Gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence

Abstract

The invention discloses a gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence, and the method comprises the steps: collecting vibration signal data of a gearbox, carrying out the preprocessing of the vibration signal data, carrying out the division according to a unified time length, obtaining a plurality of samples, and defining each sample label through a fault type; performing variable step size coarse graining operation on each sample, respectively calculating CvM distance divergence of a differential signal and reference distribution on different step sizes of each coarse graining scale, taking an average value, and compressing an original high-dimensional vibration signal sequence into a low-dimensional feature vector; taking the feature vector extracted from each signal as a data feature and the fault type as a label, training a LogitBoost model, and obtaining a multi-classification model for gearbox fault diagnosis; and utilizing the trained LogitBoost model to predict the fault type of the gearbox with an unknown operation condition. According to the method, an accurate diagnosis result can be obtained without the support of a large amount of annotated data and complicated hyper-parameter selection and training processes.

Description

Technical Field

[0001] This invention belongs to the field of gearbox condition monitoring and fault diagnosis, and in particular relates to a gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM (Cramér-von Mises) divergence. Background Technology

[0002] As a core power transmission component in industrial transmission systems, gearboxes are widely used in critical fields such as wind power equipment, metallurgical rolling mills, rail transit trains, and intelligent manufacturing machine tools. Their operational reliability directly determines the production efficiency and safety level of the entire industrial system. Under long-term high-speed, variable-load, and strong-impact operating conditions, gearboxes are prone to faults such as missing teeth, tooth cracks, tooth tip peeling, and tooth surface peeling. These faults typically manifest as weak vibration abnormalities in their early stages, gradually worsening as the fault develops, eventually leading to catastrophic failures such as gearbox seizure and shaft breakage. Therefore, achieving early and accurate diagnosis of gearbox faults is a key requirement for ensuring stable operation of industrial systems and reducing maintenance costs.

[0003] Currently, the following methods are commonly used to diagnose gearbox faults:

[0004] (1) Fault model-based diagnostic methods require the construction of fault mechanism models through gear meshing dynamics and bearing vibration theory, and the fault location is achieved by comparing the deviation between the model simulation signal and the actual monitoring signal. However, gearboxes have complex structures and are subject to factors such as multi-component coupled vibration, nonlinear friction, and operating condition disturbances (such as load fluctuations and speed changes), making it difficult to construct a mathematical model that balances accuracy and computational efficiency. At the same time, actual vibration signals are often contaminated by electromagnetic noise and mechanical noise in the industrial environment, making it difficult to distinguish between model deviation and noise interference. Therefore, fault model-based diagnostic methods are generally ineffective in fault diagnosis applications of complex systems.

[0005] (2) Diagnostic methods based on experience-based knowledge rely on the practical experience of domain experts to construct a rule base for fault characteristics and fault types, and achieve diagnosis by matching monitoring signal characteristics with the rule base. However, due to the difficulty in acquiring expert knowledge, it is usually suitable for diagnosing simpler equipment faults, but not for diagnosing faults in complex systems.

[0006] (3) Data-driven diagnostic methods are based on the collected monitoring signal data, combined with various data mining techniques to obtain the hidden deep feature information, and further determine whether the gearbox equipment is in normal condition or has a fault, so as to achieve the purpose of fault detection and diagnosis.

[0007] Currently, mainstream data-driven diagnostic methods are mainly divided into two categories: traditional feature extraction + machine learning (such as support vector machines and random forests) and deep learning (such as convolutional neural networks and long short-term memory networks). However, both have significant adaptability defects: traditional methods rely on manually designed time-domain / frequency-domain features (such as peak value, root mean square, and single-scale entropy), which are difficult to effectively separate the coupled vibration signals of gears, bearings, and shaft systems, and are not sensitive enough to the weak features of early faults; although deep learning methods can automatically extract features, the model structure is complex and training requires a large number of labeled fault samples, while early fault samples of gearboxes are scarce, which can easily lead to model overfitting; at the same time, both types of methods have the problem that hyperparameter selection depends on human experience and has weak anti-interference ability. Under the dual constraints of low signal-to-noise ratio scenarios and small sample sizes in industrial environments, their generalization ability is poor and it is difficult to accurately capture early fault features.

