolutional autoencoder.

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

[0009] Using the acquired normal working condition data of the industrial process as training data, using the acquired working condition data to be tested of the industrial process as test data, and preprocessing the training data and the test data respectively to obtain training time series data and test time series data;

[0010] Inputting the training time series data into an industrial process fault detection model for training, and obtaining a target industrial process fault detection model according to the training results, wherein the industrial process fault detection model is constructed based on a convolutional autoencoder and at least two layers of temporal convolutional networks;

[0011] Inputting the training time series data into the target industrial process fault detection model, and calculating the control limit of the squared prediction error statistic of the training time series data;

[0012] Inputting the test time series data into the target industrial process fault detection model, and calculating the square prediction error statistic of the test time series data;

[0013] Whether the operating condition data to be tested is abnormal is determined based on the control limit of the square prediction error statistic of the training time series data and the square prediction error statistic of the test time series data.

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

[0015] The encoder is used to extract the time series features of the training time series data according to the time convolution network layers with at least two different expansion coefficients, and to obtain the latent variables corresponding to the training time series data according to the extracted time series features of the training time series data;

[0016] The decoder is used to restore the latent variables output by the encoder into time series features according to a temporal convolutional network layer with at least two different expansion coefficients, and to obtain reconstructed data with the same dimension as the training time series data according to the restored time series features.

[0017] Further, 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] Among them, the expansion coefficients of the first time convolutional network layer and the sixth time convolutional network layer are 2 x , the expansion coefficient of the second time convolutional network layer and the fifth time convolutional network layer is 2 y , the expansion coefficient of the third time convolutional network layer and the sixth time convolutional network layer is 2 z , x<y<z;

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

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

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

[0024] The second one-dimensional convolutional network layer is used to obtain reconstructed data with the same dimension as the training time series data according to the restored time series features output by the sixth time 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 for weighted fusion of the time series features of the training time series data extracted by the first time convolutional network layer, the second time convolutional network layer and the third time convolutional network layer, respectively, to obtain the final time series 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 time series feature of the training time series data output by the first feature fusion module to obtain the latent variable corresponding to the training time series data;

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

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

[0032] Further, 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 according to the time series features respectively output by the first time convolutional network layer, the second time convolutional network layer, and the third time convolutional network layer;

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

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

[0036] The second fusion unit is used to obtain a second fusion feature according to the restored time series features respectively output by the fourth time convolutional network layer, the fifth time convolutional network layer, and the sixth time convolutional network layer;

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

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

[0039] Furthermore, the inputting the training time series data into the industrial process fault detection model for training, and obtaining the target industrial process fault detection model according to the training result, comprises:

[0040] Inputting the training time series data into an industrial process fault detection model for forward calculation to obtain reconstructed data corresponding to the training time series data;

[0041] Calculate the loss function value of the reconstructed data corresponding to the training time series data by using a mean square error function;

[0042] The industrial process fault detection model is gradient updated according to the loss function value to obtain a target industrial process fault detection model.

[0043] Furthermore, the step of performing gradient updating on the industrial process fault detection model according to the loss function value to obtain a target industrial process fault detection model includes:

[0044] According to the loss function value, the network weight of the industrial process fault detection model is adjusted using the Adam optimization algorithm to obtain a target network weight;

[0045] According to the loss function value, a grid search method is used to adjust the network hyperparameters of the industrial process fault detection model to obtain target hyperparameters;

[0046] A target industrial process fault detection model is obtained according to the target network weights and the target hyperparameters.

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

[0048] Further, judging whether the working condition data to be tested is abnormal according to the control limit of the square prediction error statistic of the training time series data and the square prediction error statistic of the test time series data includes:

[0049] When the square prediction error statistic of the test time series data is greater than the control limit of the square prediction error statistic of the training time series data, it is judged that the working condition data to be tested is abnormal;

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

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

[0052] A preprocessing module, used to use the acquired normal working condition data of the industrial process as training data, preprocess the training data to obtain training time series data, and used to preprocess the acquired working condition data to be tested of the industrial process to obtain test time series data;

[0053] A model training module, used for inputting the training time series data into an industrial process fault detection model for training, and obtaining a target industrial process fault detection model according to the training results, wherein the industrial process fault detection model is constructed based on a temporal convolutional network and a convolutional autoencoder;

[0054] A control limit calculation module, used for inputting the training time series data into the target industrial process fault detection model, and calculating the control limit of the square prediction error statistic of the training time series data;

[0055] A square prediction error statistic calculation module, used for inputting the test time series data into the target industrial process fault detection model, and calculating the square prediction error statistic of the test time series data;

[0056] A judgment module is used to judge whether the working condition data to be tested is abnormal based on the control limit of the square prediction error statistic of the training time series data and the square prediction error statistic of the test time series data.

