Mechanical fault diagnosis model migration training method, mechanical fault diagnosis method, device, equipment and medium

Abstract

The invention relates to a mechanical fault diagnosis model migration training method and device, a mechanical fault diagnosis method and device, equipment and a medium. The training method comprises the steps of obtaining a mechanical fault diagnosis model, a source domain sample set and a target domain sample set; outputting a sample feature and a fault prediction result of the source fault sample and a sample feature and a fault prediction result of the target fault sample by using the mechanical fault diagnosis model; determining sample classification loss according to the fault prediction results of the plurality of source fault samples and the sample labels; determining feature difference loss according to the sample features of the plurality of source fault samples and the sample features of the plurality of target fault samples; determining category confusion loss according to the fault prediction results of the multiple target fault samples; and determining target loss according to the sample classification loss, the feature difference loss and the category confusion loss, and optimizing the mechanical fault diagnosis model by using the target loss. Therefore, the accuracy of mechanical fault type diagnosis of the trained target mechanical fault diagnosis model can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of mechanical fault diagnosis, and in particular to a mechanical fault diagnosis model migration training method and a mechanical fault diagnosis method, device, equipment and medium. Background Art

[0002] Modern mechanical manufacturing systems exhibit significant characteristics such as large-scale control, high automation, and flexibility. As these systems become increasingly complex, downtime due to various faults becomes an unbearable burden. Therefore, fault diagnosis systems are considered necessary to minimize downtime due to maintenance and ensure timely action after a fault occurs. With the continuous increase in the amount of data, traditional machine learning-based fault diagnosis methods are no longer sufficient, and there is an urgent need to develop advanced intelligent fault diagnosis methods.

[0003] Deep learning uses deep hierarchical architectures to automatically extract abstract features from large-scale datasets, in contrast to traditional machine learning methods, which require manual feature engineering and impose a large workload on users. Although deep learning has achieved remarkable results in machine fault diagnosis, its performance relies heavily on the existence of large amounts of labeled data required to train diagnostic models. However, in engineering scenarios, obtaining such labeled data is usually unrealistic due to the expensive nature of fault data collection and annotation, limiting the growth potential of deep learning in this field.

[0004] Transfer learning challenges the assumption that training data and test data must follow independent and identical distribution. This paradigm motivates researchers to use transfer learning techniques to solve the problem of limited training data. In transfer learning, the model can be trained in the target domain using labeled source data and unlabeled target data, which significantly reduces the dependence on labeled data and shortens the training time in the target domain.

[0005] Existing transfer learning methods in the field of fault diagnosis mainly focus on aligning the marginal feature distributions of the source domain and the target domain. Although aligning the marginal feature distributions can help improve the diagnostic performance of the fault diagnosis model after transfer training, the fault diagnosis models generated by these methods usually focus on aligning the marginal feature distributions without considering the categories of the samples. Even if the marginal feature distributions are fully aligned and domain-invariant features are generated, there is no guarantee that these features have the discriminative ability of the target samples. This limitation has an adverse effect on the diagnostic performance of the fault diagnosis model obtained after transfer training, that is, it affects the accuracy of the fault diagnosis model obtained after transfer training. Summary of the invention

[0006] In view of this, the present disclosure proposes a mechanical fault diagnosis model migration training method and a mechanical fault diagnosis method, device, equipment and medium, which can improve the accuracy of mechanical fault diagnosis using a trained target mechanical fault diagnosis model.

[0007] According to one aspect of the present disclosure, a method for transferring and training a mechanical fault diagnosis model is provided, comprising: obtaining a mechanical fault diagnosis model to be trained and a source domain sample set and a target domain sample set; wherein the source domain sample set comprises a plurality of source fault samples and sample labels corresponding to each source fault sample, and the target domain sample set comprises a plurality of unlabeled target fault samples; the source fault sample comprises a mechanical vibration signal collected when a mechanical fault occurs under a source working condition, and the target fault sample comprises a mechanical vibration signal collected when a mechanical fault occurs under a target working condition, and the source working condition is different from the target working condition; the sample label is used to indicate the fault type corresponding to the source fault sample; wherein the mechanical fault diagnosis model comprises a feature generator and a classifier, the feature generator is used to extract sample features of the fault sample, and the classifier is used to generate a fault prediction result according to the sample features, and the fault prediction result comprises the probability that the fault sample belongs to each fault type respectively; and the mechanical fault diagnosis model is used to classify the fault sample according to the sample features based on the source fault samples in the source domain sample set and the target domain sample set. target fault samples, output the sample features and fault prediction results of each source fault sample, and the sample features and fault prediction results of each target fault sample; determine the sample classification loss according to the fault prediction results of the multiple source fault samples and the sample labels of each source fault sample, the sample classification loss characterizing the classification error of the fault type of the source fault sample by the mechanical fault diagnosis model; determine the feature difference loss according to the sample features of the multiple source fault samples and the sample features of the multiple target fault samples, the feature difference loss characterizing the feature distribution difference between the sample features of the source fault samples and the sample features of the target fault samples; determine the category confusion loss according to the fault prediction results of the multiple target fault samples, the category confusion loss characterizing the degree of confusion between different fault types by the mechanical fault diagnosis model; determine the target loss according to the sample classification loss, the feature difference loss and the category confusion loss, and use the target loss to optimize the mechanical fault diagnosis model to obtain the trained target mechanical fault diagnosis model.

[0008] In a possible implementation, the method further includes: inputting the sample features of the multiple source fault samples and the sample features of the multiple target fault samples into a domain discriminator respectively, obtaining a domain classification result of each source fault sample and a domain classification result of each target fault sample, wherein the domain classification result represents the probability that the sample feature belongs to the feature distribution space corresponding to the source domain sample set, or represents the probability that the sample feature belongs to the feature distribution space corresponding to the target domain sample set; determining a domain adversarial loss according to the domain classification results of each source fault sample and the domain classification results of each target fault sample, wherein the domain adversarial loss represents the classification error of the domain discriminator on the feature distribution space to which the sample feature belongs; wherein determining the target loss according to the sample classification loss, the feature difference loss, and the category confusion loss includes: determining the target loss according to the sample classification loss, the feature difference loss, the category confusion loss, and the domain adversarial loss.

[0009] In a possible implementation, the method further includes: adding a disturbance signal to each source fault sample in the source domain sample set to obtain a source fault disturbance sample corresponding to each source fault sample; inputting the source fault disturbance sample corresponding to each source fault sample into the mechanical fault diagnosis model to obtain a fault prediction result of the source fault disturbance sample corresponding to each source fault sample; and determining a first virtual adversarial loss based on the fault prediction results of the multiple source fault samples and the fault prediction results of the source fault disturbance sample corresponding to each source fault sample, wherein the first virtual adversarial loss characterizes the difference between the fault prediction result of the source fault sample and the fault prediction result of the corresponding source fault disturbance sample.

[0010] In a possible implementation, determining the target loss according to the sample classification loss, the feature difference loss, and the class confusion loss includes: determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, and the first virtual adversarial loss; or determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, the domain adversarial loss, and the first virtual adversarial loss; wherein the domain adversarial loss is determined according to the domain classification results of each source fault sample and the domain classification results of each target fault sample.

[0011] In a possible implementation, the method further includes: adding a disturbance signal to each target fault sample in the target domain sample set to obtain a target fault disturbance sample corresponding to each target fault sample; inputting the target fault disturbance sample corresponding to each target fault sample into the mechanical fault diagnosis model to obtain a fault prediction result of the target fault disturbance sample corresponding to each target fault sample; determining a second virtual adversarial loss based on the fault prediction results of the multiple target fault samples and the fault prediction results of the target fault disturbance sample corresponding to each target fault sample, wherein the second virtual adversarial loss characterizes the difference between the fault prediction result of the target fault sample and the fault prediction result of the corresponding target fault disturbance sample.

[0012] In a possible implementation, determining the target loss according to the sample classification loss, the feature difference loss, and the class confusion loss includes: determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, and the second virtual adversarial loss; or determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, the domain adversarial loss, and the second virtual adversarial loss; or determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, the first virtual adversarial loss, and the second virtual adversarial loss; or determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, the domain adversarial loss, the first virtual adversarial loss, and the second virtual adversarial loss; wherein the domain adversarial loss is determined according to the domain classification results of each source fault sample and the domain classification results of each target fault sample, and the first virtual adversarial loss is determined according to the fault prediction results of the multiple source fault samples and the fault prediction results of the source fault disturbance samples corresponding to each source fault sample.

[0013] According to another aspect of the present disclosure, a method for diagnosing mechanical faults is provided, comprising: acquiring fault data actually collected when a mechanical fault occurs under a target operating condition, the fault data comprising a mechanical vibration signal; determining a fault diagnosis result corresponding to the fault data based on the fault data using a target mechanical fault diagnosis model, the fault diagnosis result characterizing a fault type of the mechanical fault under the target operating condition; wherein the target mechanical fault diagnosis model is trained using the training method.

[0014] According to another aspect of the present disclosure, a mechanical fault diagnosis model migration training device is provided, comprising: an acquisition module, used to acquire a mechanical fault diagnosis model to be trained and a source domain sample set and a target domain sample set; wherein the source domain sample set includes multiple source fault samples and sample labels corresponding to each source fault sample, and the target domain sample set includes multiple unlabeled target fault samples; the source fault sample includes a mechanical vibration signal collected when a mechanical fault occurs under a source working condition, and the target fault sample includes a mechanical vibration signal collected when a mechanical fault occurs under a target working condition, and the source working condition is different from the target working condition; the sample label is used to indicate the fault type corresponding to the source fault sample; wherein the mechanical fault diagnosis model includes a feature generator and a classifier, the feature generator is used to extract sample features of the fault sample, and the classifier is used to generate a fault prediction result according to the sample features, and the fault prediction result includes the probability that the fault sample belongs to each fault type respectively; a prediction module is used to use the mechanical fault diagnosis model to output based on each source fault sample in the source domain sample set and each target fault sample in the target domain sample set. The sample features and fault prediction results of each source fault sample, as well as the sample features and fault prediction results of each target fault sample; a classification loss determination module, used to determine the sample classification loss according to the fault prediction results of the multiple source fault samples and the sample labels of each source fault sample, wherein the sample classification loss represents the classification error of the fault type of the source fault sample by the mechanical fault diagnosis model; a difference loss determination module, used to determine the feature difference loss according to the sample features of the multiple source fault samples and the sample features of the multiple target fault samples, wherein the feature difference loss represents the difference in sample feature distribution between the source domain sample set and the target domain sample set; a confusion loss determination module, used to determine the category confusion loss according to the fault prediction results of the multiple target fault samples, wherein the category confusion loss represents the degree of confusion between different fault types by the mechanical fault diagnosis model; an optimization module, used to determine the target loss according to the sample classification loss, the feature difference loss and the category confusion loss, and optimize the mechanical fault diagnosis model using the target loss to obtain a trained target mechanical fault diagnosis model.

[0015] According to another aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to implement the above method when executing the instructions stored in the memory.

[0016] According to another aspect of the present disclosure, a non-volatile computer-readable storage medium is provided, on which computer program instructions are stored, wherein the computer program instructions implement the above method when executed by a processor.

[0017] According to another aspect of the present disclosure, a computer program product is provided, including a computer-readable code, or a non-volatile computer-readable storage medium carrying the computer-readable code. When the computer-readable code runs in a processor of an electronic device, the processor in the electronic device executes the above method.

