A gearbox fault diagnosis method and device under variable working conditions based on feature granularity alignment

Abstract

The present invention relates to a gearbox state monitoring technology, specifically a gearbox fault diagnosis method and device under variable working conditions based on feature granularity alignment. The method comprises collecting and preprocessing bearing vibration data under different working conditions; using a feature extractor to extract source domain data and target domain data features respectively, and calculating the CMMD and FMMD between the two domains, as well as CFMMD; using a classifier to generate a predicted label for the source domain data; calculating the prediction loss between its original label and the predicted label, and calculating the distance loss according to the CFMMD; training a deep convolutional neural network model with minimizing the sum of losses as the objective function; using the trained feature extractor to extract target domain data features, using the trained classifier to process the target domain data features, and predicting the fault type of the target domain data. The present invention improves the fault diagnosis efficiency and accuracy of rotating machinery under variable working conditions.

Description

Technical Field

[0001] The present invention relates to a gearbox state monitoring technology, and is particularly applicable to the field of gearbox fault diagnosis under variable working conditions, and specifically relates to a gearbox fault diagnosis method under variable working conditions based on feature granularity alignment. Background Art

[0002] Gearboxes are widely used in important engineering fields such as aerospace, automobile manufacturing, rail transportation and wind power generation, which are important to the national economy and people's livelihood. They are one of the core components of mechanical equipment and play a vital role in the national economic production. Carrying out condition monitoring and diagnosis of mechanical equipment, detecting, diagnosing and predicting possible faults to "prevent them before they happen" is of great practical significance to ensure the reliable, continuous and stable operation of machinery, reduce economic losses and operating costs, and avoid major accidents.

[0003] At present, data-driven fault diagnosis methods can generally be divided into intelligent diagnosis methods based on traditional machine learning, intelligent diagnosis methods based on deep learning, and intelligent diagnosis methods based on transfer learning. With the advent of the "big data" era, big data characterized by low density, large quantity, and multiple forms has brought new challenges to bearing monitoring and diagnosis. Traditional methods based on artificial signal processing can no longer meet the needs of big data. "Data-driven" deep learning methods have emerged in the field of diagnosis and identification and have achieved a lot of results. Due to the wide variety of rolling bearing models and different operating conditions, bearing diagnosis methods based on deep learning have encountered problems such as insufficient training data or missing sample labels in practice; and the distribution of bearing status data under different working conditions and models is different, and the same network cannot be used for classification, resulting in the generalization ability of fault diagnosis models based on traditional machine learning, deep learning and other methods with the same distribution characteristics of training data and test data being reduced, or even no longer applicable, and transfer learning research is needed.

[0004] Transfer learning mainly consists of two steps: first, a deep hierarchical model is used to extract transferable features from the datasets of the source and target domains; such a model is able to learn abstract transferable features and helps reduce the cross-domain distribution differences. Second, a distance metric formula is used to measure the distribution differences of the transferable features. The results are then back-propagated to update the parameters of the deep hierarchical model to obtain features that follow a similar distribution. Therefore, the classification layer of the model trained with labeled data in the source domain can be reused to classify unlabeled data in the target domain. In summary, the choice of distance metric formula to measure the distribution difference of transferable features directly affects the transfer performance of the model. Maximum mean difference (MMD) is a widely used distance measurement loss to reduce cross-domain offsets. However, the computational complexity of MMD grows quadratically with the amount of data, which limits the generalization of MMD in large datasets. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art and improve the accuracy and reliability of fault diagnosis under variable operating conditions, the present invention proposes a gearbox fault diagnosis method and device under variable operating conditions based on feature granularity alignment, which can reduce the impact of domain drift, so that the deep learning model can well complete the fault diagnosis task under variable operating conditions with fast speed and small amount of calculation.

[0006] The present invention uses a domain adaptation method to find common features in labeled bearing data and unlabeled bearing data, alleviates the problem of data non-sharing and provides two improved MMD metrics, namely, feature compression MMD (Compressed Maximum Mean Discrepancy, CMMD) and feature-specific MMD (Feature-Specitic Maximum Mean Discrepancy, FMMD) (to alleviate the adverse effect of the calculation amount of MMD growing quadratically with the amount of data, enhance its generalization ability in large data sets and avoid cross-domain migration between the same feature distributions), forming a domain adaptation network based on CMMD and FMMD, respectively called CDAN and FDAN. By combining CMMD and FMMD, the present invention proposes a domain adaptation network based on CFMMD (Compressed and Feature-Specitic Maximum Mean Discrepancybased Domain Adap tation Network, CFDAN) to further improve the classification performance.