[0057] The beneficial effects of the present invention are:

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

[0059] (2) The industrial process fault detection method based on time convolutional autoencoder provided by the present invention adds a TCN structure to the basic convolutional autoencoder network, constructs a TCAE model, and designs a feature fusion module. The feature fusion module is combined with TCAE through jump connections to establish an industrial process fault detection model based on feature fusion and time convolutional autoencoder (TCAE-FF). The model can make full use of time series features of different scales, construct feature-rich time series data through encoding and decoding, and obtain significant feature representation of the input data, so that the model can increase the degree of differentiation of reconstruction errors of normal data and fault data, enhance the model's ability to identify fault data, and improve the accuracy of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0061] Figure 1 is a flow chart of an industrial process fault detection method based on a time convolutional autoencoder provided by an embodiment of the present invention;

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

[0063] Figure 3 is a flow chart of control limit calculation steps provided by an embodiment of the present invention;

[0064] Figure 4 is a flowchart of the fault judgment steps provided by an embodiment of the present invention;

[0065] Figure 5 is a TCAE structure diagram provided by an embodiment of the present invention;

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

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

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

[0069] Fig. 9 is a flow chart of obtaining a target industrial process fault detection model provided by an embodiment of the present invention;

[0070] Fig.10 It is a flow chart of the gradient updating steps provided by an embodiment of the present invention.

[0071] Fig.11 is a schematic diagram of a TE industrial process provided by an embodiment of the present invention;

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

[0073] Fig.13 is a schematic diagram of a TE industrial process fault detection result based on CAE provided by an embodiment of the present invention;

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

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

[0076] Fig.16 It is a schematic diagram of TE industrial process fault detection results based on TCAE-FF provided by an embodiment of the present invention;

[0077] Fig.17 It is a structural diagram of an industrial process fault detection device based on a time convolutional autoencoder provided in an embodiment of the present invention. DETAILED DESCRIPTION

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

[0079] The embodiment of the present invention provides an industrial process fault detection method based on a time convolutional autoencoder, such as Figure 1 As shown, including:

[0080] S100, using the acquired normal working condition data of the industrial process as training data, using the acquired working condition data to be tested of the industrial process as test data, and preprocessing the training data and the test data respectively to obtain training time series data and test time series data;

[0081] Specifically, the preprocessing in the embodiment of the present invention mainly includes data standardization and sliding window resampling. For the training data, the mean and standard deviation of the training data are first calculated, and then the training data is z-score standardized according to the mean and standard deviation of the training data to obtain standardized training data, and then the standardized training data is resampled using a sliding window to obtain the corresponding training time series data. Similarly, for the test data, the mean and standard deviation of the test data are first calculated, and then z-score standardized according to the mean and standard deviation of the test data, and 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, inputting the training time series data into the industrial process fault detection model for training, and obtaining a target industrial process fault detection model according to the training results, wherein the industrial process fault detection model is constructed based on a convolutional autoencoder and at least two layers of time convolutional networks;

[0083] Convolutional autoencoder is an unsupervised deep neural network model, and its basic structure consists of an encoder and a decoder. The encoder can extract the features of the input data, and its output data dimension is generally smaller than the input dimension, which can achieve data dimensionality reduction. The decoder converts the feature representation of the encoder output into the same dimension as the input data to achieve reconstruction of the input data.

[0084] The temporal convolutional network is based on a one-dimensional convolutional neural network. It adds dilated convolution and causal convolution structures to effectively extract time information. Its network structure is as follows: Figure 2 As shown in Figure 2. The temporal convolutional network adopts a causal convolution structure, which ensures that the output at each moment depends only on the current moment and the previous historical data by filling zeros, thereby avoiding the future information leakage problem that may be caused by non-causal convolution and ensuring the rationality of time series prediction. In order to handle long time series tasks, the temporal convolutional network also combines the dilated convolution structure, which adds holes in the convolution kernel by setting the dilation coefficient, thereby significantly expanding the receptive field of the convolution operation. For a one-dimensional sequence input x∈R n And convolution kernel f:{0,…,k-1}, the expression of the dilated convolution calculation F of the sequence element s is:

[0085]

[0086] In the formula, k is the convolution kernel size, d is the expansion coefficient, and x s-diis a sequence corresponding to the convolution kernel elements. Dilated convolution expands the receptive field of convolution by continuously increasing the dilation coefficient d, and the dilation coefficient of each layer in a typical temporal convolutional network structure is d=2. l (l=0,1,2,…) method. Figure 2 As shown in the figure, the d of the first convolution layer is equal to 1, and its receptive field size is 3, processing data at 3 moments at a time; the d of the second convolution layer is equal to 2, and its receptive field increases to 7; when the d of the third convolution layer increases to 4, the corresponding receptive field is 15, and feature extraction is performed on the input data at a time scale of 15 moments.