[0018] According to various aspects of the present disclosure, by training a mechanical fault diagnosis model using at least sample classification loss, feature difference loss, and category confusion loss, the feature extractor in the trained target mechanical fault diagnosis model can extract sample features with stronger domain invariance and discriminability, and the classifier can exhibit stronger classification capabilities, thereby enabling the transfer learning of the mechanical fault diagnosis model under the source working condition to the target mechanical fault diagnosis model under the target working condition, and enabling the trained target mechanical fault diagnosis model to have higher fault type diagnosis accuracy for unlabeled fault samples under the target working condition.

[0019] Further features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate exemplary embodiments, features, and aspects of the disclosure and, together with the description, serve to explain the principles of the disclosure.

[0021] Figure 1 A schematic diagram showing the comparison effect of migration training according to an embodiment of the present disclosure is shown.

[0022] Figure 2 A flowchart of a mechanical fault diagnosis model migration training according to an embodiment of the present disclosure is shown.

[0023] Figure 3 A schematic diagram showing a network structure table of a feature generator according to an embodiment of the present disclosure.

[0024] Figure 4 A schematic diagram showing an existing transfer learning training process according to the prior art.

[0025] Figure 5 A schematic diagram showing a mechanical fault vibration model migration training process according to an embodiment of the present disclosure is shown.

[0026] Figure 6 A schematic diagram showing a sample situation table of a PHM2009 data set according to an embodiment of the present disclosure.

[0027] Figure 7 A schematic diagram of an experimental result table showing the classification accuracy (%) and standard deviation of various models in the target domain on the PHM2009 dataset according to an embodiment of the present disclosure.

[0028] Figure 8 A schematic diagram of an experimental result table showing the classification accuracy (%) and standard deviation of an ablation experiment in a target domain on a PHM2009 dataset according to an embodiment of the present disclosure.

[0029] Fig. 9 A schematic diagram showing a confusion matrix of a target mechanical fault diagnosis model in a target domain trained using the training method proposed in an embodiment of the present disclosure.

[0030] Fig.10 A schematic diagram showing the confusion matrix of the DCTLN in the target domain in the prior art.

[0031] Fig.11 Schematic diagram showing the confusion matrix of the baseline model in the prior art in the target domain.

[0032] Fig.12 A schematic diagram showing a sample situation table of a CWRU2000 data set according to an embodiment of the present disclosure.

[0033] Fig.13 A schematic diagram of an experimental result table showing classification accuracy (%) and standard deviation in a target domain on a CWRU2000 dataset according to an embodiment of the present disclosure is shown.

[0034] Fig.14 A block diagram of a mechanical fault diagnosis model migration training device according to an embodiment of the present disclosure is shown.

[0035] Fig.15 A block diagram of an electronic device 1900 according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0036] Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0037] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0038] The term "and / or" herein is only a description of the association relationship of associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C can represent any one or more elements selected from the set consisting of A, B, and C. In the description of the present disclosure, "plurality" means two or more, unless otherwise clearly and specifically defined.

[0039] It should be understood that the terms "first", "second", etc. in the claims, specifications and drawings of the present disclosure are used to distinguish different objects rather than to describe a specific order. The terms "include" and "comprise" used in the specifications and claims of the present disclosure indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their collections.

[0040] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the following specific embodiments. It should be understood by those skilled in the art that the present disclosure can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present disclosure.

[0041] As mentioned above, in the field of fault diagnosis, many existing transfer learning methods mainly focus on aligning the marginal feature distributions of the source domain and the target domain. Therefore, the generated fault diagnosis model often prioritizes the invariant features of the generated domain and ignores the extraction of discriminative features of the target samples. This limitation has an adverse impact on the diagnostic performance of the generated fault diagnosis model in the target domain. Figure 1 As shown in the figure, the existing transfer learning methods in the field of fault diagnosis are aimed at Figure 1 The source fault samples and target fault samples of different domains shown in "i" can only be realized Figure 1 The classification effect shown in "ii" in the figure, that is, although the samples in different domains can be mapped to the same distribution space, because the discriminant features are not considered, the decision boundary will pass through some sample distribution areas, making the accuracy of the model in diagnosing faults not high.

[0042] In view of this, the embodiment of the present disclosure proposes a mechanical fault diagnosis model transfer training method, which aims to extract domain invariant features and focus on the discriminant properties of the features for target fault samples, so as to improve the diagnostic performance of the target mechanical fault diagnosis model after transfer training, that is, to achieve Figure 1The classification effect shown in “iii” in the figure, that is, it can map samples from different domains into the same distribution space and extract the discriminative features of the samples, so that the decision boundary is away from the dense area of ​​sample distribution, achieving more accurate fault type diagnosis.

[0043] Among them, in order to construct the mechanical fault diagnosis model transfer training method proposed in the embodiment of the present disclosure, the embodiment of the present disclosure designs a modular architecture, which mainly includes three modules, namely, a domain invariant module, a discriminant module, and a stable module, and puts forward specific requirements for each module. First, the domain invariant module is responsible for obtaining domain invariant features by aligning the edge feature distribution, and the requirement of the domain invariant module is to be able to align the edge feature distribution. Secondly, since the samples in the target domain lack labels, it is impossible to directly obtain the discriminant features for the target samples. In order to ensure that the features are not only domain invariant, but also discriminative for the target samples. Inspired by the clustering hypothesis, that is, data points located in dense areas naturally form a cluster, and samples sharing the same cluster are expected to show the same label, the discriminant module can use some unsupervised constraints to move the decision boundary away from the dense area and indirectly extract the discriminant features of the target fault sample. In other words, the requirement of the discriminant module is to be able to move the decision boundary away from the dense area. Furthermore, since some different constrained losses are added in the above-mentioned discriminant module and domain invariant module, the gradient directions of these different constrained losses cannot be guaranteed to be consistent, which may cause sudden changes in the model when predicting near certain data points and cause model crash. In order to alleviate this problem, a stabilization module is introduced to enhance the smoothness and stability of the model. Therefore, the requirement of the stabilization module is to ensure the smoothness of the model.

[0044] Next, according to the requirements of the three modules, some appropriate constraints for each module can be found to implement the mechanical fault diagnosis model transfer training method proposed in the embodiment of the present disclosure. For the requirements of the domain invariant module, a feature distribution difference measurement item is incorporated into the domain adversarial training in the mechanical fault diagnosis model transfer training method of the embodiment of the present disclosure. This combination can extract stronger domain invariant features. For the requirements of the discriminant module, the mechanical fault diagnosis model transfer training method of the embodiment of the present disclosure introduces category confusion minimization training. Intuitively speaking, the minimization of category confusion forces the classifier to show high confidence when predicting unlabeled target fault samples, thereby pushing the decision boundary away from dense areas. For the requirements of the stable module, the mechanical fault diagnosis model transfer training method proposed in the embodiment of the present disclosure introduces local Lipschitz constraints into virtual adversarial training, so that the trained target mechanical fault diagnosis model is smoother and prevents model collapse. It should be understood that each module in the above modular architecture is flexible and extensible, and any appropriate constraints in the field can be selected to constitute each module, as long as the requirements of each module are met.

[0045] In summary, in order to make the extracted sample features domain-invariant while retaining their discriminative properties, the embodiment of the present disclosure proposes a modular transfer learning architecture and puts forward specific requirements for each module. After meeting the requirements of the three modules, appropriate constraints are selected for each module to construct the mechanical fault diagnosis model transfer training method proposed in the embodiment of the present disclosure. The embodiment of the present disclosure has also been experimentally verified on a gearbox data set and a bearing fault data set. The experimental results show that the mechanical fault diagnosis model transfer training method proposed in the embodiment of the present disclosure achieves higher fault diagnosis classification accuracy. In addition, based on the above-mentioned modular architecture, the mechanical fault diagnosis model transfer training method proposed in the embodiment of the present disclosure is scalable and flexible, and any existing constraints can be replaced with other constraints according to the requirements of each module in the above-mentioned architecture.

[0046] In practical applications, the mechanical fault diagnosis model migration training method proposed in the embodiment of the present disclosure can be deployed on various terminal devices through software or hardware modification. The terminal device involved in the embodiment of the present disclosure may refer to a device with a wireless connection function and / or a wired connection function. The wireless connection function refers to the ability to connect to other devices through wireless connection methods such as wifi and Bluetooth. The terminal device involved in the embodiment of the present disclosure may also communicate with other devices through a wired connection function. The terminal device involved in the embodiment of the present disclosure may be a touch screen, a non-touch screen, or a screenless device. The touch screen can control the terminal device by clicking and sliding on the display screen with a finger or a stylus. The non-touch screen device can be connected to an input device such as a mouse, keyboard, touch panel, etc., and the terminal device can be controlled by the input device. For example, the device without a screen may be a Bluetooth speaker without a screen. For example, the terminal device of the present application may include but is not limited to user equipment (UE), mobile device, user terminal, terminal, handheld device, tablet computer, laptop computer, PDA, computing device, etc.

[0047] The mechanical fault diagnosis model migration training method of the embodiment of the present disclosure can also be deployed on a server, which can be located in the cloud or locally, and can be a physical device or a virtual device, such as a virtual machine, a container, etc., and has a wireless communication function, wherein the wireless communication function can be set in the chip (system) or other parts or components of the server. It can refer to a device with a wireless connection function, and the wireless connection function refers to the ability to connect to other servers or terminal devices through wireless connection methods such as Wi-Fi and Bluetooth. The server involved in the embodiment of the present disclosure may also have the function of communicating through a wired connection. For example, the server can receive the mechanical fault diagnosis model to be trained and the source domain sample set and the target domain sample set sent by the terminal device, and the server executes the mechanical fault diagnosis model migration training method of the embodiment of the present disclosure based on the mechanical fault diagnosis model to be trained and the source domain sample set and the target domain sample set, to obtain the trained target mechanical fault diagnosis model, and return the trained target mechanical fault diagnosis model to the terminal device, so as to feedback the trained target mechanical fault diagnosis model to the user in the terminal device.

[0048] In order to better understand the mechanical fault diagnosis model migration training method proposed in the embodiment of the present disclosure, this article introduces the relevant theoretical basis involved in the embodiment of the present disclosure as follows.

[0049] (1) Represent a domain as in represents the feature space (i.e., sample distribution space), P(X) represents the marginal probability distribution, Therefore, the source domain is represented as Similarly, the target domain is represented as Based on this, the source domain can be understood as the feature space and corresponding marginal probability distribution of labeled source fault samples collected under source conditions, and the target domain can be understood as the feature space and corresponding marginal probability distribution of unlabeled target fault samples collected under target conditions.

[0050] (2) The effectiveness of deep learning depends on the following assumptions: there is a sufficient amount of labeled data to train the diagnostic model, and the training data and test data follow the independent and identically distributed property, that is, In engineering scenarios, there are always different but related working conditions (for example, under different loads, different pressures, different speeds, etc.). ) to diagnose the types of mechanical faults that may occur under working condition A, but there is not enough labeled training data in working condition A, and working condition B (i.e. ) has enough labeled training data. In addition, due to different working conditions, the data in working condition A and the data in working condition B may not satisfy the independent and identically distributed assumption (i.e. ). Therefore, the fault diagnosis model trained with sufficient labeled data in working condition B may not be directly applied to working condition A. Transfer learning can remove this constraint to a certain extent. Transfer learning actually introduces a transfer regularization term, which prevents overfitting in the source domain using labeled data in the source domain during training by utilizing unlabeled data in the target domain. Generally, the goal of the transfer regularization term is to minimize domain differences by minimizing a metric that can measure distribution differences, so that the fault diagnosis model trained with sufficient labeled data in working condition B (i.e., source working condition) can be applied to working condition A (i.e., target working condition).

