An unbalanced fault diagnosis method based on attention perception and decision coupling
By using a deep reinforcement learning framework that generates three-channel images and asymmetric reward functions, the problems of majority class bias and single feature representation in imbalanced fault diagnosis are solved, and efficient identification and autonomous diagnosis of weak faults are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF ENERGY HEFEI COMPREHENSIVE NAT SCI CENT (ANHUI ENERGY LAB)
- Filing Date
- 2026-04-27
- Publication Date
- 2026-05-29
AI Technical Summary
When dealing with imbalanced fault diagnosis, existing technologies often suffer from majority class bias due to traditional methods, resulting in low minority class fault identification rates, limited feature representation, and a lack of adaptive reward mechanisms, making it difficult to effectively identify minor faults in dynamic environments.
A method based on attention perception and decision coupling is adopted. Three-channel images are generated through Gram angle field and Markov transfer field transformation. Combined with a deep reinforcement learning framework of channel attention mechanism and asymmetric reward function, minority class faults are autonomously diagnosed.
Without relying on manual resampling or prior threshold settings, it significantly improves the ability to identify a very small number of fault modes, providing a highly sensitive and robust intelligent diagnostic solution suitable for complex industrial scenarios.
Smart Images
Figure CN122113004A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial intelligent monitoring and fault diagnosis technology, and in particular to an unbalanced fault diagnosis method based on the coupling of attention perception and decision-making. Background Technology
[0002] With the rapid development of modern manufacturing towards intelligence and automation, health monitoring and fault diagnosis of key equipment such as rotating machinery, including fans, gearboxes, and motors, in fields such as rail transportation, wind power generation, and aerospace have become core links in ensuring industrial safety and continuous operation. However, in actual industrial scenarios, equipment operates in normal conditions for extended periods, resulting in a severe scarcity of fault samples, especially early and subtle fault samples. This leads to a significant imbalance in the collected monitoring data. Under this imbalanced data distribution, majority class (normal state) samples dominate, while minority class (fault state) samples are scarce. Traditional fault diagnosis models are prone to "majority class bias," significantly reducing the accuracy of identifying minority class faults.
[0003] Currently, fault diagnosis methods for imbalanced data are mainly divided into data-level and algorithm-level approaches. Data-level methods typically employ oversampling, undersampling, or synthetic minority oversampling (SMOTE) techniques to balance the number of samples in each class. However, these methods are prone to introducing noise or causing overfitting, especially in industrial vibration signals, where generated spurious samples may corrupt the physical properties of the original signal. Algorithm-level methods include cost-sensitive learning and ensemble learning, but they often rely on prior, manually set class weights or thresholds, lacking adaptability in dynamic environments and struggling to cope with complex operating conditions involving diverse fault modes and varying noise levels.
[0004] In terms of feature extraction, existing deep learning diagnostic models often directly use one-dimensional vibration signals or simple time-frequency diagrams (such as short-time Fourier transform spectra) as input. When dealing with extremely imbalanced data, the weak minority class fault features are easily overwhelmed by the strong energy features of the majority class samples. Although some studies have attempted to introduce image-based coding methods such as Gram angle field (GAF) or Markov transfer field (MTF), they typically only utilize a single transform, ignoring the multi-dimensional physical semantics of vibration signals, including static coordination, static differences, and dynamic evolution, resulting in insufficient feature representation capabilities.
[0005] In terms of decision-making mechanisms, deep reinforcement learning has been increasingly introduced into the field of fault diagnosis in recent years due to its autonomous exploration and sequential decision-making capabilities. However, existing reinforcement learning diagnostic frameworks are relatively weak in state representation, typically directly inputting raw signals or simple features into the agent; at the same time, their reward mechanisms are mostly statically preset fixed values, unable to provide dynamic feedback based on the model's perception of the quality of the current input features. Under extremely imbalanced data conditions, static rewards cannot compensate for the loss dilution effect caused by minority class samples, making it difficult for the agent to converge to an effective strategy for rare fault patterns.
