Industrial fault feature adaptive extraction and multi-mode detection system and method

Abstract

The invention provides an industrial fault feature adaptive extraction and multi-mode detection system and method, and belongs to the technical field of industrial fault detection. Comprising the steps of collecting multi-source data of industrial equipment, performing timestamp alignment and processing on the multi-source data to obtain a standardized feature sequence, inputting the standardized feature sequence into a dynamic convolutional neural network, extracting signal local features through a deformable convolution kernel, calculating feature weights in combination with a self-attention mechanism, and screening feature channels to obtain feature vectors of all modes; a graph structure with modals as nodes and correlation as edges is constructed, cross-modal features are aggregated through a graph attention network, and global state descriptors fused with spatio-temporal information are generated; and performing time sequence modeling on the global state descriptor through a bidirectional LSTM network, outputting fault type probability distribution, and completing industrial equipment fault detection. According to the method, the problems of insufficient single-modal information, fixed feature extraction, low efficiency of multi-modal correlation modeling and lagging model updating in traditional industrial fault detection are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial fault detection, and in particular to a system and method for adaptive extraction and multimodal detection of industrial fault features. Background Art

[0002] Fault detection for industrial equipment is a key component in ensuring safe production and reducing operation and maintenance costs. With the development of the Industrial Internet of Things and intelligent manufacturing technologies, equipment operation data has become multimodal, high-dimensional, and unstructured. Traditional fault detection methods face the following technical bottlenecks:

[0003] Limitations of single-modal detection: Existing technologies rely on a single sensor to collect data and are unable to integrate information from multiple sources. For example, motor bearing wear may manifest as increased vibration shock peaks, abnormal current harmonics, and localized temperature increases. A single modality cannot fully characterize complex fault characteristics, resulting in a high rate of missed detection.

[0004] Disadvantages of fixed feature extraction models: Existing methods often use artificially designed features or neural networks with fixed convolution kernels, which are difficult to adapt to the changes in the feature distribution of different fault modes. For example, a gearbox crack fault manifests as a low-frequency modulation feature in the early stage and evolves into a high-frequency impact feature in the later stage. Traditional fixed models cannot dynamically adjust the feature extraction strategy, resulting in significant differences in detection performance at different stages.

[0005] Inefficient multimodal data fusion: Traditional multimodal fusion methods ignore the dynamic correlations between modalities. For example, the frequency domain characteristics of the vibration signal and the harmonic components of the current signal are strongly coupled when the motor load suddenly changes, and existing methods cannot explicitly model this correlation.

[0006] Lag in online model updates: Failure modes of industrial equipment may evolve with changes in operating conditions and the environment. Existing models often require offline retraining and are unable to respond to new fault types in real time. For example, a chemical equipment developed a new bearing fault after introducing a new material. Traditional methods required shutting down the equipment to collect samples and retrain the model, resulting in an average downtime of more than 48 hours.

[0007] In summary, existing industrial fault detection technologies have significant shortcomings in terms of multimodal information fusion efficiency, feature extraction adaptability, online learning capabilities, and edge computing adaptability. Implementing dynamic correlation modeling of multimodal data, adaptive adjustment of feature extraction strategies, and lightweight online model updates are pressing technical challenges in this field. Summary of the Invention

[0008] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide a system and method for adaptive extraction of industrial fault features and multimodal detection. Through multimodal data fusion, adaptive feature extraction and time series modeling, the problems of insufficient single modal information, fixed feature extraction, inefficient multimodal association modeling and delayed model update in traditional industrial fault detection are solved, and adaptive real-time detection of industrial equipment faults is realized.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A method for adaptive extraction and multimodal detection of industrial fault features, comprising the following steps:

[0011] S1. Synchronously collect multi-source data from industrial equipment based on the multimodal data acquisition module. Then, timestamp alignment is performed on the collected multi-source data. After wavelet threshold denoising and normalization, a standardized feature sequence with unified dimension is obtained.

[0012] S2. Input the standardized feature sequence into a dynamic convolutional neural network, extract local signal features through a deformable convolution kernel, calculate feature weights in combination with a self-attention mechanism, and filter the feature channels with the top 50% weights to obtain feature vectors for each modality;

[0013] S3. Based on the feature vectors of each modality, a graph structure is constructed with modalities as nodes and correlations as edges. The cross-modal features are aggregated through a graph attention network to generate a global state descriptor that integrates spatiotemporal information.

