Device cross-domain fault diagnosis method based on multi-sensor and causal contrast decoupling

Abstract

This invention relates to the field of fault diagnosis technology and discloses a cross-domain fault diagnosis method for equipment based on multi-sensor and causal comparison decoupling. The method constructs a cross-domain fault diagnosis model, which includes a multi-channel information dynamic fusion module based on the Mamba model, a causal comparison decoupling module based on a structural causal model, and a shared classifier. It acquires raw monitoring signals collected by various sensors of the mechanical equipment under different operating conditions, thereby constructing a training set. The cross-domain fault diagnosis model is trained using the training set. The sensor signals of the mechanical equipment to be diagnosed are input into the trained cross-domain fault diagnosis model, and the fault diagnosis result of the mechanical equipment to be diagnosed is output. This invention can dynamically and efficiently fuse multi-sensor information to avoid negative fusion, and can thoroughly separate fault features and domain interference features under the guidance of causal mechanisms, thereby improving the diagnostic accuracy and generalization ability of the model under unknown operating conditions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, specifically a cross-domain fault diagnosis method for devices based on multi-sensor and causal comparison decoupling, as well as a computer terminal and computer-readable storage medium for applying this method. Background Technology

[0002] As a critical component of equipment in modern industrial production, the stability and reliability of bearings are essential for the operation of the entire system. Unexpected equipment failures can lead to severe economic losses and even personal injury or death. To comprehensively and accurately obtain the operational health status of equipment and mitigate potential risks, the industry has widely adopted multi-sensor monitoring technology. The rational use of multi-sensor monitoring data can provide richer information on equipment health status, thereby further improving the accuracy of fault diagnosis models.

[0003] However, existing multi-sensor cross-domain fault diagnosis methods still face serious technical bottlenecks in practical applications, mainly in the following two aspects: First, traditional feature fusion methods mostly employ simple concatenation or summation operations. These simple operations are insufficient to effectively characterize the specific attributes of different sensor channels and the differences between sensors, neglecting the inherent correlation of data and the specificity of fault channels. This traditional fusion framework fails to achieve a balance between data fusion and causal learning, easily leading to the loss of key diagnostic information and negative fusion phenomena, thereby severely weakening the model's generalization ability.

[0004] Second, in complex industrial environments, machinery often operates across various conditions, leading to severe distribution shifts in the target domain data. Although domain generalization techniques have been introduced to extract transferable essential features, in actual monitoring signals, fault-related information is often deeply coupled with interference information from specific domains (such as operating conditions, mechanical structure, and sensor characteristics). Existing decoupling methods, lacking causal guidance from data generation mechanisms, result in incomplete decoupling between causal and non-causal fault features. This makes diagnostic models highly susceptible to spurious correlations in the data, prone to misjudgment when faced with completely unknown target domain distributions, and unable to achieve effective cross-domain generalization.

[0005] In summary, traditional mechanical equipment fault diagnosis methods are limited in their cross-domain diagnostic performance when processing multi-sensor signals due to entangled fault causal characteristics, interference in specific domains, and information redundancy in multiple channels. This limitation urgently needs to be addressed. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention provides a cross-domain fault diagnosis method for equipment based on multi-sensor and causal comparison decoupling. This method can dynamically and efficiently fuse multi-sensor information to avoid negative fusion, and can completely separate fault features and domain interference features under the guidance of causal mechanisms, thereby improving the diagnostic accuracy and generalization ability of the model under unknown operating conditions.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention discloses a cross-domain fault diagnosis method for devices based on multi-sensor and causal comparison decoupling, including steps S1 to S4.

[0008] S1. Construct a cross-domain fault diagnosis model, which includes a multi-channel information dynamic fusion module based on the Mamba model, a causal comparison decoupling module based on the structural causal model, and a shared classifier; S2. Acquire raw monitoring signals collected by various sensors of mechanical equipment under different operating conditions, and construct a training set accordingly; S3. Train the cross-domain fault diagnosis model using the training set; The multi-channel information dynamic fusion module is used to input the original monitoring signal, dynamically perceive the differences between sensors and adaptively extract detailed features to construct a global fusion feature map that aggregates multi-channel information and retains specific channel fault features. The causal comparison decoupling module is used to decouple the global fusion feature map into fault causal features containing core fault information and non-causal features containing specific domain interference information. The shared classifier is used to predict the fault category based on the decoupled fault causal features. During the model training phase, a joint objective function is constructed based on causal comparison loss, exchange reconstruction loss, conditional independence loss, and cross-entropy loss. The cross-domain fault diagnosis model is iteratively optimized by comparing similarity and constraining feature exchange reconstruction of sample pairs containing the same fault across domains. S4. Input the sensor signals of the mechanical equipment to be diagnosed into the trained cross-domain fault diagnosis model, and output the fault diagnosis results of the mechanical equipment to be diagnosed.

[0009] As a further improvement to the above solution, the multi-channel information dynamic fusion module includes a dynamic information enhancement unit, the processing steps of which include: Convolutional layers are used to process raw signals from different sensors to obtain enhanced initial features; The enhanced initial features from different sensors are subtracted element-wise to generate differential features, and the differential weights between the initial features are calculated using global average pooling and nonlinear activation functions. The difference weights are applied to the initial multi-scale feature fusion result and the original sensor signal, respectively, to generate enhanced feature outputs for subsequent processing.

