Motor fault diagnosis method and system based on multi-dimensional data under small sample condition

Abstract

The application discloses a motor fault diagnosis method and system based on multi-dimensional data under a small sample condition, is applied to intelligent diagnosis of motor faults, and aims to solve a small sample problem. The technical scheme of the application constructs a fault diagnosis model which fuses a spherical convolution operator and a supervised branch and a self-supervised branch for collaborative training through three-dimensional point cloud and spherical mapping technology, and can realize efficient and accurate fault diagnosis of a motor system under the condition of using only a small amount of labeled samples and various complex working conditions. In the model training stage, a label is generated through a pseudo-label generation mechanism and cooperates with original data to participate in training, and a weighted loss function structure and an adaptive weight adjustment strategy are also designed to optimize the learning efficiency and generalization ability of the model under different fault categories. In conclusion, the application can realize high-precision motor fault diagnosis only by relying on a small amount of labeled samples, and effectively reduces the cost of sample labeling and fault diagnosis of an electric drive system.

Description

TECHNICAL FIELD

[0001] The present application relates to motor state monitoring and intelligent fault diagnosis technology, and in particular to a motor fault diagnosis method and system based on multi-dimensional data under small sample conditions. BACKGROUND

[0002] With the development of industrial automation and intelligent manufacturing, motors are widely used in new energy vehicles, aerospace, robots and national defense equipment as key execution components. The motor operation process is affected by multiple factors such as speed, load, environment, etc., and is prone to faults such as stator winding short circuit, rotor broken bar, bearing damage, imbalance and shaft eccentricity. Once a fault occurs, it not only causes equipment damage and economic loss, but also may cause safety problems. Traditional motor fault diagnosis methods mainly rely on signal processing techniques such as spectrum analysis and wavelet packet decomposition, combined with expert experience for feature extraction and discrimination, which has the disadvantages of relying on prior knowledge and manual features being difficult to be comprehensive and robust. In recent years, with the rise of deep learning technology, fault diagnosis methods based on convolutional neural networks (CNN) have made significant progress in feature learning and classification accuracy. However, such methods usually rely on a large amount of labeled data for supervised training, while in actual industrial scenarios, fault samples are difficult to obtain and the labeling cost is high, causing the model to be prone to overfitting and poor generalization performance. To solve the problem of insufficient labeled samples, self-supervised learning uses unlabeled data to learn useful representations through proxy tasks, which has been widely used in image and natural language fields. However, directly applying it to motor vibration signals faces challenges such as high data dimensionality, insufficient utilization of geometric characteristics, and difficulty in ensuring three-dimensional rotational invariance. SUMMARY

[0003] The technical problem to be solved by the present application is that most existing fault diagnosis methods use data-driven techniques, which require a large number of labeled samples to train the model. However, the labeling of samples requires the knowledge of domain experts, and the cost of labeling samples is high. To overcome the technical obstacles / challenges of effectively applying self-supervised learning to the field of motor vibration signal fault detection, the present application proposes a motor fault diagnosis method and system based on multi-dimensional data under small sample conditions, which is a fault diagnosis method based on the Self-SL framework. It combines the concept of multi-task learning and cooperatively trains two task branches of supervised and self-supervised, can adaptively learn feature representations from both supervised and self-supervised task branches, and optimally combines a dynamically adjusted loss function and introduces a spherical convolution model. The diagnosis method proposed in the present application can achieve high diagnostic accuracy with very few labeled samples.

[0004] Therefore, the present application provides the following technical solutions:

[0005] The application provides a motor fault diagnosis method based on multi-dimensional data under a small sample condition, comprising the following steps:

[0006] S101, data acquisition, acquiring three-axis vibration time series signals in various motor states through a three-axis vibration sensor, and constructing three-dimensional point cloud data of the motor by using the three-axis vibration time series signals, wherein the motor state is a real fault label;

[0007] S102, data enhancement and spherical mapping, performing a three-dimensional rotation operation on the three-dimensional point cloud data of step S101 to obtain rotated point cloud data and a real rotation label, and performing spherical mapping on the rotated point cloud data to obtain rotated spherical data; performing spherical mapping on the three-dimensional point cloud data of step S101 to obtain original spherical data;

[0008] S103, cooperative training of a supervised branch and a self-supervised branch, the supervised branch and the self-supervised branch adopt the same feature extraction module and direct the output to different classifiers, in the training stage, the rotated spherical data and the real rotation label are input into the self-supervised branch, and the original spherical data and the real fault label are input into the supervised branch;

[0009] S104, inputting the original spherical data corresponding to the three-axis vibration time series signals of the motor to be diagnosed into the trained supervised branch to obtain a fault diagnosis classification result.