[0014] S4. Perform time series modeling on the global state descriptor through a bidirectional LSTM network, output the probability distribution of fault types, and complete the detection of industrial equipment faults; when the confidence level of the detection result is lower than the threshold, trigger the incremental learning algorithm to update the model parameters.

[0015] Preferably, in step S1, the wavelet threshold denoising adopts db4 wavelet, and the data range after normalization processing is [-1, 1].

[0016] Preferably, in step S2, the first two layers of the dynamic convolutional neural network are used to predict the deformable convolution kernel offset, and the third layer outputs a feature map that is globally average pooled to generate a 1×256-dimensional feature vector.

[0017] Preferably, in step S2, the deformable convolution kernel predicts a two-dimensional offset through a convolution layer, thereby extracting local features, specifically:

[0018] y(x)=∑ p∈Ω ω(p)·x(x+p+Δp);

[0019] Among them, y(x) is the value of the output feature map at position x, Ω is the sampling point set of the standard convolution kernel, Δp is the two-dimensional offset, ω(p) is the weight coefficient of the standard convolution kernel at the sampling point p, and x(x+p+Δp) is the value of the input signal at the deformed sampling point (x+p+Δp).

[0020] Preferably, in step S2, the formula for calculating feature weights by the self-attention mechanism is:

[0021]

[0022] Among them, f i is the i-th eigenvector, corresponding to the one-dimensional feature channel in the standardized feature sequence; W is the weight matrix used to calculate the importance score of the eigenvector; α i is the normalized weight of the i-th eigenvector; n is the total number of eigenvectors, corresponding to the number of input feature channels; f k is the kth feature vector in the input feature sequence.

[0023] Preferably, in step S3, the calculation formula of the graph attention network is:

[0024]

[0025] in, is the feature embedding of node i in the lth layer, corresponding to the feature vector of modality i in the lth layer; N(i) is the set of neighbor nodes of node i, including all modality nodes connected to node i; α ij is the attention weight between node i and node j; W (l) is the weight matrix of the lth layer; σ is the activation function, and j is the neighbor node index, which refers to the adjacent nodes in the graph structure.

[0026] Preferably, in step S3, the graph attention network comprises 2 layers, each layer adopts an 8-head attention mechanism, and the output node embedding is spliced ​​into a 256-dimensional global feature vector.

[0027] Preferably, in step S4, the bidirectional LSTM network input is a feature sequence of 10 consecutive time steps, the hidden layer dimension is 128, and the softmax function is used to output the probability distribution of 5 types of faults.

[0028] Preferably, in step S4, the triggering condition for the incremental learning is that the confidence level of the detection result is lower than 80% or a new fault type is detected.

[0029] The present invention also provides a system for adaptive extraction and multimodal detection of industrial fault features, which is used to perform the above-mentioned method for adaptive extraction and multimodal detection of industrial fault features, including:

[0030] Multimodal data acquisition module: This module integrates vibration sensors, temperature sensors, acoustic sensors, and current sensors to synchronously collect multi-source data from industrial equipment. It then performs timestamp alignment, wavelet threshold denoising, and normalization on the multi-source data, outputting a standardized feature sequence with a unified dimension.

[0031] Adaptive feature extraction module: This module is used to input the standardized feature sequence into the dynamic convolutional neural network, extract the local features of the signal through the deformable convolution kernel, and calculate the feature weights in combination with the self-attention mechanism, filter the feature channels with the top 50% weight, and output the feature vectors of each modality;

[0032] Multimodal fusion module: This module is used to construct a graph structure with modalities as nodes and correlations as edges based on the feature vectors of each modality. It aggregates cross-modal features through a graph attention network to generate a global state descriptor that integrates spatiotemporal information.

[0033] Fault detection and classification module: This module is used to perform time series modeling on global state descriptors through a bidirectional LSTM network, output the probability distribution and confidence level of fault types, and complete real-time detection of industrial equipment faults.

[0034] Online feedback optimization module: used to trigger the incremental learning algorithm to dynamically update model parameters when the confidence level of the detection result is lower than the threshold or a new fault type is detected.

