Adaptive deep transfer fault diagnosis method and system, apparatus and medium

Abstract

Disclosed in the present invention are an adaptive deep transfer fault diagnosis method and system, an apparatus and a medium. The method comprises the following steps: S1: collecting vibration acceleration signals of industrial equipment under different working conditions, and dividing same into a source domain data set and a target domain data set; S2: building a self-tuning universal domain adaptive fault diagnosis model, which comprises a shared feature extractor, a known classifier and a plurality of unknown classifiers; S3: separately calculating a classification loss of known faults of the source domain, a discriminative loss of the plurality of unknown classifiers, a target domain soft consistency regularization loss and an information entropy maximization loss; S4: introducing a dynamic weighting strategy based on model uncertainty assessment to optimize the model parameters; and S5: using the model for diagnosis. The present invention can fully mine valid information in data, can establish reliable class decision boundaries, and in addition, uses the self-tuning dynamic update strategy to adjust weightings corresponding to different loss functions, thus allowing for quick generalization of the model to different industrial diagnosis scenarios.

Description

Adaptive deep migration fault diagnosis method, system, device and medium Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to an adaptive deep migration fault diagnosis method, system, device and medium for complex industrial scenarios. Background Art

[0002] Rotating machinery plays a vital role in the operation of mechanical systems and equipment, and has found widespread application in complex industrial scenarios, including helicopter transmissions, automotive transmissions, and marine engines. As mechanical equipment becomes increasingly complex and intelligent, automatically identifying equipment faults is crucial for improving operational reliability, reducing downtime, and increasing productivity. To address this issue, intelligent fault diagnosis technology has attracted significant attention from practitioners, researchers, and engineers. Its key concept is to utilize advanced technologies and algorithms to detect and diagnose faults during equipment operation. Therefore, applying intelligent fault diagnosis algorithms to mechanical equipment in industrial scenarios is crucial for ensuring safe and stable operation.

[0003] Among numerous methods, deep learning has attracted increasing attention due to its outstanding performance in automatically learning feature representation from data. Deep learning-driven fault diagnosis methods eliminate the need for tedious signal mechanism analysis and instead automatically extract effective features from data for fault classification through neural networks. This has driven the development and application of intelligent fault diagnosis methods in the "big data era." However, several drawbacks of deep learning-based methods have severely hindered their widespread deployment in practical applications. First, deep learning model training requires a large number of labeled samples to ensure sufficient training until convergence. However, in real industrial scenarios, equipment operates normally most of the time, and failures are rare events with extremely low probability. Furthermore, even when a large number of fault samples are generated, a labor-intensive and time-consuming labeling process is required. Furthermore, deep learning-based models have limited generalization capabilities when operating conditions change. A model trained under one operating condition must be retrained from scratch to achieve satisfactory performance under another, significantly increasing training costs and reducing diagnostic efficiency.

[0004] In the Chinese patent "A method for classifying and identifying mechanical faults under variable working conditions and partial migration conditions (CN116702002A)", Fei Zhongyang et al. used the improved local maximum mean difference as an indicator to measure the distribution difference between the subdomains of the source domain and the target domain, to solve the problem of model performance degradation under variable working conditions and inconsistent label distribution. However, the above invention only solves the domain adaptation problem of the model in the variable working condition scenario where the number of target domain faults is less than that of the source domain, lacks the ability to diagnose new faults, and cannot be generalized well to the diagnosis scenario where unknown new faults exist in the target domain. Therefore, it is necessary to study a general fault diagnosis model with strong generalization, which uses the discriminable feature knowledge of equipment faults learned and discovered from historical data, and applies it to uncertain industrial diagnosis scenarios to achieve rapid and accurate identification of industrial equipment faults, so as to adapt to the challenges brought by the dynamic changes and uncertainty of data in industrial scenarios.

[0005] Summary of the Invention

[0006] To solve the problems existing in the prior art, the present invention provides an adaptive deep migration fault diagnosis method for complex industrial scenarios. The method designs a feature extractor based on a one-dimensional convolutional neural network to extract representative features from the source domain and the target domain. Known fault categories are classified by constructing known classifiers, and multiple unknown classifiers are constructed to establish multiple decision boundaries, thereby adaptively separating known and unknown class samples. To further optimize the performance of the diagnostic model, a self-adjusting training strategy is designed to optimize the proposed diagnostic model, and a weight dynamic strategy for model uncertainty evaluation is used to adaptively assign weights to each loss function.