[0051] The following Figures 2 to 5 The mechanical fault diagnosis model migration training method proposed in the embodiment of the present disclosure is introduced in detail.

[0052] Figure 2 FIG. 1 is a flowchart of a mechanical fault diagnosis model migration training according to an embodiment of the present disclosure. The method can be applied to the above-mentioned terminal device or server and other electronic devices. Figure 2 As shown, the method includes: step S11 to step S16.

[0053] In step S11, a mechanical fault diagnosis model to be trained and a source domain sample set and a target domain sample set are obtained.

[0054] Among them, the mechanical fault diagnosis model to be trained includes a feature generator and a classifier. The feature generator is used to extract sample features of the fault sample, and the classifier is used to generate a fault prediction result based on the sample features. The fault prediction result includes the probability that the fault sample belongs to each fault type. It should be understood that those skilled in the art can customize the network structure of the feature generator and classifier in the mechanical fault diagnosis model according to actual needs. For example, the ResNet-18 network can be used as a feature generator, and the ResNet-18 network can also be adaptively adjusted (such as replacing the original version of the two-dimensional convolution with a one-dimensional convolution (for example, from 3x3 to 1x3), and removing the hierarchical structure after the global pooling layer); a fully connected layer of 256×the number of fault types can be used as a classifier, etc., and the embodiments of the present disclosure are not limited to this.

[0055] For example, Figure 3A network structure table of a feature generator is shown, including each layer (Layer) in the feature generator, namely convolution layer 1 (Conv1), pooling layer 1 (pool1), linear layers (L1-1 to L4-4), average pooling layer (AvgPool) and bottleneck layer (Bottleneck), the operator type (operator) of each layer, namely convolution operator (convolution), pooling operator (pooling), adaptive average pooling operator (Adaptive AvgPool), linear operator (Linear), the parameter size of the convolution kernel (paramere size), for example, "64×1×7,2,1" means that the size of the convolution kernel is 64×1×7, the step size is 2, the padding is 1, and so on, which will not be repeated.

[0056] Among them, the source domain sample set includes multiple source fault samples and sample labels corresponding to each source fault sample, and the target domain sample set includes multiple unlabeled target fault samples; the source fault samples include mechanical vibration signals collected when a mechanical fault occurs under a source working condition, and the target fault samples include mechanical vibration signals collected when a mechanical fault occurs under a target working condition, and the source working condition is different from the target working condition; the sample label is used to indicate the fault type corresponding to the source fault sample, wherein the fault type may include no fault (that is, normal operation) and various specific faults.

[0057] It should be understood that the machinery mentioned in the embodiments of the present disclosure may be any mechanical system (such as an engine, a motor, a gearbox, etc.) or a mechanical component in a mechanical system (such as a bearing, a gear, a rotor, etc.); the types of faults occurring in different machines are usually different. For example, the fault types of an engine may include no fault, a spark plug fault, a bearing fault, a piston fault, etc.; the fault types of an electrode may include no fault, a bearing fault, a rotor fault, etc.; the fault types of a gearbox may include no fault, a bearing fault, a gear fault, etc.; the fault types of a bearing may include no fault, an outer ring fault, an inner ring fault, a ball fault, etc.; the fault types of a rotor may include no fault, a broken rotor bar, a bent rotor, a misaligned rotor, etc., and the embodiments of the present disclosure do not impose any restrictions on this.

[0058] In practical applications, those skilled in the art may adopt any vibration signal acquisition method known in the art, for example, they may use sensors such as accelerometers or speedometers to acquire mechanical vibration signals, and this is not limited to the embodiments of the present disclosure. It should be understood that those skilled in the art may obtain source fault samples with sample labels by actually acquiring mechanical vibration signals when various faults occur in the machinery under source working conditions; or, they may obtain source fault samples with sample labels by simulating various faults that occur in the machinery under source working conditions through experimental simulation; for target fault samples without labels, they may obtain target fault samples without labels by actually acquiring mechanical vibration signals when various faults occur in the machinery under target working conditions, and this is not limited to the embodiments of the present disclosure.

[0059] As described above, the fault type may include no fault. Thus, optionally, a source fault sample with a sample label of no fault can be obtained by actually collecting or simulating the mechanical vibration signal of the machine during normal operation under the source working condition. The source domain sample set may also include source fault samples with a sample label of no fault and source fault samples with sample labels of various faults; and, the mechanical vibration signal of the machine during normal operation under the target working condition can be actually collected to obtain a target fault sample with no fault. The target domain sample set may also include target fault samples with no fault without labels and target fault samples of various faults.

[0060] In practical applications, the above-mentioned collected or simulated mechanical vibration signal can also be subjected to a fast Fourier transform to generate N Fourier coefficients for each mechanical vibration signal (for example, 1200 Fourier coefficients are generated). Since the coefficients are symmetrical, the first N / 2 Fourier coefficients can be used as fault samples (such as the first 600 Fourier coefficients are used as fault samples). That is, the first N / 2 Fourier coefficients converted from the mechanical vibration signal collected under the source working condition can be used as the source fault sample, and the first N / 2 Fourier coefficients converted from the mechanical vibration signal collected under the target working condition can be used as the target fault sample; of course, the N Fourier coefficients converted from the mechanical vibration signal can also be used as fault samples, and this is not limited to the embodiments of the present disclosure.

[0061] Considering that the model training will be carried out for multiple rounds, a total source domain sample library of a large number of source fault samples with sample labels and a total target domain sample library of a large number of unlabeled target fault samples can be generated, and according to the preset batch number, in each round of training, source fault samples and target fault samples that meet the batch number and are the same in number are randomly selected from the total source domain sample library and the total target domain sample library to obtain the source domain sample set and target domain sample set of the current batch, so as to use the source domain sample set and target domain sample set of the current batch to perform the current round of model training on the mechanical fault diagnosis model; it should be understood that the mechanical fault diagnosis model trained in the current round is the mechanical fault diagnosis model trained in the previous round. Among them, those skilled in the art can set the batch number and training rounds according to actual needs. For example, the batch number can be set to 120, the training rounds can be set to 100, etc., and the embodiments of the present disclosure are not limited to this.

[0062] It should be noted that the methods for obtaining source fault samples and target fault samples provided in the above-mentioned embodiments of the present disclosure are some possible implementation methods provided by the embodiments of the present disclosure. In fact, those skilled in the art can use any related technology known in the art to generate source fault samples and target fault samples, and construct source domain sample sets and target domain sample sets, and the embodiments of the present disclosure are not limited to this.

[0063] In step S12, the mechanical fault diagnosis model is used to output the sample features and fault prediction results of each source fault sample and the sample features and fault prediction results of each target fault sample based on each source fault sample in the source domain sample set and each target fault sample in the target domain sample set.

[0064] As described above, the number of samples in the source domain sample set of the current batch is the same as that in the target domain sample set. In practical applications, each source fault sample in the source domain sample set can be randomly spliced ​​together with each target fault sample in the target domain sample set to form a batch and input into the mechanical fault diagnosis model to obtain the sample features and fault prediction results of each source fault sample in the current batch, as well as the sample features and fault prediction results of each target fault sample. Of course, each source fault sample and each target fault sample can also be input into the mechanical fault diagnosis model respectively to obtain the sample features and fault prediction results of each source fault sample, as well as the sample features and fault prediction results of each target fault sample, which is not limited in the embodiments of the present disclosure. It should be understood that the sample features of each source fault sample and the sample features of each target fault sample are extracted by the feature generator in the mechanical fault diagnosis model, and the fault prediction results of each source fault sample and the fault prediction results of each target fault sample are generated by the classifier in the mechanical fault diagnosis model.

[0065] In step S13, a sample classification loss is determined based on the fault prediction results of multiple source fault samples and the sample labels of each source fault sample. The sample classification loss represents the classification error of the fault type of the source fault sample by the mechanical fault diagnosis model.

[0066] The sample classification loss aims to minimize the classification error of the source fault sample by the mechanical fault diagnosis model. The sample classification loss is used as a feature to enable the generator to obtain the basic loss of domain invariance and discriminant features. The loss function used to calculate the sample classification loss can be, for example, the SoftMax cross entropy loss function shown in formula (1):

[0067]

[0068] in, represents the sample classification loss, Represents multiple source fault samples in the source fault sample set Sample labels corresponding to each source fault sample x s represents any source fault sample, y s Representative source fault sample x s The sample label of G(x s ) represents the source fault sample x extracted by the feature generator G s The sample characteristics of C(G(x s )) represents the fault prediction result generated by classifier C, K represents the total number of fault types, and k represents the fault type; Represents the source fault sample x s When the sample label of is equal to k, the value is 1, otherwise the value is 0; Represents the corresponding expectations.

[0069] It should be noted that the above-mentioned calculation of the sample classification loss using the SoftMax cross entropy loss function is a possible implementation method provided by the embodiment of the present disclosure. In fact, those skilled in the art may also adopt other known classification loss functions in the art, and the embodiment of the present disclosure is not limited to this.

[0070] In step S14, a feature difference loss is determined according to sample features of a plurality of source fault samples and sample features of a plurality of target fault samples, where the feature difference loss represents the feature distribution difference between the sample features of the source fault samples and the sample features of the target fault samples.

[0071] Among them, the feature difference loss aims to make the sample features extracted by the feature generator domain invariant. After the cross-domain samples (i.e., source fault samples and target fault samples) are mapped to a shared high-level feature space using the feature generator, the feature distribution difference can be calculated using, for example, the multi-kernel maximum mean discrepancy (MK-MMD) loss function shown in formula (2).

[0072]

[0073] in, represents the feature difference loss, Sample features representing multiple source fault samples, The sample features representing multiple target fault samples, represent and The multi-core maximum mean difference between and The distance between.

[0074] It should be understood that the loss result obtained can be back-propagated subsequently by minimizing the feature difference loss to optimize the parameters of the feature generator, thereby promoting the reduction of the feature distribution space differences of the sample features extracted by the feature generator, that is, making the sample features extracted by the feature extractor domain invariant, or being able to map sample features of different feature distribution spaces to the same feature distribution space.

[0075] It should be noted that the above-mentioned calculation of feature difference loss using the MK-MMD loss function is a possible implementation method provided by the embodiment of the present disclosure. In fact, those skilled in the art may also adopt other maximum mean difference loss functions known in the art, such as the joint maximum mean difference (Joint Maximum Mean Discrepancy, JMMD), etc., which is not limited to the embodiment of the present disclosure.

[0076] In step S15, the category confusion loss is determined according to the fault prediction results of the multiple target fault samples, and the category confusion loss represents the degree of confusion between different fault types of the mechanical fault diagnosis model.

[0077] Among them, the category confusion loss aims to extract sample features with category discrimination (or distinguishing) of the unlabeled target fault samples, so that the classifier can accurately classify the unlabeled target fault samples based on the sample features with category discrimination; since the target fault samples lack labels, it is impossible to directly obtain discriminative sample features for the target fault samples. Therefore, the embodiment of the present disclosure introduces a clustering hypothesis, which assumes that the decision boundary should not intersect with high-density areas (i.e., areas where features are concentrated), thereby proposing the idea of ​​selecting a specific constraint term (i.e., loss term) corresponding to the clustering hypothesis. If the constraint term is minimized, the decision boundary will be away from the high-density area. In addition, the feature generator optimized by minimizing the constraint term will also extract discriminative sample features for the target fault samples.