[0007] In a first aspect of the present invention, the present invention provides a gearbox fault diagnosis method under variable working conditions based on feature granularity alignment, the method comprising:

[0008] Collect bearing vibration data under different working conditions, including source domain data under the current working condition and target domain data under other working conditions;

[0009] Preprocessing bearing vibration data under different working conditions, setting original labels for the source domain data, wherein the original labels correspond to the fault types of the gearbox;

[0010] Inputting the preprocessed bearing vibration data into a pre-built deep convolutional neural network model, wherein the deep convolutional neural network model includes a feature extractor and a classifier;

[0011] The first high-level features of the source domain data and the target domain data are extracted by using a feature extractor, and the overall feature distribution distance and the feature granularity distribution distance between the two domains, as well as the corresponding weighted distribution distance, are calculated based on the first high-level features of the two domains.

[0012] Using the classifier to process the first high-level features of the source domain data, generating predicted labels for the source domain data;

[0013] Calculate the prediction loss between the original label and the predicted label of the source domain data, and calculate the distance loss based on the weighted distribution distance between the two domains;

[0014] Taking minimizing the sum of losses as the objective function, optimizing the objective function using the stochastic gradient descent method, and training the deep convolutional neural network model;

[0015] The trained feature extractor is used to extract the second highest level features of the target domain data, and the trained classifier is used to process the second highest level features of the target domain data to predict the fault type of the target domain data.

[0016] In a second aspect of the present invention, the present invention further provides a gearbox fault diagnosis device under variable working conditions based on feature granularity alignment, comprising:

[0017] A data acquisition module is used to collect bearing vibration data under different working conditions, including source domain data under the current working condition and target domain data under other working conditions;

[0018] A data processing module, used for preprocessing bearing vibration data under different working conditions, setting original labels for the source domain data, wherein the original labels correspond to the fault types of the gearbox;

[0019] A deep convolutional neural network model, including a feature extractor and a classifier, is used to receive the preprocessed bearing vibration data;

[0020] A feature extractor, used to extract first high-level features of source domain data and target domain data and second high-level features of target domain data;

[0021] A classifier is used to process the first high-level features of the source domain data to generate a predicted label for the source domain data and to process the second high-level features of the target domain data to predict the fault type of the target domain data;

[0022] A distance calculation module, used to calculate the overall distribution distance of features and the distribution distance of feature granularity between the two domains, and the corresponding weighted distribution distance according to the first high-level features of the source domain data and the target domain data;

[0023] A loss calculation module, used to calculate the prediction loss based on the original label and the predicted label of the source domain data, and to calculate the distance loss based on the weighted distribution distance between the two domains;

[0024] The training module is used to minimize the loss sum as the objective function, optimize the objective function using the stochastic gradient descent method, and train the deep convolutional neural network model.

[0025] In the third aspect of the present invention, the present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a gearbox fault diagnosis method under variable operating conditions based on feature granularity alignment as described in the first aspect of the present invention is implemented.

[0026] In the fourth aspect of the present invention, the present invention also provides a computer-readable medium device having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for diagnosing gearbox faults under variable operating conditions based on feature granularity alignment as described in the first aspect of the present invention is implemented.

[0027] Beneficial effects of the present invention:

[0028] 1. In order to solve the problem that it is difficult to train an intelligent diagnosis model with strong generalization ability due to the lack of typical fault information and health mark information in the monitoring data of engineering equipment, the present invention combines the knowledge of transfer learning on the basis of deep learning, and mines the domain-invariant essential structure and characteristics between two different but related domains, so that supervised information such as labeled data can be migrated and reused between domains, which greatly enhances the robustness of the model.