[0006] In summary, how to construct an autonomous diagnostic framework that can deeply mine the heterogeneous features of vibration signals and has the ability to couple perception and decision-making, in order to overcome the technical problems of low identification rate of minority class faults, single feature representation and lack of adaptability of reward mechanism under class imbalance conditions, is a key technical problem that urgently needs to be solved in the field of industrial fault diagnosis. Summary of the Invention
[0007] The purpose of this invention is to address the technical problems of low identification rate of minority class faults, single feature representation, and lack of adaptability of reward mechanism under class imbalance conditions by constructing an autonomous diagnostic framework capable of deeply mining the heterogeneous features of vibration signals and having perception-decision coupling capability. This invention proposes an imbalance fault diagnosis method based on attention perception and decision coupling.
[0008] The technical solution of this invention: an imbalance fault diagnosis method based on the coupling of attention perception and decision-making, comprising the following steps:
[0009] S1. Acquire the one-dimensional vibration time-series observation signal of the target industrial equipment under operating conditions and preprocess it;
[0010] S2. The preprocessed time-series signal is input into the Gram angle field sinusoidal transform, Gram angle field differential transform and Markov transfer field for parallel encoding. The encoding results are spliced together to generate a three-channel image characterizing the multidimensional spatial-spectral-dynamic features of the vibration signal.
[0011] S3. Input the three-channel image into a deep neural network with a channel attention mechanism, calculate the attention weight of each feature channel, and extract the high-dimensional feature vector after attention recalibration as the state representation of the deep reinforcement learning environment.
[0012] S4. In the deep reinforcement learning framework, the average value of the channel attention weights is calculated as the perceptual confidence factor, and an asymmetric reward function dynamically guided by the perceptual confidence factor is constructed.
[0013] S5. Update the agent's policy feedback loop based on the asymmetric reward function and D3QN architecture, and output the diagnostic results for minority class faults and majority class states.
[0014] Optionally, the preprocessing in step S1 includes: normalizing the one-dimensional vibration time-series observation signal, and then using a segmented aggregation approximation method for time-series compression.
[0015] Optionally, the method for generating the three-channel image in step S2 includes:
[0016] The compressed signal segment is processed separately:
[0017] Gram angle field sinusoidal transform is used to extract static cooperative features and form the first channel;
[0018] Gram angle field differential transform is used to extract static difference features and form a second channel;
[0019] Markov transition field transformation is used to extract dynamic transition probability features and form a third channel;
[0020] The first, second, and third channels are stacked to form a three-channel heterogeneous feature image with a size of L×L×3.
[0021] Optionally, in step S3, the feature extraction process of the deep neural network with channel attention mechanism includes:
[0022] The three-channel image is spatially compressed using global average pooling to generate channel-level descriptors;
[0023] The channel-level descriptors are sequentially input into a gating mechanism that includes a fully connected layer and a Sigmoid activation function, and the output is an attention weight vector corresponding to the number of channels;
[0024] The attention weight vector is multiplied element-wise at the channel level with the three-channel image to output the recalibrated state representation.
[0025] Optionally, in step S4, the method for constructing the asymmetric reward function includes:
[0026] The attention intensity factor Ω is defined as the algebraic average of the channel attention weight vectors, which is used to characterize the model's perceptual confidence in the current state.
[0027] When the agent correctly classifies a minority of fault samples, it is given a positive reward, and the value of the positive reward increases dynamically as the attention intensity factor Ω increases.
[0028] When the agent misclassifies a minority of fault samples, a negative penalty is imposed, and the negative penalty value increases with the increase of the attention intensity factor Ω.
[0029] When the agent classifies a majority of normal samples, it is given a fixed base reward or base penalty value, regardless of whether it is correct or incorrect, for asymmetric classification incentive.