[0007] To achieve the purpose of the present invention, the present invention provides an adaptive deep migration fault diagnosis method for complex industrial scenarios, the specific steps of which are as follows:

[0008] S1: Collect vibration acceleration signals of industrial equipment under different working conditions and failure modes, and divide them into source domain dataset and target domain dataset;

[0009] S2: Build a self-tuning universal domain adaptive fault diagnosis model, including a shared feature extractor, a known classifier, and multiple unknown classifiers;

[0010] S3: Combine four loss functions for training: classification loss of known faults in the source domain, discrimination loss of multiple unknown classifiers, soft consistency regularization loss in the target domain, and information entropy maximization loss;

[0011] S4: A weighted dynamic strategy based on model uncertainty evaluation adaptively assigns weights to each loss function and optimizes the diagnosis model parameters based on the dynamically updated loss function, resulting in a self-adjusting universal domain adaptive fault diagnosis model for diagnosis;

[0012] S5: Input the data to be tested into the trained self-adjusting general domain adaptive fault diagnosis model and output the diagnosis results.

[0013] After the model is trained, its performance is tested and evaluated on unlabeled fault samples from the target domain. During the model testing phase, the decision boundary automatically constructed by the unknown classifier for each fault category is used to determine whether the sample belongs to a known or unknown class. For the known class samples, the known fault category is further determined using the known classifier, and the model's diagnostic accuracy is calculated.

[0014] Furthermore, in step S1, the collected samples are divided into source domain datasets and target domain dataset and in, represents the source domain label sample, represents the unlabeled samples in the target domain, Representative source domain samples The corresponding label, L s Represents the number of fault categories in the source domain, n s Represents the number of source domain label samples, n t Represents the number of unlabeled samples in the target domain.

[0015] Furthermore, in step S2, the constructed feature extractor G f It is designed based on a one-dimensional convolutional neural network, including a convolution layer C, a batch normalization layer B, a pooling layer P, and a fully connected layer F. The convolution layer is used to perform convolution operations between the convolution kernel and the input sample to extract deep features with time-shift invariance from the time-domain vibration data; the batch normalization layer is used to normalize the input data of each batch; the pooling layer uses a maximum pooling method to reduce the dimension of the features and the number of trainable parameters of the model; finally, the fully connected layer uses a multi-layer perceptron to map cross-domain data to a shared feature space with label information. Known classifier C k It consists of multiple batch normalization layers B and fully connected layers F. The number of neurons in the last fully connected layer is the number of source domain fault categories L s . Multiple unknown class discriminators C LS ∈{c1,c2,…,c Ls It consists of multiple binary classifiers, each of which includes a batch normalization layer B and a fully connected layer F with two neurons.

[0016] Furthermore, in step S3, four loss functions are combined for training: the classification loss of known faults in the source domain, the discrimination loss of multiple unknown classifiers, the soft consistency regularization loss in the target domain, and the information entropy maximization loss. First, the source domain classification loss is used to help the model correctly distinguish known fault categories; second, the discrimination loss of multiple unknown classifiers is designed to facilitate the generation of corresponding decision boundaries for each fault category by multiple unknown classifiers; finally, the soft consistency regularization loss and information entropy maximization loss are introduced to support the model's adaptation to the target domain. The specific implementation steps are as follows:

[0017] S31. In order to accurately determine the category of known faults in the source domain, the standard cross entropy loss function is used to calculate the known fault classification loss L1 of the diagnostic model's prediction output for source domain samples:

[0018] Where, Represents a known classifier C k The source domain samples The predicted output.

[0019] S32. To construct multiple decision boundaries capable of distinguishing known and unknown samples, a corresponding unknown discriminator is designed for each fault category. Each unknown discriminator outputs a two-dimensional vector, each vector representing the probability of being known and unknown. To ensure that these unknown classifiers can effectively establish accurate boundaries to distinguish unknown samples of the corresponding fault category, source samples are used for training to construct the category decision boundary. The discriminant loss L2 of multiple unknown classifiers can be calculated as follows:

[0020] in, Represents source domain samples Belongs to fault category L s The probability output of Represents source domain samples Not a fault category The probability output of .

[0021] S33. Considering that samples in the target domain are usually unlabeled, some effective methods choose to generate pseudo labels for these unlabeled samples to supervise the training of the model. However, some unknown samples are usually far away from the labeled known samples, which makes the pseudo labels inaccurate and misleads the learning of the neural network. To solve this problem, data augmentation technology is introduced to propagate the correct information extraction by minimizing the distance between the prediction results of two different enhanced samples. Specifically, we randomly add noise and reverse the vibration signal of length 2048 to form sample pairs for model training. The sample pairs should have consistent outputs. Therefore, the soft consistency regularization loss L3 is defined as:

[0022] in, and represents the sample pairs generated by unlabeled target domain samples through different data enhancement methods, Represents the model's response to the enhanced sample The predicted output result is Represents the model's response to the enhanced sample The predicted output results.