[0035] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0036] (1) This invention achieves accurate capture of industrial equipment fault characteristics and cross-modal correlation analysis through the collaborative processing and adaptive modeling of multimodal data. The system integrates multiple types of sensors, such as vibration, temperature, acoustic wave, and current, and constructs a standardized feature sequence through timestamp alignment and noise filtering. This breaks through the limitations of isolated single modal information and can comprehensively characterize the multi-dimensional physical characteristics of the fault.

[0037] (2) The feature extraction module of this invention, based on a dynamic convolutional neural network and a self-attention mechanism, can adaptively adjust the feature extraction strategy according to different fault modes. The deformable convolution kernel dynamically adapts to the local feature morphology of the signal by predicting the two-dimensional offset, and combines the self-attention mechanism to screen key feature channels, achieving an upgrade from "fixed pattern feature extraction" to "dynamic sensitive feature focusing", significantly improving the model's ability to express complex fault features.

[0038] (3) The present invention constructs a modal association graph structure through a graph neural network and combines it with bidirectional LSTM time series modeling to realize the spatiotemporal correlation analysis of multimodal features and dynamic prediction of fault modes; the graph attention network uses modes as nodes and correlations as edges to explicitly model the dynamic coupling relationship between different modalities and generate a global state descriptor that integrates spatiotemporal information; the bidirectional LSTM network further models the temporal dependency of feature sequences, outputs the probability distribution of fault types and triggers incremental learning to optimize model parameters, forming a closed loop of "feature extraction-fusion modeling-detection optimization", and ultimately realizing adaptive real-time detection of industrial equipment failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in 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 paying any creative work.

[0040] Figure 1 This is a flow chart of a method for adaptive extraction and multimodal detection of industrial fault features according to the present invention;

[0041] Figure 2 This is a system module block diagram of the present invention for adaptive extraction of industrial fault features and multimodal detection. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Example 1

[0045] like Figure 1 As shown, the present invention provides a method for adaptive extraction and multimodal detection of industrial fault features, comprising the following steps:

[0046] S1. Synchronously collect multi-source data from industrial equipment based on the multimodal data acquisition module. Then, timestamp alignment is performed on the collected multi-source data. After wavelet threshold denoising and normalization, a standardized feature sequence with unified dimension is obtained.

[0047] S2. Input the standardized feature sequence into a dynamic convolutional neural network, extract local signal features through a deformable convolution kernel, calculate feature weights in combination with a self-attention mechanism, and filter the feature channels with the top 50% weights to obtain feature vectors for each modality;

[0048] S3. Based on the feature vectors of each modality, a graph structure is constructed with modalities as nodes and correlations as edges. The cross-modal features are aggregated through a graph attention network to generate a global state descriptor that integrates spatiotemporal information.

[0049] S4. Perform time series modeling on the global state descriptor through a bidirectional LSTM network, output the probability distribution of fault types, and complete the detection of industrial equipment faults; when the confidence level of the detection result is lower than the threshold, trigger the incremental learning algorithm to update the model parameters.

[0050] According to the method provided above, in step S1, the wavelet threshold denoising adopts db4 wavelet, and the data range after normalization processing is [-1, 1].

[0051] In step S2, the first two layers of the dynamic convolutional neural network are used to predict the offset of the deformable convolution kernel, and the third layer outputs a feature map that is globally averaged and pooled to generate a 1×256-dimensional feature vector. The deformable convolution kernel predicts the two-dimensional offset through the convolution layer, and then extracts local features, specifically:

[0052] y(x)=∑ p∈Ω ω(p)·x(x+p+Δp);

[0053] Among them, y(x) is the value of the output feature map at position x, Ω is the sampling point set of the standard convolution kernel, Δp is the two-dimensional offset, ω(p) is the weight coefficient of the standard convolution kernel at the sampling point p, and x(x+p+Δp) is the value of the input signal at the deformed sampling point (x+p+Δp).

[0054] The formula for calculating feature weights by the self-attention mechanism is:

[0055]

[0056] Among them, f i is the i-th eigenvector, corresponding to the one-dimensional feature channel in the standardized feature sequence; W is the weight matrix used to calculate the importance score of the eigenvector; α i is the normalized weight of the i-th eigenvector; n is the total number of eigenvectors, corresponding to the number of input feature channels; f k is the kth feature vector in the input feature sequence.