[0010] Optionally, the total loss value in the network training stage is weighted by the supervised task and the self-supervised task, and a dynamic weighting mechanism is introduced:

[0011] ;

[0012] In the formula, is the total loss value, is the loss value of task i, i is 1 and 2, is the loss value of task i, that is, corresponds to is the loss value of the self-supervised task, is the loss value of the supervised task; is the uncertainty parameter of task i,

[0013] the network training process, the following update exists:

[0014] ;

[0015] In the formula, denotes the updated uncertainty parameter , and η is a proportional constant.

[0016] Optionally, the self-supervised branch is provided with a feature extraction module and a linear layer 1, and the supervised branch is provided with a feature extraction module and a linear layer 2, and the linear layer 1 and the linear layer 2 are respectively classifiers for rotation type prediction and fault category prediction;

[0017] The feature extraction module is a spherical convolution module, and the spherical convolution module is composed of a plurality of convolution layers connected in sequence, and each convolution layer is stacked by a spherical convolution layer, a nonlinear layer and a pooling layer.

[0018] Optionally, the processing process of the spherical convolution layer is as follows:

[0019] Firstly, the current spherical signal and the filter are mapped to the frequency domain by spherical Fourier transform;

[0020] ;

[0021] ;

[0022] Wherein, is the signal corresponding to the spherical Fourier transform coefficient of the order , is the spherical signal on the two-dimensional unit sphere , x is a point on the sphere, is the spherical harmonic function of the order , the filter is the neural network convolution kernel of the spherical convolution layer, which is a two-dimensional matrix of 3*3, is the filter corresponding to the spherical Fourier transform coefficient of the order ; Then, the dot product operation is completed in the frequency domain to obtain ;

[0023]

[0024] ;

[0025] Finally, the spherical inverse Fourier transform is used to map back to the original space;

[0026] ;

[0027] Wherein, is the result of the current spherical convolution layer, and b is the bandwidth.

[0028] Briefly, the spherical convolution layer is inputted with the spherical data obtained by spherical mapping and the original spherical data for feature extraction. ​​

[0029] Optionally, the process of performing spherical mapping on the rotated point cloud data or on the 3D point cloud data in step S102 is as follows:

[0030] Map rotated point cloud data or 3D point cloud data onto a 2D unit sphere. That is, to obtain the radius Geometric representation of 1 in a spherical coordinate system;

[0031] First, the center point of the rotated point cloud data or the 3D point cloud data is calculated. ;

[0032] Then according to the azimuth angle and polar angle From the center point Draw a ray from the given point and calculate the farthest intersection point between this ray and the original data (rotated point cloud data or 3D point cloud data). distance The spherical mapping can be viewed as a function, as shown in the following formula:

[0033] ;

[0034] Among them, the output The mapping value is ( ) spherical data values, This is the geometric representation in a spherical coordinate system. It is the center point farthest intersection point The distance between them; and The values ​​of are in the ranges of (0, 2π] and (0, π], respectively, and are obtained through equidistant sampling, and are uniformly distributed.

[0035] Optionally, the three-dimensional rotation operation on the three-dimensional point cloud data in step S102 is: random three-dimensional rotation operation, that is, randomly generated rotation direction and angle, thereby obtaining real rotation labels and rotated point cloud data;

[0036] The rotation directions are around the x-axis, y-axis, and z-axis.

[0037] Optionally, the process of constructing three-dimensional point cloud data in step S101 is as follows: first, the collected triaxial vibration time series signal is normalized to the [0, 1] interval using the maximum and minimum value normalization method, and then converted into a three-dimensional point cloud; then, random rotation transformation and random Gaussian noise are added for the purpose of adding interference.

[0038] The formula for calculating random rotation is:

[0039]

[0040] ;

[0041] in, The data is after random rotation transformation, where R is the rotation matrix and p is the point cloud data of the sample, such as... T is the matrix transpose. , , Let a, b, and c be the three-dimensional rotation matrices about the x-axis, y-axis, and z-axis, respectively; and let a, b, and c be the rotation matrices, respectively. , , The rotation angle, from a uniform distribution Mid-sampling, preset maximum rotation angle Take π / 18.

[0042] The present invention also provides a system based on the above-described motor fault diagnosis method, comprising:

[0043] The data acquisition module is used to acquire triaxial vibration time-series signals under various motor conditions through a triaxial vibration sensor, and to construct three-dimensional point cloud data of the motor using the triaxial vibration time-series signals, wherein the motor condition is a real fault label;

[0044] The data augmentation and spherical mapping module is used to perform three-dimensional rotation operations on the three-dimensional point cloud data to obtain rotated point cloud data and real rotation labels, and then perform spherical mapping on the rotated point cloud data to obtain rotated spherical data; and perform spherical mapping on the three-dimensional point cloud data in step S101 to obtain the original spherical data.