[0010] As a further improvement to the above scheme, the calculation formula for the enhanced feature output is as follows: ; In the formula, and These represent the raw signals from two different sensors; and They represent Corresponding enhanced feature output and The corresponding enhanced feature output; and This represents the enhanced initial features obtained through the convolutional layer; This represents the Sigmoid activation function; This indicates a global average pooling operation; This represents the preliminary feature results of multi-sensor signal sample fusion; This indicates element-wise addition. This indicates element-wise multiplication.

[0011] As a further improvement to the above solution, the multi-channel information dynamic fusion module also includes a Mamba fusion unit, the processing steps of which include: The enhanced feature output of the dynamic information enhancement unit is subjected to a depthwise separable convolution operation, and the results of different channels are mixed to generate an initial mixed feature. The initial mixed features are input into the Mamba layer to capture long-range spatial dependencies, and then an efficient channel attention operation is performed to reduce channel redundancy, resulting in the processed mixed enhancement information. The hybrid enhancement information is then added and aggregated element by element with the original signals from each sensor to obtain the final global fusion feature map.

[0012] As a further improvement to the above scheme, the calculation formulas for the hybrid enhancement information and the global fusion feature map are as follows: ; ; In the formula, This represents the hybrid enhancement information; This represents the global fused feature map; This indicates efficient channel attention operations; Indicates a linear layer; Indicates the Mamba layer; This represents the initial mixing characteristics.

[0013] As a further improvement to the above scheme, the calculation process of the causal contrast loss includes: For any given sample, samples with consistent fault labels but different domain labels are selected from the global sample space to construct fault sample pairs with the same fault but different domain labels. ;in, and These represent the global fusion feature maps extracted from the original monitoring signals of the two samples, respectively. express The domain label of the corresponding sample; express The domain label of the corresponding sample; The fault label indicates the sample; The encoder is used to select the fault sample pair. and Encode them separately as combinations of latent features and ;in, and For fault causal features that only contain cross-domain invariant attributes; and To retain only non-causal features that are sensitive to specific domains; Construct a causal contrastive loss function, expressed as follows: ; ; In the formula, This indicates the loss of causal comparison; Indicates KL divergence; This represents the posterior probability distribution induced by the encoder; Represents the Gaussian kernel function; For a fixed combination of two distributions, It is a set of different distribution combinations.

[0014] As a further improvement to the above scheme, the calculation process of the exchange reconstruction loss includes: Perform a cross-swapping operation on the fault causal features of the fault sample pairs, and transfer the fault causal features... Non-causal characteristics Combined into the first exchange feature pair , to identify the causal characteristics of the fault Non-causal characteristics Combining into a second exchange feature pair ; The first exchanged feature pair and the second exchanged feature pair are respectively input into the corresponding decoder to generate a reconstructed signal and a reconstructed label for fitting the original signal and the domain label; By constraining the reconstruction errors between the original signal, original domain label, and reconstructed signal, reconstructed label, a commutative reconstruction loss function is constructed, expressed as follows: ; In the formula, This represents the exchange reconstruction loss; This represents the conditional probability distribution of the decoder used to reconstruct the original signal; This represents the conditional probability distribution of the decoder used to reconstruct the generated domain labels.

[0015] As a further improvement to the above scheme, the formula for calculating the conditional independence loss is as follows: ; ; ; In the formula, This represents the overall optimization objective of the VAE model, i.e., the lower bound of evidence, where Parameters representing the encoder, Parameters representing the decoder, It is used for balance and The ratio between the two; express The reconstruction terms in the code are used to ensure that the decoupled features contain sufficient information. It represents the divergence value between latent variables, ensuring the degree of similarity between different latent variables; Expressing expectations; Indicates the divergence value; An approximate posterior distribution representing the causal characteristics of the fault; Represents the approximate posterior distribution of non-causal features; Indicating the causal characteristics of the fault Prior distribution; Indicating non-causal characteristics Prior distribution.

[0016] The present invention also discloses a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling as described above.

[0017] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling as described above.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a cross-domain fault diagnosis model comprising a multi-channel information dynamic fusion module, a causal comparison decoupling module, and a shared classifier. Based on an iterative optimization mechanism using joint loss functions such as causal comparison and exchange reconstruction, it achieves effective separation of fault causal features containing core fault information from specific domain interference features under heterogeneous multi-sensor data input. This strategy not only utilizes multi-dimensional sensor information for judgment but also eliminates background interference caused by specific operating conditions (such as cross-speed rotation) through a causal mechanism, thereby improving the model's fault classification accuracy and cross-domain generalization ability when facing unknown target domain distributions.

