Industrial process intelligent fault diagnosis method and system based on multi-scale depth separable convolution and cross attention fusion

Abstract

The invention provides an industrial process intelligent fault diagnosis method based on multi-scale depth separable convolution and cross attention fusion. The method comprises the following steps: preprocessing monitoring data to obtain an input feature map; parallel multi-scale two-dimensional depth separable convolution operation is adopted to carry out multi-scale feature extraction on the input feature map to obtain a multi-scale local space-time feature map; obtaining a multi-scale enhanced feature map through multi-scale channel attention; fusing the multi-scale enhanced feature map through a cross-scale cross self-attention mechanism; capturing a global long-distance time sequence dependency relationship in the fusion feature map in the industrial process through a Transform encoder; and outputting a fault classification result according to the feature vector of the global context. The invention further provides an industrial process intelligent fault diagnosis system. The method is used for solving the technical problems that feature extraction is insufficient, a multi-source information fusion mechanism is static and poor in adaptability, and global dependence modeling is insufficient in an existing method.

Description

Technical Field

[0001] This invention relates to the technical field of industrial process fault diagnosis, and in particular to an intelligent fault diagnosis method and system for industrial processes. Background Technology

[0002] In modern industrial production, such as chemical, power, and manufacturing industries, the safe, stable, and efficient operation of process control systems is crucial. Industrial processes typically consist of numerous interconnected sensors, actuators, and control loops, making the systems complex and highly nonlinear, prone to various faults. If faults are not detected and diagnosed in a timely and accurate manner, they can lead to decreased production efficiency, equipment damage, and even serious safety accidents and environmental pollution. Therefore, developing efficient and accurate fault diagnosis technologies is of great significance for ensuring industrial production.

[0003] Existing fault diagnosis technologies can be mainly divided into mechanistic model-based methods and data-driven methods. Mechanism-based methods rely on accurate mathematical and physical modeling of industrial processes, but when faced with complex and nonlinear modern industrial systems, establishing accurate mechanistic models is extremely difficult and costly.

[0004] With the development of Industrial Internet of Things (IIoT) technology, a large amount of process data has been collected, making data-driven fault diagnosis methods a research hotspot. These methods mainly include statistical methods, shallow machine learning, and deep learning methods. Statistical methods, such as Principal Component Analysis (PCA) and Partial Least Squares (PLS), are primarily suitable for linear, steady-state processes, but have limited ability to handle the nonlinear and non-stationary dynamic characteristics prevalent in modern industrial processes. Shallow machine learning methods, such as Support Vector Machines (SVM) and traditional Artificial Neural Networks (ANN), can handle certain nonlinear problems, but their feature representation capabilities are limited, often requiring complex manual feature engineering based on expert experience, and they are prone to getting trapped in local optima when processing high-dimensional data.

[0005] In recent years, deep learning methods, represented by Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), have shown great potential in the field of fault diagnosis, capable of automatically learning hierarchical feature representations from raw data. However, existing deep learning fault diagnosis methods still suffer from the following technical shortcomings in practical applications:

[0006] Multi-scale spatiotemporal feature extraction is insufficient and inefficient: Existing methods mostly employ one-dimensional convolutional networks (1D-CNN), which primarily focus on the evolution patterns of sensor data along the time axis, while neglecting the concurrent interactions between different sensor variables at the same or nearby times (i.e., spatiotemporal correlations across variables). Although standard two-dimensional convolutional networks (2D-CNN) can capture such spatiotemporal features, their high parameter count and computational complexity make it difficult to meet the real-time requirements of industrial scenarios. Furthermore, single-scale convolutional kernels cannot effectively capture the diverse manifestations of industrial faults at different time scales (such as the short-term characteristics of abrupt faults and the long-term characteristics of slowly changing faults).

[0007] Multi-source information fusion mechanisms are static and lack adaptability: In industrial processes, the contribution of measurement data from different sensors to fault diagnosis is not equal. Existing methods typically treat all sensor channels equally, leading to key fault information being easily diluted by noise or redundant information. Furthermore, when fusing features extracted from different scales, simple static fusion strategies such as splicing or averaging are often used. These strategies cannot dynamically adjust the importance of features at each scale according to the specific context of the input data, nor can they model complex dependencies between scales, thus limiting the discriminative power of the fused features.

[0008] Insufficient modeling of global temporal dependencies: Many industrial failures develop gradually, with their final failure state closely related to multiple events or states spanning a long time span in historical data. Models using CNNs as the main feature extractor, due to their inherent limitations in local receptive fields, struggle to effectively capture these long-distance temporal dependencies, resulting in insufficient diagnostic capabilities for slowly changing or complexly evolving failures.

[0009] The invention patent with publication number CN118797321A discloses an industrial process fault detection method and system based on spatiotemporal feature attention fusion. By adopting a spatiotemporal feature attention fusion model in industrial process fault detection, and combining multi-channel convolution and attention mechanism, it solves the problem of insufficient spatiotemporal feature extraction in the prior art and achieves higher fault detection accuracy and reliability. However, this invention method adopts a single-scale convolution design, which has insufficient multi-scale feature extraction capability, weak global temporal dependency modeling, and low computational efficiency. Summary of the Invention

[0010] To address the technical problems of insufficient feature extraction, static and poorly adaptable multi-source information fusion mechanisms, and inadequate global dependency modeling in existing methods for industrial process monitoring data, this invention proposes an intelligent fault diagnosis method and system for industrial processes based on multi-scale depthwise separable convolutional and cross-scale attention fusion (MS-DSC-CSAFF-Transformer, Multi-Scale Depthwise Separable Convolutional Cross-Scale Attention Feature Fusion Transformer). By introducing two-dimensional depthwise separable convolution, the comprehensiveness and efficiency of feature extraction are improved. The importance of features at each scale is dynamically adjusted through the cross-scale self-attention mechanism, effectively integrating complementary information, reducing redundancy and inconsistency of multi-scale features, and enhancing the adaptability of information fusion. The Transformer encoder is used to capture global long-distance temporal dependencies, thereby improving the accuracy and robustness of fault diagnosis.