[0045] A collaborative training module is used for the collaborative training of supervised and self-supervised branches. The supervised and self-supervised branches use the same feature extraction module and direct the output to different classifiers. During the training phase, the rotated spherical data and real rotation labels are input into the self-supervised branch, and the original spherical data and real fault labels are input into the supervised branch.

[0046] The diagnostic module is used to input the original spherical data corresponding to the triaxial vibration timing signal of the motor to be diagnosed into the trained supervised branch to obtain the fault diagnosis classification result.

[0047] The present invention also provides a computer device, comprising: one or more processors and a memory storing a computer program;

[0048] The processor invokes a computer program to implement the steps of the above-described motor fault diagnosis method.

[0049] The present invention also provides a computer-readable storage medium storing a computer program, which is invoked by a processor to implement the steps of the above-described motor fault diagnosis method.

[0050] Compared to existing technologies, the following advancements have been achieved:

[0051] (1) The technical solution of this invention can significantly reduce the dependence on labeled data and improve the diagnostic accuracy under small sample conditions. Traditional deep learning methods usually require thousands or even tens of thousands of labeled samples to achieve high diagnostic accuracy, while this invention, through self-supervised learning, can achieve a diagnostic accuracy of over 90% with only a small number of labeled samples. Specifically, the self-supervised auxiliary task automatically generates pseudo-labels (real rotated labels) using unlabeled data, enabling the model to learn more generalizable feature representations and reducing the dependence on manual labeling. In the preferred technical solution, the dynamic weight loss function adaptively adjusts the weights of the supervised classification task and the self-supervised feature learning task, ensuring that the model can still be stably optimized under limited labeled data.

[0052] (2) Strong three-dimensional feature representation capability. The technical solution of this invention utilizes three-dimensional point cloud data and spherical convolution to provide richer spatial features than traditional one-dimensional / two-dimensional data, retain the spatial distribution features of motor operating status, and improve the distinguishability of different fault samples.

[0053] (3) Improved generalization ability: The rotational isovariability of spherical convolution enhances the robustness of the model to three-dimensional data and reduces the number of samples required for training. Experiments show that the diagnostic accuracy reaches 92.83% with 48 labeled samples. By using spherical convolution (SO(3) group isovariability), the model is made robust to data rotation, avoiding performance degradation caused by differences in sensor installation angle.

[0054] (4) Multi-task collaborative optimization: The dynamic weighting mechanism in the preferred technical solution of this invention achieves a balance between self-supervised and supervised tasks, accelerates model convergence, and improves diagnostic accuracy. It is applicable to the diagnosis of various motor faults such as stator winding short circuit, rotor bar breakage, and bearing failure. Attached Figure Description

[0055] Figure 1 This is a technical roadmap of the motor fault diagnosis method provided in the embodiments of the present invention;

[0056] Figure 2 This is a flowchart of the motor fault diagnosis method provided in Embodiment 1 of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The technical features involved in the various embodiments of the invention described below can be combined with each other as long as they do not conflict with each other.

[0058] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0060] This invention provides a method and system for motor fault diagnosis based on multidimensional data under small sample conditions. It achieves motor fault diagnosis by combining multidimensional data processing of triaxial vibration data with self-supervised deep learning, even with scarce labeled samples. To enable those skilled in the art to better understand the technical solution of this invention, the invention will be further described in detail below with reference to the accompanying drawings.

[0061] Example 1

[0062] like Figure 1 and 2 As shown in the figure, an embodiment of the present invention provides a motor fault diagnosis method based on multidimensional data under small sample conditions, which includes the following steps:

[0063] Step 1: Data acquisition. Triaxial vibration time-series signals under various motor conditions are acquired using a triaxial vibration sensor, and three-dimensional point cloud data of the motor is constructed using the triaxial vibration time-series signals. The motor condition is a real fault label.

[0064] The process involves using triaxial vibration sensors to collect triaxial vibration time-series signals of the motor under normal and various fault conditions, along the X, Y, and Z orthogonal directions. The sampling frequency is no less than 10kHz, and each sample contains at least 10,000 data points per axis. After normalizing the obtained triaxial vibration time-series signals, they are mapped to three-dimensional coordinates in a spatial coordinate system, forming three-dimensional point cloud data of the motor vibration. Specifically, the data points from the three channels are used as values ​​in the three orthogonal dimensions of three-dimensional space, with the x, y, and z axes of the three-dimensional point cloud data corresponding to the three dimensions X, Y, and Z of a Cartesian coordinate system, respectively.

