A mechanical assembly fault prediction method and device based on domain generalization

Abstract

The present invention belongs to the technical field related to fault prediction and health management. It discloses a method and device for mechanical equipment fault prediction based on domain generalization, including the following steps: S1, enhancing intra-domain data and inter-domain data pairs with similar labels in a training set; S2, extracting common degradation features of the original data and enhanced data in the training set, then obtaining orthogonal degradation features for each domain, and then calculating feature orthogonalization loss; S3, calculating classification loss using a weighted adaptive classifier of orthogonal degradation features; adaptively aggregating multi-domain orthogonal degradation features; S4, applying label consistency constraints and sample pairing strategies to the aggregated features, and then calculating label consistency constraint loss, sample pairing loss, and regression loss; S5, training a mechanical fault prediction model based on the losses obtained above, and using the trained mechanical fault prediction model to perform mechanical fault prediction. The present invention improves prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field related to fault prediction and health management, and more specifically, relates to a method and device for mechanical equipment fault prediction based on domain generalization. Background Art

[0002] The primary goal of prognosis and health management (PHM) is to proactively detect potential failures and degradation trends by analyzing real-time monitoring data, thereby facilitating optimal maintenance decisions and preventing breakdowns. This is crucial for improving equipment safety and reliability and preventing catastrophic failures. As a key component of PHM, failure prediction—the ability to predict when equipment will next fail—is crucial for timely maintenance and judicious resource allocation.

[0003] In recent years, advances in sensor technology have facilitated the collection of large amounts of equipment monitoring data, providing the necessary data foundation for intelligent data-driven fault prediction technology. Compared with traditional model-based fault prediction methods, data-driven methods do not need to consider the complex degradation mechanisms of equipment, making them more suitable for deployment in complex systems. Generally, in the field of data-driven fault prediction methods, fault prediction methods based on deep learning (DL) have become a prominent competitor due to their significant advantages, such as excellent accuracy and fast response time, which has attracted considerable interest. However, due to the variability of workloads, environmental conditions, etc., mechanical equipment usually operates under different operating conditions (OC). The data distribution differences caused by variable OC (commonly referred to as domain offset) significantly reduce the predictive ability of deep prediction models.

[0004] Domain adaptation (DA) technology can effectively mitigate domain shift and has been used to address cross-domain fault prediction tasks in the past few years. DA-based fault prediction methods learn cross-domain invariant degradation feature representations by aligning the feature distributions between the source and target domains. These methods not only demonstrate robust prediction performance in the source domain but also exhibit strong generalization capabilities on the target task. However, these methods face several challenges when applied to practical cross-domain prediction tasks. First, obtaining unlabeled degradation data from the target device is crucial for DA-based fault prediction methods, as it effectively facilitates the transfer of prediction knowledge. However, given that prediction tasks often involve new devices in industrial prediction scenarios, obtaining such data in advance is often intractable. Consequently, existing DA-based fault prediction methods are unable to fully handle online cross-domain fault prediction tasks. Second, most DA-based fault prediction methods only generalize prediction knowledge from a single source domain to the target domain of interest. Consequently, when the OC of the target device changes, DA-based fault prediction methods are prone to overfitting in the original domain, resulting in inaccurate prediction results.

[0005] Domain generalization (DG) can extract robust transferable knowledge from multiple known source domains and effectively perform the target task of interest under unknown OC. Therefore, DG-based fault prediction methods have become a research hotspot. However, current DG-based fault prediction methods still have the following shortcomings: 1) These methods focus on extracting common degradation features from multiple known source domains to represent the degradation characteristics of the unknown target domain, ignoring the differences between different source domains and the unknown target domain. When the degradation patterns between some source domains and the unknown target domain differ significantly, these common degradation features may not accurately reflect the degradation trends of the unknown target domain. Therefore, new methods are needed to adaptively integrate existing degradation features based on the similarity between the unknown target domain and the known source domain, thereby enhancing the model's generalization ability to unknown target data. 2) Furthermore, these methods typically only consider marginal distribution shift when mitigating domain shift. However, the degradation data in fault prediction tasks has continuous and diverse labels. Mitigating only marginal distribution shift may lead to incorrect alignment of features with different labels, hindering the model's ability to effectively learn the mapping relationship between features and labels. Therefore, it is also necessary to align features with different labels separately, that is, to alleviate the changes in conditional distribution. Summary of the Invention

[0006] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a mechanical equipment fault prediction method and device based on domain generalization, which aims to solve the problem of low prediction accuracy in fault prediction due to the inability to obtain unknown target domain data, the failure to consider the similarity between the degradation characteristics of the source domain and the unknown target domain, and the difficulty in simultaneously reducing the edge and conditional distribution offsets between different source domains.