[0011] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0012] A method for intelligent fault diagnosis in industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion, comprising the following steps:

[0013] S1: Input multivariate time series monitoring data of industrial processes, and preprocess the monitoring data to obtain an input feature map;

[0014] S2: Parallel multi-scale two-dimensional depthwise separable convolution operations are used to extract multi-scale features from the time dimension and sensor variable dimension of the input feature map to obtain a multi-scale local spatiotemporal feature map;

[0015] S3: Enhance the feature response of important sensor channels in the multi-scale local spatiotemporal feature map by multi-scale channel attention to obtain the multi-scale enhanced feature map;

[0016] S4: By fusing multi-scale enhanced feature maps through a cross-scale self-attention mechanism, the repetitive information at different scales of sensor data is suppressed, and the scale consistent with the global fault mode is selected to dominate the fusion result, thereby obtaining the fused feature map.

[0017] S5: Capture global long-range temporal dependencies in the industrial process from the fused feature map using a Transformer encoder to obtain the feature vector of the global context;

[0018] S6: Output the fault classification result based on the feature vector of the global context.

[0019] Furthermore, the preprocessing method involves: standardizing the monitoring data; and using a sliding window to divide the standardized monitoring data into multiple two-dimensional tensors. H represents the sequence length, and W is the number of sensor variables; the reshape operation adjusts each sub-2D tensor to a format suitable for 2D convolution input, making each 2D tensor... The shape changes from H×W to H×W×C, resulting in the input feature map. C represents the number of input channels.

[0020] Furthermore, in step S2, three parallel two-dimensional depthwise separable convolution operations are employed, each applying a different kernel size to sequentially process the input feature map. Depthwise convolution and pointwise convolution are performed. Depthwise convolution extracts spatial information, while pointwise convolution performs channel mapping, ultimately yielding local spatiotemporal feature maps at three scales.

[0021] Furthermore, the method for extracting spatial information through depthwise convolution is as follows: each two-dimensional depthwise separable convolution operation is applied to the input feature map. Each channel is independently spatially convolved to extract local features, resulting in a depthwise convolution output local channel feature map X. depthwise,h,w,c The spatial dimensions of the spatial convolution include a time dimension related to the sequence length H and a variable dimension related to the number of sensor variables W.

[0022] The method of channel mapping through pointwise convolution is as follows: The local channel feature map X output by convolution at each depth is... depthwise,h,w, By applying 1×1 convolution, channel information is fused and the feature dimension is expanded to obtain the channel feature map X output by pointwise convolution. h,w,c′ By stitching the images together according to the channels, local spatiotemporal feature maps at each scale can be obtained. Where h is the time dimension direction index, w is the variable dimension direction index, and c and c' are channel indices; H', W', and C' are local spatiotemporal feature maps, respectively. The time dimension, variable dimension, and number of channels are defined, where n is the index of the two-dimensional depthwise separable convolution operation.

[0023] Furthermore, in step S3, three channel attention operations are deployed in parallel. Each channel attention operation outputs a local spatiotemporal feature map for each two-dimensional depthwise separable convolution operation. Global average pooling and global max pooling are applied to each channel respectively, and the channel description vectors of global average pooling are obtained respectively. and global max pooling channel description vector Global average pooling channel description vector and global max pooling channel description vector The inputs are fed into an MLP network and then concatenated. Channel attention weights A are generated using a sigmoid activation function. c′ ; Adjust the channel attention weight A c′ With each channel feature map X h,w,c Perform multiplication to obtain the enhanced feature map X for each channel. h,w,c ; Enhance the feature map X of each channel h,w,c Enhanced feature maps at each scale are obtained by concatenating channels.

[0024] Furthermore, in step S4, feature maps enhanced at all scales are first aligned using feature alignment. By unifying the dimensions, local spatiotemporal feature maps at three scales are obtained after feature alignment. B and T represent the time dimension and variable dimension after feature alignment, respectively, based on the local spatiotemporal feature maps at three scales. Calculate the query matrix Q csaff Key matrix K csaff Sum matrix V csaff ; Calculate context-aware feature maps R at different scales based on scaled dot product attention; Local spatiotemporal feature maps at three scales. Cross-scale fusion is performed between the two scales to obtain the cross-scale fused feature map X. fusion And utilize context-aware feature map R to fuse feature map X across scales. fusion Weighting is performed to obtain the final fused feature map F. csaff .

[0025] Furthermore, in the calculation of the query matrix Q... csaff Key matrix K csaff Sum matrix V csaff When querying, use one scale as the query and the other two scales as the key and value, respectively, for calculation;

[0026] The method for calculating the context-aware feature map R at different scales is as follows:

[0027]

[0028] Where Sotfmax(·) is the activation function, d k The scaling factor is ·, which represents the dot product.

[0029] The process of obtaining cross-scale fusion feature maps X is described above. fusion The method is as follows:

[0030]

[0031] The process of obtaining the final fused feature map F csaffThe method is as follows:

[0032] F csaff =RX fusion .