[0065] In this embodiment, step 1 is preferably performed according to the following process:

[0066] S1-1: Motor data for eight fault states under three load conditions using a three-channel sensor: Normal state (NM), Rotor bar breakage fault (BBM), Stator winding fault (WFM), Bearing fault (FBM), Rotor misalignment fault (RMM), Single-phase open circuit fault (SPM), Rotor imbalance fault (RUM), and Rotor shaft bending fault (BRM). The data is normalized to the [0, 1] interval using a maximum-minimum normalization method. The normalization calculation formula is as follows:

[0067]

[0068] in, It is the first All data points in each channel, and They are the first The maximum and minimum data point values ​​in each channel. It is the maximum value of all sampled data points. After normalization, the data points from the three channels are used as values ​​in three orthogonal dimensions in three-dimensional space to complete the data fusion and construct a three-dimensional point cloud of the motor's operating state.

[0069] S1-2: To improve the model's generalization ability, data augmentation techniques for 3D point clouds are introduced. During the model training phase, one or more minor transformations (or disturbances) are introduced into a training sample to be input into the model, without changing the label of the faulty sample. First, a random rotation transformation is applied to the sample. Assume the original sample is... T is the matrix transpose symbol, R represents the rotation matrix, and the formula for calculating random rotation transformation is shown below:

[0070]

[0071] in, The sample is after random rotation transformation. , , These are the three-dimensional rotation matrices about the x-axis, y-axis, and z-axis for sample p, respectively, where a, b, and c are the rotation matrices. , , The rotation angle, from a uniform distribution Mid-sampling, in this embodiment of the invention Take π / 18. Then add random Gaussian noise to each point in the rotated sample, as shown in the following formula:

[0072]

[0073] In the formula, This represents the sample after adding Gaussian noise. Indicates Gaussian noise. This indicates a normal distribution.

[0074] It should be understood that in this embodiment, the preprocessing steps of normalization, rotation, and Gaussian noise not only expand the training dataset but also improve the robustness of the model by introducing interference. In other feasible embodiments, the above preprocessing methods can be selectively performed or not. In this invention, step 1 is simply a matter of converting the triaxial vibration time-series signal into three-dimensional point cloud data. As for the selection of preprocessing and motor fault type, this invention does not impose specific limitations on these aspects.

[0075] Step 2: Data augmentation and spherical mapping. Perform a 3D rotation operation on the 3D point cloud data from Step 1 to obtain rotated point cloud data and real rotation labels. Then, perform spherical mapping on the rotated point cloud data to obtain rotated spherical data. Perform spherical mapping on the 3D point cloud data from Step 1 to obtain the original spherical data.

[0076] This invention's technical solution includes a supervised branch and a self-supervised branch, with different data requirements for each branch. Therefore, for the self-supervised branch, the 3D point cloud data from step 1 undergoes a random 3D rotation, and corresponding pseudo-labels (real rotation labels) are generated based on the rotation direction and angle. The rotated data and its pseudo-labels are used for the self-supervised task, while the data with real labels is used for the supervised task. Specifically, the original samples are first subjected to a random 3D rotation operation. The 3D rotation operation consists of two steps: setting the direction and angle. The direction options include the x-axis, y-axis, and z-axis, representing the selected rotation axis. Multiple angle options can be set; the number of options directly determines the number of predicted labels for the self-supervised task and the difficulty of the rotation recognition task. For example, if the rotation angle is set to π / 2, there are four possible real rotation label types: π / 2 rotation along the x-axis, π / 2 rotation along the y-axis, π / 2 rotation along the z-axis, and no rotation. The more real rotation label types there are, the greater the difficulty for the model to achieve high accuracy in the self-supervised task.

[0077] Based on this, the preferred execution process of the three-dimensional rotation operation in step 2 in this embodiment is as follows:

[0078] S2-1: First, perform a random three-dimensional rotation operation on the original sample.

[0079] S2-2: Based on the randomly generated rotation direction and angle, pseudo-labels are generated and formed with the rotated data to form sample-label pairs, which are then fed into the self-supervised task branch of the model. In this embodiment, three angle options, π / 3, π / 4, and π / 6, are selected. As the number of targets to be predicted increases, the difficulty of the rotation recognition task increases, and the test and diagnostic accuracy of the model also increases.

[0080] The process of spherical mapping for either rotated point cloud data or 3D point cloud data is as follows: mapping the rotated point cloud data or 3D point cloud data onto a two-dimensional unit sphere. That is, to obtain the radius The geometric representation of a spherical sphere in spherical coordinates. In this context, a two-dimensional unit sphere is the set of all points in three-dimensional Euclidean space that are equidistant from a fixed point; it can be viewed as the surface of a sphere with a radius of one. The two dimensions correspond to the intrinsic dimension of the sphere itself as a curved surface, rather than two Cartesian coordinate dimensions in three-dimensional space. Two-dimensional unit sphere It can be parameterized as:

[0081] ;

[0082] in, Polar angle, Azimuth angle, spherical signal It is a continuous function that represents the two-dimensional unit sphere. Points on the map are mapped to complex or real numbers: R1 represents the set of real numbers, and C represents the set of complex numbers.