[0023] S34. The concept of information in machine learning is often defined in terms of entropy, a measure of uncertainty or disorder in a system. Entropy is widely used as a guiding principle to help models extract useful information from large datasets. By introducing the entropy maximization loss, by minimizing the signal's entropy, more representative features can be extracted from the target domain, allowing the classifier to automatically adapt to the target domain. The entropy maximization loss L4 is used to train unknown classifiers and is calculated as follows:

[0024] in, Represents the jth classifier for the unlabeled target domain sample The predicted class probability output.

[0025] Furthermore, in step S4, manually adjusting the weights of each loss function is a very time-consuming and challenging task. Multi-task learning is a machine learning method in which a single model is trained to perform multiple tasks simultaneously, and the model shares some or all parameters between multiple tasks, allowing it to learn from the relationships and similarities between them (12). The optimization process of the model is similar to multi-task learning, and different learning objectives can be regarded as different tasks. In general, there are two types of uncertainty in deep model construction: epistemic uncertainty (caused by insufficient data) and arbitrary uncertainty (caused by data information limitations). Arbitrary uncertainty can be divided into homoscedastic uncertainty (data-dependent) and heteroscedastic uncertainty (task-dependent). Therefore, a weight dynamic strategy based on model uncertainty evaluation is designed, and mean square error uncertainty is introduced to help automatically determine the weight of each target. Specifically, the following steps are included:

[0026] S41. Calculate the log-likelihood of each loss function:

[0027] in, Represents the classification probability output of the cth element in the high-dimensional feature representation, and softmax represents the calculation The output after the softmax function.

[0028] S42. Calculate the log-likelihood of each loss function:

[0029] Where σ n ={σ1,σ2,σ3,σ4} are the dynamic factor parameters that can be learned by the four models. σ1 represents the dynamic factor corresponding to the loss function L1, σ2 represents the dynamic factor corresponding to the loss function L2, σ3 represents the dynamic factor corresponding to the loss function L3, and σ4 represents the dynamic factor corresponding to the loss function L4.

[0030] S43, dynamic factor σ that is continuously updated according to the model learning process n ={σ1,σ2,σ3,σ4} updates the loss function of model training, so that the model can continuously adjust the training weights on different loss functions, ensuring that the model can adaptively update the loss function according to the training process and converge faster. The dynamic update method of the loss function is as follows:

[0031] Where, L overall Represents the overall loss function of the model.

[0032] The present invention also provides an adaptive deep migration fault diagnosis system for complex industrial scenarios, which is used to implement the above method and includes the following modules:

[0033] The dataset construction module is used to collect vibration acceleration signals of industrial equipment under different working conditions and different failure modes, and divide them into source domain datasets and target domain datasets;

[0034] A model building module, used to build a self-tuning universal domain adaptive fault diagnosis model, including a shared feature extractor, a known classifier, and multiple unknown classifiers;

[0035] The training module is used to combine four loss functions for training: classification loss of known faults in the source domain, discrimination loss of multiple unknown classifiers, soft consistency regularization loss in the target domain, and information entropy maximization loss;

[0036] A weight assignment and update module is used to adaptively assign weights to each loss function based on a weight dynamic strategy for evaluating model uncertainty, and optimize the diagnostic model parameters based on the dynamically updated loss function to obtain a self-regulating universal domain adaptive fault diagnosis model for diagnosis;

[0037] The diagnosis module is used to input the test data into the trained self-adjusting general domain adaptive fault diagnosis model and output the diagnosis results.

[0038] The present invention also provides a device comprising:

[0039] at least one processor;

[0040] at least one memory for storing at least one program;

[0041] When the at least one program is executed by the at least one processor, the at least one processor implements the aforementioned method.

[0042] The present invention also provides a computer-readable storage medium storing a program executable by a processor. When the program is executed by the processor, it is used to perform the above method.

[0043] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0044] 1. Existing deep learning-driven fault diagnosis methods are typically designed for a single diagnostic scenario. The trained models are difficult to generalize to other industrial scenarios (for example, when the target domain contains new fault types that the model has not seen), which does not meet the requirements for efficient fault diagnosis in industrial applications. The present invention designs a self-adjusting intelligent fault diagnosis method that exhibits excellent intelligent fault diagnosis performance in the face of specific diagnostic scenario uncertainties. This provides a promising solution for the practical deployment of intelligent diagnostic methods in industrial environments and will help improve industrial operational efficiency and equipment reliability.

[0045] 2. The weight dynamic strategy based on model uncertainty assessment designed in the present invention introduces mean square error uncertainty, so that the model can adaptively adjust the weight distribution of the loss function, thereby achieving efficient optimization of the model and ensuring that the model can be rapidly generalized in various industrial diagnosis scenarios, providing a feasible and effective solution for model self-adjustment optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] FIG1 is an overall flow chart of an adaptive deep migration fault diagnosis method for complex industrial scenarios provided by an embodiment of the present invention.