[0008] It is evident that fault model-based methods are limited by the coupling of multiple components in the gearbox, its nonlinear characteristics, and operating condition disturbances, making it difficult to construct mathematical models that balance accuracy and efficiency. Furthermore, they are susceptible to environmental noise interference, leading to misdiagnosis. Empirical knowledge-based methods rely on expert experience and rule bases, resulting in significant knowledge transfer difficulties, poor adaptability, and an inability to handle the diverse fault types in complex systems. While data-driven methods do not require pre-setting fault mechanisms, their feature extraction lacks specificity, model robustness struggles to match the actual operating conditions of coupled vibration and low signal-to-noise ratio in gearboxes, and they require a large number of labeled samples. In summary, none of the existing diagnostic methods can efficiently address the need for accurate gearbox fault diagnosis. Therefore, there is an urgent need for a novel gearbox fault diagnosis method that can adaptively extract multi-scale fault features in small-sample scenarios and possesses strong anti-interference capabilities. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence, which can be used for gearbox fault mode recognition based on vibration signals in small sample scenarios.

[0010] The technical solution of the present invention is as follows:

[0011] A gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence includes:

[0012] S1: Collect vibration signal data of the gearbox and preprocess it. Divide it into multiple samples according to a uniform time length and define the label of each sample according to the fault type.

[0013] S2: Perform variable-step coarsening on each sample, at each coarsening scale. At different step sizes, the slope CvM distance divergence between the differential signal and the reference distribution is calculated, and the average slope CvM distance divergence at the same coarse-grained scale is taken to compress the original high-dimensional vibration signal sequence into... 3D feature vector;

[0014] S3: Using the feature vector extracted from each signal as data features and the fault type as a label, train the LogitBoost model to obtain a multi-classification model for gearbox fault diagnosis;

[0015] S4: For gearboxes with unknown operating conditions, collect their vibration signals, calculate their variable step size multi-scale slope CvM divergence index, input it into the trained multi-classification model, and output the fault type.

[0016] Furthermore, the frequency response parameter of the vibration acceleration sensor that collects vibration signal data of the gearbox should be no less than 1kHz, and the sampling frequency of the vibration signal sequence should be no less than 2.56kHz; when dividing a single sample under a uniform time length, the overlap between samples should not be higher than 50%, and multiple vibration signal samples with a time length of no less than 0.10s should be obtained.

[0017] Furthermore, S2 includes the following sub-steps:

[0018] S2.1: Given a series of values ​​for the coarsening scale, perform a coarsening operation with a variable step size at each coarsening scale;

[0019] S2.2: For each coarse-grained subsequence, a series of subsequences with an embedding dimension of m are first reconstructed;

[0020] S2.3: For each element in each subsequence of dimension m, calculate its backward difference to obtain the difference numerator sequence;

[0021] S2.4: Select an m-1 dimensional probability distribution function as a reference probability distribution;

[0022] S2.5: For each difference numerator sequence, the distribution characteristics are estimated using the empirical cumulative distribution function to obtain the corresponding empirical distribution. Then, the CvM distance is used to measure the difference between the empirical distribution and the reference probability distribution to obtain the corresponding slope CvM distance divergence.

[0023] S2.6: Take the average of the slope CvM distance divergence at the same coarse-grained scale to obtain the variable step-size slope CvM divergence of the vibration signal sequence at that scale. This allows us to construct the variable step-size multi-scale slope CvM divergence of the vibration signal sequence at all scales, which serves as the feature vector extracted for each signal.

[0024] Furthermore, the reference probability distribution is a Gaussian distribution, wherein the mean of the Gaussian distribution is an m-1 dimensional zero vector; the covariance matrix of the Gaussian distribution is selected as the average covariance matrix of the normal samples in the training set that have not been coarsened.