[0019] 2. This invention achieves adaptive aggregation of multi-channel monitoring signals by combining dynamic information enhancement and Mamba fusion technology in a multi-channel dynamic information fusion module. The dynamic information enhancement unit amplifies the sensitivity differences of different sensors to specific fault responses by calculating feature difference weights and reapplying them to the original signal and multi-scale features, thus promoting the extraction and enhancement of complementary detail features. By combining the Mamba layer with an efficient channel attention mechanism, the modeling of the spatial dependence of long signal sequences and the reduction of channel redundancy are completed while maintaining linear time complexity. Furthermore, the final residual addition aggregation strategy effectively preserves the unique fault characteristics of specific sensor channels, avoiding the loss of multi-source data information and negative fusion phenomena caused by direct splicing.

[0020] 3. This invention achieves deep quantitative constraints on the potential feature space by introducing causal comparison loss based on cross-domain sample pairs with the same fault, exchange reconstruction loss based on feature crossover, and conditional independence loss based on lower bounds of evidence. Causal comparison constraint forces features of the same type of fault to remain highly similar and significantly different from domain interference features, anchoring core diagnostic information within the causal feature space. The exchange reconstruction operation further verifies the decoupling process by reconstructing the signals from the exchanged features, ensuring that the decoupled causal features do not omit key fault attributes, and that non-causal features fully absorb interference biases from specific domains. Conditional independence constraint utilizes dynamic balance parameters to prevent feature space overlap or information redundancy between causal and non-causal features. The synergistic effect of multi-dimensional loss constraints improves the sufficiency of feature decoupling and reduces the risk of cross-class misclassification between different types of faults. Attached Figure Description

[0021] Figure 1This is a flowchart of the cross-domain fault diagnosis method for devices based on multi-sensor and causal comparison decoupling in Embodiment 1 of the present invention.

[0022] Figure 2 This is a physical diagram of the experimental equipment used in Embodiment 1 of the present invention.

[0023] Figure 3 This is an architectural diagram of the dynamic information enhancement unit in Embodiment 1 of the present invention. Figure 3 In this context, conv represents a convolutional layer, GAP represents global average pooling, and sigmoid represents normalization.

[0024] Figure 4 This is an architectural diagram of the Mamba fusion unit in Embodiment 1 of the present invention. Figure 4 In this context, LN stands for layer normalization, and Linear stands for fully connected layer.

[0025] Figure 5 This is a schematic diagram of the structural causal model in Embodiment 1 of the present invention.

[0026] Figure 6 These are the original vibration signals from different sensors in Embodiment 1 of the present invention.

[0027] Figure 7 The accuracy of different methods on TR1-TR5 tasks in Embodiment 1 of the present invention.

[0028] Figure 8 These are the F1 values ​​for different methods in Embodiment 1 of the present invention.

[0029] Figure 9 This is the confusion matrix of different methods on the TR 2 task in Embodiment 1 of the present invention; Figure 9 In the diagram, (a) represents the CCDG method, (b) represents the MDGP method, (c) represents the CCN method, (d) represents the CDDG method, (e) represents the DDDG method, and (f) represents the method proposed in this invention; where "B" on the coordinate axis indicates ball bearing failure, "I" indicates inner race failure, "N" indicates normal, and "O" indicates outer race failure.

[0030] Figure 10 This is the confusion matrix of the ablation experiment on the TR 2 task in Embodiment 1 of the present invention; Figure 10 (a) represents the baseline method, (b) represents a single sensor, (c) represents the summation of signals from different sensors, and (d) represents the method with missing data. (e) is missing (f) is the method proposed in this invention.

[0031] Figure 11 This is to ensure the accuracy of the ablation experiments on TR 1-TR 5 tasks in Embodiment 1 of the present invention.

[0032] Figure 12 This is a schematic diagram of the structure of the computer terminal in Embodiment 2 of the present invention. Detailed Implementation

[0033] 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.

[0034] Example 1 Please see Figure 1 This embodiment provides a device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling, including steps S1 to S4.

[0035] S1. Construct a cross-domain fault diagnosis model, which includes a multi-channel information dynamic fusion module based on the Mamba model, a causal comparison decoupling module based on the structural causal model, and a shared classifier.

[0036] S2. Acquire raw monitoring signals from various sensors of the mechanical equipment under different operating conditions, and construct a training set accordingly.

[0037] Mechanical equipment (such as rotating machinery, bearings, gearboxes, etc.) typically operates under complex and variable conditions (such as different speeds and loads) in real industrial scenarios. Due to limitations in installation location and sensitive frequency bands, a single sensor cannot comprehensively reflect the true health status of the equipment. Therefore, such as Figure 2 As shown, multiple sensors such as acceleration, temperature, and pressure can be deployed to collect multi-source heterogeneous data, which can capture richer and more complementary health status information, thereby breaking through the performance limit of a single sensor.

[0038] It should be noted that in this invention, "domain" specifically refers to the specific operating conditions or data distribution environment of the equipment, such as different rotational speeds, load states, or operating temperatures. Correspondingly, "cross-domain fault diagnosis" refers to the process of directly applying a diagnostic model, after training it under several known operating conditions (i.e., the source domain, for example, training the model on samples collected when the equipment speed is between 500 rpm and 875 rpm), to a completely unknown new operating condition (i.e., the target domain, for example, directly diagnosing the equipment state at a speed of 1000 rpm) with significant data distribution shifts, to identify faults. The core advantage of this invention lies in eliminating the specific domain interference characteristics caused by different operating conditions through a causal comparison decoupling mechanism, thereby achieving highly reliable cross-domain diagnosis across different operating environments.