[0047] FIG2 is a flow chart of a model optimization strategy based on a dynamically updated loss function in an embodiment of the present invention.

[0048] FIG3 is a schematic diagram of the diagnostic accuracy of the method proposed in an embodiment of the present invention on known fault categories.

[0049] FIG4 is a schematic diagram showing the diagnostic accuracy of the method proposed in an embodiment of the present invention for unknown fault categories. DETAILED DESCRIPTION

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] Referring to FIG1 , the present invention provides an adaptive deep migration fault diagnosis method for complex industrial scenarios, which specifically includes the following steps:

[0052] Step 1: Collect vibration acceleration signals of industrial equipment under different working conditions and different failure modes, and divide the collected samples into source domain data sets and target domain dataset and in, represents the source domain label sample, represents the unlabeled samples in the target domain, Representative source domain samples The corresponding label, L s Represents the number of fault categories in the source domain, n s Represents the number of source domain label samples, n t Represents the number of unlabeled samples in the target domain.

[0053] Step 2: Build a self-tuning universal domain adaptive fault diagnosis model, including a shared feature extractor, a known classifier, and multiple unknown classifiers.

[0054] The constructed feature extractor G f It is designed based on a one-dimensional convolutional neural network, including a convolutional layer C, a batch normalization layer B, a pooling layer P, and a fully connected layer F. The convolutional layer is used to perform convolution operations between the convolution kernel and the input samples to extract deep features with time-shift invariance from the time-domain vibration data; the batch normalization layer is used to normalize the input data of each batch, which helps to maintain the stable distribution of the input of each layer, thereby accelerating the convergence of the network; the pooling layer adopts the maximum pooling method to reduce the dimension of the features and reduce the number of trainable parameters of the model; finally, the fully connected layer uses a multi-layer perceptron to map cross-domain data to a shared feature space with label information.

[0055] Known classifier C kIt includes multiple batch normalization layers B and fully connected layers F. The batch normalization layer B helps to accelerate the training of the neural network. The function of the fully connected layer F is to connect all the nodes of the input layer or the previous layer with each node of the current layer. Each connection has a weight, so that the complex relationship of the input data can be learned. The number of neurons in the last fully connected layer of the fully connected layer F is the number of source domain fault categories L. s . Multiple unknown class discriminators C LS ∈{c1,c2,…,c Ls}It includes multiple binary classifiers for establishing independent discrimination boundaries for each fault category, and each binary classifier includes a batch normalization layer B and a fully connected layer F with two neurons.

[0056] Step 3: Calculate the classification loss of known faults in the source domain, the discrimination loss of multiple unknown classifiers, the soft consistency regularization loss in the target domain, and the information entropy maximization loss respectively.

[0057] First, the source domain classification loss is used to help the model correctly distinguish known fault categories. Second, the discriminant loss of multiple unknown classifiers is designed to facilitate the generation of corresponding decision boundaries for each fault category. Finally, the soft consistency regularization loss and information entropy maximization loss are introduced to support the model's adaptation to the target domain. The specific implementation steps are as follows:

[0058] S31. In order to accurately determine the category of known faults in the source domain, the standard cross entropy loss function is used to calculate the known fault classification loss L1 of the self-adjusting universal domain adaptive fault diagnosis model for the source domain samples:

[0059] Where, Represents a known classifier C k For source domain samples The predicted output.

[0060] S32. To construct multiple decision boundaries capable of distinguishing known and unknown samples, a corresponding unknown discriminator is designed for each fault category. Each unknown discriminator outputs a two-dimensional vector, each vector representing the probability of being known and unknown. To ensure that these unknown classifiers can effectively establish accurate boundaries to distinguish unknown samples of the corresponding fault category, source samples are used for training to construct the category decision boundary. The discriminant loss L2 of multiple unknown classifiers can be calculated as follows:

[0061] in, Represents source domain samples Belongs to fault category L s The probability output of Represents source domain samples Not a fault category The probability output of .

[0062] S33. Considering that samples in the target domain are usually unlabeled, some effective methods choose to generate pseudo labels for these unlabeled samples to supervise the training of the model. However, some unknown samples are usually far away from the labeled known samples, which makes the pseudo labels inaccurate and misleads the learning of the neural network. In order to solve this problem, the present invention introduces data enhancement technology to propagate correct information extraction by minimizing the distance between the prediction results of two different enhanced samples. Specifically, the present invention randomly adds noise and inverts the entire vibration signal (it can also be used for vibration signals of other lengths, such as inverting vibration signals of length 2048) to form sample pairs for model training. The sample pairs should have consistent outputs. Therefore, the soft consistency regularization loss L3 is defined as:

[0063] in, and represents the sample pairs generated by unlabeled target domain samples through different data enhancement methods, Represents the model's response to the enhanced sample The predicted output result is Represents the model's response to the enhanced sample The predicted output results.