[0033] Furthermore, in step S5, the final fused feature map is... Flattening into a tensor tensor The query matrix Q, key matrix K, and value matrix V are linearly projected; the self-attention weights of h heads are calculated through a single Transformer encoding operation; and the self-attention weights of the outputs of each head are concatenated to obtain the multi-head attention weight MultiHead(X). csaff MultiHead(X) attention weights csaff ) and the input tensor X csaff After performing residual connections and layer normalization, the feature tensor F is obtained. csaff1 The feature vector of the global context is obtained based on the feedforward network:

[0034] F csaff2 =LayerNorm(FFN(F csaff1 )+F csaff1 )

[0035] Where LayerNorm(·) is layer normalization and FFN(·) is feedforward network;

[0036] After layer normalization following the Transformer encoding operation, two Dropout layers with a dropout rate of 0.5 were configured.

[0037] Furthermore, an intelligent fault diagnosis system for industrial processes includes, in sequence, a data preprocessing and sample construction module, a multi-scale spatiotemporal feature extraction module, a multi-scale channel attention module, a cross-scale cross-self-attention fusion module, a global time-dependent modeling module, and a fault classification module, wherein:

[0038] Data preprocessing and sample construction module: used to input multivariate time series monitoring data of industrial processes, and to preprocess the monitoring data to obtain input feature maps;

[0039] Multi-scale spatiotemporal feature extraction module: used to extract multi-scale features from the time dimension and sensor variable dimension of the input feature map by using parallel multi-scale two-dimensional depthwise separable convolution operations to obtain multi-scale local spatiotemporal feature maps;

[0040] Multi-scale channel attention module: used to enhance the feature responses of important sensor channels in multi-scale local spatiotemporal feature maps through multi-scale channel attention to obtain multi-scale enhanced feature maps;

[0041] Cross-scale self-attention fusion module: It is used to fuse multi-scale enhanced feature maps through a cross-scale self-attention mechanism to suppress repetitive information at different scales of sensor data and select the scale consistent with the global fault mode to dominate the fusion result and obtain fused feature maps.

[0042] Global Temporal Dependency Modeling Module: Used to capture global long-range temporal dependencies in industrial processes from the fused feature map using a Transformer encoder, and obtain the feature vector of the global context.

[0043] Fault classification module: Used to output fault classification results based on the feature vector of the global context.

[0044] Furthermore, the multi-scale spatiotemporal feature extraction module includes three parallel-deployed two-dimensional depthwise separable convolution modules to implement the three parallel two-dimensional depthwise separable convolution operations.

[0045] The multi-scale channel attention module includes three channel attention modules deployed in parallel to implement the three channel attention operations.

[0046] The beneficial effects of this invention are as follows:

[0047] The invention improves the comprehensiveness and efficiency of feature extraction: It adopts multi-scale two-dimensional depthwise separable convolution (2D-DSC), which can not only fully capture spatiotemporal features, but also ensure the coverage of fault features at different time scales through multi-scale design. At the same time, compared with standard 2D-CNN, it greatly reduces the amount of computation and parameters, improves the efficiency of feature extraction, and provides a comprehensive and high-quality feature foundation for subsequent fault diagnosis.

[0048] Enhanced attention to key information: By introducing a channel attention mechanism (CAM) after each scale, this invention can adaptively identify and dynamically enhance the feature response of sensor channels containing key fault information, suppress noise and redundant information, and improve the signal-to-noise ratio and discriminative power of the features.

[0049] This invention achieves dynamic and efficient multi-scale information fusion: The Cross-Self-Attention Feature Fusion (CSAFF) module can dynamically assign weights to features at each scale and fuse them based on the interdependencies between fault features in complex industrial processes at different scales. This overcomes the limitations of traditional static fusion methods, effectively integrates complementary information, reduces redundancy, generates more robust and discriminative fused features, and solves the problems of redundancy and inconsistency among multi-scale features.

[0050] This invention enhances the ability to model global temporal dependencies: Based on the effective extraction and fusion of multi-scale local features, it utilizes a Transformer encoder to perform global context modeling on the fused features, which can accurately capture the correlation of fault modes across long time intervals, further improve the understanding of fault features from a global perspective, and make up for the shortcomings of traditional CNN and RNN in long-distance dependency modeling.

[0051] Improved fault diagnosis performance: Through the synergistic effect of multi-scale feature extraction, key information enhancement, dynamic feature fusion and global temporal dependency modeling, this invention achieves comprehensive feature learning and fusion from local to global and from spatiotemporal details to long-range dependencies. Experimental verification shows that it can significantly improve the accuracy and robustness of fault diagnosis in complex industrial processes. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0053] Figure 1 This is a schematic diagram of the MS-DSC-CSAFF-Transformer model structure of the present invention.

[0054] Figure 2 This is a T-SNE visualization result of the TE process before and after classification in this invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1

[0057] A method for intelligent fault diagnosis in industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion, such as... Figure 1 As shown, the steps include:

[0058] S1: Input multivariate time series monitoring data of industrial processes, and preprocess the monitoring data to obtain input feature maps.

[0059] The preprocessing method is as follows:

[0060] First, 52 process monitoring variables were selected from the multivariate time series monitoring data as input features, i.e., the variable dimension was 52, and the monitoring data was standardized using the Z-score standardization method.

[0061] Segmentation: The standardized multivariate time series monitoring data is segmented into multiple two-dimensional tensors according to the window length H (H=5 in this embodiment). The shape is (5, 52); where the window length H is the sequence length of each two-dimensional tensor, i.e., the time dimension, and W is the number of sensor variables, i.e., the variable dimension. By segmenting the data, it is made suitable for window-based feature extraction, providing a data input basis for subsequently solving the gradient vanishing problem that RNN / LSTM may encounter when processing ultra-long sequences.

[0062] Shape transformation: Transform each 2D tensor using the reshape operation. Adjusting the format to suit 2D convolution input involves adding a new dimension (channel dimension) to make the 2D tensor... Shape becomes Thus, the input feature map is obtained. As input data for subsequent operations, where C is the number of channels, initially set to 1, therefore the input feature map in this embodiment... The shape is (5,52,1).