[0039] S3. The cross-domain fault diagnosis model is trained using the training set.

[0040] The multi-channel information dynamic fusion module is used to input the original monitoring signal, dynamically perceive the differences between sensors and adaptively extract detailed features, and construct a global fusion feature map that aggregates multi-channel information and retains the fault features of specific channels.

[0041] The multi-channel information dynamic fusion module includes a dynamic information enhancement unit and a Mamba fusion unit.

[0042] Please see Figure 3 The processing steps of the dynamic information enhancement unit include: Convolutional layers are used to process raw signals from different sensors to obtain enhanced initial features. The purpose of using convolutional layers for initial processing is to amplify the micro-detail information in the signal, thereby enhancing the model's ability to perceive fine-grained features in the sensor signal. The enhanced initial features from different sensors are subtracted element-wise to generate differential features, and the differential weights between the initial features are calculated using global average pooling and nonlinear activation functions. The difference weights are applied to the initial multi-scale feature fusion result and the original sensor signal, respectively, to generate enhanced feature outputs for subsequent processing.

[0043] The core mechanism of the dynamic information enhancement unit of this invention lies in deliberately amplifying and utilizing the sensitivity differences in the response to the same fault among different heterogeneous sensors through feature subtraction and attention weight allocation. Applying the calculated difference weights back to the original signal and the preliminary fusion result effectively promotes the extraction of complementary detail information and feature enhancement in the signal.

[0044] The formula for calculating the enhanced feature output is as follows: ; In the formula, and These represent the raw signals from two different sensors; and They represent Corresponding enhanced feature output and The corresponding enhanced feature output; and This represents the enhanced initial features obtained through the convolutional layer; This represents the Sigmoid activation function; This indicates a global average pooling operation; This represents the preliminary feature results of multi-sensor signal sample fusion; This indicates element-wise addition. This indicates element-wise multiplication.

[0045] Please see Figure 4 The processing steps of the Mamba fusion unit include: Enhanced feature output of dynamic information enhancement unit and Perform depthwise separable convolution operations and analyze the results for different channels. and Perform mixing to generate initial mixed features ; Initial blending features The input is fed into a Mamba layer to capture long-range spatial dependencies, followed by efficient channel attention (ECA) operations to reduce channel redundancy, resulting in processed hybrid augmented information. ; Traditional sequence models such as Transformer suffer from quadratic growth in computational complexity due to their self-attention mechanisms when processing long-term, high-sampling-rate multi-sensor signals, resulting in extremely high computational costs. This invention introduces a Mamba layer based on a state-space model, which adjusts state updates based on relevant dynamic parameters, achieving efficient dependency modeling of long sequences with linear time complexity. Subsequently, an efficient channel attention mechanism is incorporated, effectively reducing channel information redundancy during multi-channel feature fusion and refining the features.

[0046] Hybrid Enhancement Information Raw signals from each sensor and Element-wise aggregation is performed to obtain the final global fused feature map. Finally, residual connections are used to aggregate the processed features with the original signal across layers. This method not only effectively preserves the unique fault characteristics of specific sensor channels but also completely avoids the negative fusion phenomenon caused by direct and crude splicing of features in traditional multi-source data fusion methods, thus achieving efficient utilization of multi-sensor signals.

[0047] The calculation formulas for the hybrid enhancement information and the global fusion feature map are as follows: ; ; In the formula, This represents the global fused feature map; This indicates efficient channel attention operations; Indicates a linear layer; Indicates the Mamba layer; This represents the initial mixed characteristics.

[0048] The causal comparison decoupling module is used to decouple the global fused feature map into causal features containing core fault information and non-causal features containing interference information in a specific domain. The shared classifier is used to predict the fault category based on the decoupled causal features.

[0049] In actual diagnostics, data often exhibits spurious correlations. If a model mistakenly uses background noise unique to a specific speed as a diagnostic criterion, serious misdiagnosis can occur when the device switches to an unknown speed (crossing domains). Therefore, the causal decoupling module of this invention addresses this issue. Figure 5 Designed under the guidance of the structural causal model shown ( Figure 5 In this context, D represents the enhancement information corresponding to different sensor signals. This refers to the causal characteristics of the fault (used for diagnosis). (where y is the fault label, d is the domain label, and x1 and x2 are the original signal samples) aims to forcibly extract the physical essence of the fault (causal features) from the complex signals and eliminate the background interference (non-causal features) that fluctuates due to changes in operating conditions.

[0050] The causal decoupling module consists of multiple encoders and decoders, where the encoders are used to decouple the causal characteristics of faults. Non-causal characteristics The decoder is used to enable fault causal characteristics. Non-causal characteristics To preserve as much valid information as possible from the samples, the classifier is used for final fault diagnosis. For convenience, Abbreviated as .

[0051] During the training phase of the cross-domain fault diagnosis model, causal contrastive loss is used. Exchange and reconstruction losses Loss of conditional independence and cross-entropy loss Constructing a joint objective function By performing similarity comparison and feature exchange to reconstruct constraints on sample pairs containing the same fault across domains, the cross-domain fault diagnosis model is iteratively optimized.