[0057] In step S3, the calculation formula of the graph attention network is:

[0058]

[0059] in, is the feature embedding of node i in the lth layer, corresponding to the feature vector of modality i in the lth layer; N(i) is the set of neighbor nodes of node i, including all modality nodes connected to node i; α ij is the attention weight between node i and node j; W (l) is the weight matrix of layer l; σ is the activation function, and j is the neighbor node index, referring to the adjacent nodes in the graph structure. The graph attention network consists of two layers, each using an 8-head attention mechanism, and the output node embeddings are concatenated into a 256-dimensional global feature vector.

[0060] In step S4, the bidirectional LSTM network input is a feature sequence of 10 consecutive time steps, the hidden layer dimension is 128, and a softmax function is used to output the probability distribution of five types of faults. The trigger condition for incremental learning is that the confidence level of the detection result is less than 80% or a new fault type is detected.

[0061] Reference Figure 2 , which provides a system for adaptive extraction and multimodal detection of industrial fault features based on the above method, including:

[0062] Multimodal data acquisition module: This module integrates vibration sensors, temperature sensors, acoustic sensors, and current sensors to synchronously collect multi-source data from industrial equipment. It then performs timestamp alignment, wavelet threshold denoising, and normalization on the multi-source data, outputting a standardized feature sequence with a unified dimension.

[0063] Adaptive feature extraction module: This module is used to input the standardized feature sequence into the dynamic convolutional neural network, extract the local features of the signal through the deformable convolution kernel, and calculate the feature weights in combination with the self-attention mechanism, filter the feature channels with the top 50% weight, and output the feature vectors of each modality;

[0064] Multimodal fusion module: This module is used to construct a graph structure with modalities as nodes and correlations as edges based on the feature vectors of each modality. It aggregates cross-modal features through a graph attention network to generate a global state descriptor that integrates spatiotemporal information.

[0065] Fault detection and classification module: This module is used to perform time series modeling on global state descriptors through a bidirectional LSTM network, output the probability distribution and confidence level of fault types, and complete real-time detection of industrial equipment faults.

[0066] Online feedback optimization module: used to trigger the incremental learning algorithm to dynamically update model parameters when the confidence level of the detection result is lower than the threshold or a new fault type is detected.

[0067] Among them, the process of dynamic updating of model parameters by the incremental learning algorithm is as follows: when the confidence of the detection result output by the fault detection and classification module is lower than the configurable threshold of 80% for three consecutive times or a new fault type that is not in the existing classification label set is detected, the incremental learning process is triggered. First, the parameters of the first two layers of the dynamic convolutional neural network, namely the deformable convolution kernel offset prediction layer, are frozen, and only the parameter update permission of the third and subsequent layers is opened. The current suspicious samples or new fault samples are extracted to construct an incremental learning sample set containing standardized feature sequences and manually labeled labels; then the incremental sample set is subjected to the same timestamp alignment, wavelet denoising, and normalization preprocessing process as S1 to generate an incremental standardized feature sequence, and then the local features of the incremental samples are extracted through the first two layers of the dynamic convolutional neural network with frozen parameters. After screening the channels with the top 50% of the weights through the self-attention mechanism, an incremental feature vector set is generated; at the same time, the Fidelity of the existing model parameters on the old task is calculated based on the elastic weight consolidation (EWC) algorithm. Sher information matrix is ​​used to quantify the importance of each parameter to the old task, and EWC regularization constraints containing the regularization hyperparameter λ are imposed on the hidden layer parameters and classification layer weights of the bidirectional LSTM network to prevent the forgetting of old knowledge; then the online gradient descent (OGD) algorithm is used to perform local parameter updates only on the third-layer feature map output layer of the dynamic convolutional neural network, the second-layer weight matrix of the graph attention network, the hidden layer to output layer weights of the bidirectional LSTM network, and the classification layer bias. The incremental feature vector set is trained for 5-10 rounds using mini-batch gradient descent. The learning rate is set to 1e-4 and momentum optimization is applied to accelerate convergence. The parameter update scale does not exceed 20% of the total parameters. Finally, a validation set containing 10% old samples and new samples is used to evaluate the model performance. If the accuracy of the old fault type drops by <5% and the new fault detection accuracy is >90%, the updated parameters are accepted and solidified in the online feedback optimization module, and the data samples in the adaptive feature extraction module are updated synchronously.