[0064] S34. The concept of information in machine learning is often defined in terms of entropy, a measure of uncertainty or disorder in a system. Entropy is widely used as a guiding principle to help models extract useful information from large datasets. By introducing the entropy maximization loss, by minimizing the signal's entropy, more representative features can be extracted from the target domain, allowing the classifier to automatically adapt to the target domain. The entropy maximization loss L4 is used to train unknown classifiers and is calculated as follows:

[0065] in, Represents the jth unknown class classifier for the unlabeled target domain sample The predicted class probability output.

[0066] Step 4: A weighted dynamic strategy based on model uncertainty evaluation adaptively assigns weights to each loss function and optimizes the diagnostic model parameters based on the dynamically updated loss function.

[0067] Manually adjusting the weights of each loss function is a very time-consuming and challenging task. Multi-task learning is a machine learning method in which a single model is trained to perform multiple tasks simultaneously, and the model shares some or all parameters between multiple tasks, allowing it to learn from the relationships and similarities between them. The optimization process of the model is similar to multi-task learning, and different learning objectives can be regarded as different tasks. In general, there are two types of uncertainty in deep model construction: epistemic uncertainty (caused by insufficient data) and arbitrary uncertainty (caused by data information limitations). Arbitrary uncertainty can be divided into homoscedastic uncertainty (data-dependent) and heteroscedastic uncertainty (task-dependent). Therefore, a weight dynamic strategy based on model uncertainty evaluation is designed, and mean square error uncertainty is introduced to help automatically determine the weight of each objective. Specifically, the following steps are included:

[0068] S51. Calculate the log-likelihood of each loss function:

[0069] in, Represents the classification probability output of the cth element in the high-dimensional feature representation, and softmax represents the calculation The output after the softmax function.

[0070] S52. Calculate the log-likelihood of each loss function:

[0071] Where σ n ={σ1,σ2,σ3,σ4} are the dynamic factor parameters that can be learned by the four loss functions. σ1 represents the dynamic factor corresponding to the known fault classification loss function L1, σ2 represents the dynamic factor corresponding to the discriminant loss function L2 of multiple unknown classifiers, σ3 represents the dynamic factor corresponding to the soft consistency regularization loss function L3, and σ4 represents the dynamic factor corresponding to the information entropy maximization loss function L4. Represents the classification probability output of the cth element in the high-dimensional feature representation.

[0072] S53, according to the dynamic factor parameters that are continuously updated during the model learning process to σ n ={σ1,σ2,σ3,σ4} updates the loss function of model training, so that the model can continuously adjust the training weights on different loss functions, ensuring that the model can adaptively update the loss function according to the training process and converge faster.

[0073] The dynamic update method of the loss function is as follows:

[0074] Where, L overall Represents the overall loss function of the model.

[0075] The trained fault diagnosis model is tested and evaluated on unlabeled fault samples from the target domain. During the model testing phase, the decision boundary automatically constructed by the unknown classifier for each fault category is used to determine whether the sample belongs to a known or unknown class. For the known class samples, the known fault category is further determined using the known classifier, and the model's diagnostic accuracy is calculated.

[0076] Step 5: Input the test data into the trained self-adjusting general domain adaptive fault diagnosis model and output the diagnosis results.

[0077] In some embodiments of the present invention, an adaptive deep migration fault diagnosis system for complex industrial scenarios is provided, comprising the following modules:

[0078] The dataset construction module is used to collect vibration acceleration signals of industrial equipment under different working conditions and different failure modes, and divide them into source domain datasets and target domain datasets;

[0079] A model building module, used to build a self-tuning universal domain adaptive fault diagnosis model, including a shared feature extractor, a known classifier, and multiple unknown classifiers;

[0080] The training module is used to combine four loss functions for training: classification loss of known faults in the source domain, discrimination loss of multiple unknown classifiers, soft consistency regularization loss in the target domain, and information entropy maximization loss;

[0081] A weight assignment and update module is used to adaptively assign weights to each loss function based on a weight dynamic strategy for evaluating model uncertainty, and optimize the diagnostic model parameters based on the dynamically updated loss function to obtain a self-regulating universal domain adaptive fault diagnosis model for diagnosis;

[0082] The diagnosis module is used to input the test data into the trained self-adjusting general domain adaptive fault diagnosis model and output the diagnosis results.