[0007] To achieve the above objectives, according to one aspect of the present invention, a mechanical assembly fault prediction method based on domain generalization is provided, the prediction method comprising the following steps:

[0008] S1, linearly interpolate the intra-domain data and inter-domain data pairs with similar labels in the training set to obtain intra-domain and inter-domain enhanced data; the training set includes a portion of the full life cycle sensor data of multiple mechanical assemblies under different working conditions;

[0009] S2, using the public feature extractor F CF Extract the common degradation features of the original data and enhanced data in the training set, and then input the common degradation features of the enhanced data in the domain and the original data in the training set into multiple different orthogonal degradation feature extractors F ODFk∈{1,…,K} To obtain the orthogonal degenerate features of each domain, and then calculate the covariance matrix between any two source domain orthogonal degenerate features, and then calculate the L2 norm of the covariance matrix, the L2 norm is the feature orthogonalization loss;

[0010] S3, based on the common degradation features, learns an orthogonal degradation feature weighted adaptive classifier that can identify the data source, and then calculates the classification loss. According to the classification results of the orthogonal degradation feature weighted adaptive classifier, multi-domain orthogonal degradation features are adaptively aggregated to accurately and effectively represent the degradation characteristics of the test data.

[0011] S4 applies label consistency constraints and sample pairing strategies to the aggregated features, promoting the aggregation of features with similar labels and the separation of features with different labels in the aggregate space. Label consistency constraint loss and sample pairing loss are then calculated. A regressor for predicting fault occurrence time is learned based on the aggregated features and their corresponding labels, and the regression loss is then calculated.

[0012] S5, based on feature orthogonalization loss, classification loss, label consistency constraint loss, sample pairing loss and regression loss, simultaneously train the common feature extractor F CF , multiple orthogonal degenerate feature extractors F ODFk∈{1,…,K} , orthogonal degradation feature weighted adaptive classifier C and regressor R, and use the trained mechanical fault prediction model to predict mechanical faults.

[0013] Furthermore, the Hungarian algorithm is used to pair the intra-domain data and inter-domain data with similar labels, and then the paired data are interpolated to generate intra-domain and inter-domain enhanced data.

[0014] Furthermore, the augmented data are interpolations of data pairs from the same and different source domains, respectively. These interpolations are calculated as follows:

[0015]

[0016]

[0017]

[0018]

[0019]

[0020] in represents the in-domain augmented data generated using the in-domain data from the k-th source domain, is the remaining time label for the next failure of the corresponding mechanical equipment, is its corresponding domain label; represents the inter-domain enhanced data generated by the inter-domain data pair consisting of the k-th source domain and the l-th source domain, is the remaining time label of the next failure of the corresponding mechanical equipment; γ is sampled from the beta distribution and is used to control the ratio between the two samples during the interpolation process.

[0021] Furthermore, given multiple source raw data and in-domain augmented data Feature orthogonalization is achieved by optimizing the covariance matrix between any two data to a zero matrix. The calculation formula of the orthogonalization loss is:

[0022]

[0023] Where Cov(·) represents the covariance of the matrix.

[0024] Furthermore, the output of the orthogonal degraded feature weighted adaptive classifier is the probability that each input data belongs to each source domain, and the input is the common degraded feature. The cross entropy loss between the probability distribution of the output of the orthogonal degraded feature weighted adaptive classifier and the true domain label is calculated. The cross entropy loss is the classification loss; the input of the regressor is the aggregated feature, and the output is a value between 0 and 1; the true label is also a value between 0 and 1, and the square of the difference between the output of the regressor and the true label is the regression loss.

[0025] Furthermore, given the input data xk,i , its aggregate feature z k,i Expressed as:

[0026]

[0027] Among them, ω k,i represents the kth source domain The weight reflects the i-th input data x k,i With the kth source domain similarity; in addition, ω k,i >0, and

[0028] Furthermore, the label consistency constraint LCC is defined as the absolute difference between the similarity of a cross-domain data pair in its original space and the degree of difference between them in the aggregated feature space, as follows:

[0029]

[0030] In the formula, for any two data from different source domains and The similarity between them is defined as in and They represent the remaining time labels for the next failure of the mechanical equipment corresponding to these data; and are the data from the kth source domain respectively The corresponding aggregate features and data from the lth source domain corresponding aggregate features; and are the in-domain augmented data of the k-th source domain The corresponding aggregated features and the in-domain enhanced data of the lth source domain The corresponding aggregation features.

[0031] Further, according to The value of determines the degree of difference between the features of any two data from different source domains, where Represents the data of the kth source domain The label and data of the lth source domain The difference between the labels of , and the Euclidean distance d(·) is used to measure the difference between the aggregated features; the label consistency constraint loss is:

[0032] The sample pairing loss is:

[0033]

[0034] Where nb is the batch size; represents the i-th data from the k-th source domain corresponding aggregate features; represents the jth data from the lth source domain The corresponding aggregation features.

[0035] The present invention also provides a mechanical equipment fault prediction system based on domain generalization, the system including a memory and a processor, the memory storing a computer program, and the processor executing the mechanical equipment fault prediction method based on domain generalization as described above when executing the computer program.

[0036] The present invention also provides a computer-readable storage medium, which stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the mechanical equipment fault prediction method based on domain generalization as described above.

[0037] In general, compared with the prior art, the above technical solutions conceived by the present invention provide a method and device for predicting mechanical equipment faults based on domain generalization, which has the following beneficial effects:

[0038] 1. During training, this method pairs batches of similarly labeled raw operating parameter measurement signals from the same and different source domains, then performs intra-domain and inter-domain data augmentation to increase the richness of the training data. Using similarly labeled data pairs for data augmentation avoids generating augmented data that misrepresents actual degradation patterns. Furthermore, intra-domain data augmentation captures the various degradation patterns inherent in each specific source data, ensuring that the prediction model extracts more representative domain-specific degradation features. Inter-domain data augmentation, by leveraging data from different source domains, mitigates overfitting of the mechanical fault prediction model within a specific source domain, thereby improving prediction accuracy.