[0029] 2. Compared with the prior art, the present invention proposes two improved MMD indicators, namely feature compression MMD (CMMD) and specific feature MMD (FMMD), which respectively calculate the overall feature distribution distance and feature particle size distribution distance, effectively reducing the amount of calculation in the training process, alleviating the adverse effect of the MMD calculation amount growing quadratically with the amount of data, and enhancing its generalization ability in large data sets. By forming CFMMD through the weighted combination of CMMD and FMMD and using CFMMD as the distance measurement formula, it is possible to suppress the adverse effects caused by the overall cross-domain migration of transferable features while ignoring the cross-domain migration of individual features. Combining the CFMMD distance metric with a deep convolutional neural network, the present invention proposes a domain adaptive network model based on CFMMD that can effectively reduce the impact of domain drift, thereby improving the fault diagnosis efficiency and accuracy of rotating machinery under variable working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of a gearbox fault diagnosis method under variable working conditions based on feature granularity alignment according to an embodiment of the present invention;

[0031] Figure 2 It is a structural diagram of a gearbox fault diagnosis device under variable working conditions based on feature granularity alignment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] Figure 1 is a flow chart of a gearbox fault diagnosis method under variable working conditions based on feature granularity alignment according to an embodiment of the present invention, such as Figure 1 As shown, the method includes:

[0034] 101. Collect bearing vibration data under different working conditions, including source domain data under current working conditions and target domain data under other working conditions;

[0035] In an embodiment of the present invention, considering that there are many types of rolling bearings and their operating conditions vary, traditional bearing diagnosis methods based on deep learning have encountered problems such as insufficient training data or missing sample labels in practice; and the distribution of bearing status data under different working conditions and models is different, and the same network cannot be used for classification. Based on this, the present invention aligns the feature granularity of bearing vibration signals in different fields, and uses deep transfer learning to optimize bearing diagnosis. In an embodiment of the present invention, bearing vibration data in different health states under the same working condition can be used as a data domain, and source domain samples and target domain samples are selected from different data domains; data sets under different working conditions can be source domains or target domains for each other; for example Assume that there are bearing vibration data of different health states under four different working conditions, namely working condition A, working condition B, working condition C and working condition D, in the data domain. In some embodiments, the bearing vibration data of working conditions A to working condition C can be used as source domain data, and the bearing vibration data of working condition D can be used as target domain data. In other embodiments, the bearing vibration data of working conditions B to working condition D can be used as source domain data, and the bearing vibration data of working condition A can be used as target domain data. Assuming that training is required to diagnose the bearing vibration data of working condition A, then working conditions B to working condition D are used as source domain data under the current working conditions, and working condition A is used as target domain data under other working conditions, and so on. Other examples are not listed one by one.

[0036] In the embodiment of the present invention, the motor current signal in the electromechanical system where the gear box is located can be obtained by collecting the stator current of the motor through a current transformer at a certain sampling frequency. At this time, the current signal obtained is a time series with dimensions. The dimension represents the number of data points in the current signal, and the number of dimensions is related to the sampling frequency and the sampling duration.

[0037] Exemplarily, when the current transformer collects a 4s current signal at a sampling frequency of 64kHz, the number of dimensions of the obtained current signal is 256,000, that is, a total of 256,000 data points can be collected by collecting a 4s current signal at a sampling frequency of 64kHz.

[0038] Exemplarily, the sensor collects fault signals of the rotating machinery under different working conditions, and for each fault signal, 1024 sampling points are used as a sample length.

[0039] 102. Preprocess the bearing vibration data under different working conditions, and set an original label for the source domain data, wherein the original label corresponds to a fault type of the gearbox;

[0040] In the embodiment of the present invention, the one-dimensional bearing diagnosis data needs to be preprocessed, and the original label needs to be added to the source domain sample. The original label corresponds to the fault type. For example, assuming that the fault location and fault degree of the source domain sample A1 of the working condition A are IR07, then IR07 can be used as the original label of the source domain sample A1. Specifically, assuming that the labeled source domain data and unlabeled target domain data D s and D t The data domains representing the source domain and the target domain respectively; The i-th sample representing the source domain has a category of represents the jth sample in the target domain; C represents the total number of categories; M and N represent the number of samples in the source domain and the number of samples in the target domain, respectively.