[0087] The embodiment of the present invention further introduces a temporal convolutional network on the basis of the convolutional autoencoder, designs a new temporal convolutional autoencoder (TCAE), and creates an industrial process fault detection model based on the new temporal convolutional autoencoder. After at least two stacked temporal convolutional networks, the convolution receptive field can be gradually expanded to change the time scale of time series feature extraction, more accurately implement time series modeling, and improve fault response capability and fault detection accuracy. Specifically, improving the accuracy of time series modeling means improving the accuracy of the model's reconstruction of 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 normal data from fault data. In general, the more accurate the model is for normal data reconstruction, the smaller the corresponding SPE statistic is, and the smaller the control limit is. When fault data is input, the model reconstruction error is larger, and the corresponding SPE statistic is larger, and it is easier to exceed the control limit, so the fault data can be detected in time.

[0088] S300, inputting the training time series data into the target industrial process fault detection model, and calculating the control limit of the square prediction error statistic of the training time series data;

[0089] Specifically, Figure 3 As shown, step S300 also includes:

[0090] S310, calculating the squared prediction error statistic (SPE) of the training time series data according to the reconstruction error of the target industrial process fault detection model, as shown in the following formula (2):

[0091]

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

[0093] S320, calculating the control limit of the squared prediction error statistic of the training time series data through kernel density estimation.

[0094] The specific formula is as shown in formula (3):

[0095]

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

[0097] S400, inputting the test time series data into the target industrial process fault detection model, and calculating the square prediction error statistics of the test time series data;

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

[0099] S500. Determine whether the working condition data to be tested is abnormal based on the control limit of the square prediction error statistic of the training time series data and the square prediction error statistic of the test time series data.

[0100] Specifically, Figure 4 As shown, step S500 also includes:

[0101] S510, when the square prediction error statistic of the test time series data is greater than the control limit of the square prediction error statistic of the training time series data, it is determined that the working condition data to be tested is abnormal;

[0102] S520. When the square prediction error statistic of the test time series data is less than or equal to the control limit of the square prediction error statistic of the training time series data, it is determined that the working condition data to be tested is normal.

[0103] In summary, the embodiment of the present invention further introduces a temporal convolutional network on the basis of a convolutional autoencoder, designs a new temporal convolutional autoencoder (TCAE), and creates and trains an industrial process fault detection model based on the new temporal convolutional autoencoder for fault detection of industrial process-related data. After at least two stacked temporal convolutional networks, the convolution receptive field can be gradually expanded to change the time scale of temporal feature extraction of industrial process-related data, enrich the time scale of extracted temporal features, 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.

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

[0105] The encoder is used to extract the time series features of the training time series data according to the time convolution network layers with at least two different expansion coefficients, and to obtain the latent variables corresponding to the training time series data according to the extracted time series features of the training time series data;

[0106] The decoder is used to restore the latent variables output by the encoder into time series features according to a temporal convolutional network layer with at least two different expansion coefficients, and to obtain reconstructed data with the same dimension as the training time series data according to the restored time series features.

[0107] In addition to extracting and restoring the time series features of the training time series data at different time scales by setting at least two time convolution network layers with different expansion coefficients in the encoder and decoder respectively, enriching the time scale of extracting time series features and realizing time series modeling more accurately, the embodiment of the present invention also introduces residual connections in the time convolution network to form a residual block structure. Specifically, a complete residual block structure includes two sets of expanded causal convolutions, weight normalization, activation functions and random inactivation links. The time convolution network with residual connections can retain a larger receptive field on the basis of fewer network layers, fully explore the deep-level features of time series data, and at the same time, the residual connections can effectively suppress problems such as overfitting and gradient explosion in the process of increasing network depth.