[0030] Optionally, the asymmetric reward function is specifically expressed as:
[0031] ;
[0032] in, For a moment The reward value, The preset sensitivity coefficient, For a moment Attention intensity factor For the current action, For real labels, To fix the base reward and penalty values, Represents the majority class. Indicates a minority class.
[0033] Optionally, the update strategy for the D3QN architecture described in step S5 includes:
[0034] A dual-Q network mechanism is used to decouple action selection and target value evaluation;
[0035] The action value function is decomposed into a state value function and an advantage function using a competitive network architecture;
[0036] An experience replay buffer is constructed to store the state transition sequence, and mini-batch training is performed using random sampling.
[0037] A soft update strategy is adopted to periodically synchronize online network parameters to the target network.
[0038] Optionally, the data sources for the one-dimensional vibration time-series observation signal include: measured vibration signals collected from a self-built fault simulation test bench, and / or publicly available gearbox or planetary gearbox vibration datasets.
[0039] Compared with the prior art, this application includes the following beneficial technical effects:
[0040] This invention proposes an imbalanced fault diagnosis method based on the coupling of attention perception and decision-making. Through the synergistic effect of heterogeneous feature fusion, adaptive channel attention recalibration, and an asymmetric reward mechanism guided by perceptual confidence, it achieves, for the first time, deep coupling of perceptual intensity and decision feedback within a deep reinforcement learning framework. This method, without relying on manual resampling or prior threshold settings, can autonomously capture weak fault fingerprints submerged in strong background noise. While ensuring the diagnostic reliability of most classes of samples, it significantly improves the recognition upper limit for critical fault patterns in a very small number of classes. This invention provides a highly sensitive, robust, and human-intervention-free intelligent diagnostic solution for complex industrial scenarios with high reliability requirements and difficult fault data acquisition, such as wind power, aviation, and rail transportation. Attached Figure Description
[0041] Figure 1 The flowchart for the SCD-Net diagnostic agent execution process.
[0042] Figure 2 The flowchart shows the process of heterogeneous feature fusion of publicly available WT-gearbox dataset samples using GAF-MTF.
[0043] Figure 3 The flowchart shows the process of building a PT600 dataset for a private laboratory and fusing heterogeneous features using GAF-MTF.
[0044] Figure 4 Visualize the t-SNE plot comparing the PT600 dataset before and after model training. Detailed Implementation
[0045] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.
[0046] Example
[0047] like Figure 1This invention discloses an unbalanced fault diagnosis method based on attention perception and decision coupling. For a one-dimensional vibration time-series observation signal data source, a three-channel image (SSDM) generation method is performed sequentially, followed by a feature extraction process using a deep neural network with a channel attention mechanism, and an asymmetric reward function construction method dynamically guided by a perception confidence factor. Finally, an optimal fault diagnosis model is constructed through a D3QN architecture update strategy. The optimal fault diagnosis model is then used to classify the operating status of the device under test in real time and output the fault diagnosis results.
[0048] Measured vibration signals were acquired from a self-built PT600 fault simulation test bench, a high-precision, modular dedicated experimental platform. By replacing the standardized fault module motor, it included three operating conditions and six working states: healthy, bearing fault, rotor eccentricity fault, rotor imbalance fault, rotor bar fault, and voltage imbalance fault. A WT-planetary gearbox dataset for benchmark performance comparison was also obtained, containing five working states: gear healthy, gear damaged, tooth surface wear, tooth root fracture, and missing tooth. The above two types of observed signals were then subjected to unified preprocessing.
[0049] Original signal , It is the i-th data point. Min-Max normalization is performed to map the amplitude to the [0,1] interval, followed by Piecewise Aggregate Approximation (PAA) compression to divide the original signal of length n into segments. Divide the data into equal-length segments, calculate the average amplitude within each segment, and obtain a length of... The compressed sequence is given by the following formula:
[0050]
[0051] in It is the i-th data point after compression. is the j-th data point in the original signal, and i is the index of the compressed sequence, ranging from 1 to L. L×L determines the resolution of the target image.