[0083] The technical solution of this invention can reduce the dimensionality of the original 3D data through spherical mapping, avoiding the computational cost of directly processing 3D data, while retaining its original 3D characteristics. The specific process is as follows:

[0084] First, calculate the center point of the rotated point cloud data or the 3D point cloud data.

[0085] Then according to the azimuth angle and polar angle From the center point A ray is drawn from the given point, and the farthest intersection point between this ray and the original data (including rotated point cloud data and 3D point cloud data) is calculated. distance The distance value of the farthest intersection point As the mapping value, the spherical mapping can be viewed as a function, as shown in the formula below:

[0086]

[0087] Among them, the output The mapping value is ( ) spherical data values, It is the center point farthest intersection point The distance between them; and The ranges are (0, 2π] and (0, π], respectively, and are obtained through equidistant sampling, resulting in a uniform distribution. The sampling interval is a user-defined hyperparameter; a larger sampling interval results in higher resolution and finer detail, but also higher data storage and computation costs. The specific value of the interval is not limited in this invention.

[0088] Step 3: Co-training of supervised and self-supervised branches. The spherical data is input into a fault diagnosis model trained on a dual-branch task. One branch of the model utilizes limited labeled data to complete the main fault diagnosis task; the other branch generates rotation samples and real rotation label pairs through 3D rotation to complete the designed rotation recognition task, learning feature representations beneficial to the main task from the task.

[0089] In this embodiment, both the supervised and self-supervised branches use the same feature extraction module, but direct the output to different classifiers. The self-supervised branch includes a feature extraction module and a linear layer 1, while the supervised branch includes a feature extraction module and a linear layer 2. Linear layer 1 and linear layer 2 are classifiers for predicting rotation type and fault category, respectively. The feature extraction module incorporates a spherical convolution module, which consists of several convolutional layers connected in series. Each convolutional layer is composed of a spherical convolutional layer, a nonlinear layer, and a pooling layer stacked together. The main function of the spherical convolutional layer is to extract features from the input spherical data.

[0090] The spherical convolution module proposed in this embodiment has four layers, each including spherical convolution, nonlinearity, and pooling. The feature maps output by each convolutional layer in this embodiment are 16, 32, 64, and 128, respectively. The feature vector output by the last convolutional layer will be used as input data for a classifier to complete different tasks afterwards.

[0091] A spherical convolutional layer is defined as a convolution on SO(3), where SO(3) is the set of all rotations in three-dimensional Euclidean space, represented as a 3×3 orthogonal matrix with determinant 1. Spherical convolution takes into account the algebraic properties of SO(3). Applied to this invention, spherical convolution is also equivalent to three-dimensional rotations, i.e., group elements and group operations of SO(3). Furthermore, spherical convolution also satisfies Fourier's theorem. This means that spherical convolution can be implemented in the frequency domain. The formulas for calculating the Spherical Fourier Transform (SFT) and its inverse Spherical Fourier Transform (ISFT) are shown below:

[0092]

[0093] (6)

[0094] Where b is the bandwidth of f, and is the l-th order m-th spherical harmonic function. The dot product formula for a spherical signal in the frequency domain is shown below:

[0095]

[0096] in, yes Step The coefficients of the spherical Fourier transform. It is defined on a two-dimensional sphere The original signal on the sphere, where x is a point on the sphere, and the filter... This refers to the convolution kernel of a neural network, which is a 3×3 two-dimensional matrix. For filter correspond Step Spherical Fourier transform coefficients .

[0097] In short, the implementation of spherical convolution first involves transforming the original spherical signal using a spherical Fourier transform. and filter The data is mapped to the spectral domain, then the dot product operation is performed in the spectral domain, and finally mapped back to the original space through the inverse spherical Fourier transform.

[0098] The rotated spherical data obtained by spherical mapping and the original spherical data are input into the spherical convolutional layer for feature extraction.

[0099] Furthermore, the technical solution of this invention introduces a dynamic weighting mechanism based on uncertainty, which combines supervised task loss and self-supervised task loss, and optimizes model parameters by balancing the training weights of the two branches through regularization terms.

[0100] In this model, the original spherical data and corresponding self-supervised labels are used in a self-supervised task to allow the model to learn feature representations beneficial to the main task. This learning process is completed autonomously. Through the designed architecture, the self-supervised task can assist the model in autonomously learning useful data representations. In this embodiment, it is assumed that the rotated spherical data X and its corresponding pseudo-label Y are jointly trained using the same feature extraction module, and the output results are directed to different classifiers. The goal of the self-supervised task is to progressively reduce the value of the loss function L and train the model. This can be expressed as a formula:

[0101] (8)

[0102] In the formula, N is the number of samples on the self-supervised branch. It is the set of trainable parameters of the model. ) is the loss function for a single sample.