[0068] The following is implemented on a certain model of industrial motor fault detection platform. The hardware uses Intel Xeon E5-2678 processor and NVIDIA T4 GPU, and the software is developed based on the PyTorch 1.10 framework. The implementation process is as follows:

[0069] First, vibration sensors, infrared temperature sensors, acoustic wave sensors, and current sensors were deployed around industrial motors. The vibration sensor had a sampling rate of 10 kHz, the infrared temperature sensor had an acquisition accuracy of ±0.5°C, the acoustic wave sensor had a frequency response of 20 Hz-20 kHz, and the current sensor had a range of 0-50 A. Multi-source data synchronization was achieved using the IEEE 1588 protocol, with measured timestamp alignment errors of ≤0.5 ms. The original signal was subjected to a 5-layer DB4 wavelet basis decomposition and hard threshold denoising, improving the vibration signal-to-noise ratio from 23.4 dB to 42.0 dB. The temperature signal was normalized to the [-1, 1] range using a piecewise linear transformation, and the current signal was Z-score normalized to eliminate dimensionality differences. Ultimately, a 4×512-dimensional normalized feature sequence was generated.

[0070] Next, a three-layer dynamic convolutional network is constructed: the first two layers use 3×3 deformable convolution kernels with 64 / 128 output channels, and the third layer generates a 256-dimensional feature vector through global average pooling. The deformable convolution kernel dynamically adjusts the sampling position through the offset prediction module, using the formula:

[0071] y(x)=∑ p∈Ω ω(p)·x(x+p+Δp);

[0072] Experimental results show that this structure captures 91.7% of the local features of early-stage gearbox cracks, a 32.7% improvement compared to fixed convolution kernels. After filtering the top 50% of high-weight channels using a self-attention mechanism, 89.3% of valid information is retained in bearing fault detection tasks, reducing computational time from 58ms to 34ms.

[0073] Finally, a two-layer graph attention network (GAT) was used for feature fusion: each layer was equipped with an eight-head attention mechanism, the node dimension was expanded from 128 to 256, and the activation function was LeakyReLU (α = 0.2). In the detection of motor winding short circuits, cross-modal fusion increased the accuracy from 83.2% to 96.8%, outputting a 256-dimensional global state descriptor.

[0074] Finally, the global features from 10 consecutive time steps were fed into a bidirectional LSTM network, which then output a five-category fault probability distribution using a softmax function. Testing showed a classification accuracy of 98.2±1.3%, a significant improvement over traditional SVMs. When detection confidence falls below 80% or a new fault is discovered, Elastic Weight Conservation (EWC) incremental learning is triggered. After adding 500 new samples, the model's accuracy increased from 68.4% to 95.4% in 37 minutes, while occupying only 2.7MB of memory.

[0075] Therefore, the above-mentioned system and method for adaptive extraction and multimodal detection of industrial fault features is adopted. Through multimodal data fusion, adaptive feature extraction and time series modeling, the problems of insufficient single modal information, fixed feature extraction, inefficient multimodal correlation modeling and delayed model update in traditional industrial fault detection are solved, and adaptive real-time detection of industrial equipment faults is realized.

[0076] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for adaptive extraction and multimodal detection of industrial fault features, characterized in that: The following steps are involved: S1. Synchronously collect multi-source data from industrial equipment based on the multimodal data acquisition module. Then, timestamp alignment is performed on the collected multi-source data. After wavelet threshold denoising and normalization, a standardized feature sequence with unified dimension is obtained. S2. Input the standardized feature sequence into a dynamic convolutional neural network, extract local signal features through a deformable convolution kernel, calculate feature weights in combination with a self-attention mechanism, and filter the feature channels with the top 50% weights to obtain feature vectors for each modality; S3. Based on the feature vectors of each modality, a graph structure is constructed with modalities as nodes and correlations as edges. The cross-modal features are aggregated through a graph attention network to generate a global state descriptor that integrates spatiotemporal information. S4. Perform time series modeling on the global state descriptor through a bidirectional LSTM network, output the probability distribution of fault types, and complete the detection of industrial equipment faults; when the confidence level of the detection result is lower than the threshold, trigger the incremental learning algorithm to update the model parameters.