[0052] The preprocessed normalized one-dimensional vibration signal sequence X is projected into polar coordinate space, and the angles between each time point are calculated. A Gram angle field is constructed using correlation. The first channel (GASF) employs Gram angle and field transformation, extracting the static cooperative features of the signal by calculating the cosine of the sum of paired angles. This channel captures the global correlation and amplitude envelope structure of the vibration signal in the time domain, reflecting the steady-state characteristics of equipment operation. The second channel (GADF) employs Gram angle difference field transformation, extracting the static difference features of the signal by calculating the sine of the difference between paired angles. This channel is highly sensitive to phase shifts and local transient fluctuations in the signal, used to characterize abrupt impact signals when the equipment suffers minor damage. The calculation formulas for both are as follows:
[0053]
[0054] in, and These are the pixel values in the i-th row and j-th column of the two matrices (i.e., the final generated image), respectively. and They represent the first The inverse cosine angle values of the data points after standardization are used in the third channel (MTF): Markov transfer field transform is employed. First, the amplitude space of the normalized signal is divided into Q quantile intervals (i.e., state components). The time transition probabilities of the signal between different states are statistically analyzed and mapped to a Markov transfer reference matrix of size Q×Q. This channel extracts the dynamic transition probability characteristics of the signal by characterizing the transition frequency and temporal evolution of the signal in the state space. The formula is shown below:
[0055]
[0056] in It is the pixel value in the i-th row and j-th column of the MTF matrix. These are specific values in the Markov transition matrix, representing the data transitioning from state to state. Transition to state The probability, and These are the specific values at time points i and j after compression, respectively. This indicates the conditions that need to be met.
[0057] Finally, the three single-channel matrices representing different dimensional features are stacked in RGB color channel order to form a three-channel heterogeneous feature image of size L×L×3. (SSDM), see the formula below.
[0058]
[0059] in, Represents the real number field. This represents a deep stacking and fusion operation of data, enabling the network to achieve the same pixel coordinates. The system simultaneously reads information from GASF, GADF, and MTF.
[0060] This image transforms the implicit "spectral-spatial-dynamic" information in the original one-dimensional signal into high-fidelity image texture. Figure 2 and Figure 3 This provides an observable state space with multidimensional physical semantics for subsequent deep reinforcement learning agents.
[0061] The feature extraction process of a deep neural network with channel attention mechanism: First, spatial information compression is performed. The spatial dimension aggregation of the input three-channel SSDM image is performed using a global average pooling operator. This process compresses the distribution characteristics of L×L pixels in each feature channel into a scalar value, thereby obtaining a channel-level feature descriptor with a global receptive field. The formula is shown below:
[0062]
[0063] in, It is the spatial size of the feature map. Indicates at a point in time Time, No The pixel value at coordinates (i,j) in each feature channel. It performs a summation process on all pixels within the channel.
[0064] Then, channel dependency modeling is performed, and the generated descriptors are... The input consists of a gating mechanism comprised of two fully connected (FC) layers. The first fully connected layer performs dimensionality reduction and compression on the channel features, combining them with the ReLU activation function to achieve non-linear combination of features, uncovering the intrinsic evolutionary relationships between different channels, and realizing non-linear mapping. The second fully connected layer restores the features to the original channel dimension and passes them through the Sigmoid normalized activation function, outputting a set of adaptive attention weight vectors W with values ranging from [0,1]. The calculation formula is as follows:
[0065]
[0066] in, It is the Sigmoid activation function. It is the ReLU activation function. , These are the weight matrices of two fully connected layers.