[0039] 2. This invention enhances the diversity of source domain degradation features through orthogonalization, rather than extracting common degradation features from multiple source domains. The application of feature orthogonalization reduces the redundant information between multi-source degradation features and can comprehensively capture various degradation features presented in the original feature space of different source domains, thereby providing diversified feature support for unknown target data in the subsequent testing phase.

[0040] 3. By learning an orthogonal degradation feature weighted adaptive classifier C that can identify the data source in the training phase, the multi-source orthogonal degradation features can be optimally fused through weight selection in the testing phase, thereby accurately representing the degradation features of the unknown target data. In this way, the present invention can represent and learn the degradation features of the unknown target based on the similarity between the unknown target data and the known source domain data, thereby enhancing the model's generalization ability for unknown target data.

[0041] 4. The present invention can simultaneously reduce the marginal and conditional distribution differences between features in different source domains through semantically consistent orthogonalization, thereby promoting the learning of cross-domain invariant degradation knowledge, which helps to improve the generalization ability of the mechanical fault prediction model and then improve the fault prediction accuracy of the mechanical fault prediction model on unknown target numbers. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of a mechanical assembly fault prediction method based on domain generalization provided by the present invention;

[0043] Figure 2 (a), (b), (c), and (d) are schematic diagrams of the fault prediction results of the full life cycle data of bearing P1_1, the fault prediction results of the full life cycle data of bearing P1_2, the fault prediction results of the full life cycle data of bearing X1_1, and the fault prediction results of the full life cycle data of bearing X1_2, respectively;

[0044] Figure 3 (a), (b), (c), (d), (e), and (f) are, respectively, the edge distribution maps of the multi-source orthogonal degradation features extracted using the baseline in Task 1, the edge distribution maps of the multi-source orthogonal degradation features extracted using EMODF-V1 in Task 1, the edge distribution maps of the multi-source orthogonal degradation features extracted using EMODF in Task 1, the edge distribution maps of the multi-source orthogonal degradation features extracted using the baseline in Task 7, the edge distribution maps of the multi-source orthogonal degradation features extracted using EMODF-V1 in Task 7, and the edge distribution maps of the multi-source orthogonal degradation features extracted using EMODF in Task 7;

[0045] Figure 4(a), (b), (c), (d), (e), and (f) are the conditional distribution graphs of the multi-source orthogonal degradation features extracted using the baseline in Task 1, the conditional distribution graphs of the multi-source orthogonal degradation features extracted using EMODF-V1 in Task 1, the conditional distribution graphs of the multi-source orthogonal degradation features extracted using EMODF in Task 1, the conditional distribution graphs of the multi-source orthogonal degradation features extracted using the baseline in Task 7, the conditional distribution graphs of the multi-source orthogonal degradation features extracted using EMODF-V1 in Task 7, and the conditional distribution graphs of the multi-source orthogonal degradation features extracted using EMODF in Task 7, respectively.

[0046] Figure 5 (a) and (b) are schematic diagrams of the similarity between the test data and different source domains in Task 1, and schematic diagrams of the fault prediction results of the bearing P1_1 full life cycle data using the orthogonal degradation features of a single source domain and the adaptive fusion of multi-domain orthogonal degradation features in Task 1, respectively. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0048] See also Figure 1 The present invention provides a mechanical assembly fault prediction method based on domain generalization, which mainly includes the following steps:

[0049] S1, collects sensor data from multiple mechanical assemblies under different working conditions throughout their life cycle, and divides the collected data into training and test sets.

[0050] Among them, sampling is performed at a certain time interval and frequency, so as to obtain the measurement signal data of the operating parameters of the mechanical equipment throughout its life cycle.

[0051] S2, linearly interpolates the intra-domain data and inter-domain data pairs with similar labels in the training set to obtain intra-domain and inter-domain enhanced data.

[0052] Specifically, the Hungarian algorithm is used to pair intra-domain data and inter-domain data with similar labels, and then the paired data are interpolated to generate intra-domain and inter-domain enhanced data.

[0053] Intra-domain and inter-domain data enhancement is performed on the measurement signal data of the full life cycle operating parameters collected under different working conditions. Intra-domain enhancement refers to the use of data in the same source domain for enhancement. It aims to capture the various degradation patterns inherent in each specific source data to ensure that the mechanical fault prediction model can extract more representative degradation features of specific domains. At the same time, inter-domain data enhancement uses data from different source domains for data enhancement to reduce the overfitting of the mechanical fault prediction model on a specific source domain. By sharing data between different source domains, inter-domain enhancement can accurately obtain a generalized degradation feature representation from each source domain, thereby improving the generalization ability of the model when encountering unknown data. When enhancing intra-domain and inter-domain data, samples with similar labels are used for data enhancement, thereby avoiding the use of samples with too large label differences to generate enhanced samples that do not conform to the true degradation pattern. The specific instructions are as follows:

[0054] Taking inter-domain data enhancement as an example, for two batches of data from different source domains and in, and Denote batches of data from the kth and lth source domains, respectively. First, the label difference matrix between them is calculated and optimized using the Hungarian algorithm to ensure that all data are uniquely paired and the sum of label differences of all data pairs is minimized. The calculation formula for the label difference matrix between all samples of these two batches of data is:

[0055]

[0056] Among them, n bm =n bn =n b , n b is the batch size.