[0041] In the embodiment of the present invention, the intuitive one-dimensional original vibration signal data cannot reveal the fault problem, so it is necessary to analyze the bearing vibration signal of the gearbox, that is, firstly, the one-dimensional bearing vibration data signal is converted into a single-channel grayscale image by format conversion; the single-channel grayscale image is copied twice and converted into a three-channel image to be diagnosed. For example, the one-dimensional data (1024) is converted into a grayscale image with a pixel size of 1*32*32 by format conversion, and the converted vibration signal is a single-channel grayscale image; the vibration signal is converted into a three-channel RGB format of 3*32*32 by copying; the RGB image data is sent to a parameter-initialized deep convolutional neural network model for diagnosis and detection. Of course, the image here can also be an image of other formats. The above examples are only for ease of understanding and are not limitations of the present invention.

[0042] 103. Input the preprocessed bearing vibration data into a pre-built deep convolutional neural network model, wherein the deep convolutional neural network model includes a feature extractor and a classifier;

[0043] In an embodiment of the present invention, the existing basic deep convolutional neural network model is first directly adopted, and the deep convolutional neural network model includes a feature extractor and a classifier; then the classifier is modified, and the number of neurons in the last layer of the classifier is adjusted so that the number of neurons in the last layer of the classifier is equal to the number of types of the original labels, thereby ensuring that each diagnostic type corresponds to a neuron.

[0044] 104. Using a feature extractor to extract first high-level features of the source domain data and the target domain data respectively, and calculating the overall feature distribution distance and feature granularity distribution distance between the two domains and the corresponding weighted distribution distance based on the first high-level features of the two domains;

[0045] In an embodiment of the present invention, during the training process, the source domain data and the target domain data are respectively input into the feature extractor, and the feature extractor can respectively extract the first high-level features of the source domain data and the first high-level features of the target domain data, and in the subsequent process, operations can be performed based on the first high-level features extracted from the two.

[0046] In an embodiment of the present invention, the present invention uses a domain adaptation method to find common features in labeled bearing data and unlabeled bearing data, thereby alleviating the problem of data non-sharing, and thus provides two improved MMD metrics, namely, feature compression MMD (Compressed Maximum Mean Discrepancy, CMMD) and feature-specific MMD (Feature-Specitic Maximum Mean Discrepancy, FMMD), to alleviate the adverse effect of the calculation amount of MMD growing quadratically with the amount of data, enhance its generalization ability in large data sets, and avoid cross-domain migration between the same feature distributions.

[0047] Among them, the calculation formula of the overall distribution distance of the feature is expressed as:

[0048]

[0049] Where Φ(·) is a nonlinear mapping function in the reconstruction kernel Hilbert space; M represents the number of source domain data samples, represents the i-th sample of the source domain data, N represents the number of target domain data samples, represents the jth sample of the target domain data; H represents the Hilbert space, and k(·,·) is the Gaussian kernel function used to map the source domain and target domain data to the Hilbert space.

[0050] Among them, the calculation formula of the distribution distance of characteristic particle size is expressed as:

[0051]

[0052] Among them, R is the total number of features, that is, the number of neurons in the last layer of the feature extractor; n is the number of samples in the source domain and the target domain selected for one training; is the set of r-th features of n samples in the source domain, is the set of r-th features of n samples in the target domain, n is the number of samples in the source domain and target domain selected for one training; Represents the rth feature of the nth sample of the source domain data; Represents the rth feature of the nth sample of the target domain data.

[0053] Among them, the Gaussian kernel function is expressed as:

[0054]

[0055] Where δ is the bandwidth, represents the i-th sample in the source domain, represents the jth sample in the target domain.

[0056] It can be understood that in the embodiment of the present invention, CMMD uses two empirical estimates of distribution errors as the distance between two data distributions in the Hilbert space, and FMMD uses two empirical estimates of distribution errors as the distance between two data feature distributions in the Hilbert space; the present invention calculates the distribution distance between different fields from the data dimension and the feature dimension, which can reduce the amount of calculation in the training process, and alleviate the adverse effect of the calculation amount of MMD growing quadratically with the amount of data, thereby enhancing its generalization ability in large data sets.