[0067] Finally, feature recalibration and adaptive enhancement are performed: The weight vector W is reapplied to the original three-channel SSDM image using a channel-level element-wise multiplication method. Based on the current input fault characteristics, the weights of channels containing high-frequency random noise or redundant information (such as the MTF channel severely affected by environmental interference) are dynamically reduced, while weak feature channels containing key fault signs (such as the GADF channel representing minority types of impact signals) are forcibly amplified. The final output is the recalibrated state representation. This enables deep reinforcement learning agents to filter out irrelevant background and focus on highly discriminative non-equilibrium fault feature fingerprints. The formula is as follows:
[0068]
[0069] in, This is the final feature map output after recalibration. This represents multiplication by channel. It is the original input SSDM image data. This means applying the first weight to the first channel (GASF), and so on. , , These represent the original feature channels of GASF, GADF, and MTF, respectively.
[0070] The method for constructing an asymmetric reward function dynamically guided by the perceptual confidence factor includes: extracting the weight vector output by the channel attention module. The attention intensity factor is obtained by calculating its algebraic average. The calculation formula is as follows:
[0071]
[0072] in, For a moment Attention intensity factor Indicates the first The weighting coefficient for each channel, C=3, This reflects the deep neural network's overall evaluation of the saliency of features in the current input SSDM image: a high score indicates that the current sample contains clear and typical fault identification features, and the model's perception confidence is high; a low score indicates that the current sample features are blurry or severely affected by noise, and the model's perception is in an uncertain state. Then, a dynamic asymmetric reward mapping based on perception confidence is performed. When the agent accurately identifies a minority class fault sample, the system awards the sum of the base reward value and the attention gain term; if the identification is incorrect, a heavy penalty is imposed. This mechanism ensures that the more confident the model is in identifying a fault, the higher the reward for correct identification and the heavier the cost of incorrect identification, thus forcing the policy network to deeply mine the marginal features of minority class samples. For the majority class normal samples with a large amount of data, regardless of whether the classification is correct or incorrect, only a fixed base reward value or base penalty value is given. The formula is shown below:
[0073]
[0074] in, For a moment The asymmetric reward function, The preset sensitivity coefficient, For a moment Attention intensity factor For the current action, For real labels, To fix the base reward and penalty values, Represents the majority class. This represents the minority class. Through the aforementioned asymmetric design, the agent is endowed with risk-preference characteristics during training iterations: when dealing with an extremely skewed state space, the agent will actively tend to converge towards minority class features due to the high dynamic rewards brought by minority class samples; combined with the channel attention mechanism's enhancement effect on weak signals, this reward function effectively compensates for the loss function dilution problem caused by the scarcity of samples, significantly improving the model's G-mean and F1-score under extremely imbalanced conditions.
[0075] The D3QN architecture update strategy specifically includes: First, implementing a dual-network decoupling evaluation mechanism: adopting a dual-network architecture of an online network and a target network, aiming to eliminate overestimation bias in traditional Q-Learning. During the update process, the current online network is utilized... Determine the optimal diagnostic action and utilize the target network. Evaluate the Q-value of this action. Its target value. The calculation formula is as follows:
[0076]
[0077] in, For a moment The asymmetric reward function, It is a discount factor. and This represents the state at the next time step and the possible candidate actions at the next time step. Then, a Dueling eigenvalue decomposition architecture is performed: the terminal output of the deep convolutional neural network is decomposed into two independent streams: the state-value stream and the value stream. and action advantage flow The former extracts the state value function V(s) to learn the inherent value of the current SSDM image state itself; the latter extracts the action advantage function A(s,a) to measure the additional benefit of taking a specific diagnostic action in the current state relative to average performance. Finally, the action value function is reconstructed by combining V(s) and A(s,a) and subtracting the average of the advantage functions. The aggregation and reconstruction formula is as follows:
[0078]
[0079] in, and These are network parameters unique to the advantage stream and the value stream, respectively. It indicates the size of the action space. Represents the action space. These represent the current state and the action, respectively. This indicates the action to be taken at the next moment. Finally, a loop experience replay buffer is established to store the state transition sequence generated by the interaction between the agent and the SSDM image environment. , The system represents the state at the next moment. Training is performed by randomly sampling small batches of samples from a buffer. During experience replay, an asymmetric reward function is used to enable the agent to repeatedly learn from stored rare, minority-class fault experiences, preventing the learned fault identification strategy from being overwritten by massive amounts of normal-state data during real-time training. A soft update strategy is employed, periodically updating the parameters of the online network. Synchronize to target network After the D3QN architecture training converges, the optimized online network is extracted as a fault diagnosis classifier. The single-source vibration signal of the device under test collected in real time is converted into an SSDM image and input into the classifier. The classifier outputs the corresponding fault category index, thereby realizing accurate identification and early warning of the state of industrial equipment under unbalanced working conditions.