[0078] It can be known that category confusion refers to the phenomenon that when a classifier makes a category prediction, there is confusion between the correct category and the ambiguous category. The purpose of Minimize Category Confusion (MCC) is to force the classifier to show confidence in the category prediction of the unlabeled target fault sample, thereby moving the decision boundary away from the high-density area. Since the labels of the target fault samples cannot be obtained, MCC cannot calculate the true category confusion, that is, the error matrix. However, the degree of confusion can be reflected by the inner product between the probability predicted by the classifier and its transpose, so that it can be calculated using the probability predicted by the classifier.

[0079] Exemplarily, the MCC loss function shown in formula (3) can be used to calculate the category confusion loss:

[0080]

[0081] Among them, x t represents any target fault sample, represents the target domain sample set (i.e., a batch containing B target fault samples), K represents the total number of fault types, j and j′ represent the jth fault type and the j′th fault type respectively, and j′≠j, C jj′ represents the degree of confusion between each pair of the jth fault type and the j′th fault type, and is used to evaluate and measure the confusion between different categories. The transpose of the matrix representing the probability that each target fault sample among the B target fault samples in each batch is assigned to the jth fault type, The matrix representing the probability that each target fault sample in the B target fault samples in each batch is assigned to the j′th fault type. W represents the importance of each target fault sample in the target domain sample set, which can be quantified as a probability weight, C jj′W is used in order to obtain a probability distribution that assigns higher probabilities to samples that are more confident in their category predictions. W is represented by the corresponding diagonal matrix. The importance of the i-th target fault sample in W is W ii It can be calculated using formula (4):

[0082]

[0083] in, represents the fault prediction result of the i-th target fault sample among B target fault samples in each batch, represents the fault prediction result of the i′th target fault sample among the B target fault samples in each batch, exp represents an exponential function with a natural constant as the base, and H represents entropy, which can measure uncertainty. It can be expressed as formula (5-1), It can be expressed as formula (5-2):

[0084]

[0085] in, represents the probability that the i-th target fault sample in the fault prediction result of the i-th target fault sample belongs to the j-th fault type, Represents the probability that the i′th target fault sample in the fault prediction result of the i′th target fault sample belongs to the jth fault type.

[0086] It should be noted that the above-mentioned use of the MCC loss function to calculate the category confusion loss is a possible implementation method provided by the embodiment of the present disclosure. In fact, those skilled in the art may also adopt other category confusion loss functions known in the art, and the embodiment of the present disclosure is not limited to this.

[0087] In step S16, the target loss is determined according to the sample classification loss, the feature difference loss and the category confusion loss, and the mechanical fault diagnosis model is optimized using the target loss to obtain a trained target mechanical fault diagnosis model.

[0088] In practical applications, the sample classification loss, feature difference loss and category confusion loss can be accumulated to obtain the target loss. Alternatively, the sample classification loss, feature difference loss and category confusion loss can be weighted and summed using a weighting coefficient to obtain the target loss. For example, the target loss can be expressed as formula (6):

[0089]

[0090] Wherein, λ represents a weighting coefficient, and λ can be gradually adjusted from 0 to 1 according to λ=2 / (1+exp(-10·p))-1, where p represents the current training round.

[0091] Among them, the target loss is used to optimize the mechanical fault diagnosis model. For example, the parameters in the mechanical fault diagnosis model can be optimized by using a model optimization method such as back propagation based on the target loss. Among them, the feature difference loss It can mainly optimize the parameters of the feature generator and the category classification loss and class confusion loss It can be used to optimize the parameters of the classifier and feature generator. It should be understood that those skilled in the art can use any known model optimization method in the art to optimize the mechanical fault diagnosis model using target loss, and the embodiments of the present disclosure are not limited to this.

[0092] As described above, the training of the mechanical fault diagnosis model can be performed in multiple rounds, and the mechanical fault diagnosis model optimized in the current round can be used as the mechanical fault diagnosis model trained in the next round. Those skilled in the art can set the training end conditions, such as the target loss converges or is set to 0, the iterative rounds reach a preset total number of rounds, etc., so as to use the mechanical fault diagnosis model optimized when the training end conditions are met as the target mechanical fault diagnosis model after training. The embodiments of the present disclosure are not limited to this.

[0093] According to the training method of the embodiment of the present disclosure, by training the mechanical fault diagnosis model using at least sample classification loss, feature difference loss and category confusion loss, the feature extractor in the trained target mechanical fault diagnosis model can extract sample features with domain invariance and discriminability, and the classifier can exhibit stronger classification ability, thereby realizing the transfer learning of the mechanical fault diagnosis model under the source working condition to the target mechanical fault diagnosis model under the target working condition, and the trained target mechanical fault diagnosis model has higher fault type diagnosis accuracy for unlabeled fault samples under the target working condition.

[0094] In order to further enhance the domain invariance of the sample features extracted by the feature generator and align the distribution of edge features, in a possible implementation, the domain invariance of the sample features extracted by the feature generator may be improved by referring to a domain adversarial training method. Specifically, the method may further include:

[0095] Step S17, inputting the sample features of multiple source fault samples and the sample features of multiple target fault samples into the domain discriminator respectively, obtaining the domain classification results of each source fault sample and the domain classification results of each target fault sample, the domain classification results representing the probability that the sample features belong to the feature distribution space corresponding to the source domain sample set, or representing the probability that the sample features belong to the feature distribution space corresponding to the target domain sample set;

[0096] Step S18, determining the domain adversarial loss according to the domain classification results of each source fault sample and the domain classification results of each target fault sample, where the domain adversarial loss represents the classification error of the domain discriminator on the feature distribution space to which the sample features belong.

[0097] It should be understood that those skilled in the art may adopt a domain discriminator known in the art, or may customize the network structure of the domain discriminator. For example, a domain discriminator composed of multiple hidden fully connected layers may be adopted. For example, a domain discriminator mainly comprising 3 hidden fully connected layers may be adopted. The convolution kernel sizes of each hidden fully connected layer may be 256×1024, 1024×1024, and 1024×2, respectively. The embodiments of the present disclosure are not limited to this.

[0098] Among them, the domain discriminator takes the sample features extracted by the feature generator as input and distinguishes these sample features as source domain or target domain. It can be seen that domain adversarial training can be divided into two steps. First, the feature generator G is fixed, that is, the feature generator G is not optimized, and the sample features G(x s ) or the sample feature G(x t ) is regarded as the input of the domain discriminator D, the classification error of the domain discriminator D is minimized to optimize the domain classifier D, and then the domain discriminator D is fixed, and the classification error of the domain discriminator D is maximized to optimize the feature generator G. Intuitively, if a good domain discriminator cannot distinguish the sample features extracted by the feature generator as the source domain or the target domain, then these sample features can be considered to be almost domain invariant. Among them, due to the existence of the gradient reversal layer, these two steps can actually be further simplified. Therefore, the domain adversarial training loss function shown in (7-1) can be used to calculate the above domain adversarial loss:

[0099]

[0100] in, represents the domain adversarial loss, sup represents the supremum function, which can be understood as the maximization function. Represents finding a domain discriminator such that Reach the supremum; Represents the expectation of the domain classification results of each source fault sample in the source domain sample set, represents the expected domain classification results of each target fault sample in the target domain sample set, D(G(x s )) represents the probability that the sample feature of the source fault sample belongs to the feature distribution space corresponding to the source domain sample set (that is, the probability that the sample feature belongs to the source domain), D(G(x t)) represents the probability that the sample feature of the target fault sample belongs to the feature distribution space corresponding to the source domain sample set. For example, it can be set to a probability greater than 0.5 to belong to the source domain sample set, otherwise it belongs to the target domain sample set.

[0101] As mentioned above, domain adversarial training is divided into two steps. First, the feature generator G is fixed to optimize the domain classifier D. Based on the above formula (7-1), it can be considered that it can maximize is a good domain discriminator, where the first term is maximized It is to make the domain discriminator better classify the source fault samples (the first sample comes from the source domain sample set, ), because the second There is a minus sign in , so maximizing the second term is actually minimizing the second term, that is, making the probability of the target fault sample from the target domain sample set being classified as belonging to the source domain sample set as small as possible, so as to obtain a good domain discriminator. The second step is to find a feature generator that can minimize the domain adversarial loss, that is, to optimize the feature generator so that the domain discriminator obtained in the previous step cannot determine whether the sample feature comes from the source domain or the target domain. In this way, the sample features generated by the feature generator can have domain invariance.

[0102] As mentioned above, the domain classification result can also characterize the probability that the sample feature belongs to the feature distribution space corresponding to the target domain sample set. Therefore, the domain adversarial training loss function shown in formula (7-2) can be used to calculate the above domain adversarial loss:

[0103]

[0104] It should be understood that the explanation of formula (7-2) can refer to the above formula (7-1) and will not be repeated here.

[0105] It should be noted that using the above formula (7-1) or formula (7-2) to calculate the domain adversarial loss is a possible implementation method provided by the embodiment of the present disclosure. In fact, those skilled in the art may also adopt other domain adversarial training loss functions known in the art, and the embodiment of the present disclosure is not limited to this.

[0106] Based on the domain adversarial loss, the step S16 determines the target loss according to the sample classification loss, feature difference loss, and category confusion loss, and may include: determining the target loss according to the sample classification loss, feature difference loss, category confusion loss, and domain adversarial loss. The target loss may then be used to train a mechanical fault diagnosis model to obtain a trained target mechanical fault diagnosis model.

[0107] In practical applications, the sample classification loss, feature difference loss, class confusion loss and domain adversarial loss can be accumulated to obtain the target loss. Alternatively, the sample classification loss, feature difference loss, class confusion loss and domain adversarial loss can be weighted and summed using a weighting coefficient to obtain the target loss. For example, the target loss can be expressed as formula (8):

[0108]

[0109] As mentioned above, λ represents a weighting coefficient, and λ can be gradually adjusted from 0 to 1 according to λ=2 / (1+exp(-10·p))-1, where p represents the current training round.

[0110] Then, the target loss shown in formula (8) can be used to train the mechanical fault diagnosis model to obtain the trained target mechanical fault diagnosis model. Domain adversarial loss You can mainly optimize the parameters of the feature generator and the category classification loss and class confusion loss It can be used to optimize the parameters of the classifier and feature generator. It should be understood that those skilled in the art can use any known model optimization method in the art to optimize the mechanical fault diagnosis model using target loss, and the embodiments of the present disclosure are not limited to this.

[0111] According to the training method of the embodiment of the present disclosure, by training the mechanical fault diagnosis model using at least sample classification loss, feature difference loss, category confusion loss and domain adversarial loss, the feature extractor in the trained target mechanical fault diagnosis model can extract sample features with stronger domain invariance and discriminability, and the classifier can exhibit stronger classification ability, thereby realizing the transfer learning of the mechanical fault diagnosis model under the source working condition to the target mechanical fault diagnosis model under the target working condition, and enabling the trained target mechanical fault diagnosis model to have higher fault type diagnosis accuracy for unlabeled fault samples under the target working condition.