[0052] The expression for the joint objective function is: ; in, , and These are the weight parameters.

[0053] The calculation process for causal contrast loss includes: For any given sample, samples with consistent fault labels but different domain labels are selected from the global sample space to construct fault sample pairs with the same fault but different domain labels. ;in, and These represent the global fusion feature maps extracted from the original monitoring signals of the two samples, respectively. express The domain label of the corresponding sample; express The domain label of the corresponding sample; The fault label indicates the sample.

[0054] The encoder is used to select the fault sample pair. and Encode them separately as combinations of latent features and The expression is: ; in, and For fault causal features that only contain cross-domain invariant attributes; and Non-causal features that retain only sensitive information in specific domains. The same fault labeling requirements. There are strong similarities between them; different domain labels require (or )and Between and and There are significant differences between them.

[0055] ; in, Indicates the expected value; To represent arbitrary divergence, this embodiment uses average divergence for ease of implementation. The calculation formula is: ; To transform these qualitative constraints into quantitative indicators, this embodiment designs a causal contrast loss function for iterative optimization, the formula of which is shown below: ; ; In the formula, Indicates KL divergence; This represents the posterior probability distribution induced by the encoder; Represents the Gaussian kernel function; For a fixed combination of two distributions, It is a set of different distribution combinations.

[0056] The purpose of constructing the causal contrast loss function in this invention is to force the causal features of samples with the same fault label to maintain a high degree of similarity (by minimizing the divergence of the first term of the formula), while forcing the fault features to maintain a significant difference from the non-causal features containing information of a specific domain (by inversely maximizing the distance between them through the Gaussian kernel function), thereby anchoring the true fault information in the causal feature space.

[0057] To ensure that the decoupled causal features fully encompass the core fault-related information, while the non-causal features fully retain domain-specific interference information, this embodiment introduces an exchange-reconstruction loss. Through feature exchange and signal reconstruction, the model's ability to extract cross-domain invariant fault features is further enhanced. The calculation process of the exchange-reconstruction loss includes: Perform a cross-swapping operation on the fault causal features of the fault sample pairs, and transfer the fault causal features... Non-causal characteristics Combined into the first exchange feature pair , to identify the causal characteristics of the fault Non-causal characteristics Combining into a second exchange feature pair ; The first and second exchanged feature pairs are input into their respective decoders to generate a reconstructed signal and reconstructed label for fitting the original signal and domain label, expressed as follows: ; In the formula, For decoders; For The original signal reconstructed from latent variables; For The original signal is reconstructed from the latent variables.

[0058] By constraining the reconstruction errors between the original signal, original domain label, and reconstructed signal, reconstructed label, a commutative reconstruction loss function is constructed, expressed as follows: ; In the formula, This represents the conditional probability distribution of the decoder used to reconstruct the original signal; This represents the conditional probability distribution of the decoder used to reconstruct the generated domain labels.

[0059] The purpose of introducing the exchange reconstruction loss in this invention is to deeply verify and supervise the thoroughness of model decoupling by forcibly inserting fault features belonging to domain A into the background interference features of domain B for signal restoration. This ensures that the decoupled causal features do not miss any key fault attributes, while the non-causal features fully encompass domain interference bias, thus improving the model's generalization ability in unknown domains.

[0060] The formula for calculating the conditional independence loss is: ; ; ; In the formula, This represents the overall optimization objective of the VAE model, i.e., the lower bound of evidence, where Parameters representing the encoder, Parameters representing the decoder, It is used for balance and The ratio between the two; express The reconstruction terms in the code are used to ensure that the decoupled features contain sufficient information. It represents the divergence value between latent variables, ensuring the degree of similarity between different latent variables; Expressing expectations; Indicates the divergence value; An approximate posterior distribution representing the causal characteristics of the fault; Represents the approximate posterior distribution of non-causal features; Indicating the causal characteristics of the fault Prior distribution; Indicating non-causal characteristics Prior distribution.

[0061] In a decoupling framework, if causal and non-causal features overlap, the classification network will still utilize domain interference information as a shortcut. Therefore, given the samples and labels, it is necessary to force... and Strict conditional independence is satisfied. This invention uses the β-VAE lower bound of evidence (ELBO) variational objective for constraint, which, while enforcing conditional independence between the two, ensures the high-fidelity reconstruction capability of the signal by the feature.

[0062] S4. Input the sensor signals of the mechanical equipment to be diagnosed into the trained cross-domain fault diagnosis model, and output the fault diagnosis results of the mechanical equipment to be diagnosed.

[0063] To evaluate the effectiveness of the diagnostic method proposed in this invention, this embodiment also conducted multiple comparative tests and ablation experiments.

[0064] A. Domain Generalization Comparison Experiment The experimental data used in this embodiment comes from a bearing failure testing platform. Figure 2 The bearing fault location was marked, and a structural diagram of the test bench was shown. This test bench consists of a drive motor, a loading motor, rolling bearings, a flywheel, a gearbox, and various sensors, including nine types of sensors such as acceleration sensors, pressure sensors, and displacement sensors. The vibration data collected in the experiment were all from a 6205 type deep groove ball bearing. In the bearing dataset, "Nomal" indicates the bearing's normal condition, "OF" indicates an outer ring fault, "IF" indicates an inner ring fault, and "BF" indicates a ball fault. In addition, the dataset also includes data on operating conditions, such as speed and temperature. It should be noted that the bearing damage recorded in this dataset is 0.2 mm.