2. The method for adaptive extraction and multimodal detection of industrial fault features according to claim 1, characterized in that: In step S1, the wavelet threshold denoising adopts db4 wavelet, and the data range after normalization processing is [-1, 1].

3. The method for adaptive extraction and multimodal detection of industrial fault features according to claim 1, characterized in that: In step S2, the first two layers of the dynamic convolutional neural network are used to predict the offset of the deformable convolution kernel, and the third layer outputs a feature map that is globally average pooled to generate a 1×256-dimensional feature vector.

4. The method for adaptive extraction and multimodal detection of industrial fault features according to claim 3, characterized in that: In step S2, the deformable convolution kernel predicts the two-dimensional offset through the convolution layer, and then extracts local features, specifically: y(x)=∑ p∈Ω ω(p)·x(x+p+Δp); Among them, y(x) is the value of the output feature map at position x, Ω is the sampling point set of the standard convolution kernel, Δp is the two-dimensional offset, ω(p) is the weight coefficient of the standard convolution kernel at the sampling point p, and x(x+p+Δp) is the value of the input signal at the deformed sampling point (x+p+Δp).

5. The method for adaptive extraction and multimodal detection of industrial fault features according to claim 1, characterized in that: In step S2, the formula for calculating feature weights by the self-attention mechanism is: Among them, f i is the i-th eigenvector, corresponding to the one-dimensional feature channel in the standardized feature sequence; W is the weight matrix used to calculate the importance score of the eigenvector; α i is the normalized weight of the i-th eigenvector; n is the total number of eigenvectors, corresponding to the number of input feature channels; f k is the kth feature vector in the input feature sequence.

6. The method for adaptive extraction and multimodal detection of industrial fault features according to claim 1, characterized in that: In step S3, the calculation formula of the graph attention network is: in, is the feature embedding of node i in the lth layer, corresponding to the feature vector of modality i in the lth layer; N(i) is the set of neighbor nodes of node i, including all modality nodes connected to node i; α ij is the attention weight between node i and node j; W (l) is the weight matrix of the lth layer; σ is the activation function, and j is the neighbor node index, which refers to the adjacent nodes in the graph structure.

7. The method for adaptive extraction and multimodal detection of industrial fault features according to claim 6, characterized in that: In step S3, the graph attention network consists of 2 layers, each layer adopts an 8-head attention mechanism, and the output node embedding is spliced ​​into a 256-dimensional global feature vector.

8. The method for adaptive extraction and multimodal detection of industrial fault features according to claim 1, characterized in that: In step S4, the bidirectional LSTM network input is a feature sequence of 10 consecutive time steps, the hidden layer dimension is 128, and the softmax function is used to output the probability distribution of 5 types of faults.

9. The method for adaptive extraction and multimodal detection of industrial fault features according to claim 1, characterized in that: In step S4, the triggering condition for the incremental learning is that the confidence level of the detection result is lower than 80% or a new fault type is detected.

10. A system for adaptive extraction and multimodal detection of industrial fault features, characterized in that: A method for performing adaptive extraction and multimodal detection of industrial fault features according to any one of claims 1 to 9, comprising: Multimodal data acquisition module: This module integrates vibration sensors, temperature sensors, acoustic sensors, and current sensors to synchronously collect multi-source data from industrial equipment. It then performs timestamp alignment, wavelet threshold denoising, and normalization on the multi-source data, outputting a standardized feature sequence with a unified dimension. Adaptive feature extraction module: This module is used to input the standardized feature sequence into the dynamic convolutional neural network, extract the local features of the signal through the deformable convolution kernel, and calculate the feature weights in combination with the self-attention mechanism, filter the feature channels with the top 50% weight, and output the feature vectors of each modality; Multimodal fusion module: This module is used to construct a graph structure with modalities as nodes and correlations as edges based on the feature vectors of each modality. It aggregates cross-modal features through a graph attention network to generate a global state descriptor that integrates spatiotemporal information. Fault detection and classification module: This module is used to perform time series modeling on global state descriptors through a bidirectional LSTM network, output the probability distribution and confidence level of fault types, and complete real-time detection of industrial equipment faults. Online feedback optimization module: used to trigger the incremental learning algorithm to dynamically update model parameters when the confidence level of the detection result is lower than the threshold or a new fault type is detected.

Images