[0112] Considering that directly using the above sample classification loss, feature difference loss, category confusion loss and domain adversarial loss to train the mechanical fault diagnosis model may cause the model to crash, in order to reduce the risk of model crash, virtual adversarial loss can be introduced as an additional constraint item added to the target loss, which helps to make the model training smoother. Virtual adversarial training can act as a lubricant in the entire model training. Virtual adversarial training can be applied to the target domain sample set and / or the source domain sample set. Specifically, when applied to the source domain sample set, the method may also include:

[0113] Step S19, adding a disturbance signal to each source fault sample in the source domain sample set to obtain a source fault disturbance sample corresponding to each source fault sample;

[0114] Step S20, inputting the source fault disturbance samples corresponding to each source fault sample into the mechanical fault diagnosis model to obtain the fault prediction results of the source fault disturbance samples corresponding to each source fault sample;

[0115] Step S21, determining a first virtual adversarial loss according to the fault prediction results of multiple source fault samples and the fault prediction results of source fault disturbance samples corresponding to each source fault sample, wherein the first virtual adversarial loss represents the difference between the fault prediction result of the source fault sample and the fault prediction result of the corresponding source fault disturbance sample.

[0116] Among them, the disturbance signal can be a smaller vibration signal, which can be selected from a smaller signal range and randomly added to each source fault sample in the source domain sample set to generate a source fault disturbance sample corresponding to each source fault sample, and then the mechanical fault diagnosis model can be used to output the fault prediction result of the source fault disturbance sample corresponding to each source fault sample; it should be understood that the difference between the source fault sample and the source fault disturbance sample after the source fault sample is disturbed is small, or the fault prediction results of the two are similar or the same (that is, the distance is close), thus, the mechanical fault diagnosis model can be trained by the first virtual adversarial loss, which can make the mechanical fault diagnosis model more stable and less prone to collapse, avoiding the model from causing a large change in the fault prediction result just because a little disturbance is added to the sample, or, if the mechanical fault diagnosis model can classify the source fault sample and the source fault disturbance sample with a smaller disturbance added into the same category, then the mechanical fault diagnosis model is considered to be a smooth model with a lower risk of collapse.

[0117] Exemplarily, the virtual adversarial training loss function shown in formula (9) can be used to calculate the first virtual adversarial loss:

[0118]

[0119] in, represents the first virtual adversarial loss, x s Representative source fault samples, represents the source fault sample set, r represents the disturbance signal, ∥r∥≤∈ represents that the disturbance signal is less than a small threshold ∈, x s +r represents the source fault disturbance sample corresponding to the source fault sample, G(x s +r) represents the sample characteristics of the source fault disturbance sample, C(G(x s +r)) represents the fault prediction result of the source fault disturbance sample, D KL represents the KL divergence function, DKL (C(G(x s ))∥C(G(x s +r))) is equivalent to calculating C(G(x s )) and C(G(x s +r)), It can be understood as finding a disturbance signal r that makes D KL maximum, Represents the calculation of each source fault sample corresponding to Among them, the use of Training the model enables the classifier in the model to ensure consistency within the norm ball neighborhood of each source fault sample.

[0120] It should be noted that using the above formula (9) to calculate the first virtual adversarial loss is a possible implementation method provided by the embodiment of the present disclosure. In fact, those skilled in the art may also adopt other known virtual adversarial training loss functions in the art, and the embodiment of the present disclosure does not limit this.

[0121] Based on the first virtual adversarial loss, the step S16 determines the target loss according to the sample classification loss, the feature difference loss and the category confusion loss, and may further include:

[0122] Determine the target loss based on the sample classification loss, feature difference loss, category confusion loss, and the first virtual adversarial loss; or,

[0123] Determine the target loss based on sample classification loss, feature difference loss, category confusion loss, domain adversarial loss, and first virtual adversarial loss;

[0124] As mentioned above, the domain adversarial loss is determined based on the domain classification results of each source fault sample and the domain classification results of each target fault sample.

[0125] In practical applications, the sample classification loss, feature difference loss, category confusion loss and the first virtual adversarial loss can be accumulated to obtain the target loss, or the sample classification loss, feature difference loss, category confusion loss and the first virtual adversarial loss can be weighted and summed using a weighting coefficient to obtain the target loss; thus, the target loss can also be expressed as formula (10-1):

[0126]

[0127] Alternatively, the sample classification loss, feature difference loss, category confusion loss, domain adversarial loss and first virtual adversarial loss can be accumulated to obtain the target loss. Alternatively, the sample classification loss, feature difference loss, category confusion loss, domain adversarial loss and first virtual adversarial loss can be weighted and summed using a weighting coefficient to obtain the target loss. Thus, the target loss can also be expressed as formula (10-2):

[0128]

[0129] Then, the target loss shown in formula (10-1) or formula (10-2) can be used to train the mechanical fault diagnosis model to obtain the trained target mechanical fault diagnosis model. Domain adversarial loss You can mainly optimize the parameters of the feature generator and the category classification loss Class Confusion Loss and the first virtual adversarial loss It can be used to optimize the parameters of the classifier and feature generator. It should be understood that those skilled in the art can use any known model optimization method in the art to optimize the mechanical fault diagnosis model using target loss, and the embodiments of the present disclosure are not limited to this.

[0130] According to the training method of the embodiment of the present disclosure, by training the mechanical fault diagnosis model using at least sample classification loss, feature difference loss, category confusion loss, domain adversarial loss and first virtual adversarial loss, the feature extractor in the trained target mechanical fault diagnosis model can extract sample features with stronger domain invariance and discriminability, and the classifier can exhibit stronger classification ability, and the target mechanical fault diagnosis model can be made more stable, smooth and less prone to collapse, thereby realizing the transfer learning of the mechanical fault diagnosis model under the source working condition to the target mechanical fault diagnosis model under the target working condition, and the trained target mechanical fault diagnosis model has higher fault type diagnosis accuracy for unlabeled fault samples under the target working condition.

[0131] As described above, virtual adversarial training may also be applied to the target domain sample set. Thus, in a possible implementation, the method may further include:

[0132] Step S22, adding a disturbance signal to each target fault sample in the target domain sample set to obtain a target fault disturbance sample corresponding to each target fault sample;

[0133] Step S23, inputting the target fault disturbance samples corresponding to each target fault sample into the mechanical fault diagnosis model to obtain the fault prediction results of the target fault disturbance samples corresponding to each target fault sample;

[0134] Step S24, determining a second virtual adversarial loss based on the fault prediction results of multiple target fault samples and the fault prediction results of target fault disturbance samples corresponding to each target fault sample, wherein the second virtual adversarial loss represents the difference between the fault prediction result of the target fault sample and the fault prediction result of the corresponding target fault disturbance sample.

[0135] Among them, the disturbance signal can be a smaller vibration signal, which can be selected from a smaller signal range and randomly added to each target fault sample in the target domain sample set to generate a target fault disturbance sample corresponding to each target fault sample, and then the mechanical fault diagnosis model can be used to output the fault prediction result of the target fault disturbance sample corresponding to each target fault sample; it should be understood that the difference between the target fault sample and the target fault disturbance sample after the target fault sample is disturbed is small, or the fault prediction results of the two are similar or the same (that is, the distance is close), thus, the mechanical fault diagnosis model can be trained by using the second virtual adversarial loss, which can make the mechanical fault diagnosis model more stable and less prone to collapse, avoiding the model from causing a large change in the fault prediction result just because a little disturbance is added to the sample, or, if the mechanical fault diagnosis model can classify the target fault sample and the target fault disturbance sample with added disturbance into the same category, then the mechanical fault diagnosis model is considered to be a smooth model with a lower risk of collapse.

[0136] Exemplarily, the second virtual adversarial loss may be calculated using the virtual adversarial loss function shown in formula (11):

[0137]

[0138] in, represents the second virtual adversarial loss, x t represents the target fault sample, represents the target fault sample set, r represents the disturbance signal, ∥r∥≤∈ represents that the disturbance signal is less than a small threshold v, x t +r represents the target fault disturbance sample corresponding to the target fault sample, G(x t +r) represents the sample characteristics of the target fault disturbance sample, C(G(x t +r)) represents the fault prediction result of the target fault disturbance sample, D KL represents the KL divergence function, D KL (C(G(x t ))∥C(G(x t +r))) is equivalent to calculating C(G(x t )) and C(G(x t +r)), It can be understood as finding a disturbance signal r that makes DKL maximum, Represents the calculation of each target fault sample corresponding to Among them, the use of The model is trained to make the classifier consistent within the norm ball neighborhood of each target fault sample.

[0139] It should be noted that using the above formula (11) to calculate the second virtual adversarial loss is a possible implementation method provided by the embodiment of the present disclosure. In fact, those skilled in the art may also adopt other known virtual adversarial training loss functions in the art, and the embodiment of the present disclosure does not limit this.

[0140] Based on the second virtual adversarial loss, the step S16 determines the target loss according to the sample classification loss, the feature difference loss and the category confusion loss, including:

[0141] Determine the target loss based on the sample classification loss, feature difference loss, category confusion loss, and the second virtual adversarial loss; or,

[0142] Determine the target loss based on the sample classification loss, feature difference loss, category confusion loss, domain adversarial loss, and second virtual adversarial loss; or,

[0143] Determine the target loss according to the sample classification loss, the feature difference loss, the category confusion loss, the first virtual adversarial loss, and the second virtual adversarial loss; or,

[0144] Determine the target loss based on sample classification loss, feature difference loss, category confusion loss, domain adversarial loss, first virtual adversarial loss, and second virtual adversarial loss;

[0145] As described above, the domain adversarial loss is determined based on the domain classification results of each source fault sample and the domain classification results of each target fault sample, and the first virtual adversarial loss is determined based on the fault prediction results of multiple source fault samples and the fault prediction results of the source fault disturbance samples corresponding to each source fault sample.

[0146] In practical applications, the sample classification loss, feature difference loss, category confusion loss and the second virtual adversarial loss can be accumulated to obtain the target loss, or the sample classification loss, feature difference loss, category confusion loss and the second virtual adversarial loss can be weighted and summed using a weighting coefficient to obtain the target loss; thus, the target loss can also be expressed as formula (12-1):

[0147]

[0148] Alternatively, the sample classification loss, feature difference loss, category confusion loss, domain adversarial loss and the second virtual adversarial loss can be accumulated to obtain the target loss, or the sample classification loss, feature difference loss, category confusion loss, domain adversarial loss and the second virtual adversarial loss can be weighted and summed using a weighting coefficient to obtain the target loss; thus, the target loss can also be expressed as formula (12-2):

[0149]

[0150] Alternatively, the sample classification loss, feature difference loss, category confusion loss, first virtual adversarial loss and second virtual adversarial loss may be accumulated to obtain the target loss. Alternatively, the sample classification loss, feature difference loss, category confusion loss, first virtual adversarial loss and second virtual adversarial loss may be weighted and summed using a weighting coefficient to obtain the target loss. Thus, the target loss may also be expressed as formula (12-3):

[0151]

[0152] Alternatively, the sample classification loss, feature difference loss, category confusion loss, domain adversarial loss, first virtual adversarial loss and second virtual adversarial loss may be accumulated to obtain the target loss. Alternatively, the sample classification loss, feature difference loss, category confusion loss, domain adversarial loss, first virtual adversarial loss and second virtual adversarial loss may be weighted and summed using a weighting coefficient to obtain the target loss. Thus, the target loss may also be expressed as formula (12-4):

[0153]

[0154] Then, any target loss in the above formula (12-1) to formula (12-4) can be used to train the mechanical fault diagnosis model to obtain the trained target mechanical fault diagnosis model. Domain adversarial loss You can mainly optimize the parameters of the feature generator and the category classification loss Class Confusion Loss First virtual confrontation loss and the second virtual adversarial loss It can be used to optimize the parameters of the classifier and feature generator. It should be understood that those skilled in the art can use any known model optimization method in the art to optimize the mechanical fault diagnosis model using target loss, and the embodiments of the present disclosure are not limited to this.