[0057] The steps to achieve maximum pairing using the Hungarian algorithm are:

[0058] Step 1: From L d Subtract the smallest number from each row of , and create a new matrix to contain the result;

[0059] Step 2: Subtract the smallest number in each column of the new matrix from that column, creating another matrix to contain the results.

[0060] Step 3: Check the possibility of the best allocation and try to cross out all the zeros in the matrix with the least number of lines; if the number of lines is greater than or equal to the number of rows in the matrix, go to step 6; otherwise, go to step 4;

[0061] Step 4: For the matrix elements not covered in step 3, subtract the smallest uncovered element value from all uncovered elements in the matrix elements; if any element in step 3 is repeatedly covered by a line, add the smallest uncovered element value to the repeatedly covered element;

[0062] Step 5: Repeat steps 3 and 4 until the optimal matrix is ​​obtained;

[0063] Step 6: Start assigning, starting with the row or column that has only one zero, match the items that have zero, use only one match per row and column, delete the rest of the cells.

[0064] For the paired inter-domain data pairs and domain data pairs in and They represent the i-th sample from the k-th source domain, the j-th sample from the k-th source domain, and the j-th sample from the l-th source domain respectively; and They represent the remaining time labels for the next failure of their corresponding mechanical equipment; d k and d l Represent their domain labels respectively.

[0065] The augmented data are interpolations of data pairs from the same and different source domains, respectively. These interpolations can be calculated as follows:

[0066]

[0067]

[0068]

[0069]

[0070]

[0071] in represents the in-domain augmented data generated using the in-domain data from the k-th source domain, is the remaining time label for the next failure of the corresponding mechanical equipment, is its corresponding domain label; represents the inter-domain enhanced data generated by the inter-domain data pair consisting of the k-th source domain and the l-th source domain, The remaining time label for the next failure of the corresponding mechanical equipment. Given that the inter-domain augmented data is generated from data in different source domains, no domain label is assigned to the inter-domain augmented data. In addition, γ is sampled from a beta distribution to control the ratio between the two samples during the interpolation process.

[0072] S3, using the public feature extractor F CF Extract the common degradation features of the original data and enhanced data in the training set, and then input the common degradation features of the in-domain enhanced data and the original data in the training set into multiple different orthogonal degradation feature extractors To obtain the orthogonal degenerate features of each domain, and then calculate the covariance matrix between any two source domain orthogonal degenerate features, and then calculate the L2 norm of the covariance matrix, the L2 norm is the feature orthogonalization loss.

[0073] Orthogonality is used to reduce the similarity between orthogonal degradation features from different source domains, thereby reducing the redundancy of degradation information. Simultaneously, unrelated multi-domain orthogonal degradation features help enrich the diversity of degradation features, facilitating the effective representation of degradation features of unknown targets during the subsequent testing phase, thereby improving the model's fault prediction accuracy under unknown operating conditions.

[0074] In order to provide diverse degradation feature support, we focus on enhancing the diversity of degradation features rather than extracting common degradation features. To this end, we apply feature orthogonalization methods to ensure that the aggregated features comprehensively capture the various degradation features presented in the original feature space. In other words, the goal is to alleviate the redundancy between multi-source degradation features and reduce the correlation between them. More formally, given multi-source raw data and in-domain augmented data Feature orthogonalization is achieved by optimizing the covariance matrix between any two data (original or in-domain augmented data) to a zero matrix. The orthogonalization loss can be calculated as:

[0075]

[0076] Where Cov(·) represents the covariance of the matrix. Uncorrelated multi-domain orthogonal degradation features help enrich the diversity of degradation features, making it easier to effectively represent the degradation features of unknown targets during the subsequent testing phase, thereby improving the fault prediction accuracy of the model under unknown working conditions.

[0077] S4, based on the common degradation features, learns an orthogonal degradation feature weighted adaptive classifier that can identify the source of the data, and then calculates the classification loss; according to the classification results of the orthogonal degradation feature weighted adaptive classifier, adaptively aggregates multi-domain orthogonal degradation features to accurately and effectively represent the degradation characteristics of the test data.

[0078] The output of the orthogonal degraded feature weighted adaptive classifier is the probability that each input data belongs to each source domain. The input is the common degraded feature. The cross entropy loss between the probability output by the orthogonal degraded feature weighted adaptive classifier and the true domain label is calculated. The cross entropy loss is the classification loss.

[0079] The input of the regressor is the aggregated feature, and the output is a value between 0 and 1; the true label is also a value between 0 and 1, and the square of the difference between the output of the regressor and the true label is the regression loss.

[0080] The orthogonal degradation feature weighted adaptive classifier C is used to evaluate the similarity between the test data and the source domain data. Then, based on the obtained similarity, the multi-domain orthogonal degradation features are adaptively aggregated to construct the degradation representation of the test data.