[0057] In a preferred embodiment of the present invention, the present invention also performs weighted combination of the distribution distance CMMD of the feature as a whole and the distribution distance FMMD of the feature size to form a weighted distribution distance CFMMD, which is expressed as:

[0058] CFMMD(X s ,X t )=μ·CMMD(X s ,X t )+(1-μ)·FMMD(X s ,X t )

[0059] μ∈(0,1)

[0060] Among them, CMMD (X s ,X t ) is the source domain data X s and target domain data X tThe overall distribution distance of the characteristics between s ,X t ) is the source domain data X s and target domain data X t The distribution distance of the characteristic granularity between them; μ is the weight between the distribution distance of the overall feature and the distribution distance of the characteristic granularity.

[0061] This embodiment uses CFMMD as a distance metric formula, which can suppress the adverse effects caused by the overall cross-domain migration of transferable features while ignoring the cross-domain migration of individual features, thereby improving the fault diagnosis efficiency and accuracy of rotating machinery under variable working conditions.

[0062] 105. Process the first high-level features of the source domain data using the classifier to generate a predicted label for the source domain data;

[0063] In the embodiment of the present invention, during the training process, the first high-level features of the source domain data are input into the classifier, and the classifier can process the first high-level features to obtain the predicted labels of the source domain data.

[0064] 106. Calculate the prediction loss between the original label and the predicted label of the source domain data, and calculate the distance loss according to the weighted distribution distance between the two domains;

[0065] In the embodiment of the present invention, the label loss between the predicted label of the source domain sample output by the classifier and the real label of the source domain sample is calculated, and the formula used is expressed as follows:

[0066]

[0067] Among them, Y s represents the original label of the source domain data, f(X s ) represents the predicted label of the source domain data, J(·,·) represents the cross entropy loss function,

[0068] In the embodiment of the present invention, the weighted distribution distance between the two domains is used as the distance loss, that is, the domain loss, and the adopted formula is expressed as:

[0069] loss dis (X s ,X t )=μ·CMDD(X s ,X t )+(1-μ)·FMDD(X s ,X t )

[0070] Among them, CMMD (X s ,X t ) is the source domain data Xs and target domain data X t The overall distribution distance of the characteristics between s ,X t ) is the source domain data X s and target domain data X t The distribution distance of the characteristic granularity between them; μ is the weight between the distribution distance of the overall feature and the distribution distance of the characteristic granularity.

[0071] Therefore, the final loss function is expressed as:

[0072]

[0073] 107. Taking minimizing the loss sum as the objective function, optimizing the objective function by using the stochastic gradient descent method, and training the deep convolutional neural network model;

[0074] In the embodiment of the present invention, the goal is to minimize the sum of losses, that is, to minimize the distance loss and minimize the prediction loss, which is expressed as:

[0075]

[0076] where J(·,·) represents the cross entropy loss function, λ>0 is a hyperparameter, steps is the total number of training times, and step is the current training step number.

[0077] The objective function is optimized using the stochastic gradient descent method, and the deep convolutional neural network model is trained until the model meets the accuracy requirements. The training can be stopped. At this time, the parameters of the trained deep convolutional neural network model can be fixed, and the trained deep convolutional neural network model can be used to detect and identify the target domain data.

[0078] 108. The second highest level features of the target domain data are extracted using the trained feature extractor, and the second highest level features of the target domain data are processed using the trained classifier to predict the fault type of the target domain data.

[0079] In an embodiment of the present invention, it is necessary to use a trained feature extractor to re-extract the second highest level features of the target domain data, and use a trained classifier to process the second highest level features of the target domain data to predict the fault type of the target domain data.

[0080] Figure 2 is a structural diagram of a gearbox fault diagnosis device under variable working conditions based on feature granularity alignment according to an embodiment of the present invention, such as Figure 2 As shown, the device comprises:

[0081] A data acquisition module is used to collect bearing vibration data under different working conditions, including source domain data under the current working condition and target domain data under other working conditions;

[0082] The data acquisition module can use the bearing vibration data of different health states under the same working condition as a data domain, and the source domain samples and the target domain samples are selected from different data domains; the data sets under different working conditions can be each other's source domains or target domains; for example, assuming that there are bearing vibration data of different health states under four different working conditions, namely working condition A, working condition B, working condition C and working condition D, in the data domain, in some embodiments, the bearing vibration data of working condition A-working condition C can be used as source domain data, and the bearing vibration data of working condition D can be used as target domain data; in other embodiments, the bearing vibration data of working condition B-working condition D can be used as source domain data, and the bearing vibration data of working condition A can be used as target domain data. Assuming that training is required to diagnose the bearing vibration data of working condition A, then working condition B-working condition D is used as the source domain data under the current working condition, and working condition A is used as the target domain data under other working conditions, and so on. Other examples are not listed one by one.