[0108] Furthermore, if 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 which are connected in sequence;

[0109] The decoder includes 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 connected in sequence;

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

[0111] Among them, the expansion coefficient of the first time convolutional network layer TCN1 and the sixth time convolutional network layer TCN4 is 2 x , the expansion coefficient of the second time convolutional network layer TCN2 and the fifth time convolutional network layer TCN5 is 2 y , the expansion coefficient of the third time convolutional network layer TCN3 and the fourth time convolutional network layer TCN6 is 2 z , x<y<z;

[0112] The first time convolution network layer TCN1, the second time convolution network layer TCN2, and the third time convolution network layer TCN3 are all used to extract the time series features of the training time series data at the corresponding time scale according to the corresponding expansion coefficients;

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

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

[0115] The second one-dimensional convolutional network layer Cov2 is used to obtain reconstructed data with the same dimension as the training time series data according to the restored time series features output by the sixth time convolutional network layer TCN6.

[0116] In the embodiment of the present invention, each layer of TCN is a complete residual block structure, and its corresponding expansion coefficient d changes in the form of an exponential 2. Specifically, in the embodiment of the present invention, set x=0, y=1, z=2, that is, the expansion coefficients corresponding to the first time convolution network layer TCN1, the second time convolution network layer TCN2, and the third time convolution network layer TCN3 in the encoder are set to 1, 2, and 4 respectively, and the expansion coefficients of the fourth time convolution network layer TCN4, the fifth time convolution network layer TCN5, and the sixth time convolution network layer TCN6 in the decoder are set to 4, 2, and 1 respectively. For high-dimensional time series data, the encoder continuously expands the receptive field through the three-layer TCN structure, extracts the time series features of the training time series data at a larger time scale, and then further compresses the feature dimension through the first one-dimensional convolution layer Cov1 to obtain the corresponding latent variable. The decoder takes the latent variable as input, continuously reduces the receptive field through the three-layer TCN structure in turn, gradually reduces the time scale to restore the corresponding time series features, and finally uses the second one-dimensional convolution layer Cov2 to convert it into reconstructed data of the same dimension as the training time series data.

[0117] The temporal convolutional autoencoder provided in the embodiment of the present invention maintains the typical encoder and decoder structure as a whole, uses the TCN residual block to replace the fully connected layer of the autoencoder, and adds a one-dimensional convolutional network to change the dimension of the data. The addition of TCN can keep the model at a low computational complexity while flexibly changing the receptive field of the convolution, capturing the timing information of the input data at different time intervals, thereby improving the accuracy of the model's reconstruction of the data.

[0118] Furthermore, if Figure 6As shown, the encoder further includes a first feature fusion module 10, and the decoder further 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 time series features of the training time series data extracted by the first time convolutional network layer TCN1, the second time convolutional network layer TCN2 and the third time convolutional network layer TCN3, so as to obtain the final time series 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 time series feature of the training time series data output by the first feature fusion module 10 to obtain the latent variables corresponding to the training time series data;

[0123] The second feature fusion module 20 is used to weightedly fuse the time series 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 time series features of the restored training time series data;

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

[0125] In order to further enhance the ability of the industrial process fault detection model to extract temporal features of industrial process data, an embodiment of the present invention combines TCAE with a feature fusion module to construct an industrial process fault detection model based on TCAE and feature fusion, that is, 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, that is, between the third temporal convolutional network layer TCN3 and the first one-dimensional convolutional network layer Cov1, and a second feature fusion module 20 is set between the last TCN layer and the one-dimensional convolutional network layer of the decoder, that is, between the sixth temporal convolutional network layer TCN6 and the second one-dimensional convolutional network layer Cov2. Figure 6 shown.

[0127] For the encoder, through the jump connection structure, the training time series features of the outputs of the first time convolutional network layer TCN1, the second time convolutional network layer TCN2, and the third time convolutional network layer TCN3 are copied respectively as the input of the first fusion module 10 for weighted fusion; for the decoder, through the jump connection structure, the restored training time series features of the outputs of the fourth time convolutional network layer TCN4, the fifth time convolutional network layer TCN5, and the sixth time convolutional network layer TCN6 are copied respectively as the input of the second fusion module 20 for weighted fusion to obtain the final time series features of the restored training time series data.

[0128] In summary, the present invention adds the TCN structure to the basic convolutional autoencoder network, constructs the TCAE model, and designs a feature fusion module. The feature fusion module is combined with TCAE through jump connections to establish an industrial process fault detection model based on feature fusion and time convolutional autoencoder (TCAE-FF). The model can make full use of time series features of different scales, construct feature-rich time series data through encoding and decoding, and obtain significant feature representation of the input data, so that the model can increase the degree of differentiation of reconstruction errors of normal data and fault data, enhance the model's ability to identify fault data, and improve the accuracy of fault detection.