[0081] Considering the complexity and diversity of actual operating conditions, there may be differences in the degradation features between the unknown target domain and a single known source domain. However, as the number of available source domains increases, the unknown target domain may show similarities with one or more known source domains. Therefore, using ODFs from multiple known source domains has the potential to improve prediction accuracy. To this end, an ODF weighted adaptive classifier C is designed to optimally fuse multiple source ODFs by weight selection to obtain an aggregated unified feature representation. More formally, given the input data x k,i , its aggregate feature z k,i It can be expressed as:

[0082]

[0083] Among them, ω k,i represents the kth source domain The weight reflects the i-th input data x k,i With the kth source domain In addition, ω k,i >0, and

[0084] The above feature aggregation process can be interpreted as a form of ensemble learning, where each feature extraction component is trained into a robust model that performs well on a specific source domain. Specifically, each feature extraction component consists of two components: a common feature extractor F CF and an orthogonal degradation feature extractor for each source domain The design of this feature extraction method is inspired by previous research on deep transfer learning networks, where lower layers typically learn low-level and general features, while higher layers specialize in domain-specific features.

[0085] To obtain the weight ω for each source domain k,i, the designed ODF weighted adaptive classifier C takes the common degradation features as input and applies the Softmax function to convert its output into a probability distribution. In the model training phase, since each source data are accompanied by a domain label d k,i , so the similarity ω k,i Just set it as follows:

[0086]

[0087] Therefore, the classification loss of C can be calculated as:

[0088]

[0089] Among them, L CE represents the cross entropy loss function.

[0090] When testing on the target machine, without access to the domain labels of the unknown target data, the similarity of each target data to all source domains can be measured by the trained C as follows:

[0091]

[0092] Among them, ω j =[ω 1,j ,ω 2,j ,,…,ω K,j ] T . Similarly, ω k,j represents the similarity between the jth target data and the kth source domain. Then, the degradation characteristics of the target data can be expressed as:

[0093]

[0094] In this way, MODFE is able to represent and learn the degradation features of unknown targets based on the similarity between the unknown target data and the known source domain data. Furthermore, the aggregated degradation features are mapped to the remaining time value of the next failure of the mechanical equipment through the regressor R. Therefore, the prediction loss can be calculated as follows:

[0095]

[0096] Among them, L task represents the mean absolute error function which is widely used in fault prediction.

[0097] S5 applies label consistency constraints and sample pairing strategies to the aggregated features, thereby promoting the aggregation of features with similar labels and the separation of features with different labels in the aggregate space, and then calculates the label consistency constraint loss and sample pairing loss; at the same time, a regressor for predicting faults is learned based on the aggregated features and their corresponding labels, and then the regression loss is calculated.

[0098] according to The value of determines the degree of difference between the features of any two data from different source domains, and the Euclidean distance d(·) is used to measure the difference between the aggregated features; the label consistency constraint loss is:

[0099] The sample pairing loss is:

[0100]

[0101] Considering the presence of marginal and conditional distribution shifts between orthogonal degenerate features from different source domains, a semantic consistency regularization strategy is used to constrain the aggregated features before fault prediction. Specifically, the semantic consistency regularization strategy consists of two parts: a label consistency constraint and a sample pairing strategy. The label consistency constraint forces the similarity between different features to be consistent with the similarity between their corresponding labels. The sample pairing strategy encourages features with similar labels to be as similar as possible. Based on this, degenerate features with similar labels can be aggregated, while degenerate features with different labels can be separated, thereby simultaneously mitigating marginal and conditional distribution shifts between different domains.

[0102] A semantic consistency regularization term is established to constrain the aggregated features. Specifically, the semantic consistency regularization term consists of two parts: label consistency constraints (LCC) and sample pairing strategy (SPS). LCC and SPS are designed to encourage data with similar labels to be close in the feature space, while data with inconsistent labels are kept away from each other, thereby mitigating both marginal and conditional distribution shifts.

[0103] Label consistency constraint (LCC) is based on the basic principle that the similarity between data features depends on the similarity between data in the original space. Therefore, for any two data from different source domains, and The similarity between them is defined as in and They represent the remaining time labels of the next failure of the mechanical equipment corresponding to these data. As an explanation of the above similarity definition, if and With different labels, The value of will increase, indicating that the two data are not similar. and have the same label, then This shows and Highly similar. The goal is to The value of d(·) is used to determine the degree of difference between the features of any two data from different source domains. In this embodiment, the Euclidean distance d(·) is used to measure the difference between the aggregated features.

[0104] In summary, LCC is defined as the absolute difference between the similarity of a cross-domain data pair in its original space and the degree of difference between them in the aggregated feature space, as follows:

[0105]

[0106] Assume that two data samples form a cross-domain pair The feature values ​​of are different, but their labels are the same. In this case, P k (X|Y)≠P l (X|Y). Based on L lcc , and will have the same eigenvalues, resulting in Therefore, in the aggregate space, since the differences between features are eliminated, P k (X|Y)=P l (X|Y). In another case, suppose and The features are the same, but have different labels, which means but if and are aligned in the aggregate feature space, which will lead to negative transfer. However, L lcc will be based on and Limit the difference between labels and Therefore, when learning cross-domain invariant features, L lcc This can avoid the misalignment of some data by the deep model. lcc To a certain extent, it bridges the marginal and conditional distribution differences between multiple source domains in the aggregated feature space, thereby enhancing the generalization ability of the model.