[0063] Traditional 1D convolutions can only handle the temporal dimension, ignoring the interactions between variables; while standard 2D convolutions can capture spatiotemporal features but are computationally expensive. Here, converting the data into a 2D shape provides structured input for subsequent use of 2D depthwise separable convolutions to simultaneously extract temporal and variable-dimensional features. Additionally, adding a channel dimension (initially 1) allows the model to expand feature channels in later steps.

[0064] S2: Parallel multi-scale two-dimensional depthwise separable convolution operations are used to extract multi-scale features from the time dimension and sensor variable dimension of the input feature map to obtain a multi-scale local spatiotemporal feature map.

[0065] Three parallel two-dimensional depthwise separable convolutional operations are employed, each applying a different kernel size to the input feature map. Depthwise convolution and pointwise convolution are performed. Spatial information is extracted through depthwise convolution, and channel mapping is performed through pointwise convolution, ultimately resulting in local spatiotemporal feature maps at three scales.

[0066] Depthwise convolution: on the input feature map Each channel is independently convolved spatially (time and variable dimensions) to extract local features:

[0067]

[0068] Among them, W depthwise,i,j,c Let be the weight parameters in the depthwise convolution kernel, i be the temporal direction index in the kernel, j be the variable direction index in the kernel, c be the channel index of the input feature map X0, and the kernel size be k×k (3×3, 5×5, or 7×7 in this embodiment). h+i,w+j,c For the element value in the input feature map X0, corresponding to the convolutional kernel offset by i in the time dimension H and offset by j in the variable dimension W, with channel position c, X depthwise,h,w,c The local channel feature map is output by the depthwise convolution, where h is the time dimension orientation index and w is the variable dimension orientation index.

[0069] Pointwise convolution: The local channel feature map X output by convolution at each depth. depthwise,h,w,c Applying 1x1 convolutions to fuse channel information and expand feature dimensions:

[0070]

[0071] Among them, W pointwise,c,c′ Here, c' represents the weight parameters in the pointwise convolution kernel, and c' represents the index of the output channel number. h,w,c′ The channel feature maps output by pointwise convolution are concatenated along the channels to obtain local spatiotemporal feature maps at each scale. H', W', and C' are local spatiotemporal feature maps, respectively. The time dimension, variable dimension, and number of channels are the indexes of the two-dimensional depthwise separable convolution operations.

[0072] The local spatiotemporal feature maps obtained in this embodiment at three scales are as follows: small-scale local spatiotemporal feature map Mesoscale local spatiotemporal feature map Large-scale local spatiotemporal feature maps

[0073] Industrial fault characteristics often manifest at different time scales (such as short-term abrupt changes or long-term drifts). Multi-scale design ensures comprehensive coverage of various fault modes. Parallel two-dimensional depthwise separable convolution operations extract features from local to macroscopic levels (using three different scale convolution kernels: 3×3 extracts fine-grained local features, 5×5 captures a wider range of local patterns, and 7×7 provides more global feature representation capabilities), avoiding feature omissions caused by a single scale and improving the comprehensiveness of feature extraction. Depthwise convolution focuses on the spatiotemporal dimension (reducing computational complexity), and pointwise convolution processes channel relationships, reducing computation and parameter count compared to standard 2D convolution, achieving efficient feature extraction.

[0074] S3: Enhance the feature response of important sensor channels in the multi-scale local spatiotemporal feature map by multi-scale channel attention to obtain the multi-scale enhanced feature map.

[0075] Since fault features often exhibit non-uniform distribution across different channels, the model needs to focus on channels containing important information while suppressing redundant or interfering signals. Therefore, this invention introduces a channel attention mechanism combining Global Average Pooling (GAP) and Global Max Pooling (GMP). This mechanism improves the response of key features through adaptive channel weighting, optimizes feature selection capabilities, and thus enhances the discriminative performance of the classification model.

[0076] Three-channel attention operations are deployed in parallel, with each channel attention operation generating a local spatiotemporal feature map for each depthwise separable convolution operation. Each channel is subjected to Global Average Pooling (GAP) and Global Max Pooling (GMP) respectively, and the channel description vectors obtained by Global Average Pooling are obtained respectively. and global max pooling channel description vector

[0077]

[0078] Where a is the local spatiotemporal feature map X h,w,c′ In the time dimension, the directional index is b, which is the local spatiotemporal feature map X. h,w,c′ Dimensional index of variables.

[0079] Furthermore, the global average pooling channel description vector is... and global max pooling channel description vector The inputs are fed into an MLP network and then concatenated. Channel attention weights A are generated using a sigmoid activation function. c′ :

[0080]

[0081] Furthermore, the channel attention weight A c′ With each channel feature map X h,w,c Perform multiplication to obtain the enhanced feature map for each channel:

[0082] X′ h,w,c′ =A c′ ·X h，w，c′ .

[0083] Furthermore, the enhanced feature map X′ for each channel is... h,w,c′Enhanced feature maps at each scale are obtained by concatenating channels.

[0084] The enhanced feature maps obtained in this embodiment at three scales are as follows: small-scale local spatiotemporal feature map. Mesoscale local spatiotemporal feature map Large-scale local spatiotemporal feature maps

[0085] In industrial data, different sensor channels contribute unevenly to fault diagnosis (some sensors contain critical information, while others may be noisy). Channel attention mechanisms adaptively enhance important channels, enabling the model to focus on features more meaningful for fault diagnosis, suppressing noise, and improving feature discriminative power. The design of multi-scale depthwise separable convolutions allows the model to capture local features of time-series data at different scales, while the channel attention mechanism adaptively weights the importance of features from different channels, allowing the model to focus on features more meaningful for fault diagnosis. This effectively filters and enhances data during the feature extraction stage, laying the foundation for subsequent time-series modeling and classification tasks.