[0083] The system of this embodiment can execute an adaptive deep migration fault diagnosis method for complex industrial scenarios provided by the method embodiment of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0084] In some embodiments of the present invention, there is provided an apparatus comprising:

[0085] at least one processor;

[0086] at least one memory for storing at least one program;

[0087] When the at least one program is executed by the at least one processor, the at least one processor implements the method provided in the aforementioned embodiment.

[0088] In some embodiments of the present invention, a computer-readable storage medium is provided, in which a processor-executable program is stored, characterized in that the processor-executable program is used to perform the method provided in the aforementioned embodiment when executed by the processor.

[0089] The present invention is further described below with reference to the accompanying drawings and experimental cases.

[0090] In some embodiments of the present invention, multiple fault diagnosis experiments under variable working conditions are designed on a public PU bearing fault dataset (a rolling bearing dataset provided by the University of Paderborn, Germany) to verify the effectiveness of the method proposed in the present invention.

[0091] PU bearing datasets are often used to verify the performance of mechanical fault diagnosis technology. The test bench includes the following modules: test motor, measurement shaft, bearing module, flywheel, and load motor. The experimental data was obtained by installing ball bearings with different damage types in the test module of the test bench. Faulty bearings are divided into two types: artificial damage and actual damage. Seven different types of artificially damaged faulty bearings were selected here, as shown in Table 1. They have different fault locations (inner ring and outer ring), different damage degrees (level 1 and level 2) and different fault injection methods (electroengraving, drilling and electrospark machining). The effectiveness of the method proposed in this invention was tested and evaluated. In the experimental operating conditions, the spindle speed, load torque and bearing radial force are the three main variables. Among them, the spindle runs at two speeds of 1500rpm and 900rpm respectively, the load torque has two working conditions of 0.1Nm and 0.7Nm, and the bearing radial forces are 400N and 1000N respectively. Therefore, there are a total of four different operating conditions.

[0092] Table 1 Main fault categories of the PU bearing fault dataset

[0093] Based on the aforementioned PU bearing fault dataset, we designed a total of 10 transfer diagnosis tasks for different scenarios, including two CDA tasks, two PDA tasks, three ODA tasks, and three OPDA tasks. The detailed diagnostic tasks are shown in Table 2. Each diagnostic task includes two different working conditions, each representing a source domain and a target domain. Among them, "N15_M07_F10" represents the working condition of 1500rpm, load torque of 0.7Nm, and bearing radial force of 1000N; "CDA" represents the diagnosis scenario in which the fault categories of the source domain and the target domain are consistent; "PDA" represents the diagnosis scenario in which there is a private fault category in the source domain; "ODA" represents the diagnosis scenario in which there is an unknown new fault category in the target domain; "ODA(2)" indicates that there are two unknown fault categories in the target domain, which means that fault diagnosis under this setting will be more difficult; "OPDA" indicates the diagnosis scenario in which there is a private fault category in the source domain and the target domain respectively; similarly, "OPDA(2)" indicates that there are two unknown fault categories in the target domain and there is one private fault category in the source domain. This scenario is also designed to verify the robustness of the proposed AUDAN.

[0094] Table 2 Cross-domain fault diagnosis experiment designed based on PU bearing fault dataset

[0095] The Adam optimizer was used as the optimization algorithm to adjust the model parameters, with weight decay set to 0.0005. The learning rates of the feature extractor, classifier, and adaptive weighting module were set to 0.05, 0.001, and 0.0001, respectively. To reduce the influence of random factors on the experimental results, the final results were averaged over five trials.

[0096] In order to further verify the superiority of the method of this embodiment, several classic fault diagnosis algorithms are used for comparison with the method of the present invention:

[0097] Comparison method 1: DACNN, the network consists of a feature extractor, a label predictor, and a domain classifier. The feature extractor structure is consistent with the proposed method, and domain-universal features of the source and target domains are extracted through an adversarial training strategy.

[0098] Comparison method 2: MWDTN, an open set domain adaptive diagnosis algorithm based on multi-source weighted deep transfer network.

[0099] Comparison method 3: DSDAN, a general domain adaptation diagnosis algorithm based on deep self-supervised domain adaptation network.