[0107] In order to avoid the situation where a small number of closely labeled sample features from different source domains cannot be clustered together well, a sample pairing strategy (SPS) is used to compare samples with similar labels in different source domains. SPS helps to tightly cluster samples with similar labels in the aggregated feature space, similar to local clustering in classification tasks. SPS uses the Hungarian algorithm to maximize the pairing of samples from different source domains in a random process and make the features of labeled samples as close as possible. For successfully paired cross-domain sample pairs, SPS calculates the Euclidean distance between their features The SPS loss in batch processing is defined as follows:

[0108]

[0109] S6, based on feature orthogonalization loss, classification loss, label consistency constraint loss, sample pairing loss and regression loss, simultaneously train the common feature extractor F CF , multiple orthogonal degenerate feature extractors The orthogonal degradation feature weighted adaptive classifier C and regressor R use the trained mechanical fault prediction model to predict mechanical faults.

[0110] The mechanical failure prediction model includes a common feature extractor F CF , multiple orthogonal degenerate feature extractors Orthogonal degenerate feature weighted adaptive classifier C and regressor R.

[0111] Combine all the above loss functions to train the common feature extractor F at the same time CF , multiple orthogonal degenerate feature extractors Orthogonal degenerate feature weighted adaptive classifier C and regressor R. The total loss is:

[0112] L=L r +k t (L ort +αL c +βL lcc +γL sps ),

[0113] in, are time-varying coefficients; q and q m Denote the qth iteration and the maximum number of iterations, respectively. The trained model is deployed on the test data to perform fault prediction. The true value of the fault prediction time is the percentage of the remaining time until the next failure of the mechanical equipment in the entire life cycle. The true fault prediction time label for the tth sampling point is:

[0114]

[0115] Among them, T t and T0 represent the order t of the sampling points and the total number of sampling points in the entire life cycle, respectively.

[0116] The following uses bearings as a specific object and combines specific examples to further explain the method for predicting mechanical equipment faults based on domain generalization of the present invention. The specific steps are as follows:

[0117] (1) The measurement signal data of the bearing's full life cycle operating parameters were extracted. The accelerated degradation test bearing dataset of the IEEE PHM 2012 Challenge and the XJTU-SY bearing dataset were used for training and testing, respectively. Both datasets contain two orthogonally mounted accelerometers. The sampling frequency of the accelerometer in the first dataset is 25.6kHz, the single sampling time is 0.1s, and the sampling interval is 10s. Therefore, each sensor can obtain 2560 sampling points per sampling. The sampling frequency of the accelerometer in the second dataset is 25.6kHz, the single sampling time is 1.28s, and the sampling interval is 1min. Therefore, each sensor can obtain 32768 sampling points per sampling. The specific details of the two bearing datasets and the designed multi-source domain generalized fault prediction tasks are shown in Tables 1 and 2, respectively.

[0118] Table 1 PHM 2012 and XJTU-SY bearing datasets

[0119]

[0120]

[0121] Table 2 Multi-source domain generalized fault prediction tasks

[0122]

[0123] (2) During the training process, the Hungarian algorithm is used to pair batches of full-lifecycle operational parameter measurement signal data from the same source domain with batches of full-lifecycle operational parameter measurement signal data from different source domains. Intra-domain data pairs and cross-domain data pairs with similar labels are used to perform intra-domain data augmentation and inter-domain data augmentation, respectively, to expand the diversity of the training data. The number of out-of-domain augmented samples generated is consistent with the batch size, and the number of in-domain augmented samples generated is half the batch size.

[0124] (3) Construct a common feature extractor model and an orthogonal degradation feature extractor model. The network structures of the common feature extractor and the orthogonal degradation feature extractor are shown in Table 3. The degradation features of the data are extracted using a common feature extractor with shared parameters and an orthogonal degradation feature extractor with unshared parameters. These features are then orthogonalized to reduce redundant degradation information. Specifically, the orthogonalization loss is calculated, which is the L2 norm of the covariance matrix between any two source domain degradation features.

[0125] (4) Construct a weighted adaptive classifier and regressor. During the training phase, the orthogonal degradation feature weighted adaptive classifier is used to classify the data of each source domain, that is, to accurately determine the source of each data. The classification cross entropy loss is calculated based on the classification results output by the orthogonal degradation feature weighted adaptive classifier and the true domain label of the data; then the similarity between the training data and each source domain is determined, and the multi-source orthogonal degradation features are aggregated by weight selection. The aggregated features are input into the regressor for fault prediction. The regression loss is calculated by the predicted value output by the regressor and the true label of the data. During the testing phase, the well-trained weighted adaptive classifier can identify the similarity between the unknown target data and the multi-source domain data, and then adaptively integrate the multi-source orthogonal degradation features based on the obtained similarity, thereby effectively representing the degradation characteristics of the unknown target data.