[0086] S4: By fusing multi-scale enhanced feature maps through a cross-scale self-attention mechanism, the repetitive information at different scales of sensor data is suppressed, and the scale consistent with the global fault mode is selected to dominate the fusion result, thereby obtaining the fused feature map.

[0087] Fault diagnosis fusion features are formed by integrating multi-scale features. In industrial process monitoring, features at different scales often contain complementary discriminative information. Traditional feature fusion methods, such as simple concatenation or weighted averaging, struggle to fully uncover the intrinsic connections between cross-scale features, limiting the model's expressive power and diagnostic performance. To address these issues, this invention proposes a cross-scale cross-self-attention fusion (CSAFF) module, which dynamically fuses multi-scale features extracted from depthwise separable convolutions at different scales and assigns corresponding weights based on their importance in specific fault modes, achieving adaptive fusion of multi-scale features.

[0088] Feature alignment: Since feature maps enhanced at different scales may have different sizes, all feature maps are first made to be the same size by adjustment (such as padding or upsampling).

[0089] In this embodiment, the feature-aligned small-scale local spatiotemporal feature map is used as the query, the feature-aligned medium-scale local spatiotemporal feature map is used as the key, and the feature-aligned large-scale local spatiotemporal feature map is used as the value. Alternatively, different combinations of the three scales can be used, with any scale as the query and the other two scales as the key and value, respectively. The calculation methods for the query matrix, key matrix, and value matrix are as follows:

[0090] Q csaff =W q X′1,K csaff =W k X′2,V csaff =W v X′3;

[0091] in, It is a learnable linear transformation matrix. This represents local spatiotemporal feature maps at different scales after feature alignment, where B and T represent the time dimension and variable dimension after feature alignment, respectively. These are the features after a linear transformation.

[0092] Furthermore, context-aware feature maps at different scales are computed based on scaled dot product attention:

[0093]

[0094] Where Sotfmax(·) is the activation function, d k is a scaling factor to prevent gradient vanishing; ● represents the dot product.

[0095] Furthermore, the local spatiotemporal feature maps at different scales after feature alignment are weighted and fused using cross-attention weights between different scales to obtain the final fused feature map F. csaff :

[0096]

[0097] F csaff =R⊙X fusion ;

[0098] Among them, X fusion For cross-scale fusion feature maps, Π represents element-wise product and ⊙ represents element-wise multiplication.

[0099] The CSAFF module introduces a cross-attention mechanism across multiple scales, adaptively allocating weights based on the correlation of features at different scales. This effectively integrates multi-scale features and suppresses redundant information from different scales in sensor data through attention weights, avoiding information redundancy that may occur with traditional stitching methods. It resolves inconsistencies by selecting a scale consistent with the global fault mode to dominate the fusion result. For example, a small-scale feature might capture an instantaneous pulse from a sensor (potentially noise), while a large-scale feature shows the channel is stable over a long period. The CSAFF module suppresses pulse interference through attention weights, making the fused features more discriminative and robust, better reflecting the inherent connections and changing patterns of time-series data at different scales, and further improving the model's ability to identify fault modes. Furthermore, for instance, a small-scale feature might be a temperature anomaly peak at t=100 (a short-term event), while a large-scale feature might be a slow pressure increase from t=50 to t=150 (a long-term trend). The CSAFF module uses a dot product of Q and K to address these inconsistencies. T The correlation between small-scale and large-scale data is calculated. If the correlation is high, the pressure features are weighted and fused into the final representation to establish a cross-time fault clue association of "early pressure rise → later temperature peak".

[0100] S5: Capture global long-distance temporal dependencies in the industrial process from the fused feature map using a Transformer encoder to obtain the feature vector of the global context.

[0101] In each Transformer encoding layer, feature maps are fused. Flattening into a tensor tensor The query matrix Q, key matrix K, and value matrix V are linearly projected; further, the self-attention weights for each head are calculated:

[0102]

[0103] in, W i Q W i K W i V Let Q be the learnable linear transformation matrix for the i-th head. i Let K be the query matrix for the i-th head. i The key matrix of the i-th head, V i Let be the value matrix of the i-th head.

[0104] Furthermore, the self-attention weights of the outputs of each head are concatenated to obtain the multi-head attention weights:

[0105] MultiHead(Xcsaff =ConCat(head1,...,head) h W O ;

[0106] Among them, W O It is the output weight, and ConCat(·) is the concatenation operation.

[0107] Furthermore, the multi-head attention weights are related to the input tensor X. csaff After performing residual connections and layer normalization, the feature vector F of the global context is obtained based on the feedforward network. csaff2 :

[0108] F csaff1 =LayerNorm(X csaff +MultiHead(X csaff ));

[0109] F csaff2 =LayerNorm(FFN(F csaff1 )+F csaff1 );

[0110] Where LayerNorm(·) is the layer normalization, FFN(F csaff1 The network is a feedforward network. Batch normalization (BN) is performed after the Transformer encoder, and two Dropout layers with a dropout rate of 0.5 are configured.

[0111] S6: Output the fault classification result based on the feature vector of the global context.

[0112] out = softmax(FC3(F csaff2 ));

[0113] FC3 is a fully connected layer.

[0114] The introduction of the Transformer encoder provides the model with powerful global temporal dependency modeling capabilities. Its self-attention mechanism dynamically allocates weights based on the correlation between different time steps, capturing long-distance dependencies in temporal data. This is particularly important for fault detection tasks, as fault occurrences are often closely related to a series of changes in historical data. This invention introduces a multi-head self-attention (MHSA) mechanism, enabling the model to capture global features over long time spans, thereby improving its ability to model complex fault modes. Furthermore, residual connections and layer normalization (LN) further improve gradient propagation, facilitating the training of deeper fault diagnosis models.