[0083] A data processing module, used for preprocessing bearing vibration data under different working conditions, setting original labels for the source domain data, wherein the original labels correspond to the fault types of the gearbox;

[0084] In an embodiment of the present invention, the data processing module can use the bearing vibration data of different health states under the same working condition as a data domain, and the source domain samples and the target domain samples are selected from different data domains; the data sets under different working conditions can be each other's source domains or target domains; for example, assuming that there are bearing vibration data of different health states under four different working conditions, namely working condition A, working condition B, working condition C and working condition D, in the data domain, in some embodiments, the bearing vibration data of working condition A-working condition C can be used as source domain data, and the bearing vibration data of working condition D can be used as target domain data; in other embodiments, the bearing vibration data of working condition B-working condition D can be used as source domain data, and the bearing vibration data of working condition A can be used as target domain data. Assuming that training is required to diagnose the bearing vibration data of working condition A, then working condition B-working condition D is used as the source domain data under the current working condition, and working condition A is used as the target domain data under other working conditions, and so on. Other examples are not listed one by one.

[0085] A deep convolutional neural network model, including a feature extractor and a classifier, is used to receive the preprocessed bearing vibration data;

[0086] In an embodiment of the present invention, the basic model of the deep convolutional neural network model can directly use the existing model, but the present invention needs to adjust the number of neurons in the last layer of the classifier so that the number of neurons in the last layer of the classifier is equal to the number of types of the original labels, thereby ensuring that each diagnostic type corresponds to a neuron.

[0087] A feature extractor, used to extract first high-level features of source domain data and target domain data and second high-level features of target domain data;

[0088] In some embodiments, the feature extractor includes ResNet-50 and a fully connected layer, with a single convolution layer at the beginning, followed by four groups of blocks, each of which has 3, 4, 6, and 3 blocks, and each block has three layers of networks, which are responsible for performing convolution operations on the image, extracting image features, and acting as an identity mapping layer. Finally, a single fully connected layer reduces the number of features after convolution to the specified number of features.

[0089] In addition, the feature extractor may also adopt other existing designs, and those skilled in the art can make corresponding adjustments according to the actual situation of the present invention.

[0090] A classifier is used to process the first high-level features of the source domain data to generate a predicted label for the source domain data and to process the second high-level features of the target domain data to predict the fault type of the target domain data;

[0091] In some embodiments, the classifier includes a fully connected layer and a Softmax layer, wherein the number of nodes in the fully connected layer is 128, and the number of neurons in the Softmax layer is set to the number of labels of the source domain data.

[0092] A distance calculation module, used to calculate the overall distribution distance of features and the distribution distance of feature granularity between the two domains, and the corresponding weighted distribution distance according to the first high-level features of the source domain data and the target domain data;

[0093] The distance calculation module includes a first distribution distance calculation unit, a second distribution distance calculation unit and a third distribution distance calculation unit.

[0094] The first distribution distance calculation unit is used to calculate the distribution distance CMMD of the overall feature according to the distance between the source domain data sample and the target domain data sample in the Hilbert space.

[0095] The second distribution distance calculation unit is used to calculate the distribution distance FMMD of the feature granularity according to the distance between the source domain data sample feature and the target domain data sample feature in the Hilbert space.

[0096] The third distribution distance calculation unit is used to perform weighted combination based on the distribution distance of the feature as a whole calculated by the first distribution distance calculation unit and the distribution distance of the feature granularity calculated by the second distribution distance calculation unit to obtain a weighted distribution distance CFMMD.

[0097] It can be understood that, in the embodiment of the present invention, the first distribution distance calculation unit performs distance calculation on the features of all samples of the source domain data and the features of all samples of the target domain data, and the second distribution distance calculation unit performs distance calculation on the single feature of the sample of the source domain data and the corresponding single feature of the sample of the target domain data. Therefore, the first distribution distance calculation unit can mine the overall distribution distance of the sample, and the second distribution distance calculation unit can mine the distribution distance of the feature. Through the combination of the two, the adverse effects caused by the overall cross-domain migration of the transferable features and ignoring the cross-domain migration of individual features can be suppressed.