[0126] (5) Constrain the aggregated features of different source domains, that is, semantically restrict the aggregated features through label consistency constraints and sample pairing strategies. Specifically, in the aggregated feature space, degraded features of different source domains with similar labels are brought closer together, and degraded features of different source domains with different labels are separated, thereby effectively reducing the marginal and conditional distribution differences between multi-domain degraded features, and calculating the label consistency loss and the loss of the sample pairing strategy.

[0127] (6) Combine all the losses calculated above to simultaneously train the common feature extractor F CF , multiple orthogonal degenerate feature extractors Orthogonal degenerate feature weighted adaptive classifier C and regressor R. In the back propagation process, the optimization algorithm is Adam. The specific hyper parameters during the model training process are shown in Table 4. After the model training is completed, the test data set is input for testing. First, use F CF To obtain the common degradation features of the test data. The common degradation features are then fed into To obtain multi-source orthogonal degradation features, the common degradation features are fed into the parallel C to determine the similarity between the test data and the multi-source data. Based on the obtained similarity, the multi-source orthogonal degradation features are adaptively aggregated to accurately obtain the degradation representation of the unknown target data. Finally, the aggregated degradation features are fed into R, which outputs the predicted remaining time to next failure of the mechanical equipment in the test data.

[0128] Table 3: EMODF network structure

[0129]

[0130]

[0131] (7) Visualization of prediction results. Taking tasks 1, 2, 7, and 8 as examples, the test bearings are P1_1, P1_2, X1_1, and X1_2. Mean absolute error (MAE), root mean square error (RMSE), and score are used to analyze the experimental results. Their calculation methods are as follows:

[0132]

[0133]

[0134]

[0135]

[0136]

[0137] in, with y i Represent the predicted value and the true value respectively; Er i Indicates the percentage error. The smaller the MAE and RMSE, the better the model prediction performance. The larger the Score, the better the model prediction performance. The fault prediction results are as follows Figure 2 As shown in Table 5, the MAE, RMSE and Score values ​​of all tasks are given. Figure 2 As can be seen from Table 5, the method proposed in this embodiment performs well in solving the fault prediction problem in cross-domain scenarios with unknown target domain data, which verifies the effectiveness of the present invention.

[0138] Table 5: Fault prediction results of MODEF on all tasks

[0139]

[0140] Table 6: Fault prediction results of different methods on Task 1 and Task 7

[0141]

[0142]

[0143] In order to further verify the effectiveness of the proposed method in simultaneously alleviating the differences between the edge and conditional distributions of different source domains, the multi-source ODFs extracted by different methods are visualized using the t-distributed random neighbor embedding technique. Specifically, Figure 3 and Figure 4 The results of mapping multi-source ODF into one-dimensional and two-dimensional spaces are shown respectively. These figures illustrate the marginal and conditional distributions of multi-source ODF using different methods on Task 1 and Task 7. The compared methods include: (1) baseline, (2) MODFE-V1, which is similar to MODFE, but the label consistency constraint and sample pairing strategy are replaced by Maximum Mean Discrepancy (MMD), and (3) MODFE. The prediction results of these methods on Task 1 and Task 7 are shown in Table 6. Figure 3 and Figure 4 It can be seen from the above that the present invention can simultaneously alleviate the differences between the edge and conditional distributions of different source domains. Figure 3 、 Figure 4 As can be seen from Table 6, simultaneously alleviating marginal and conditional variations between multiple existing domains and learning general predictive knowledge is a promising approach to improve the generalization ability of prognostic models.

[0144] In order to further verify the effectiveness of the method proposed in the present invention to adaptively fuse multi-domain orthogonal degradation features based on the similarity between the test data and multi-source data to represent the degradation features of the test data, Task 1 is selected to illustrate the advantages of the above-mentioned similarity-based aggregation features. Figure 5 The similarity between the test data and the source domain data and the prediction results of the remaining time for the next failure of the mechanical equipment are shown. Figure 5 As can be seen from the figure, the similarity-adaptive aggregation of multi-domain orthogonal degraded features can better represent the degraded features of the unknown target domain. By adaptively selecting the degraded features of the source domain data that are closer to the target data, the generalization ability of the model can be improved.

[0145] To further verify the superiority of the present invention in fault prediction results, the present invention is compared with several related methods, including a multi-scale convolutional neural network (MSCNN) based on deep learning, a double-convolutional neural network (DCNN), and Mixup based on domain generalization, representation self-challenging (RSC), and adversarial out-domain augmentation (AOA). As can be seen from Table 7, compared with MSCNN, DCNN, Mixup, RSC, and AOA, the present invention achieves the best prediction performance, which verifies the effectiveness of the EMODF model established by the present invention.

[0146] Table 7: Comparison of average prediction results of EMODF and related methods on all tasks

[0147]

[0148] The present invention also provides a mechanical equipment fault prediction system based on domain generalization, the system including a memory and a processor, the memory storing a computer program, and the processor executing the mechanical equipment fault prediction method based on domain generalization as described above when executing the computer program.

[0149] The present invention also provides a computer-readable storage medium, which stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the mechanical equipment fault prediction method based on domain generalization as described above.