[0115] Finally, this invention uses the Adam optimizer to train the model, and the loss function is categorical cross-entropy:

[0116]

[0117] Among them, y m This indicates whether the sample belongs to the m-th category; 1 indicates belonging and 0 indicates not belonging. This represents the model's predicted probability that a sample belongs to that category, where M is the number of categories.

[0118] The entire model architecture is designed around the core idea of ​​hierarchical feature extraction and fusion, gradually building a comprehensive representation of time-series data from local to global and from low-dimensional to high-dimensional. Multi-scale deep separable convolution and channel attention mechanisms first extract multi-scale features and enhance channel-level features from the input data, providing the model with a rich foundation of local spatiotemporal features. Building upon this, the CSAFF module adaptively fuses multi-scale features through a cross-scale attention mechanism, further optimizing the discriminativeness and robustness of the features. The Transformer encoder models the fused features in the global temporal dimension, capturing long-distance dependencies in the time-series data, thus enabling the model to effectively identify and classify fault modes in the time-series data from different scales and dimensions.

[0119] Example 2

[0120] An intelligent fault diagnosis system for industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion, comprising:

[0121] Data preprocessing and sample construction module: used to input multivariate time series monitoring data of industrial processes, and to preprocess the monitoring data to obtain input feature maps.

[0122] A multi-scale spatiotemporal feature extraction module is used to extract multi-scale local spatiotemporal feature maps by performing parallel multi-scale two-dimensional deep separable convolution operations on the time dimension and sensor variable dimension of the input feature map. The multi-scale spatiotemporal feature extraction module deploys three two-dimensional deep separable convolution modules in parallel, each connected to the data preprocessing module, to implement the three parallel two-dimensional deep separable convolution operations.

[0123] Multi-scale channel attention module: This module enhances the feature responses of important sensor channels in the multi-scale local spatiotemporal feature map by using multi-scale channel attention to obtain the multi-scale enhanced feature map. The multi-scale channel attention module comprises three parallel channel attention modules, each connected to a corresponding two-dimensional depthwise separable convolutional module to implement the three channel attention operations.

[0124] Cross-self-attention fusion module: Connected to the three channel attention modules, it fuses multi-scale enhanced feature maps through a cross-self-attention mechanism to suppress repetitive information at different scales of sensor data and selects the scale consistent with the global fault mode to dominate the fusion result, thereby obtaining the fused feature map.

[0125] Global temporal dependency modeling module: Connected to the cross-self-attention fusion module, it is used to capture global long-distance temporal dependencies in the industrial process in the fused feature map through the Transformer encoder and obtain the feature vector of the global context.

[0126] Fault classification module: Connected to the Transformer encoder module, it is used to output fault classification results based on the feature vector of the global context.

[0127] Analysis of experimental results:

[0128] To verify the effectiveness of this invention, the internationally recognized Tennessee-Eastman (TE) process was used as the simulation platform to validate the effectiveness of the proposed method. The TE process is a complex nonlinear dynamic simulation model based on a real chemical plant and has become a standard benchmark dataset for evaluating the performance of industrial process monitoring and fault diagnosis algorithms.

[0129] This dataset mainly contains the following characteristics:

[0130] Process variables: The simulation system contains 52 process observation variables, including 41 measured variables and 11 manipulated variables. These high-dimensional, strongly coupled variables comprehensively reflect the operating state of the process.

[0131] Fault Types: The TE process predefines 21 different fault types, including normal operating conditions (Fault0) and 20 abnormal operating conditions (Fault1 to Fault20). These fault types are diverse, covering a variety of common industrial fault modes such as step changes, random fluctuations, slow drift, and equipment stickiness, which can comprehensively test the robustness and generalization ability of the diagnostic model.

[0132] Data Structure: Each fault type contains independent training and test sets. In this invention, the training set for each fault contains 480 samples, and the test set contains 960 samples, providing sufficient data for model training and validation. The optimal parameter configuration of the MS-DSC-CSAFF-Transformer model proposed in this invention is obtained based on a series of comparative experiments and empirical optimization.

[0133] Model architecture parameters: Multi-scale Deeply Separable Convolutional (MCDSC) module: Employs three parallel 2D deeply separable convolutions with kernel sizes of 3x3, 5x5, and 7x7, respectively. The number of output channels (i.e., the number of filters) per branch is set to 128. Cross-scale Attention Feature Fusion (CSAFF) module: Uses a single layer. Transformer encoder module: Employs a single-layer standard Transformer encoder, with the number of heads (h) for the multi-head self-attention mechanism set to 6. It uses a fully connected layer with an output dimension of 128 and configures two dropout layers with a dropout rate of 0.5 to prevent overfitting.

[0134] Training parameters: The sliding window length for the time series was set to 5, the deep learning framework used was Keras 2.13 and Python 3.9, all models were trained using the Adam optimizer, the initial learning rate was set to 0.001, the batch size was 64, and the total number of training epochs was 100.

[0135] TE diagnostic performance comparison:

[0136] This invention utilizes the TE chemical engineering dataset and employs various deep learning models for fault diagnosis, comparing and analyzing their performance through experiments. The experimental environment and parameter settings were kept consistent to ensure fair comparison. The experiments primarily evaluated the diagnostic performance of models such as MLP, ResCNN, Transform-encoder, and DSCNN-ATT-encoder, focusing on the precision (P) and recall (R) of each model across different fault categories.