[0155] According to the training method of the embodiment of the present disclosure, by using sample classification loss, feature difference loss, category confusion loss, domain adversarial loss, first virtual adversarial loss and / or second virtual adversarial loss to train the mechanical fault diagnosis model, the feature extractor in the trained target mechanical fault diagnosis model can extract sample features with stronger domain invariance and discriminability, and the classifier can exhibit stronger classification ability, and the target mechanical fault diagnosis model can be made more stable, smooth and not prone to collapse, thereby realizing the transfer learning of the mechanical fault diagnosis model under the source working condition to the target mechanical fault diagnosis model under the target working condition, and the trained target mechanical fault diagnosis model has higher fault type diagnosis accuracy for unlabeled fault samples under the target working condition.

[0156] It can be seen that in the existing fault diagnosis field, many existing popular transfer methods can generally be summarized into two categories. The first category focuses on aligning marginal feature distributions and generally includes four main steps. First, a feature generator is used to map cross-domain data to a shared feature space. Subsequently, various distance metrics are used to quantify the differences in marginal feature distributions. In addition, a classifier is used to classify the labeled source features. Finally, the classification results and distribution differences are back-propagated to optimize the model parameters. The second category has a similar architectural framework to the first category, with the only difference being the second step. In the second category, the second step acts as a transfer regularization term to measure the difference in marginal feature distributions by distinguishing the loss of the domain classifier of the source and target features. These two methods can also be combined to better align marginal distributions and extract more domain-invariant features. These methods achieve the same goal, namely aligning marginal feature distributions, by different means, such as Figure 4 The following is an existing transfer learning training process. Figure 4 As shown, for source fault samples (source) and target fault samples (target), the prior art uses the sample features extracted by the feature generator and the classifier to calculate the domain adversarial loss (domain adversarial) or MK-MMD loss to optimize the feature generator, and calculates the cross entropy loss (cross entropy) based on the fault prediction results output by the classifier to optimize the feature generator and the classifier.

[0157] Based on the migration training method proposed in the above-mentioned embodiment of the present disclosure, the embodiment of the present disclosure can achieve the following Figure 5The training process shown is used to train the feature generator and classifier in the mechanical fault diagnosis model. Among them, MK-MMD is selected to calculate the feature difference loss, the domain discriminator is used to calculate the domain adversarial loss, MCC is used to calculate the category confusion loss, and the virtual adversarial training loss function (Virtual Adversarial Training, VAT) is selected to calculate the virtual adversarial loss. Specifically, the feature generator G is responsible for mapping cross-domain data into high-level features (i.e., extracting sample features of fault samples), and using MK-MMD to calculate the feature difference loss based on the sample features of the source fault samples and the sample features of the target fault samples. And the domain classification results of the sample features of the source fault sample and the sample features of the target fault sample generated by the domain discriminator are used to calculate the domain adversarial loss, that is, Using feature difference loss Domain adversarial loss The feature generator is optimized by back propagation to measure the difference in marginal feature distribution using feature difference loss and domain adversarial loss. The classifier C runs on the source domain and the target domain, takes the features extracted by the feature generator as input, and outputs the fault prediction result. For the fault prediction result of the source fault sample, the cross entropy loss is calculated to obtain the category classification loss. And the first virtual adversarial loss And can use category classification loss And the first virtual adversarial loss Back propagation is used to optimize the classifier and feature generator to ensure that the classifier remains consistent within the canonical sphere neighborhood of each source fault sample and reduce the classification error of the source fault sample. For the fault prediction results of the target fault sample, MCC can be used to calculate the category confusion loss And use VAT to calculate the second virtual confrontation failure And use the class confusion loss And the second virtual adversarial training loss The back-propagation method is used to optimize the classifier and feature generator to ensure that the classifier remains consistent within the canonical sphere neighborhood of each target fault sample and to improve the classification accuracy of the target fault samples.

[0158] Based on the mechanical vibration model training method proposed in the above embodiment of the present disclosure, the embodiment of the present disclosure also provides a mechanical fault diagnosis method, including:

[0159] Step S31, acquiring fault data actually collected when a mechanical fault occurs under a target working condition, the fault data including a mechanical vibration signal;

[0160] Step S32, using the target mechanical fault diagnosis model to determine the fault diagnosis result corresponding to the fault data based on the fault data, the fault diagnosis result characterizes the fault type of the mechanical failure under the target working condition; wherein the target mechanical fault diagnosis model is trained by the training method.

[0161] In step S31, those skilled in the art may adopt a vibration signal acquisition method known in the art to acquire a mechanical vibration signal when a mechanical failure occurs under a target working condition as fault data. As described above, the mechanical vibration signal may also be subjected to a fast Fourier transform to obtain N or N / 2 Fourier coefficients as fault data, which is not limited to this embodiment of the present disclosure.

[0162] In step S32, the fault data can be input into the target mechanical fault diagnosis model, and the probability of the fault data belonging to various fault types can be output. Then, the fault type with the highest probability can be selected as the fault diagnosis result corresponding to the mechanical fault, which is not limited to this embodiment of the present disclosure.

[0163] According to the diagnostic method of the embodiment of the present disclosure, it is possible to use the target machine fault diagnosis model to obtain an accurate fault type diagnosis when a machine fault occurs under the target working condition.

[0164] In order to intuitively express the beneficial effects of the mechanical fault diagnosis model transfer training method proposed in the embodiment of the present disclosure, the embodiment of the present disclosure also conducted experimental verification on the PHM2009 dataset and the CWRU2000 dataset for the transfer training method proposed in the embodiment of the present disclosure and some existing transfer learning methods. The experimental process and experimental results are described below.

[0165] Among them, the fault data in the PHM2009 dataset (PHM09, 2009) is a typical industrial gearbox data sample. Among them, the gearbox configuration includes two gear types: spur gears and spiral cut (threaded) gears. The gearbox consists of three shafts, four gears and six bearings. Synchronous data acquisition is adopted, and the accelerometer is mounted on the retaining plates of the input and output shafts. In addition, a tachometer is used to generate 10 pulses per revolution to provide accurate zero-crossing information. Data acquisition is carried out at a sampling frequency of 66.67kHz, capturing vibration signals at shaft speeds of 30Hz, 35Hz, 40Hz, 45Hz and 50Hz under high load and low load conditions. The experimental dataset contains six different fault types, each of which originates from different component failures within the gearbox. As Figure 6The sample case table of the PHM2009 dataset shown in the figure can be used to evaluate the model performance by using the vibration signals of the helical gearbox running at 30 Hz, 35 Hz, 40 Hz and 45 Hz under high load conditions. The four sample sets (i.e., four domains) corresponding to these four frequencies can be marked as A, B, C and D respectively. Under different working conditions, the number of samples of each fault type in the six fault types (i.e., No. of Category 0 to 5) at each load frequency of 30 Hz, 35 Hz, 40 Hz and 45 Hz is 222, that is, the dataset contains a total of 5328 data samples, each sample length is 1200 data points.

[0166] Based on the above PHM2009 data set, the embodiment of the present disclosure evaluated the target mechanical fault diagnosis model (i.e., Proposed) obtained by the training method proposed in the embodiment of the present disclosure in 12 transfer fault diagnosis experiments, and compared it with the diagnosis models obtained by other nine existing transfer learning methods (i.e., CDAN (Conditional Adversarial Domain Adaptation), JAN (Joint Adaptation Network), DCTLN (Deep Convolutional Transfer Learning Network), DANN (Domain-Adversarial Neural Network), DAN (Deep Adaptation Network), MCD (Maximum Classifier Discrepancy), MDD (Maximum Classifier Discrepancy), CORAL (Correlation Alignment), and baseline model Basis), and obtained the following results: Figure 7 The experimental results table of model classification accuracy (%) and standard deviation is shown in the figure. Figure 7 As shown in the figure, o→p represents a migration fault diagnosis experiment from the source domain sample set o to the target domain sample set p, o≠p, o,p∈A,B,C,D, for example, A→B represents the migration from the source domain sample set A to the target domain sample set B. Among them, the training dataset includes samples from the source domain sample set and the target domain sample set, while the test dataset can only include samples from the target domain sample set. For fair comparison, the model structures adopted by different methods remain the same, that is, the same network structure can be used to construct the feature generators and classifiers of different methods. Under the above configuration, each migration fault diagnosis experiment can be repeated five times to ensure reliable results and robustness.

[0167] As described above, by comparing the classification accuracy achieved by the training method proposed in this public embodiment and the models trained by the above nine existing migration methods in the target domain to evaluate their performance, it can be obtained that Figure 7 The experimental results table is shown, in which ↑ indicates how much higher the accuracy of the training method of the embodiment of the present disclosure is than the second place transfer method, and ↓ indicates how much lower the accuracy of the training method of the embodiment of the present disclosure is than the first place transfer method. The baseline model refers to a convolutional neural network (CNN) trained only on labeled data in the source domain. DANN obtains good transfer performance through a domain discriminator. DAN relies on the MK-MMD distance metric. MDD and DCORAL also rely on some very effective metrics for measuring distribution differences. DCTLN further incorporates MMD into domain adversarial training and better aligns marginal feature distributions. CDAN and JAN are also methods that attempt to obtain discriminative features for target samples. MCD enhances the discrimination performance using two classifiers. As Figure 7 As shown, the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure achieved competitive results. In 9 out of 12 experiments, the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure exceeded other existing methods in terms of migration fault diagnosis accuracy. This means that even in the absence of labeled data, the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure can effectively identify the fault condition of the gearbox.

[0168] Compared with the baseline, not only the training method proposed in the embodiment of the present disclosure, but also all other popular transfer methods perform well in classification accuracy. The key difference between the transfer method and the baseline model trained only by the source data is that the transfer regularization utilizes the labeled data of the source domain and the unlabeled data of the target domain. Transfer regularization is generally used to align marginal feature distributions by minimizing the metric that measures the distribution difference. In the training method proposed in the embodiment of the present disclosure, transfer regularization refers to the domain invariant module, more specifically, it is a combination of feature difference loss and domain adversarial loss. As the results show, due to the addition of transfer regularization, transfer learning can do what traditional deep learning cannot do, and further promote intelligent mechanical fault diagnosis.

[0169] For methods such as DAN, DANN, DCORAL, MDD and DCTLN, these methods are first analyzed. DAN through MK-MMD and DANN through domain discriminator are popular migration methods, and according to the results, their performance is similar. Both DCORAL and MDD aim to obtain domain invariant features by reducing distribution differences, but their performance is generally average. DCTLN can actually be regarded as a combination of DAN and DANN, and performs better in most migration tasks. Even with the integration of DAN and DANN, DCTLN still focuses on aligning the marginal feature distribution and does not solve the problem of extracting target distinguishing features. Fundamentally, these methods are only equivalent to the domain invariant module in the training method proposed in the embodiment of the present disclosure. However, the training method proposed in the embodiment of the present disclosure not only includes the domain invariant module, but also integrates the discriminant module and the stabilization module. The category confusion loss in the discriminant module helps to extract discriminative features for the target sample, while the virtual adversarial loss in the stabilization module is responsible for ensuring the smoothness of the model. As shown in the results, the training method proposed in the embodiment of the present disclosure beat the existing methods in 9 out of 12 experiments. To some extent, this means that simply aligning marginal feature distributions is far from enough and MK-MCC and VAT have accomplished their intended work goals.