[0100] Table 3 shows the average accuracy of our method and the comparative methods in different cross-domain diagnosis scenarios. Overall, our method achieves the highest accuracy in most diagnostic tasks and can effectively detect unknown fault types in the target domain. It can be seen that in the CDA scenario (Tasks A1-A2), all methods achieve satisfactory diagnostic performance. This is because CDA diagnostic tasks only require addressing the data distribution discrepancy between the source and target domains, which existing methods can effectively address. By observing the PDA scenario (Tasks A3-A4), we notice a slight decrease in the accuracy of the DACNN. This is attributed to the presence of some private fault types in the source domain in the PDA scenario, which leads to negative transfer. In the ODA and OPDA scenarios, the DACNN method's diagnostic accuracy significantly decreases due to its inability to diagnose unknown fault types in the target domain. For example, in Task A10, the DACNN achieves an accuracy of only 59.27%. Therefore, across all diagnostic tasks, the average accuracy of the DACNN is only 76.89%. The MWDTN, which incorporates an unknown fault type diagnosis mechanism, achieves an accuracy of 85.75% and an h-score of 73.29%. However, MWDTN's performance in OPDA scenarios is less than impressive. This is primarily due to its inability to mitigate negative transfer when both the source and target domains contain private classes. As an effective method for UDA scenarios, DSDAN achieves competitive classification performance across multiple diagnostic tasks. However, its performance degrades when the target domain contains multiple unknown fault types. In contrast, the proposed method achieves a diagnostic performance of 99.83%, demonstrating strong superiority.

[0101] Table 3 Model fault diagnosis accuracy (%) under different cross-domain fault diagnosis scenarios

[0102] The diagnostic accuracy of the present invention for known fault categories and unknown fault categories in different migration diagnosis tasks are shown in Figures 3 and 4, respectively. It can be seen that the present invention can achieve high diagnostic accuracy both for known fault categories that have been learned in the source domain and for unknown fault categories that have newly appeared in the target domain.

[0103] In summary, the present invention aims to address the problem of poor diagnostic accuracy caused by the inability to adapt the source domain training model due to unknown diagnostic scenarios when performing fault diagnosis in actual industrial scenarios. This method designs an adaptive deep migration fault diagnosis method for complex industrial scenarios, which realizes the rapid adaptation and accurate fault diagnosis of the model in multiple different diagnostic tasks, and has high application value in actual industrial production.

[0104] As for the adaptive deep migration fault diagnosis system, device and medium for complex industrial scenarios disclosed in the embodiment, since they correspond to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0105] Finally, it should be noted that although the implementation of the present invention has been described in detail with reference to examples, it is easy for those skilled in the art to understand that any modifications, substitutions and improvements made without departing from the spirit and principles of the present invention as described in the appended claims should be included in the scope of protection of the present invention.

Claims

1. An adaptive deep transfer fault diagnosis method for complex industrial scenarios, characterized in that It includes the following steps: S1: Collect the vibration acceleration signals of industrial equipment in different fault modes under different working conditions, and divide them into a source domain dataset and a target domain dataset; S2: Build a self-regulating general domain adaptive fault diagnosis model, including a shared feature extractor, a known classifier, and multiple unknown classifiers; S3: Combine four loss functions, namely the classification loss of known faults in the source domain, the discrimination loss of multiple unknown classifiers, the target domain soft consistency regularization loss, and the information entropy maximization loss, for training; S4: Adaptively assign weights to each loss function based on the weight dynamic strategy of model uncertainty evaluation, and optimize the diagnostic model parameters based on the dynamically updated loss function to obtain a self-regulating general domain adaptive fault diagnosis model for diagnosis; S5: Input the data to be tested into the trained self-regulating general domain adaptive fault diagnosis model and output the diagnostic results.

2. An adaptive depth transfer fault diagnosis method for complex industrial scenarios according to claim 1, characterized in that In step S1, the collected samples are divided into a source domain dataset and a target domain dataset and where represents the source domain label samples, represents the target domain unlabeled samples, represents the source domain samples the corresponding label, L s represents the number of fault categories existing in the source domain, n s represents the number of source domain label samples, n t represents the number of target domain unlabeled samples.

3. An adaptive deep transfer fault diagnosis method for complex industrial scenarios according to claim 1, characterized in that The feature extractor includes a convolutional layer, a batch normalization layer, a pooling layer, and a fully connected layer. The convolutional layer is used to extract deep features with time shift invariance from the time-domain vibration data. The batch normalization layer is used to normalize the input data of each batch. The pooling layer is used to reduce the dimension of the features. The fully connected layer is used to map the cross-domain data to a shared feature space with label information. The known classifier includes a batch normalization layer and a fully connected layer. The batch normalization layer is used to accelerate the training of the model. The fully connected layer is used to connect all nodes of the input layer or the previous layer to each node of the current layer. The unknown class discriminator includes multiple binary classifiers, and the binary classifier is used to establish an independent discrimination boundary for each fault category.

4. An adaptive depth transfer fault diagnosis method for complex industrial scenarios according to claim 3, characterized in that In step S2, each binary classifier includes a batch normalization layer and a fully connected layer with two neurons.