[0137] As shown in Table 1, the results indicate that MLP performed the worst, ResCNN showed some improvement, and the Transformer Encoder model further improved accuracy and recall. The proposed MS-DSC-CSAFF-Transformer model achieved the best performance. In terms of mean precision and mean recall, the MLP model had P and R values ​​of 0.80 and 0.79, respectively, indicating its weak ability to identify complex faults, especially performing poorly on fault categories 3, 9, and 15, demonstrating its poor generalization ability to difficult-to-classify faults. ResCNN, on the other hand, had average P and R values ​​of 0.88, a significant improvement over MLP, indicating that CNNs with residual structures can more effectively extract fault features. However, some fault categories still showed inaccurate identification, such as the low recall rates for faults 9 and 15. The Transformer Encoder's average P and R values ​​increased to 0.91, demonstrating the advantage of its global attention mechanism in extracting long-term dependencies, further enhancing its ability to identify most fault categories, especially showing better performance than ResCNN on faults 3, 9, and 15. However, the optimal model, MS-DSC-CSAFF-Transformer, achieved average P and R values ​​of 0.97 and 0.96, respectively, significantly outperforming other models. This indicates that the model combining multi-scale deep separable convolution and Transformer can fully extract local and global features, and the channel attention mechanism enhances the selection ability of key features. This results in significantly better performance on difficult-to-classify faults (such as faults 3, 9, and 15) compared to other models, with recall rates exceeding 0.79 and significantly improved accuracy. In summary, MS-DSC-CSAFF-Transformer demonstrates the best fault diagnosis performance on the TE fault dataset, showcasing its superiority in complex fault scenarios and making it suitable for industrial fault diagnosis tasks requiring high accuracy and high recall.

[0138] This invention also includes a visualization analysis of t-distributed stochastic neighborhood embedding (T-SNE), showing that: Figure 2As shown, the data distribution before training was relatively chaotic, with significant overlap between categories, indicating low discriminative power of the original features in the low-dimensional space. After training, however, different categories exhibited a clear clustering structure in the feature space, the distribution of data points became more compact, and the boundaries between categories became clearer. This demonstrates that the model successfully learned discriminative deep features and could effectively extract key information about fault modes. Overall, the model effectively extracted key features, improved data separability, and significantly enhanced fault diagnosis capabilities on the TE dataset. Compared to traditional methods, the architecture proposed in this invention, which combines multi-scale deep separable convolution, channel attention mechanism, cross-scale feature fusion (CSAFF), and Transformer encoder, fully utilizes local spatiotemporal information, global feature relationships, and cross-scale information fusion, making the model more robust in handling complex fault modes.

[0139] Experimental results demonstrate that the proposed method exhibits superior performance in fault detection tasks. Specifically, multi-scale depthwise separable convolution effectively extracts local spatiotemporal features at different scales while reducing computational complexity; the channel attention mechanism further enhances the model's selectivity for key features and improves feature representation capabilities; the cross-scale feature fusion (CSAFF) mechanism effectively integrates information from different scales by introducing an adaptive attention strategy, reducing redundancy and improving the model's generalization ability. Furthermore, the Transformer encoder utilizes a self-attention mechanism to model long-distance temporal dependencies, making the model more robust when handling complex fault modes. In summary, the proposed method fully leverages the advantages of convolutional neural networks in local feature extraction, while combining attention mechanisms and Transformer to enhance global information modeling capabilities, thereby improving the accuracy and stability of fault diagnosis.

[0140] Table 1 Performance Comparison Results

[0141]

[0142]

[0143] 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 intelligent fault diagnosis in industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion, characterized in that, The steps are as follows: S1: Input multivariate time series monitoring data of industrial processes, and preprocess the monitoring data to obtain an input feature map; S2: Parallel multi-scale two-dimensional depthwise separable convolution operations are used to extract multi-scale features from the time dimension and sensor variable dimension of the input feature map to obtain a multi-scale local spatiotemporal feature map; S3: Enhance the feature response of important sensor channels in the multi-scale local spatiotemporal feature map by multi-scale channel attention to obtain the multi-scale enhanced feature map; S4: By fusing multi-scale enhanced feature maps through a cross-scale self-attention mechanism, the repetitive information at different scales of sensor data is suppressed, and the scale consistent with the global fault mode is selected to dominate the fusion result, thereby obtaining the fused feature map. S5: Capture global long-range temporal dependencies in the industrial process from the fused feature map using a Transformer encoder to obtain the feature vector of the global context; S6: Output the fault classification result based on the feature vector of the global context.

2. The intelligent fault diagnosis method for industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion as described in claim 1, characterized in that, The preprocessing method is as follows: standardizing the monitoring data; and using a sliding window to divide the standardized monitoring data into multiple two-dimensional tensors. H represents the sequence length, and W is the number of sensor variables; the reshape operation adjusts each sub-2D tensor to a format suitable for 2D convolution input, making each 2D tensor... The shape changes from H×W to H×W×C, resulting in the input feature map. C represents the number of input channels.

3. The intelligent fault diagnosis method for industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion as described in claim 2, characterized in that, In step S2, three parallel two-dimensional depthwise separable convolution operations are used, each applying a different kernel size to sequentially process the input feature map. Depthwise convolution and pointwise convolution are performed. Depthwise convolution extracts spatial information, while pointwise convolution performs channel mapping, ultimately yielding local spatiotemporal feature maps at three scales.