[0129] Further, 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 according to the time series features respectively output by the first time convolutional network layer TCN1, the second time convolutional network layer TCN2 and the third time convolutional network layer TCN3;

[0131] The first weight calculation unit 12 is used to calculate the weights corresponding to the first time convolution network layer TCN1, the second time convolution network layer TCN2, and the third time convolution network layer TCN3 respectively according to the first fusion feature through an efficient channel attention mechanism;

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

[0133] like Figure 8 As shown, the second fusion unit 21 is used to obtain the second fusion feature according to the restored time series features respectively output by the fourth time convolutional network layer TCN4, the fifth time convolutional network layer TCN5 and the sixth time convolutional network layer TCN6;

[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 according to the second fusion feature through an efficient channel attention mechanism;

[0135] The second weighted output unit 23 is used to weightedly fuse the time series features output by the fourth time convolutional network layer TCN4, the fifth time convolutional network layer TCN5 and the sixth time convolutional network layer TCN6 according to the weights corresponding to the fourth time convolutional network layer TCN4, the fifth time convolutional network layer TCN5 and the sixth time convolutional network layer TCN6, so as to obtain the final time series features of the restored training time series data.

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

[0137] First, the output features of the corresponding TCN layer are copied through the skip connection structure and fused into feature U by adding the corresponding elements. Then, the efficient channel attention mechanism is used to calculate the corresponding weights, and the fused feature U is globally averaged and pooled to obtain the compressed feature S. One-dimensional convolution is further used for 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 the weighted features are fused by element-wise addition to obtain the final output features.

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

[0139]

[0140] in Indicates the addition of corresponding elements. Then, the ECA structure is used to perform global average pooling on U to compress the information of each channel from the time dimension:

[0141]

[0142] Among them, i represents the corresponding channel position, C is the feature dimension, L represents the time series length of the fused feature U, and t represents the corresponding time series position. The compressed representation S is convolved in one dimension to realize channel interaction and generate the weight matrices Z1, Z2, and Z3 of the corresponding branches. In the selection stage, unlike the ECA module which uses sigmoid to calculate weights, the feature fusion module uses softmax to convert the obtained weight matrices Z1, Z2, and Z3 into the weights of the corresponding branches:

[0143]

[0144] Where K j Represents the attention weight of the corresponding branch j. After obtaining each weight, the time series features extracted by the corresponding branch are weighted, and finally the weighted features are fused to obtain the final time series feature V:

[0145]

[0146] In the formula, represents the multiplication of corresponding elements, Indicates the addition of corresponding elements. This feature fusion module weights the time series features of different scales in the channel direction and obtains effective fusion features, which can focus on the important time series information and enhance the richness of the overall features.

[0147] Furthermore, if Fig. 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 according to the training results, including:

[0148] S210, inputting the training time series data into the industrial process fault detection model for forward calculation to obtain 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, the network model is trained using the time-based back propagation algorithm, and 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, calculating the loss function value of the reconstructed data corresponding to the training time series data through a mean square error function;

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

[0152]

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

[0154] S230. Performing gradient update on the industrial process fault detection model according to the loss function value to obtain a target industrial process fault detection model.

[0155] like Fig.10 As shown in Figure 2, the gradient update process specifically includes:

[0156] S231. According to the loss function value, the network weight of the industrial process fault detection model is adjusted using the Adam optimization algorithm to obtain a target network weight;

[0157] 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 cycle, and requires less computing space. The target network weight here specifically refers to the network weight that minimizes the loss function value through training.

[0158] S232. According to the loss function value, a grid search method is used to adjust the network hyperparameters of the industrial process fault detection model to obtain target hyperparameters;

[0159] The grid search (GS) method is used to tune the network hyperparameters. First, the candidate confidence values ​​of different hyperparameters are determined based on prior knowledge to generate a grid to be searched. Then, the GS method is used to tune the hyperparameters of the industrial process fault detection model and select the optimal target hyperparameter combination. The target hyperparameter here refers to the hyperparameter that minimizes the loss function value after training. Finally, according to the target hyperparameter combination, the appropriate training round is selected to complete the model training and save the network parameters.

[0160] S233. Obtain a target industrial process fault detection model according to the target network weights and target hyperparameters.