[0098] A loss calculation module, used to calculate the prediction loss based on the original label and the predicted label of the source domain data, and to calculate the distance loss based on the weighted distribution distance between the two domains;

[0099] The loss calculation module includes a first loss calculation unit, a second loss calculation unit and a third loss unit.

[0100] The first loss calculation unit is used to calculate the label loss according to the predicted label of the source domain sample and the true label of the source domain sample.

[0101] The second loss calculation unit is used to calculate the distance loss according to the weighted distribution distance CFMMD.

[0102] The third distribution distance unit is used to perform weighted combination according to the label loss calculated by the first loss calculation unit and the distance loss calculated by the second loss calculation unit to obtain a final loss.

[0103] The training module is used to minimize the loss sum as the objective function, optimize the objective function using the stochastic gradient descent method, and train the deep convolutional neural network model.

[0104] In an embodiment of the present invention, a computer device is further included, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described in steps 101 to 108 when executing the computer program.

[0105] The memory (random-access memory, RAM); the memory may also include a non-volatile memory (non-volatile memory), such as a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD); the memory may also include a combination of the above types of memory.

[0106] Optionally, the memory is also used to store program instructions. The processor can call the program instructions stored in the memory to implement the method shown in the above embodiment of the present invention.

[0107] The processor may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and a NP.

[0108] The processor may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD) or a combination thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL) or any combination thereof.

[0109] In addition, the present invention further provides a computer-readable medium device on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the method described in steps 101 to 108 are implemented.

[0110] In order to verify the migration effect of the method of the present invention under variable load conditions, the faulty bearing data of Case Western Reserve University was used for verification. Experiments were conducted on the data with a sampling frequency of 12kHz at the driving end of the rolling bearing, and the vibration signal data was classified into 10 categories of states. 1024 points of the vibration signal were taken for each sample, and 100 samples of rolling bearing vibration signal data under different working conditions were obtained. Working condition A represents a sample set of 0hp-1797r / min, working condition B represents a sample set of 1hp-1772r / min, working condition C represents a sample set of 2hp-1750r / min, and working condition D represents a sample set of 3hp-1730r / min. The composition of rolling bearing data samples under specific different working conditions is shown in Table 1 below. The migration was performed under different loads (0, 1, 2, 3hp) to verify the effectiveness of the method of the present invention, and the specific diagnostic results are shown in Table 2.

[0111] Table 1 Composition of rolling bearing data samples under different working conditions

[0112]

[0113] Table 2 Fault diagnosis results

[0114]

[0115] For the diagnosis of bearing faults under different loads, when MMD is used as the distance measurement formula for the difference in the distribution of transferable features, the average recognition rate is 89.48%. When the CFMMD method is used for migration, the average recognition accuracy is improved by 1.33%. This shows that the present invention can better realize the correct classification of faults, improve the accuracy and reliability of fault diagnosis and reduce the amount of calculation compared with MMD, greatly improve the versatility and generalization of deep learning models, and can well handle the influence of domain drift under variable working conditions on fault diagnosis methods based on traditional deep learning.