[0065] To preserve all the complex details and features of the original signal, this embodiment does not perform any form of data cleaning, denoising, normalization, or standardization on the original data, and uses two vibration acceleration sensors in different directions.

[0066] Figure 6 This displays some of the raw vibration signals under different health conditions, with the horizontal axis representing time and the vertical axis representing the amplitude of each signal. Specifically, the first and second rows show the signals collected by different sensors, namely sensor 1 and sensor 2.

[0067] Evaluation Criteria: To quantitatively compare the performance of different models in intelligent fault diagnosis tasks, this embodiment selects three core evaluation metrics: accuracy, precision, and F1 score. ; ; ; In the formula, TP, FP, TN and FN represent true positive, false positive, true negative and false negative, respectively.

[0068] This embodiment compares the diagnostic method proposed in this invention with five popular methods, including CCDG, MDGP, GCN, CDDG, and DDDG.

[0069] (1) Dataset Construction The speed-crossing dataset contains five different rotational speeds: 500 rpm, 625 rpm, 750 rpm, 875 rpm, and 1000 rpm, labeled 1, 2, 3, 4, and 5, respectively, while all other experimental conditions remain constant. Each rotational speed represents a different domain, and each domain contains four bearing conditions: normal, inner race fault, outer race fault, and ball fault, as shown in Table 1. In this experimental setup, this embodiment follows the "leave one domain out" rule, that is, one domain is considered as the target domain, and the other four domains are considered as source domains, thus forming a total of five domain generalization tasks (denoted as TR n, where n represents the target domain number).

[0070] Table 1: Details of different domains

[0071] In Table 1, DR n represents different domains.

[0072] (2) Comparison results Table 2 Figure 7 and Figure 8 The diagnostic performance of the proposed method and other fault diagnosis methods in different generalized tasks is demonstrated, and the average accuracy of each method is calculated.

[0073] Table 2: Comparative Analysis of the Accuracy of Diagnostic Results between the Method Proposed in This Invention and Other Methods

[0074] Overall, the proposed method outperforms all compared algorithms across TR1 to TR4 tasks and in terms of average accuracy, demonstrating its superior generalization ability for fault diagnosis in unknown domain scenarios. Experimental data also show that the model's diagnostic accuracy on TR2, TR3, and TR4 tasks is significantly higher than on the other generalization tasks, indicating a higher similarity between the target domain data distribution and the source domain distribution for these three tasks. In some cases, the target domain distribution even falls within the source domain distribution range, effectively reducing the distribution deviation between the source and target domains.

[0075] Specifically, domain generalization algorithms such as MDGP and CCDG exhibit poor overall diagnostic performance, particularly in the TR5 task where the target domain conditions differ significantly from the source domain distribution. This indicates that simply constraining globally invariant features is insufficient to effectively address the complex and variable operating conditions in bearing fault diagnosis scenarios. While algorithms like CDDG, DDDG, and CCN optimize model performance through methods such as aggregating causal factors and de-entanglement of multi-scale features, the performance improvement is limited, suggesting that these methods still fail to fully characterize the unique properties of both causal and non-causal fault features. Furthermore, all of these algorithms are based on single-sensor data, failing to uncover the complementary information contained in multi-sensor signals, further limiting the model's generalization level.

[0076] Figure 9 The confusion matrix of diagnostic results from different methods is shown.

[0077] In summary, the diagnostic method proposed in this invention effectively alleviates the problems of dispersed intra-class features and blurred inter-class feature boundaries, and demonstrates excellent classification accuracy and generalization performance in bearing fault diagnosis, thereby reducing the misdiagnosis rate of unseen faults.

[0078] B. Ablation test To quantitatively analyze the functional roles and contributions of each core module in the proposed framework, this paper conducts systematic ablation experiments. The experiments employ a single-module removal verification method. While ensuring that the remaining network structures remain completely consistent with the original model, the following ablation control groups were set up: removing the Mamba fusion module; replacing the original fusion structure with a simplified element-wise addition fusion module (denoted as "addition"); removing the causal contrastive unentanglement loss function; and setting the hyperparameter corresponding to the exchange reconstruction loss to 0 to eliminate the constraint effect of this loss.

[0079] This paper also constructs a baseline model that eliminates all additional constraints and uses only feature stitching to fuse multi-sensor signals. In addition, a single-sensor input control group is set up, selecting the sensor signal with the highest average diagnostic accuracy as the model input. All ablation experiments follow the data configuration standards specified in the dataset construction section.

[0080] As can be seen from Table 3 and the ablation experiment results, after removing any one core module, the average diagnostic accuracy of the model decreased to varying degrees compared to the complete architecture, proving that each module in the framework of this paper can positively improve the domain generalization fault diagnosis performance.