[0170] Compared with CDAN, JAN and MCD, the training method proposed in the embodiment of the present disclosure performs better, although these three methods also attempt to obtain discriminative features. The training architecture of the training method proposed in the embodiment of the present disclosure consists of three modules, and specific requirements are designed for each module. In this case, the most appropriate constraints can be customized or selected according to the requirements of each module. However, these three methods are not modular and cannot reach a level comparable to the training method proposed in the embodiment of the present disclosure in all three aspects (domain invariance, distinguishability and stability). This also explains to a certain extent why the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure performs better than CDAN, JAN and MCD.

[0171] In order to verify whether the virtual adversarial loss and the category confusion loss really work, the disclosed embodiment also conducted a comprehensive ablation analysis and obtained the following results: Figure 8 The classification accuracy (%) and standard deviation of the ablation experiment in the target domain on the PHM2009 dataset are shown in Figure 8As shown, ↑ in the third and fourth columns indicates an improvement in accuracy compared to the second column, and ↓ indicates a decrease. The addition of virtual adversarial loss is indicated by "+VAT", and the addition of category confusion loss is indicated by "+MCC". The experimental results show that DANN+MK-MMD+MCC is sufficient to outperform DANN+MK-MMD in eight out of twelve experiments. MCC further improves the ability of DANN+MK-MMD, which proves the effectiveness of MCC. In addition, when VAT is integrated into the model, the performance is further improved. Ultimately, in eleven of the twelve experiments, both VAT and MCC played an important role in achieving the best performance. In general, the ablation experiments further verified the effectiveness of each loss in the training method proposed in the embodiments of the present disclosure.

[0172] The disclosed embodiment also provides a visualization diagram to illustrate the influence of extracting the discriminative features of the target domain. Specifically, the migration fault diagnosis experiment C→A is used as an example, and the results are presented in Fig. 9 , Fig.10 and Fig.11 In the confusion matrix shown. Among them, Fig. 9 , Fig.10 and Fig.11 The confusion matrices of the target mechanical fault diagnosis model, DCTLN, and baseline model trained by the training method proposed in the embodiment of the present disclosure in the target domain are shown respectively. In the confusion matrix, the six numbers on the diagonal represent different results in six fault types. The larger the number and the darker the color, the closer the fault type predicted by the model is to the actual distribution of the actual fault type. By comparing Fig. 9 , Fig.10 and Fig.11 , it can be found that the baseline model cannot distinguish different categories well in the target domain, which indicates that the baseline cannot extract discriminative features for the target samples, but the migration method can work well, and the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure shows a stronger classification performance, thus verifying its ability to extract discriminative features more effectively.

[0173] In the experiment on the CWRU2000 dataset, the embodiments of the present disclosure used bearing data obtained from the CWRU2000 dataset. The bearing data was obtained from a mechanical system driven by an electric motor, and the bearing type used in the experiment was SKF6205. The system operated under four different loads with a sampling frequency of 12kHz. The bearing dataset includes four different fault conditions: (1) normal condition (NC), (2) outer race fault (OF), (3) inner race fault (IF), and (4) roller fault (RF). Each fault condition is associated with a specific fault diameter: IF is 0.18 mm, OF is 0.36 mm, and RF is 0.54 mm. As Fig.12 A sample situation table of a CWRU2000 data set is shown. The fault samples are obtained at three different load levels (0Hp, 1Hp and 2Hp), forming three sample sets A, B and C. Sample sets A, B and C contain 10 different fault types (i.e., No. of Category 0 to 9), respectively, and the number of samples of each fault type in each sample set when the load is 0Hp, 1Hp and 2Hp is also shown. For example, the number of samples of No. of Category 0 in sample set A is 203, and so on. Each sample in the data set may have 1200 data points.

[0174] Based on the above CWRU2000 dataset, six migration fault diagnosis experiments were used to evaluate the effectiveness of the training method proposed in the embodiment of the present disclosure. In these six experiments, the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure directly uses the original data as input while maintaining consistency with the experimental settings used in the previous PHM2009 experiment. The training method proposed in the embodiment of the present disclosure was compared with the baseline, DAN, DANN, DCTLN, CDAN, and JAN, and the following results were obtained: Fig.13 The experimental result table shown, that is, the classification accuracy (%) and standard deviation of each model in the target domain on the CWRU dataset, ↑ indicates how much higher the training method proposed in the embodiment of the present disclosure is than the method with the second highest accuracy, ↓ indicates how much lower the training method proposed in the embodiment of the present disclosure is than the method with the highest accuracy. As shown in the results, in the six migration fault diagnosis experiments, the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure outperformed the other methods in three of the experiments. In the other three experiments, one of the other methods was slightly better than the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure. But in general, the training method proposed in the embodiment of the present disclosure is no worse than any other method and performs well in the CWRU dataset. This shows that the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure can work under the conditions of the original data. In addition, although the above Fig.13 There is a slight imbalance between fault type 0 (normal type) and fault types 1-9 in the CWRU2000 data set shown, but the training method proposed in the embodiment of the present disclosure can still be effectively executed. These observations reflect to a certain extent the generalization ability of the target mechanical fault diagnosis model trained by the training method proposed in the embodiment of the present disclosure on unbalanced data and different input types.

[0175] In the above experiment, it is shown that the currently popular migration method for fault diagnosis has an inherent defect. Even if the model can obtain domain-invariant features, it cannot guarantee the acquisition of discriminative sample features. The mechanical fault diagnosis model training method proposed in the embodiment of the present disclosure can achieve good fault diagnosis effects on the gearbox dataset and the bearing fault dataset. The experimental results conducted on the gearbox dataset and the bearing fault dataset show that the target mechanical fault diagnosis model trained by the mechanical fault diagnosis model training method proposed in the embodiment of the present disclosure has achieved better fault diagnosis classification accuracy under various working conditions, exceeding various current mainstream fault diagnosis methods. In addition, additional ablation experiments further verified the effectiveness of the mechanical fault diagnosis model training method proposed in the embodiment of the present disclosure.

[0176] Fig.14 A block diagram of a mechanical fault diagnosis model migration training device according to an embodiment of the present disclosure is shown. Fig.14 As shown, the device comprises:

[0177] An acquisition module 141 is used to acquire a mechanical fault diagnosis model to be trained and a source domain sample set and a target domain sample set; wherein the source domain sample set includes multiple source fault samples and sample labels corresponding to each source fault sample, and the target domain sample set includes multiple unlabeled target fault samples; the source fault sample includes a mechanical vibration signal collected when a mechanical fault occurs under a source working condition, and the target fault sample includes a mechanical vibration signal collected when a mechanical fault occurs under a target working condition, and the source working condition is different from the target working condition; the sample label is used to indicate the fault type corresponding to the source fault sample; wherein the mechanical fault diagnosis model includes a feature generator and a classifier, the feature generator is used to extract sample features of the fault sample, and the classifier is used to generate a fault prediction result according to the sample features, and the fault prediction result includes the probability that the fault sample belongs to each fault type respectively;

[0178] A prediction module 142, configured to output sample features and fault prediction results of each source fault sample and sample features and fault prediction results of each target fault sample based on each source fault sample in the source domain sample set and each target fault sample in the target domain sample set by using the mechanical fault diagnosis model;

[0179] A classification loss determination module 143, configured to determine a sample classification loss according to the fault prediction results of the plurality of source fault samples and the sample labels of the respective source fault samples, wherein the sample classification loss represents a classification error of the fault type of the source fault sample by the mechanical fault diagnosis model;

[0180] A difference loss determination module 144 is used to determine a feature difference loss according to the sample features of the multiple source fault samples and the sample features of the multiple target fault samples, wherein the feature difference loss represents the difference in sample feature distribution between the source domain sample set and the target domain sample set;

[0181] A confusion loss determination module 145 is used to determine a category confusion loss according to the fault prediction results of the plurality of target fault samples, wherein the category confusion loss represents the degree of confusion between different fault types by the mechanical fault diagnosis model;

[0182] The optimization module 146 is used to determine the target loss according to the sample classification loss, the feature difference loss and the category confusion loss, and optimize the mechanical fault diagnosis model using the target loss to obtain a trained target mechanical fault diagnosis model.

[0183] In a possible implementation, the device also includes: a domain adversarial loss determination module, which is used to: input the sample features of the multiple source fault samples and the sample features of the multiple target fault samples into the domain discriminator respectively, and obtain the domain classification results of each source fault sample and the domain classification results of each target fault sample, wherein the domain classification results represent the probability that the sample feature belongs to the feature distribution space corresponding to the source domain sample set, or represent the probability that the sample feature belongs to the feature distribution space corresponding to the target domain sample set; determine the domain adversarial loss according to the domain classification results of each source fault sample and the domain classification results of each target fault sample, wherein the domain adversarial loss represents the classification error of the domain discriminator on the feature distribution space to which the sample feature belongs.

[0184] In a possible implementation manner, determining the target loss according to the sample classification loss, the feature difference loss, and the category confusion loss includes: determining the target loss according to the sample classification loss, the feature difference loss, the category confusion loss, and the domain adversarial loss.

[0185] In a possible implementation, the device also includes: a first virtual adversarial loss determination module, which is used to: add a disturbance signal to each source fault sample in the source domain sample set to obtain a source fault disturbance sample corresponding to each source fault sample; input the source fault disturbance sample corresponding to each source fault sample into the mechanical fault diagnosis model to obtain a fault prediction result of the source fault disturbance sample corresponding to each source fault sample; determine a first virtual adversarial loss based on the fault prediction results of the multiple source fault samples and the fault prediction results of the source fault disturbance sample corresponding to each source fault sample, wherein the first virtual adversarial loss characterizes the difference between the fault prediction result of the source fault sample and the fault prediction result of the corresponding source fault disturbance sample.

[0186] In a possible implementation, determining a target loss according to the sample classification loss, the feature difference loss, and the class confusion loss includes: determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, and the first virtual adversarial loss; or determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, the domain adversarial loss, and the first virtual adversarial loss; wherein the domain adversarial loss is determined according to the domain classification results of each source fault sample and the domain classification results of each target fault sample.

[0187] In a possible implementation, the device also includes: a second virtual adversarial loss determination module, which is used to: add a disturbance signal to each target fault sample in the target domain sample set to obtain a target fault disturbance sample corresponding to each target fault sample; input the target fault disturbance sample corresponding to each target fault sample into the mechanical fault diagnosis model to obtain a fault prediction result of the target fault disturbance sample corresponding to each target fault sample; determine a second virtual adversarial loss based on the fault prediction results of the multiple target fault samples and the fault prediction results of the target fault disturbance sample corresponding to each target fault sample, wherein the second virtual adversarial loss characterizes the difference between the fault prediction result of the target fault sample and the fault prediction result of the corresponding target fault disturbance sample.

[0188] In a possible implementation, determining the target loss according to the sample classification loss, the feature difference loss, and the class confusion loss includes: determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, and the second virtual adversarial loss; or determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, the domain adversarial loss, and the second virtual adversarial loss; or determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, the first virtual adversarial loss, and the second virtual adversarial loss; or determining the target loss according to the sample classification loss, the feature difference loss, the class confusion loss, the domain adversarial loss, the first virtual adversarial loss, and the second virtual adversarial loss; wherein the domain adversarial loss is determined according to the domain classification results of each source fault sample and the domain classification results of each target fault sample, and the first virtual adversarial loss is determined according to the fault prediction results of the multiple source fault samples and the fault prediction results of the source fault disturbance samples corresponding to each source fault sample.