[0116] In the description of the present invention, it is necessary to understand that the terms "coaxial", "bottom", "one end", "top", "middle", "the other end", "upper", "one side", "top", "inside", "outside", "front", "center", "both ends" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0117] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0118] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A gearbox fault diagnosis method under variable working conditions based on feature granularity alignment, characterized in that: The following steps are involved: Collect bearing vibration data under different working conditions, including source domain data under the current working condition and target domain data under other working conditions; Preprocessing bearing vibration data under different working conditions, setting original labels for the source domain data, wherein the original labels correspond to the fault types of the gearbox; Inputting the preprocessed bearing vibration data into a pre-built deep convolutional neural network model, wherein the deep convolutional neural network model includes a feature extractor and a classifier; The first high-level features of the source domain data and the target domain data are extracted by using a feature extractor, and the overall feature distribution distance and the feature granularity distribution distance between the two domains, as well as the corresponding weighted distribution distance, are calculated based on the first high-level features of the two domains. The calculation formula of the overall distribution distance of the feature is expressed as: Where Φ(·) is a nonlinear mapping function in the reconstruction kernel Hilbert space; M represents the number of source domain data samples, x i s represents the i-th sample of the source domain data, N represents the number of target domain data samples, x t j represents the jth sample of the target domain data; H represents the Hilbert space, and k(·,·) is the Gaussian kernel function; The calculation formula of the distribution distance of characteristic particle size is expressed as: Among them, R is the total number of features, that is, the number of neurons in the last layer of the feature extractor; is the set of r-th features of n samples in the source domain, is the set of r-th features of n samples in the target domain, n is the number of samples in the source domain and target domain selected for one training; Represents the rth feature of the nth sample of the source domain data; Represents the rth feature of the nth sample of the target domain data; The calculation formula of the weighted distribution distance between two fields is expressed as: CFMD(X s ,X t )=μ·CMMD(X s ,X t )+(1-μ)·FMMD(X s ,X t ) μ∈(0,1) Among them, CMMD (Xs, Xt) is the source domain data X s and target domain data X t The overall distribution distance of the characteristics between s ,X t ) is the source domain data X s and target domain data X t The distribution distance of the characteristic granularity between them; μ is the weight between the distribution distance of the overall feature and the distribution distance of the characteristic granularity; Using the classifier to process the first high-level features of the source domain data, generating predicted labels for the source domain data; Calculate the prediction loss between the original label and the predicted label of the source domain data, and calculate the distance loss based on the weighted distribution distance between the two domains; use the weighted distribution distance between the two domains as the distance loss, also known as the domain loss; Taking minimizing the sum of losses as the objective function, optimizing the objective function using the stochastic gradient descent method, and training the deep convolutional neural network model; The trained feature extractor is used to extract the second highest level features of the target domain data, and the trained classifier is used to process the second highest level features of the target domain data to predict the fault type of the target domain data.

2. The gearbox fault diagnosis method under variable working conditions based on feature granularity alignment according to claim 1 is characterized in that: The preprocessing of the bearing vibration data includes converting the one-dimensional bearing vibration data signal into a single-channel grayscale image by format conversion; and duplicating the single-channel grayscale image twice to convert it into a three-channel image.

3. The gearbox fault diagnosis method under variable working conditions based on feature granularity alignment according to claim 1 is characterized in that: The pre-construction process of the deep convolutional neural network model includes building a feature extractor and a classifier, and adjusting the number of neurons in the last layer of the classifier so that the number of neurons in the last layer of the classifier is equal to the number of types of the original labels.

4. A gearbox fault diagnosis device under variable working conditions based on feature granularity alignment, characterized in that: A gearbox fault diagnosis method under variable working conditions based on feature granularity alignment is used to implement any one of claims 1 to 3, comprising: A data acquisition module is used to collect bearing vibration data under different working conditions, including source domain data under the current working condition and target domain data under other working conditions; A data processing module, used for preprocessing bearing vibration data under different working conditions, setting original labels for the source domain data, wherein the original labels correspond to the fault types of the gearbox; A deep convolutional neural network model, including a feature extractor and a classifier, is used to receive the preprocessed bearing vibration data; A feature extractor, used to extract first high-level features of source domain data and target domain data and second high-level features of target domain data; A classifier is used to process the first high-level features of the source domain data to generate a predicted label for the source domain data and to process the second high-level features of the target domain data to predict the fault type of the target domain data; A distance calculation module, used to calculate the overall distribution distance of features and the distribution distance of feature granularity between the two domains, and the corresponding weighted distribution distance according to the first high-level features of the source domain data and the target domain data; A loss calculation module, used to calculate the prediction loss based on the original label and the predicted label of the source domain data, and to calculate the distance loss based on the weighted distribution distance between the two domains; The training module is used to minimize the loss sum as the objective function, optimize the objective function using the stochastic gradient descent method, and train the deep convolutional neural network model.

5. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for diagnosing gearbox faults under variable operating conditions based on feature granularity alignment as described in any one of claims 1 to 3 is implemented.

6. A computer readable medium device having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, a gearbox fault diagnosis method under variable working conditions based on feature granularity alignment is implemented as described in any one of claims 1 to 3.

Images