[0081] Table 3: Ablation Study Results of Different Modules

[0082] Analysis of Mamba Fusion Module Ablation Results The baseline model exhibits the lowest equilibrium accuracy, indicating that relying solely on sensor signal stitching and shallow latent space constraints is insufficient to meet the domain-generalized fault diagnosis requirements in complex scenarios. Compared to multi-sensor input schemes, the diagnostic performance of single-sensor mode is significantly lower, demonstrating that relying solely on single sensor information makes it difficult to achieve stable and reliable fault identification in unknown domains. Furthermore, direct stitching of multi-sensor signals can lead to negative fusion issues.

[0083] Compared to the simple fusion model that adds elements one by one, the multi-sensor fusion module based on the Mamba architecture can fully explore the bearing health status features carried in multi-source signals, effectively avoid the drawbacks of negative fusion in the signal fusion process, and further improve the generalization accuracy and diagnostic effect of the model in unknown domain scenarios.

[0084] Analysis of the results of causal comparison, entanglement resolution, and ablation Causal comparison loss With exchange and reconstruction loss All of these methods significantly enhance the model's feature deentanglement capability and domain generalization performance. The causal contrast loss contributes more significantly to the overall model; removing this loss results in a greater decrease in model accuracy, demonstrating that fault causal feature deentanglement is a core element of domain-generalized fault diagnosis. The commutative reconstruction loss can further strengthen the model's feature deentanglement effect, guiding the model to learn fault features with higher discriminative power, thereby optimizing the final diagnostic performance.

[0085] Example 2 This embodiment provides a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling as described in Embodiment 1.

[0086] like Figure 12 As shown, the computer terminal provided in this embodiment includes: at least one processor 101, and a memory 102 connected to at least one processor 101. This embodiment does not limit the specific connection medium between the processor 101 and the memory 102. Figure 12 The example shown is the connection between processor 101 and memory 102 via bus 100. Bus 100 is... Figure 12 The connections between other components are shown in bold lines and are for illustrative purposes only, not as limiting information. Bus 100 can be divided into address bus, data bus, control bus, etc., for ease of representation. Figure 12 The bus is represented by a single thick line, but this does not indicate that there is only one bus or one type of bus. Alternatively, the processor 101 may also be called a controller; there is no restriction on the name.

[0087] In this embodiment, the memory 102 stores instructions that can be executed by at least one processor 101. The at least one processor 101 can execute the aforementioned method by executing the instructions stored in the memory 102.

[0088] The processor 101 is the control center of the device. It can connect to various parts of the control device through various interfaces and lines. By running or executing instructions stored in memory 102 and calling data stored in memory 102, the processor can perform various functions and process data, thereby monitoring the device as a whole.

[0089] In one possible design, processor 101 may include one or more processing units. Processor 101 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into processor 101. In some embodiments, processor 101 and memory 102 may be implemented on the same chip; in some embodiments, they may also be implemented on separate chips.

[0090] Processor 101 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit, field-programmable gate array or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling disclosed in Embodiment 1 can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules in processor 101.

[0091] Memory 102, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 102 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage, magnetic disk, optical disk, etc. Memory 102 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In this embodiment, memory 102 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0092] By designing and programming the processor 101, the code corresponding to the device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling described in the foregoing embodiments can be embedded into the chip, thereby enabling the chip to execute the code during runtime. Figure 1 The steps of the device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling are shown. How to design and program the processor 101 is a technique well-known to those skilled in the art and will not be described further here.

[0093] Example 3 This embodiment provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the steps of the device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling as described in Embodiment 1.

[0094] The computer-readable storage medium may include flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the storage medium may be an internal storage unit of a computer device, such as the hard disk or memory of the computer device. In other embodiments, the storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., provided on the computer device. Of course, the storage medium may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0095] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for cross-domain fault diagnosis of equipment based on multi-sensor and causal contrast decoupling, characterized in that, include: S1. Construct a cross-domain fault diagnosis model, which includes a multi-channel information dynamic fusion module based on the Mamba model, a causal comparison decoupling module based on the structural causal model, and a shared classifier; S2. Acquire raw monitoring signals collected by various sensors of mechanical equipment under different operating conditions, and construct a training set accordingly; S3. Train the cross-domain fault diagnosis model using the training set; The multi-channel information dynamic fusion module is used to input the original monitoring signal, dynamically perceive the differences between sensors and adaptively extract detailed features to construct a global fusion feature map that aggregates multi-channel information and retains specific channel fault features. The causal comparison decoupling module is used to decouple the global fusion feature map into fault causal features containing core fault information and non-causal features containing specific domain interference information. The shared classifier is used to predict the fault category based on the decoupled fault causal features. During the model training phase, a joint objective function is constructed based on causal comparison loss, exchange reconstruction loss, conditional independence loss, and cross-entropy loss. The cross-domain fault diagnosis model is iteratively optimized by comparing similarity and constraining feature exchange reconstruction of sample pairs containing the same fault across domains. S4. Input the sensor signals of the mechanical equipment to be diagnosed into the trained cross-domain fault diagnosis model, and output the fault diagnosis results of the mechanical equipment to be diagnosed.