[0103] By utilizing self-generated pseudo-labels and continuously training, the model can learn useful feature representations without relying on external, manually generated data labels.

[0104] This invention employs a combined supervised and self-supervised training approach, essentially a multi-task learning problem. The supervised branch performs the primary task, aiming to diagnose motor faults using a limited set of labeled samples. The self-supervised branch performs the secondary task, rotation recognition, aiming to learn feature representations that improve the accuracy of the primary task. The following details how to balance these two tasks during training to avoid favoring one over the other. Assume the loss value for the self-supervised task is... The loss value for supervised tasks is The total loss is the weighted sum of all losses. , and These are the weight values ​​for the two tasks, and the magnitude of the weight value determines the proportion of the loss value of each task to the total loss value, and thus determines the importance of each task.

[0105] In some embodiments, the total loss value can be expressed as:

[0106] (9)

[0107] In other embodiments, considering that manually setting weight combinations and continuously experimenting to verify the optimal weight combination is inefficient, the present invention preferably introduces a dynamic weighting mechanism based on uncertainty in the setting of the loss function to more effectively balance the weights of the two tasks during training.

[0108] That is, a regularization term is introduced into the total loss. , The expression is as shown in the formula:

[0109] (10)

[0110] Introducing the uncertainty parameter δ, which is inversely proportional to the -2 power of the weight W, i.e.:

[0111] (11)

[0112] Among them, the uncertainty parameter δ is a temperature parameter in the Gibbs distribution that determines the flatness of the final probability distribution. Fluctuations in δ affect the learning difficulty of each loss and reflect the uncertainty of a specific task. Therefore, by simplification, the regularization term... It can also be expressed through calculation as follows:

[0113] (12)

[0114] Therefore, the total loss value with the introduction of a dynamic weighting mechanism can be expressed as:

[0115] (13)

[0116] Uncertainty parameters It is also a learnable parameter. In this embodiment, the model is updated with a low learning rate during training, as expressed by the following formula:

[0117] (14)

[0118] In summary, the feature extraction modules of the self-supervised and supervised branches of the present invention sample the same structure, and the model parameters are synchronously adjusted under the guidance of the total loss function.

[0119] After comparing the experimental results of the self-supervised branching model and the unsupervised branching model, it can be clearly found that, under both sample size conditions, the test accuracy of the unsupervised branching model is significantly lower than that of the self-supervised branching model. This indicates that self-supervised branching is of great significance in motor fault diagnosis methods based on multidimensional data processing and self-supervised learning.

[0120] Step 4: Input the original spherical data corresponding to the triaxial vibration timing signal of the motor to be diagnosed into the trained supervised branch to obtain the fault diagnosis classification result.

[0121] It should be understood that, following the aforementioned steps, the triaxial vibration time-series signal of the motor to be diagnosed is first converted into three-dimensional point cloud data, then spherically mapped to spherical data, and then input into the trained supervised branch to obtain the fault diagnosis classification result.

[0122] Example 2

[0123] This invention also provides a system based on the above-described motor fault diagnosis method, comprising a data acquisition module, a data augmentation module, a spherical mapping module, a collaborative training module, and a diagnosis module connected in sequence or to each other.

[0124] The data acquisition module is used to collect triaxial vibration time-series signals under various motor conditions through a triaxial vibration sensor, and to construct three-dimensional point cloud data of the motor using the triaxial vibration time-series signals. The motor condition is a real fault label.

[0125] The data augmentation and spherical mapping module is used to perform three-dimensional rotation operations on the three-dimensional point cloud data to obtain rotated point cloud data and real rotation labels, and then perform spherical mapping on the rotated point cloud data to obtain rotated spherical data; and perform spherical mapping on the three-dimensional point cloud data in step S101 to obtain the original spherical data.

[0126] The collaborative training module is used for the collaborative training of supervised and self-supervised branches. The supervised and self-supervised branches use the same feature extraction module and direct the output to different classifiers. During the training phase, the rotated spherical data and real rotation labels are input into the self-supervised branch, while the original spherical data and real fault labels are input into the supervised branch.

[0127] The diagnostic module is used to input the original spherical data corresponding to the triaxial vibration timing signal of the motor to be diagnosed into the trained supervised branch to obtain the fault diagnosis classification result.

[0128] It should also be understood that the specific implementation process of each module is described in the above method. This invention will not repeat it here. The above division of functional modules is only for illustrative purposes. In some embodiments, some functional modules can be combined and some functional modules can be separated. Each functional module can be implemented in software, hardware, or a combination of software and hardware. The software and hardware devices include, but are not limited to, general-purpose computer equipment, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0129] Example 3

[0130] This invention also provides a computer device, including: one or more processors and a memory storing a computer program; wherein the processor calls the computer program to implement the steps of the above-described motor fault diagnosis method.