[0080] Finally, the feasibility of this method was verified through multiple random seed experiments, as detailed in [link to relevant documentation]. Figure 4Before training (left figure), all fault samples were highly mixed in two-dimensional space, without forming clear boundaries or clusters. After model training (right figure), the data points were rearranged in the feature space, forming five clear, independent, and compact clusters. The points within each cluster are highly clustered, and the distance between different clusters is large with clear boundaries. This proves that the model not only distinguishes fault types but also successfully decouples fault features from operating condition features, improving the model's generalization ability.
[0081] It is worth noting that this invention constructs an asymmetric reward function dynamically guided by the perceptual confidence factor, which provides a positive reward that increases with the attention intensity for the correct identification of minority class fault samples and imposes an increasing penalty for incorrect identification. This forces the deep reinforcement learning agent to actively converge to minority class features in an extremely skewed state space, effectively overcoming the fault underreporting problem caused by the "majority class bias" in traditional methods and significantly improving the G-mean and F1-score indicators.
[0082] This invention simultaneously performs Gram angle field sinusoidal transform, Gram angle field differential transform, and Markov transfer field transform on the same-dimensional vibration signal, splicing them to generate a spatial-spectral-dynamic three-channel image (SSDM). It also extracts the signal's static cooperative features, static differential features, and dynamic transfer probability features. Compared to single-feature or simple time-frequency diagrams, this provides richer and more complementary equipment state information, laying a high-quality observable state space for agent decision-making. Furthermore, this invention introduces a channel attention mechanism into the deep neural network. Through global average pooling and gating units, it adaptively calculates the weights of each feature channel, amplifying channels containing key fault symptoms (such as the shock-sensitive GADF channel) and suppressing high-noise or redundant channels. This allows the model to capture weak minority class fault fingerprints even under strong background noise, improving the robustness of feature extraction.
[0083] Furthermore, this invention innovatively defines the average value of channel attention weights as the perceptual confidence factor and directly embeds it into the reward function. This allows the agent's policy update to depend not only on the action outcome but also on the current feature quality. The more confident the model is in the fault characteristics, the higher the reward for correct identification and the greater the cost of incorrect identification, thus achieving synergistic optimization of perception and decision-making and overcoming the lack of adaptive capability in traditional reinforcement learning's fixed reward mechanism. This invention employs a dual-Q network to decouple action selection and target value evaluation, eliminating the overestimation bias of traditional Q-Learning. Simultaneously, it uses a competitive network structure to decompose state value and action advantage, combined with experience replay and soft update strategies, enabling the agent to repeatedly learn rare fault experiences, preventing it from being overwhelmed by a large number of normal samples, and improving training convergence and policy generalization ability under extremely imbalanced data. This invention solves the class imbalance problem entirely at the algorithm level, without relying on data preprocessing methods such as oversampling and undersampling, avoiding the problems of generating false samples that destroy the physical properties of the original signal or discarding scarce sample information, thus maintaining the authenticity and generalization ability of the diagnostic model.
[0084] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.