[0161] Another embodiment of the present invention provides an industrial process fault detection method based on a time convolutional autoencoder for Tennessee Eastman (TE), wherein the TE process consists of a reactor, a condenser, a circulating compressor, a separator, and a stripping tower, etc. Fig.11 , the method comprising:

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

[0163] This embodiment uses simulation experiment data, which is collected in the TE process simulation system. The TE process includes 12 operating variables, 22 continuous measurement values ​​and 19 component variables. The collected data includes 22 sub-datasets, one of which is data collected under normal conditions, and the remaining 21 are fault data samples. Each data set is sampled once every 3 minutes, and the total time is two days, that is, each data set includes 960 samples. In order to generate a fault data set, the corresponding fault is introduced after 8 hours of normal operation, so each fault data set contains 160 normal samples and 800 fault samples in turn.

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

[0165] By analyzing and preprocessing the sample data (including missing values, abnormal values, and constant value variable processing, etc.), and combining process mechanism information with relevant expert experience, a candidate input variable set consisting of 11 operating variables and 22 measurement variables was finally determined. The corresponding variable descriptions are shown in Table 1. At the same time, 18 commonly used fault sub-datasets were selected for detection, excluding the 3rd, 9th, and 15th faults. All fault types are shown in Table 2.

[0166] Table 1 Candidate input variables for TE process

[0167]

[0168]

[0169] Table 2TE process fault description

[0170]

[0171] Before fault detection, the collected data is standardized and resampled by sliding window. First, the normal working condition data in the sample data is used as training data, and its mean and standard deviation are calculated, and then z-score standardization is performed. The fault data in the sample data is used as test data, and z-score standardization is performed on it in combination with the mean and standard deviation of the training data. Finally, the standardized training data and test data are resampled using a sliding window of the same width to obtain time series data for training and testing of the temporal convolutional autoencoder.

[0172] S3: Building and training an industrial process fault detection model for Tennessee Eastman:

[0173] 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 back propagation algorithm. The specific training process mainly includes three steps:

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

[0175] (2) Reverse calculation: Calculate the 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 status of the network. The smaller the MSE value, the better the reconstruction effect of the data set.

[0176] (3) Gradient update: 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 cycle, and requires less computational space.

[0177] The grid search (GS) method is used to tune the network hyperparameters. First, the candidate confidence values ​​of different hyperparameters are determined based on prior knowledge to generate a grid to be searched; then the GS method is used to tune the TCAE-FF network hyperparameters and select the optimal hyperparameter combination. Finally, according to the optimal hyperparameter combination, the appropriate training round is selected to complete the model training and save the network parameters.

[0178] S4: Calculate the statistics and corresponding control limits of normal samples:

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

[0180] After obtaining the SPE statistic, use the kernel density estimation to calculate the corresponding control limits. For details, refer to formula (3)

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

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

[0183] S6: Perform fault detection:

[0184] The control limit calculated from the normal sample statistics is used as the judgment standard, and the statistics corresponding to the test data are 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] In order to analyze the fault detection performance of the method of the present application from multiple angles, the present embodiment is compared with the fault detection method based on kernelized principal component analysis (KPCA), the fault detection method based on CAE, the fault detection method based on long short-term memory autoencoder (LSTM Auto-Encoder, LSTM-AE) and the fault detection method based on TCAE. At the same time, the present application adopts the following two parameters as the evaluation indicators of each method for fault detection, 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] Among them, N1 and N2 are the number of fault samples and the number of normal samples respectively, N 1,f and N 2,f are the number of samples detected as faulty samples and normal samples, respectively.

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

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

[0193] Table 3 Detection results of five methods for TE process

[0194]

[0195] Comparing the detection results of the above five methods, from a general point of view, under the premise that the average false alarm rate is not much different, the fault detection methods based on TCAE and TCAE-FF provided by the present invention have good performance, and the TCAE-FF method performs better. The average alarm rate of the TCAE-FF method is 25.09% higher than that of the traditional KPCA method, indicating that the fault detection accuracy can be effectively improved by adding TCN to extract the input time series features and fusing the extracted time series features through the feature fusion module. At the same time, the TCAE-FF method is 14.59%, 10.08%, and 4.59% higher than the same type of deep learning methods CAE, LSTM-AE, and TCAE, respectively, which further reflects the effectiveness of the feature fusion module, can more effectively extract the time series features of different scales of the input data, enhance the model's ability to distinguish normal samples from fault samples, and improve the fault detection rate. Observing the detection results of various types of faults, the TCAE-FF method has obvious improvements in the detection of faults 5, 10, 16, 19, and 20 compared with the other four methods. Among them, the SPE statistic curve corresponding to fault 10 is shown as follows. Figures 12 to 16 shown.