[0131] Specific implementation:

[0132] S101, Data Acquisition: The triaxial vibration time-series signal of the motor under various motor conditions is acquired through a triaxial vibration sensor, and the three-dimensional point cloud data of the motor is constructed using the triaxial vibration time-series signal. The motor condition is the actual fault label.

[0133] S102, Data augmentation and spherical mapping: Perform a 3D rotation operation on the 3D point cloud data from step S101 to obtain rotated point cloud data and real rotation labels, then perform spherical mapping on the rotated point cloud data to obtain rotated spherical data; perform spherical mapping on the 3D point cloud data from step S101 to obtain original spherical data.

[0134] S103, supervised branch and self-supervised branch are trained together. The supervised branch and self-supervised branch use the same feature extraction module and direct the output to different classifiers. During the training phase, the rotated spherical data and real rotation labels are input into the self-supervised branch, and the original spherical data and real fault labels are input into the supervised branch.

[0135] S104: Input the original spherical data corresponding to the triaxial vibration timing signal of the motor to be diagnosed into the trained supervised branch to obtain the fault diagnosis classification result.

[0136] The processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0137] The memory can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory and the processor calls and executes the algorithm program of the motor fault diagnosis method of the embodiments of this invention.

[0138] Example 4

[0139] This invention also provides a computer-readable storage medium storing a computer program, which is invoked by a processor to implement the steps of the above-described motor fault diagnosis method.

[0140] Specific implementation:

[0141] S101, Data Acquisition: The triaxial vibration time-series signal of the motor under various motor conditions is acquired through a triaxial vibration sensor, and the three-dimensional point cloud data of the motor is constructed using the triaxial vibration time-series signal. The motor condition is the actual fault label.

[0142] S102, Data augmentation and spherical mapping: Perform a 3D rotation operation on the 3D point cloud data from step S101 to obtain rotated point cloud data and real rotation labels, then perform spherical mapping on the rotated point cloud data to obtain rotated spherical data; perform spherical mapping on the 3D point cloud data from step S101 to obtain original spherical data.

[0143] S103, supervised branch and self-supervised branch are trained together. The supervised branch and self-supervised branch use the same feature extraction module and direct the output to different classifiers. During the network training stage, the rotated spherical data and real rotation labels are input into the self-supervised branch, and the original spherical data and real fault labels are input into the supervised branch.

[0144] S104: Input the original spherical data corresponding to the triaxial vibration timing signal of the motor to be diagnosed into the trained supervised branch to obtain the fault diagnosis classification result.

[0145] For details on the implementation of each step, please refer to the description in the aforementioned diagnostic method embodiment.

[0146] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the hardware and software device described in any of the foregoing embodiments, such as the hard drive or memory of the controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart MediaCard (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both internal storage units and external storage devices of the controller. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0147] Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application refers to flowchart illustrations and / or instructions executed by a processor of a method, apparatus (system), and computer program product according to embodiments of this application to create means for implementing the functions specified in one or more flowchart illustrations and / or one or more block diagrams. These computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowchart illustrations and / or one or more block diagrams. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more blocks of a block diagram.

[0149] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also within the protection scope of this invention.

Claims

1. A method for motor fault diagnosis based on multidimensional data under small sample conditions, characterized in that: Includes the following steps: S101, Data Acquisition: The triaxial vibration time-series signal of the motor under various motor conditions is acquired through a triaxial vibration sensor, and the three-dimensional point cloud data of the motor is constructed using the triaxial vibration time-series signal. The motor condition is the actual fault label. S102, Data augmentation and spherical mapping: Perform a 3D rotation operation on the 3D point cloud data from step S101 to obtain rotated point cloud data and real rotation labels, then perform spherical mapping on the rotated point cloud data to obtain rotated spherical data; perform spherical mapping on the 3D point cloud data from step S101 to obtain original spherical data. S103, supervised branch and self-supervised branch are trained together. The supervised branch and the self-supervised branch use the same feature extraction module and direct the output to different classifiers. During the training phase, the rotated spherical data and real rotation labels are input into the self-supervised branch, and the original spherical data and real fault labels are input into the supervised branch. S104: Input the original spherical data corresponding to the triaxial vibration timing signal of the motor to be diagnosed into the trained supervised branch to obtain the fault diagnosis classification result; The self-supervised branch has a feature extraction module and a linear layer 1, and the supervised branch has a feature extraction module and a linear layer 2. Linear layer 1 and linear layer 2 are classifiers for the purpose of rotation type prediction and fault category prediction, respectively. The feature extraction module is a spherical convolution module, which is composed of several convolutional layers connected in series. Each convolutional layer consists of a spherical convolutional layer, a nonlinear layer, and a pooling layer stacked together. The processing procedure for the spherical convolutional layer is as follows: First, the current spherical signal and the filter are combined using a spherical Fourier transform. Mapped to the spectral domain; ； ； in, It is a signal correspond Step The coefficients of the spherical Fourier transform. It is on a two-dimensional unit sphere The spherical signal is given by x, where x is a point on the sphere. yes Step Spherical harmonic functions of the first order; filters For neural network convolution kernels of spherical convolutional layers, For filter correspond Step The coefficients of the spherical Fourier transform of the order of 1; Then, the dot product operation is performed in the frequency domain to obtain... ; ； Finally, the original space is mapped back through the inverse Fourier transform of the sphere; ； in, b represents the result of the current spherical convolutional layer, where b is the bandwidth.