5. The adaptive depth transfer fault diagnosis method for complex industrial scenarios according to claim 1, wherein In step S3, combine four loss functions, namely the classification loss of known faults in the source domain, the discrimination loss of multiple unknown classifiers, the target domain soft consistency regularization loss, and the information entropy maximization loss, for training. First, use the source domain classification loss to help the model correctly distinguish known fault categories. Secondly, design the discrimination loss of multiple unknown classifiers to facilitate multiple unknown classifiers to generate corresponding decision boundaries for each fault category. Finally, introduce the soft consistency regularization loss and the information entropy maximization loss to support the model to adapt to the target domain.

6. The adaptive deep transfer fault diagnosis method for complex industrial scenarios according to claim 4, wherein The specific steps of step S3 include: S31. Use the standard cross-entropy loss function to calculate the known fault classification loss L1 of the self-regulating general domain adaptive fault diagnosis model for the predicted output of the source domain samples: In the formula, Denote the known classifier C k for source domain samples The predicted output, Represent source domain samples The corresponding label, n s represents the number of source domain label samples; S32. Corresponding an unknown discriminator to each fault category, each unknown discriminator outputs a two-dimensional vector, and each vector represents the possibility of known and unknown respectively. Among them, the source samples are used for training to construct the class decision boundary, and the discrimination loss L2 of multiple unknown classifiers is calculated as follows: Among them, Indicates the source domain samples Belonging to failure category L s Probability output Indicates source domain samples Does not belong to the fault category Probability output; S33. Introduce data augmentation technology to spread correct information extraction by minimizing the distance between the prediction results of two different augmented samples. Therefore, define the soft consistency regularization loss L3 as follows: Among them, And Indicates sample pairs generated from unlabeled target domain samples through different data augmentation methods, The representation model for the enhanced samples The predicted output result, The representation model for the enhanced samples The predicted output result, Denote the unlabeled samples in the target domain, L s Denote the number of fault categories existing in the source domain; S34. Introduce the information entropy maximization loss. By minimizing the information entropy of the signal, more representative features are extracted from the target domain to enable the classifier to automatically adapt to the target domain. The information entropy maximization loss L4 is used to train the unknown classifier, and its calculation formula is as follows: Among them, Denote the j-th unknown class classifier for unlabeled target domain samples Predicted class probability output.

7. An adaptive deep transfer fault diagnosis method for complex industrial scenarios according to any one of claims 1-6, characterized in that, The steps of step S4 include: S41. Calculate the log-likelihood of each loss function: Among them, Represents the classification probability output of the c-th element in the high-dimensional feature representation, and softmax represents the calculation The output after passing through the softmax function Represent the source domain label samples, Representing unlabeled samples in the target domain; S42. Calculate the log-likelihood of each loss function: where, σ n ={σ1, σ2, σ3, σ4} are dynamic factor parameters that can be learned by four loss functions. σ1 represents the dynamic factor corresponding to the known fault classification loss function L1, σ2 represents the dynamic factor corresponding to the discriminant loss function L2 of multiple unknown classifiers, σ3 represents the dynamic factor corresponding to the soft consistency regularization loss function L3, and σ4 represents the dynamic factor corresponding to the information entropy maximization loss function L4. Indicating the classification probability output of the c-th element in the high-dimensional feature representation; S43. The dynamic factor σ updated continuously according to the model learning process n ={σ1, σ2, σ3, σ4} updates the loss function for model training. The dynamic update method for the loss function is as follows: Where, L overall represents the overall loss function of the model.

8. An adaptive deep transfer fault diagnosis system for complex industrial scenarios, characterized in that, For implementing the method according to any one of claims 1-7, it includes the following modules: A dataset construction module for collecting the vibration acceleration signals of industrial equipment in different fault modes under different working conditions and dividing them into a source domain dataset and a target domain dataset; A model building module for building a self-regulating general domain adaptive fault diagnosis model, including a shared feature extractor, a known classifier, and multiple unknown classifiers; A training module for training by combining four loss functions, namely, the classification loss of known faults in the source domain, the discrimination losses of multiple unknown classifiers, the target domain soft consistency regularization loss, and the information entropy maximization loss; A weight allocation and update module for adaptively allocating weights to each loss function based on a weight dynamic strategy for model uncertainty evaluation, and optimizing the diagnostic model parameters based on the dynamically updated loss functions to obtain a self-regulating general domain adaptive fault diagnosis model; A diagnosis module for inputting the data to be measured into the trained self-regulating general domain adaptive fault diagnosis model and outputting a diagnosis result.

9. A device, characterized in that, Comprising: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1-7.

10. A computer-readable storage medium storing a program executable by a processor, characterized in that, The program executable by the processor is used to execute the method according to any one of claims 1-7 when executed by the processor.

Images