4. The intelligent fault diagnosis method for industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion as described in claim 3, characterized in that, The method for extracting spatial information through depthwise convolution is as follows: each two-dimensional depthwise separable convolution operation is applied to the input feature map. Each channel is independently spatially convolved to extract local features, resulting in a depthwise convolution output local channel feature map X. depthwise,h,w,c The spatial dimensions of the spatial convolution include a time dimension related to the sequence length H and a variable dimension related to the number of sensor variables W. The method of channel mapping through pointwise convolution is as follows: The local channel feature map X output by convolution at each depth is... depthwise,h,w,c By applying 1×1 convolution, channel information is fused and the feature dimension is expanded to obtain the channel feature map X output by pointwise convolution. h,w,c′ By stitching the images together according to the channels, local spatiotemporal feature maps at each scale can be obtained. Where h is the time dimension direction index, w is the variable dimension direction index, and c and c' are channel indices; H', W', and C' are local spatiotemporal feature maps, respectively. The time dimension, variable dimension, and number of channels are defined, where n is the index of the two-dimensional depthwise separable convolution operation.

5. The intelligent fault diagnosis method for industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion as described in claim 4, characterized in that, In step S3, three channel attention operations are deployed in parallel. Each channel attention operation outputs a local spatiotemporal feature map for each two-dimensional depthwise separable convolution operation. Global average pooling and global max pooling are applied to each channel respectively, and the channel description vectors of global average pooling are obtained respectively. and global max pooling channel description vector Global average pooling channel description vector and global max pooling channel description vector The inputs are fed into an MLP network and then concatenated. Channel attention weights A are generated using a sigmoid activation function. c′ ; Adjust the channel attention weight A c′ With each channel feature map Perform multiplication to obtain the enhanced feature map for each channel. Enhanced feature map for each channel Enhanced feature maps at each scale are obtained by concatenating channels.

6. The intelligent fault diagnosis method for industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion as described in claim 5, characterized in that, In step S4, feature maps enhanced at all scales are first aligned using feature alignment. By unifying the dimensions, local spatiotemporal feature maps at three scales are obtained after feature alignment. B and T represent the time dimension and variable dimension after feature alignment, respectively, based on the local spatiotemporal feature maps at three scales. Calculate the query matrix Q csaff Key matrix K csaff Sum matrix V csaff ; Calculate context-aware feature maps R at different scales based on scaled dot product attention; Local spatiotemporal feature maps at three scales. Cross-scale fusion is performed between the two scales to obtain the cross-scale fused feature map X. fusion And utilize context-aware feature map R to fuse feature map X across scales. fusion Weighting is performed to obtain the final fused feature map F. csaff .

7. The intelligent fault diagnosis method for industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion as described in claim 6, characterized in that, In the calculation query matrix Q csaff Key matrix K csaff Sum matrix V csaff hour, The query is performed using one of the two scales, with the other two scales representing the key and value, respectively. The method for calculating the context-aware feature map R at different scales is as follows: Where Sotfmax(·) is the activation function, d k The scaling factor is ·, which represents the dot product. The process of obtaining cross-scale fusion feature maps X is described above. fusion The method is as follows: The process of obtaining the final fused feature map F csaff The method is as follows: F csaff =R⊙X fusion 。 8. The intelligent fault diagnosis method for industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion as described in claim 6 or 7, characterized in that, In step S5, the final fused feature map is... Flattening into a tensor tensor The query matrix Q, key matrix K, and value matrix V are linearly projected; the self-attention weights of h heads are calculated through a single Transformer encoding operation; and the self-attention weights of the outputs of each head are concatenated to obtain the multi-head attention weights. MultiHead(X) attention weight csaff ) and the input tensor X csaff After performing residual connections and layer normalization, the feature tensor F is obtained. csaff1 The feature vector of the global context is obtained based on the feedforward network: F csaff2 =LayerNorm(FFN(F csaff1 )+F csaff1 ) Where LayerNorm(·) is layer normalization and FFN(·) is feedforward network; After layer normalization following the Transformer encoding operation, two Dropout layers with a dropout rate of 0.5 were configured.

9. An intelligent fault diagnosis system for industrial processes, employing the intelligent fault diagnosis method for industrial processes based on multi-scale depthwise separable convolution and cross-attention fusion as described in any one of claims 1 to 8, characterized in that, It includes, in sequence, a data preprocessing and sample construction module, a multi-scale spatiotemporal feature extraction module, a multi-scale channel attention module, a cross-scale self-attention fusion module, a global time-dependent modeling module, and a fault classification module, wherein: Data preprocessing and sample construction module: used to input multivariate time series monitoring data of industrial processes, and to preprocess the monitoring data to obtain input feature maps; Multi-scale spatiotemporal feature extraction module: Used to extract multi-scale features from the time dimension and sensor variable dimension of the input feature map by using parallel multi-scale two-dimensional depthwise separable convolution operations to obtain multi-scale local spatiotemporal feature maps; Multi-scale channel attention module: used to enhance the feature responses of important sensor channels in multi-scale local spatiotemporal feature maps through multi-scale channel attention to obtain multi-scale enhanced feature maps; Cross-scale self-attention fusion module: It is used to fuse multi-scale enhanced feature maps through a cross-scale self-attention mechanism to suppress repetitive information at different scales of sensor data and select the scale consistent with the global fault mode to dominate the fusion result and obtain fused feature maps. Global Temporal Dependency Modeling Module: Used to capture global long-range temporal dependencies in industrial processes from the fused feature map using a Transformer encoder, and obtain the feature vector of the global context. Fault classification module: Used to output fault classification results based on the feature vector of the global context.

10. The intelligent fault diagnosis system for industrial processes according to claim 9, characterized in that, The multi-scale spatiotemporal feature extraction module includes three parallel two-dimensional depthwise separable convolution modules to implement the three parallel two-dimensional depthwise separable convolution operations. The multi-scale channel attention module includes three channel attention modules deployed in parallel to implement the three channel attention operations.

Images