[0150] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A mechanical assembly fault prediction method based on domain generalization, characterized by: The prediction method includes the following steps: S1, linearly interpolate the intra-domain data and inter-domain data pairs with similar labels in the training set to obtain intra-domain and inter-domain enhanced data; the training set includes a portion of the full life cycle sensor data of multiple mechanical assemblies under different working conditions; S2, using the public feature extractor F CF Extract the common degradation features of the original data and enhanced data in the training set, and then input the common degradation features of the enhanced data in the domain and the original data in the training set into multiple different orthogonal degradation feature extractors F ODFk∈{1,…,K} To obtain the orthogonal degenerate features of each domain, and then calculate the covariance matrix between any two source domain orthogonal degenerate features, and then calculate the L2 norm of the covariance matrix, the L2 norm is the feature orthogonalization loss; S3, based on the common degradation features, learns an orthogonal degradation feature weighted adaptive classifier that can identify the source of the data, and then calculates the classification loss; according to the classification results of the orthogonal degradation feature weighted adaptive classifier, adaptively aggregates the multi-domain orthogonal degradation features; S4 applies label consistency constraints and sample pairing strategies to the aggregated features, promoting the aggregation of features with similar labels and the separation of features with different labels in the aggregate space. Label consistency constraint loss and sample pairing loss are then calculated. A regressor for predicting fault occurrence time is learned based on the aggregated features and their corresponding labels, and the regression loss is then calculated. S5, based on feature orthogonalization loss, classification loss, label consistency constraint loss, sample pairing loss and regression loss, simultaneously train the common feature extractor F CF , multiple orthogonal degenerate feature extractors F ODFk∈{1,…,K} , orthogonal degradation feature weighted adaptive classifier C and regressor R, and use the trained mechanical fault prediction model to predict mechanical faults.

2. The mechanical assembly fault prediction method based on domain generalization according to claim 1, characterized in that: The Hungarian algorithm is used to pair the intra-domain data and inter-domain data with similar labels, and then the paired data are interpolated to generate intra-domain and inter-domain enhanced data.

3. The mechanical assembly fault prediction method based on domain generalization according to claim 1, characterized in that: The augmented data are interpolations of data pairs from the same and different source domains, respectively. These interpolations are calculated as follows: in represents the in-domain augmented data generated using the in-domain data from the k-th source domain, is the remaining time label for the next failure of the corresponding mechanical equipment, is its corresponding domain label; represents the inter-domain enhanced data generated by the inter-domain data pair consisting of the k-th source domain and the l-th source domain, is the remaining time label of the next failure of the corresponding mechanical equipment; γ is sampled from the beta distribution and is used to control the ratio between the two samples during the interpolation process.

4. The mechanical assembly fault prediction method based on domain generalization according to claim 3, characterized in that: Given multiple source raw data and in-domain augmented data Feature orthogonalization is achieved by optimizing the covariance matrix between any two data to a zero matrix. The calculation formula of the orthogonalization loss is: Where Cov(·) represents the covariance of the matrix.

5. The mechanical assembly fault prediction method based on domain generalization according to claim 1, characterized in that: The output of the orthogonal degenerate feature weighted adaptive classifier is the probability that each input data belongs to each source domain. The input is the common degenerate feature. The cross entropy loss between the probability output by the orthogonal degenerate feature weighted adaptive classifier and the true domain label is calculated. The cross entropy loss is the classification loss. The input of the regressor is the aggregated feature, and the output is a value between 0 and 1. The true label is also a value between 0 and 1. The square of the difference between the regressor output and the true label is the regression loss.

6. The mechanical assembly fault prediction method based on domain generalization according to claim 1, characterized in that: Given input data x k,i , its aggregate feature z k,i Expressed as: Among them, ω k,i represents the kth source domain The weight reflects the i-th input data x k,i With the kth source domain similarity; in addition, ω k,i >0, and 7. The mechanical assembly fault prediction method based on domain generalization according to claim 1, characterized in that: The label consistency constraint LCC is defined as the absolute difference between the similarity of a cross-domain data pair in its original space and the degree of difference between them in the aggregated feature space, as follows: In the formula, for any two data from different source domains and The similarity between them is defined as in and They represent the remaining time labels for the next failure of the mechanical equipment corresponding to these data; and are the data from the kth source domain respectively The corresponding aggregate features and data from the lth source domain corresponding aggregate features; and are the in-domain augmented data of the k-th source domain The corresponding aggregated features and the in-domain enhanced data of the lth source domain The corresponding aggregation features.

8. The mechanical assembly fault prediction method based on domain generalization according to claim 7, characterized in that: according to The value of determines the degree of difference between the features of any two data from different source domains, where Represents the data of the kth source domain The label and data of the lth source domain The difference between the labels of , and the Euclidean distance d(·) is used to measure the difference between the aggregated features; the label consistency constraint loss is: The sample pairing loss is: Where n b is the batch size; represents the i-th data from the k-th source domain corresponding aggregate features; represents the jth data from the lth source domain The corresponding aggregation features.

9. A mechanical equipment fault prediction system based on domain generalization, characterized by: The system includes a memory and a processor, the memory stores a computer program, and the processor executes the mechanical equipment fault prediction method based on domain generalization according to any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by the processor, the machine-executable instructions prompt the processor to implement the mechanical equipment fault prediction method based on domain generalization according to any one of claims 1 to 8.

Images