2. The multi-sensor based and causally contrast-decoupled device cross-domain fault diagnosis method according to claim 1, characterized in that, The multi-channel information dynamic fusion module includes a dynamic information enhancement unit, the processing steps of which include: Convolutional layers are used to process raw signals from different sensors to obtain enhanced initial features; The enhanced initial features from different sensors are subtracted element-wise to generate differential features, and the differential weights between the initial features are calculated using global average pooling and nonlinear activation functions. The difference weights are applied to the initial multi-scale feature fusion result and the original sensor signal, respectively, to generate enhanced feature outputs for subsequent processing.

3. The multi-sensor based and causally contrast-decoupled device cross-domain fault diagnosis method according to claim 2, characterized in that, The formula for calculating the enhanced feature output is as follows: In the formula, and These represent the raw signals from two different sensors; and They represent Corresponding enhanced feature output and The corresponding enhanced feature output; and This represents the enhanced initial features obtained through the convolutional layer; This represents the Sigmoid activation function; This indicates a global average pooling operation; This represents the preliminary feature results of multi-sensor signal sample fusion; This indicates element-wise addition. This indicates element-wise multiplication.

4. The cross-domain fault diagnosis method for equipment based on multi-sensor and causal comparison decoupling as described in claim 2, characterized in that, The multi-channel information dynamic fusion module also includes a Mamba fusion unit, whose processing steps include: The enhanced feature output of the dynamic information enhancement unit is subjected to a depthwise separable convolution operation, and the results of different channels are mixed to generate an initial mixed feature. The initial mixed features are input into the Mamba layer to capture long-range spatial dependencies, and then an efficient channel attention operation is performed to reduce channel redundancy, resulting in the processed mixed enhancement information. The hybrid enhancement information is then added and aggregated element by element with the original signals from each sensor to obtain the final global fusion feature map.

5. The cross-domain fault diagnosis method for equipment based on multi-sensor and causal comparison decoupling as described in claim 4, characterized in that, The calculation formulas for the hybrid enhancement information and the global fusion feature map are as follows: In the formula, This represents the hybrid enhancement information; This represents the global fused feature map; and These represent the raw signals from two different sensors. This indicates element-wise addition; This indicates efficient channel attention operations; Indicates a linear layer; Indicates the Mamba layer; This represents the initial mixing characteristics.

6. The multi-sensor based and causally contrast-decoupled device cross-domain fault diagnosis method according to claim 1, characterized in that, The calculation process for the causal contrast loss includes: For any given sample, samples with consistent fault labels but different domain labels are selected from the global sample space to construct fault sample pairs with the same fault but different domain labels. ;in, and These represent the global fusion feature maps extracted from the original monitoring signals of the two samples, respectively. express The domain label of the corresponding sample; express The domain label of the corresponding sample; The fault label indicates the sample; The encoder is used to select the fault sample pair. and Encode them separately as combinations of latent features and ;in, and For fault causal features that only contain cross-domain invariant attributes; and To retain only non-causal features that are sensitive to specific domains; Construct a causal contrastive loss function, expressed as follows: wherein denotes the causal contrast loss; denotes the KL divergence; denotes the posterior probability distribution induced by the encoder; denotes the Gaussian kernel function; is a fixed combination of two distributions, is a set of different distribution combinations.

7. The multi-sensor based and causally contrast-decoupled device cross-domain fault diagnosis method according to claim 6, characterized in that, The calculation process for the exchange reconstruction loss includes: Perform a cross-swapping operation on the fault causal features of the fault sample pairs, and transfer the fault causal features... Non-causal characteristics Combined into the first exchange feature pair , to identify the causal characteristics of the fault Non-causal characteristics Combining into a second exchange feature pair ; The first exchanged feature pair and the second exchanged feature pair are respectively input into the corresponding decoder to generate a reconstructed signal and a reconstructed label for fitting the original signal and the domain label; By constraining the reconstruction errors between the original signal, original domain label, and reconstructed signal, reconstructed label, a commutative reconstruction loss function is constructed, expressed as follows: In the formula, This represents the exchange reconstruction loss; This represents the conditional probability distribution of the decoder used to reconstruct the original signal; This represents the conditional probability distribution of the decoder used to reconstruct the generated domain labels.

8. The cross-domain fault diagnosis method for equipment based on multi-sensor and causal comparison decoupling as described in claim 6, characterized in that, The formula for calculating the conditional independence loss is as follows: In the formula, This represents the overall optimization objective of the VAE model, i.e., the lower bound of evidence, where Parameters representing the encoder, Parameters representing the decoder, It is used for balance and The ratio between the two; express The reconstruction terms in the code are used to ensure that the decoupled features contain sufficient information. It represents the divergence value between latent variables, ensuring the degree of similarity between different latent variables; Expressing expectations; Indicates the divergence value; An approximate posterior distribution representing the causal characteristics of the fault; Represents the approximate posterior distribution of non-causal features; Indicating the causal characteristics of the fault Prior distribution; Indicating non-causal characteristics Prior distribution; This represents the conditional probability distribution of the decoder used to reconstruct the original signal.

9. A computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the device cross-domain fault diagnosis method based on multi-sensor and causal comparison decoupling as described in any one of claims 1 to 8.

Images