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

[0197] A preprocessing module 31 is used to use the acquired normal working condition data of the industrial process as training data, preprocess the training data to obtain training time series data, and to preprocess the acquired working condition data to be tested of the industrial process to obtain test time series data;

[0198] A model training module 32, used for inputting the training time series data into the industrial process fault detection model for training, and obtaining a target industrial process fault detection model according to the training results, wherein the industrial process fault detection model is constructed based on a temporal convolutional network and a convolutional autoencoder;

[0199] A control limit calculation module 33, used to input the training time series data into the target industrial process fault detection model, and calculate the control limit of the square prediction error statistic of the training time series data;

[0200] A square 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 square prediction error statistic of the test time series data;

[0201] The judgment module 35 is used to judge whether the working condition data to be tested is abnormal according to the control limit of the square prediction error statistic of the training time series data and the square prediction error statistic of the test time series data.

[0202] An industrial process fault detection device 30 based on a temporal convolutional autoencoder provided in an embodiment of the present invention further introduces a temporal convolutional network on the basis of a convolutional autoencoder, designs a new temporal convolutional autoencoder (TCAE), and creates and trains an industrial process fault detection model based on the new temporal convolutional autoencoder for fault detection of industrial process related data. After at least two stacked temporal convolutional networks, the convolution receptive field can be gradually expanded to change the time scale of temporal feature extraction of industrial process related data, enrich the time scale of extracted temporal features, 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 an industrial process fault detection device 30 based on a time convolutional autoencoder provided in an embodiment of the present invention can refer to the above-mentioned principle description of an industrial process fault detection method based on a time convolutional autoencoder, which will not be repeated here.

[0204] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD 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 principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An industrial process fault detection method based on a temporal convolutional autoencoder, characterized in that: include: Using the acquired normal working condition data of the industrial process as training data, using the acquired working condition data to be tested of the industrial process as test data, and preprocessing the training data and the test data respectively to obtain training time series data and test time series data; Inputting the training time series data into an industrial process fault detection model for training, and obtaining a target industrial process fault detection model according to the training results, wherein the industrial process fault detection model is constructed based on a convolutional autoencoder and at least two layers of temporal convolutional networks; Inputting the training time series data into the target industrial process fault detection model, and calculating the control limit of the squared prediction error statistic of the training time series data; Inputting the test time series data into the target industrial process fault detection model, and calculating the square prediction error statistic of the test time series data; Whether the operating condition data to be tested is abnormal is determined based on the control limit of the square prediction error statistic of the training time series data and the square prediction error statistic of the test time series data.

2. The method according to claim 1, characterized in that: The industrial process fault detection model includes an encoder and a decoder; the output end of the encoder is connected to the input end of the decoder; The encoder is used to extract the time series features of the training time series data according to the time convolution network layers with at least two different expansion coefficients, and to obtain the latent variables corresponding to the training time series data according to the extracted time series features of the training time series data; The decoder is used to restore the latent variables output by the encoder into time series features according to a temporal convolutional network layer with at least two different expansion coefficients, and to obtain reconstructed data with the same dimension as the training time series data according to the restored time series features.

3. The method according to claim 2, characterized in that 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; 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; The first one-dimensional convolutional network layer is connected to the fourth temporal convolutional network layer; Among them, the expansion coefficients of the first time convolutional network layer and the sixth time convolutional network layer are 2 x , the expansion coefficient of the second time convolutional network layer and the fifth time convolutional network layer is 2 y , the expansion coefficient of the third time convolutional network layer and the sixth time convolutional network layer is 2 z , x<y<z; The first time convolutional network layer, the second time convolutional network layer and the third time convolutional network layer are all used to extract the time series features of the training time series data at the corresponding time scale according to the corresponding expansion coefficient; The first one-dimensional convolutional network layer is used to compress the feature dimension of the time series features of the training time series data output by the third time convolutional network layer to obtain latent variables corresponding to the training time series data; The fourth time convolutional network layer, the fifth time convolutional network layer and the sixth time convolutional network layer are all used to restore the latent variables output by the first one-dimensional convolutional network layer to time series features according to the corresponding expansion coefficients at the corresponding time scales; The second one-dimensional convolutional network layer is used to obtain reconstructed data with the same dimension as the training time series data according to the restored time series features output by the sixth time convolutional network layer.