2. The motor fault diagnosis method according to claim 1, characterized in that: The total loss during network training is a weighted average of supervised and self-supervised tasks, with a dynamic weighting mechanism introduced: ； In the formula, This is the total loss value. Let i be the loss value for task i, where i takes the values ​​1 and 2, corresponding to The loss value for the self-supervised task. This represents the loss value for the supervised task. Let be the uncertainty parameter for task i. Network training process, The following updates will be made: ； In the formula, Indicates the updated uncertainty parameter η is a proportionality constant.

3. The motor fault diagnosis method according to claim 1, characterized in that: The process of performing spherical mapping on the rotated point cloud data or the spherical mapping on the 3D point cloud data in step S102 is as follows: Map rotated point cloud data or 3D point cloud data onto a 2D unit sphere. That is, to obtain the radius Geometric representation of 1 in a spherical coordinate system; First, the center point of the rotated point cloud data or the 3D point cloud data is calculated. ; Then according to the azimuth angle and polar angle From the center point Draw a ray from the given point and calculate the farthest intersection point between the ray and the rotated point cloud data or the 3D point cloud data. distance The distance value of the farthest intersection point As the mapping value, the spherical mapping can be viewed as a function, as shown in the formula below: ； Among them, the output For the mapped value, that is ( ) spherical data values, It is the center point farthest intersection point The distance between them; and The values ​​of are in the ranges of (0, 2π] and (0, π], respectively, and are obtained through equidistant sampling, and are uniformly distributed.

4. The motor fault diagnosis method according to claim 1, characterized in that: In step S102, the 3D rotation operation on the 3D point cloud data is: random 3D rotation operation, that is, randomly generated rotation direction and angle, thereby obtaining real rotation labels and rotated point cloud data; The rotation directions are around the x-axis, y-axis, and z-axis.

5. The motor fault diagnosis method according to claim 1, characterized in that: The process of constructing three-dimensional point cloud data in step S101 is as follows: First, the collected triaxial vibration time series signal is normalized to the [0, 1] interval using the maximum and minimum value normalization method and then converted into a three-dimensional point cloud; then, random rotation transformation and random Gaussian noise are added for the purpose of adding interference. The formula for calculating random rotation is: ； ； in, Let R be the data after random rotation transformation, where R is the rotation matrix and p is the point cloud data of the sample. , , Let a, b, and c be the three-dimensional rotation matrices about the x-axis, y-axis, and z-axis, respectively; and let a, b, and c be the rotation matrices, respectively. , , The rotation angle, from a uniform distribution Mid-sampling, This is the preset maximum rotation angle.

6. A system based on the motor fault diagnosis method according to any one of claims 1-5, characterized in that: include: The data acquisition module is used to acquire triaxial vibration time-series signals under various motor conditions through a triaxial vibration sensor, and to construct three-dimensional point cloud data of the motor using the triaxial vibration time-series signals, wherein the motor condition is a real fault label; The data augmentation and spherical mapping module is used to perform three-dimensional rotation operations on the three-dimensional point cloud data to obtain rotated point cloud data and real rotation labels, and then perform spherical mapping on the rotated point cloud data to obtain rotated spherical data; and perform spherical mapping on the three-dimensional point cloud data in step S101 to obtain the original spherical data. A collaborative training module is used for the collaborative training of supervised and self-supervised branches. The supervised and self-supervised branches use the same feature extraction module and direct the output to different classifiers. During the training phase, the rotated spherical data and real rotation labels are input into the self-supervised branch, and the original spherical data and real fault labels are input into the supervised branch. The diagnostic module is used to input the original spherical data corresponding to the triaxial vibration timing signal of the motor to be diagnosed into the trained supervised branch to obtain the fault diagnosis classification result.

7. A computer device, characterized in that: include: One or more processors; A memory that stores computer programs; The processor invokes a computer program to achieve the following: The steps of the motor fault diagnosis method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: The computer program is stored and is invoked by the processor to implement: The steps of the motor fault diagnosis method according to any one of claims 1-5.

Images