[0189] According to the training device of the embodiment of the present disclosure, by utilizing sample classification loss, feature difference loss, category confusion loss, domain adversarial loss, first virtual adversarial loss and / or second virtual adversarial loss to train a mechanical fault diagnosis model, the feature extractor in the trained target mechanical fault diagnosis model can extract sample features with stronger domain invariance and discriminability, and the classifier can exhibit stronger classification ability, and the target mechanical fault diagnosis model can be made more stable, smooth and less prone to collapse, thereby realizing the transfer learning of the mechanical fault diagnosis model under the source working condition to the target mechanical fault diagnosis model under the target working condition, and the trained target mechanical fault diagnosis model can have higher fault type diagnosis accuracy for unlabeled fault samples under the target working condition.

[0190] Based on the above training device, the embodiment of the present disclosure further provides a mechanical fault diagnosis device, including:

[0191] A data acquisition module, used to acquire fault data actually collected when a mechanical fault occurs under a target working condition, wherein the fault data includes a mechanical vibration signal;

[0192] A fault diagnosis module is used to determine a fault diagnosis result corresponding to the fault data based on the fault data using a target mechanical fault diagnosis model, wherein the fault diagnosis result represents the fault type of the mechanical failure under the target working condition; wherein the target mechanical fault diagnosis model is obtained by training using the training device.

[0193] According to the diagnostic device of the embodiment of the present disclosure, it is possible to use the target machine fault diagnosis model to obtain an accurate fault type diagnosis when a machine fault occurs under the target working condition.

[0194] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0195] The embodiment of the present disclosure also provides a computer-readable storage medium on which computer program instructions are stored, and the computer program instructions implement the above method when executed by a processor. The computer-readable storage medium can be a volatile or non-volatile computer-readable storage medium.

[0196] An embodiment of the present disclosure further proposes an electronic device, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to implement the above method when executing the instructions stored in the memory.

[0197] The embodiments of the present disclosure also provide a computer program product, including a computer-readable code, or a non-volatile computer-readable storage medium carrying the computer-readable code. When the computer-readable code runs in a processor of an electronic device, the processor in the electronic device executes the above method.

[0198] Fig.15 1 is a block diagram of an electronic device 1900 according to an embodiment of the present disclosure. For example, the electronic device 1900 may be provided as a server or a terminal device. Fig.15 , the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932 for storing instructions executable by the processing component 1922, such as an application. The application stored in the memory 1932 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 1922 is configured to execute instructions to perform the above method.

[0199] The electronic device 1900 may also include a power supply component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input / output interface 1958 (I / O interface). The electronic device 1900 may operate based on an operating system stored in the memory 1932, such as Windows Server 2003. TM , Mac OS X TM , Unix TM ,Linux TM , FreeBSD TM or similar.

[0200] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions, which can be executed by the processing component 1922 of the electronic device 1900 to perform the above method.

[0201] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0202] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0203] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0204] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0205] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0206] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0207] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0208] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of the module, program segment or instruction includes one or more executable instructions for realizing the specified logical function. In some alternative implementations, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of special hardware and computer instructions.

[0209] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A mechanical fault diagnosis model migration training method, characterized in that: include: Obtain a mechanical fault diagnosis model to be trained and a source domain sample set and a target domain sample set; wherein the source domain sample set includes multiple source fault samples and sample labels corresponding to each source fault sample, and the target domain sample set includes multiple unlabeled target fault samples; the source fault sample includes a mechanical vibration signal collected when a mechanical fault occurs under a source working condition, and the target fault sample includes a mechanical vibration signal collected when a mechanical fault occurs under a target working condition, and the source working condition is different from the target working condition; the sample label is used to indicate the fault type corresponding to the source fault sample; wherein the mechanical fault diagnosis model includes a feature generator and a classifier, the feature generator is used to extract sample features of the fault sample, and the classifier is used to generate a fault prediction result according to the sample features, and the fault prediction result includes the probability that the fault sample belongs to each fault type respectively; Using the mechanical fault diagnosis model, based on each source fault sample in the source domain sample set and each target fault sample in the target domain sample set, output sample features and fault prediction results of each source fault sample, as well as sample features and fault prediction results of each target fault sample; Determine a sample classification loss according to the fault prediction results of the plurality of source fault samples and the sample label of each source fault sample, wherein the sample classification loss represents a classification error of the fault type of the source fault sample by the mechanical fault diagnosis model; Determine a feature difference loss according to the sample features of the multiple source fault samples and the sample features of the multiple target fault samples, wherein the feature difference loss represents a feature distribution difference between the sample features of the source fault samples and the sample features of the target fault samples; Determining a category confusion loss according to the fault prediction results of the plurality of target fault samples, wherein the category confusion loss represents the degree of confusion between different fault types by the mechanical fault diagnosis model; A target loss is determined according to the sample classification loss, the feature difference loss, and the category confusion loss, and the mechanical fault diagnosis model is optimized using the target loss to obtain a trained target mechanical fault diagnosis model.

2. The method according to claim 1, characterized in that The method further comprises: Inputting the sample features of the multiple source fault samples and the sample features of the multiple target fault samples into the domain discriminator respectively, obtaining the domain classification results of each source fault sample and the domain classification results of each target fault sample, wherein the domain classification results represent the probability that the sample features belong to the feature distribution space corresponding to the source domain sample set, or represent the probability that the sample features belong to the feature distribution space corresponding to the target domain sample set; Determine a domain adversarial loss according to the domain classification results of each source fault sample and the domain classification results of each target fault sample, wherein the domain adversarial loss represents a classification error of the domain discriminator on a feature distribution space to which the sample feature belongs; The determining of the target loss according to the sample classification loss, the feature difference loss and the category confusion loss includes: A target loss is determined according to the sample classification loss, the feature difference loss, the category confusion loss, and the domain adversarial loss.

3. The method according to claim 1 or 2, characterized in that: The method further comprises: Adding a disturbance signal to each source fault sample in the source domain sample set to obtain a source fault disturbance sample corresponding to each source fault sample; Inputting source fault disturbance samples corresponding to each source fault sample into the mechanical fault diagnosis model to obtain fault prediction results of source fault disturbance samples corresponding to each source fault sample; According to the fault prediction results of the multiple source fault samples and the fault prediction results of the source fault disturbance samples corresponding to each source fault sample, a first virtual adversarial loss is determined, wherein the first virtual adversarial loss represents the difference between the fault prediction result of the source fault sample and the fault prediction result of the corresponding source fault disturbance sample.

4. The method according to claim 3, characterized in that The determining the target loss according to the sample classification loss, the feature difference loss and the category confusion loss includes: Determine a target loss according to the sample classification loss, the feature difference loss, the category confusion loss and the first virtual adversarial loss; or, Determining a target loss according to the sample classification loss, the feature difference loss, the category confusion loss, the domain adversarial loss, and the first virtual adversarial loss; The domain adversarial loss is determined according to the domain classification results of each source fault sample and the domain classification results of each target fault sample.

5. The method according to claim 1 or 2, characterized in that: The method further comprises: Adding a disturbance signal to each target fault sample in the target domain sample set to obtain a target fault disturbance sample corresponding to each target fault sample; Inputting the target fault disturbance samples corresponding to each target fault sample into the mechanical fault diagnosis model to obtain the fault prediction results of the target fault disturbance samples corresponding to each target fault sample; According to the fault prediction results of the multiple target fault samples and the fault prediction results of the target fault disturbance samples corresponding to each target fault sample, a second virtual adversarial loss is determined, wherein the second virtual adversarial loss characterizes the difference between the fault prediction result of the target fault sample and the fault prediction result of the corresponding target fault disturbance sample.

6. The method according to claim 5, characterized in that The determining the target loss according to the sample classification loss, the feature difference loss and the category confusion loss includes: Determine a target loss according to the sample classification loss, the feature difference loss, the category confusion loss, and the second virtual adversarial loss; or, Determine a target loss according to the sample classification loss, the feature difference loss, the category confusion loss, the domain adversarial loss, and the second virtual adversarial loss; or, Determine a target loss according to the sample classification loss, the feature difference loss, the category confusion loss, the first virtual adversarial loss, and the second virtual adversarial loss; or, Determining a target loss according to the sample classification loss, the feature difference loss, the category confusion loss, the domain adversarial loss, the first virtual adversarial loss, and the second virtual adversarial loss; Among them, the domain adversarial loss is determined according to the domain classification results of each source fault sample and the domain classification results of each target fault sample, and the first virtual adversarial loss is determined according to the fault prediction results of the multiple source fault samples and the fault prediction results of the source fault disturbance samples corresponding to each source fault sample.

7. A mechanical fault diagnosis method, characterized in that: include: Acquire fault data actually collected when a mechanical fault occurs under target working conditions, wherein the fault data includes a mechanical vibration signal; Determine a fault diagnosis result corresponding to the fault data based on the fault data using a target machine fault diagnosis model, wherein the fault diagnosis result represents a fault type of the machine under the target working condition; Wherein, the target mechanical fault diagnosis model is trained using the training method described in any one of claims 1 to 6.

8. A mechanical fault diagnosis model migration training device, characterized in that: include: An acquisition module, used to acquire a mechanical fault diagnosis model to be trained and a source domain sample set and a target domain sample set; wherein the source domain sample set includes multiple source fault samples and sample labels corresponding to each source fault sample, and the target domain sample set includes multiple unlabeled target fault samples; the source fault sample includes a mechanical vibration signal collected when a mechanical fault occurs under a source working condition, and the target fault sample includes a mechanical vibration signal collected when a mechanical fault occurs under a target working condition, and the source working condition is different from the target working condition; the sample label is used to indicate the fault type corresponding to the source fault sample; wherein the mechanical fault diagnosis model includes a feature generator and a classifier, the feature generator is used to extract sample features of the fault sample, and the classifier is used to generate a fault prediction result according to the sample features, and the fault prediction result includes the probability that the fault sample belongs to each fault type respectively; A prediction module, configured to output sample features and fault prediction results of each source fault sample and sample features and fault prediction results of each target fault sample based on each source fault sample in the source domain sample set and each target fault sample in the target domain sample set by using the mechanical fault diagnosis model; A classification loss determination module, configured to determine a sample classification loss according to the fault prediction results of the plurality of source fault samples and the sample labels of the respective source fault samples, wherein the sample classification loss represents a classification error of the fault type of the source fault sample by the mechanical fault diagnosis model; A difference loss determination module, configured to determine a feature difference loss according to sample features of the plurality of source fault samples and sample features of the plurality of target fault samples, wherein the feature difference loss represents a difference in sample feature distribution between the source domain sample set and the target domain sample set; A confusion loss determination module, used to determine a category confusion loss according to the fault prediction results of the plurality of target fault samples, wherein the category confusion loss represents the degree of confusion between different fault types by the mechanical fault diagnosis model; The optimization module is used to determine the target loss according to the sample classification loss, the feature difference loss and the category confusion loss, and optimize the mechanical fault diagnosis model using the target loss to obtain a trained target mechanical fault diagnosis model.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to implement the method described in any one of claims 1 to 7 when executing the instructions stored in the memory.

10. A non-volatile computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images