4. The method according to claim 2, characterized in that: The encoder further includes a first feature fusion module, and the decoder further includes a second feature fusion module; 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 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; The first feature fusion module is used for weighted fusion of the time series features of the training time series data extracted by the first time convolutional network layer, the second time convolutional network layer and the third time convolutional network layer, respectively, to obtain the final time series features of the training time series data; The first one-dimensional convolutional network layer is used to compress the feature dimension of the final time series feature of the training time series data output by the first feature fusion module to obtain the latent variable corresponding to the training time series data; The second feature fusion module is used for weighted fusion of the time series features restored by the fourth time convolutional network layer, the fifth time convolutional network layer and the sixth time convolutional network layer, respectively, to obtain the final time series features of the restored training time series data; The second one-dimensional convolutional network layer is used to obtain reconstructed data with the same dimension as the training time series data according to the restored final time series features of the training time series data.

5. The method according to claim 4, characterized in that 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 according to the time series features respectively output by the first time convolutional network layer, the second time convolutional network layer, and the third time convolutional network layer; The first weight calculation unit is used to calculate the weights corresponding to the first time convolutional network layer, the second time convolutional network layer, and the third time convolutional network layer respectively according to the first fusion feature through an efficient channel attention mechanism; The first weighted output unit is used to perform weighted fusion on the time series features respectively output by the first time convolutional network layer, the second time convolutional network layer, and the third time convolutional network layer according to the weights respectively corresponding to the first time convolutional network layer, the second time convolutional network layer, and the third time convolutional network layer, so as to obtain the final time series features of the training time series data; The second fusion unit is used to obtain a second fusion feature according to the restored time series features respectively output by the fourth time convolutional network layer, the fifth time convolutional network layer, and the sixth time convolutional network layer; The second weight calculation unit is used to calculate the weights corresponding to the fourth time convolutional network layer, the fifth time convolutional network layer, and the sixth time convolutional network layer respectively according to the second fusion feature through an efficient channel attention mechanism; The second weighted output unit is used to perform weighted fusion on the time series features output by the fourth time convolutional network layer, the fifth time convolutional network layer and the sixth time convolutional network layer respectively according to the weights corresponding to the fourth time convolutional network layer, the fifth time convolutional network layer and the sixth time convolutional network layer respectively, so as to obtain the final time series features of the restored training time series data.

6. The method according to any one of claims 1 to 5, 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 according to the training result, comprises: Inputting the training time series data into an industrial process fault detection model for forward calculation to obtain reconstructed data corresponding to the training time series data; Calculate the loss function value of the reconstructed data corresponding to the training time series data by using a mean square error function; The industrial process fault detection model is gradient updated according to the loss function value to obtain a target industrial process fault detection model.

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

8. The method according to claim 1, characterized in that The preprocessing includes data normalization and sliding window resampling.

9. The method according to claim 1, characterized in that: The determining whether the operating condition data to be tested is abnormal according to the control limit of the square prediction error statistic of the training time series data and the square prediction error statistic of the test time series data includes: When the square prediction error statistic of the test time series data is greater than the control limit of the square prediction error statistic of the training time series data, it is judged that the working condition data to be tested is abnormal; When the square prediction error statistic of the test time series data is less than or equal to the control limit of the square prediction error statistic of the training time series data, it is judged that the working condition data to be tested is normal.

10. An industrial process fault detection device based on a time convolutional autoencoder, applied to the industrial process fault detection method based on a time convolutional autoencoder according to any one of claims 1 to 9, characterized in that: include: A preprocessing module, used to use the acquired normal working condition data of the industrial process as training data, preprocess the training data to obtain training time series data, and used to preprocess the acquired working condition data to be tested of the industrial process to obtain test time series data; A model training module, used for inputting the training time series data into an industrial process fault detection model for training, and obtaining a target industrial process fault detection model according to the training results, wherein the industrial process fault detection model is constructed based on a temporal convolutional network and a convolutional autoencoder; A control limit calculation module, used for inputting the training time series data into the target industrial process fault detection model, and calculating the control limit of the square prediction error statistic of the training time series data; A square prediction error statistic calculation module, used for inputting the test time series data into the target industrial process fault detection model, and calculating the square prediction error statistic of the test time series data; A judgment module is used to judge whether the working condition data to be tested is abnormal based on the control limit of the square prediction error statistic of the training time series data and the square prediction error statistic of the test time series data.

Images