Claims
1. An imbalance fault diagnosis method based on the coupling of attention perception and decision-making, characterized in that, Includes the following steps: S1. Acquire the one-dimensional vibration time-series observation signal of the target industrial equipment under operating conditions and preprocess it; S2. The preprocessed time-series signal is input into the Gram angle field sinusoidal transform, Gram angle field differential transform and Markov transfer field for parallel encoding. The encoding results are spliced together to generate a three-channel image characterizing the multidimensional spatial-spectral-dynamic features of the vibration signal. S3. Input the three-channel image into a deep neural network with a channel attention mechanism, calculate the attention weight of each feature channel, and extract the high-dimensional feature vector after attention recalibration as the state representation of the deep reinforcement learning environment. S4. In the deep reinforcement learning framework, the average value of the channel attention weights is calculated as the perceptual confidence factor, and an asymmetric reward function dynamically guided by the perceptual confidence factor is constructed. S5. Update the agent's policy feedback loop based on the asymmetric reward function and D3QN architecture, and output the diagnostic results for minority class faults and majority class states.
2. The imbalance fault diagnosis method based on the coupling of attention perception and decision-making according to claim 1, characterized in that, The preprocessing in step S1 includes: normalizing the one-dimensional vibration time-series observation signal, and then using a segmented aggregation approximation method for time-series compression.
3. The imbalance fault diagnosis method based on the coupling of attention perception and decision-making according to claim 1, characterized in that, The method for generating the three-channel image in step S2 includes: The compressed signal segment is processed separately: Gram angle field sinusoidal transform is used to extract static cooperative features and form the first channel; Gram angle field differential transform is used to extract static difference features and form a second channel; Markov transition field transformation is used to extract dynamic transition probability features and form a third channel; The first, second, and third channels are stacked to form a three-channel heterogeneous feature image with a size of L×L×3.
4. The imbalance fault diagnosis method based on the coupling of attention perception and decision-making according to claim 1, characterized in that, In step S3, the feature extraction process of the deep neural network with channel attention mechanism includes: The three-channel image is spatially compressed using global average pooling to generate channel-level descriptors; The channel-level descriptors are sequentially input into a gating mechanism that includes a fully connected layer and a Sigmoid activation function, and the output is an attention weight vector corresponding to the number of channels; The attention weight vector is multiplied element-wise at the channel level with the three-channel image to output the recalibrated state representation.
5. The imbalance fault diagnosis method based on the coupling of attention perception and decision-making according to claim 1, characterized in that, In step S4, the method for constructing the asymmetric reward function includes: The attention intensity factor Ω is defined as the algebraic average of the channel attention weight vectors, which is used to characterize the model's perceptual confidence in the current state. When the agent correctly classifies a minority of fault samples, it is given a positive reward, and the value of the positive reward increases dynamically as the attention intensity factor Ω increases. When the agent misclassifies a minority of fault samples, a negative penalty is imposed, and the negative penalty value increases with the increase of the attention intensity factor Ω. When the agent classifies a majority of normal samples, it is given a fixed base reward or base penalty value, regardless of whether it is correct or incorrect, for asymmetric classification incentive.
6. The imbalance fault diagnosis method based on the coupling of attention perception and decision-making according to claim 5, characterized in that, The asymmetric reward function is specifically expressed as follows: ； in, For a moment The reward value, The preset sensitivity coefficient, For a moment Attention intensity factor For the current action, For real labels, To fix the base reward and penalty values, Represents the majority class, Indicates a minority class.
7. The imbalance fault diagnosis method based on the coupling of attention perception and decision-making according to claim 1, characterized in that, The update strategy for the D3QN architecture described in step S5 includes: A dual-Q network mechanism is used to decouple action selection and target value evaluation; The action value function is decomposed into a state value function and an advantage function using a competitive network architecture; An experience replay buffer is constructed to store the state transition sequence, and mini-batch training is performed using random sampling. A soft update strategy is adopted to periodically synchronize online network parameters to the target network.
8. The imbalance fault diagnosis method based on the coupling of attention perception and decision-making according to claim 1, characterized in that, The data sources for the one-dimensional vibration time-series observation signals include: measured vibration signals collected from a self-built fault simulation test bench, and / or publicly available gearbox or planetary gearbox vibration datasets.