[0025] Furthermore, the types of gearbox failures tagged include missing teeth, tooth cracks, tooth surface peeling, and tooth tip peeling.

[0026] Furthermore, the hyperparameters of the LogitBoost model are as follows:

[0027] The CART regression tree is selected as the weak learner, with the maximum depth of the CART regression tree set to no more than 4, the minimum number of sample splits no less than 5, and the minimum number of leaf node samples no less than 2; the number of iterations of the weak learner is set to no more than 50; and the log-likelihood loss function is selected as the loss function.

[0028] Furthermore, the training process of the LogitBoost model is as follows:

[0029] Initialize the cumulative score function for all samples and categories, and initialize the probabilities after converting the scores using the Softmax function, iterating a total of R times. For each iteration r, traverse each category and train a CART regression tree for each category. The specific process is as follows:

[0030] (1) For each sample i, based on the current probability p ik Calculate weights and target value z ik ;

[0031] (2) Using a CART regression tree as a weak learner, the inputs are samples, target values, and sample weights; the CART regression tree aims to minimize the weighted mean squared error, finding the optimal split point and leaf node values ​​to obtain the first... Categories in round iteration The regression tree model; in the first After traversing all categories and obtaining K CART regression trees in each iteration, decentralization is first performed to ensure that the sum of the score functions is zero, then the cumulative score function is updated; finally, the updated cumulative score function F is used. k (x) Recalculate the probability of all samples. This is used for the next iteration.

[0032] Furthermore, the specific reasoning process of the multi-classification model that has completed training in S4 is as follows:

[0033] For each category k, the predicted values ​​of all historical CART regression trees are summed up, and the category with the highest score is selected as the prediction result.

[0034] A gearbox vibration fault diagnosis device based on variable step size multi-scale slope CvM divergence includes one or more processors for implementing a gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence.

[0035] A computer-readable storage medium is characterized in that it stores a program thereon, which, when executed by a processor, implements a gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence.

[0036] The beneficial effects of this invention are as follows:

[0037] This invention addresses the shortcomings of existing gearbox fault diagnosis and analysis techniques by proposing a method primarily based on vibration acceleration signal analysis and processing. It extracts a series of variable-step-size multi-scale differences from the phase space of the original vibration signal sequence, calculates the CvM distance with a reference distribution to derive a divergence index, constructs a feature vector, and finally inputs it into a LogitBoost model trained based on historical source domain data. This process diagnoses and outputs the specific fault category of the gearbox, thus providing crucial technical support for the safe and stable operation of the equipment. This process achieves accurate diagnostic results without requiring extensive labeled data or complex hyperparameter selection and training, which is of great significance for practical engineering applications. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence according to an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of variable step size multi-scale coarsening according to an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of the gearbox fault simulation test bench and vibration sensor arrangement in an embodiment of the present invention.

[0041] Figure 4 The faulty gear is used in the gearbox fault simulation experiment in this embodiment of the invention.

[0042] Figure 5 This is a time-domain diagram of the vibration signal sequence of the gearbox in different states in an embodiment of the present invention.

[0043] Figure 6 Error bar plots of the slope CvM divergence of the gearbox vibration signal in each state on 10 scales.

[0044] Figure 7 This is a graph showing the change of the loss function value of the LogitBoost multi-class classification model during training with the number of iterations in an embodiment of the present invention.

[0045] Figure 8 This is a confusion matrix diagram of the output results of the LogitBoost multi-classification model after diagnosing the input test set gearbox vibration signal in an embodiment of the present invention. Detailed Implementation

[0046] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0047] This invention is based on vibration signal analysis. It extracts a series of variable-step-size multi-scale differences from the phase space of the original vibration signal sequence, calculates the CvM distance with a reference distribution to obtain a divergence index, constructs a feature vector, and finally inputs it into a LogitBoost model trained on historical data from the source domain. This process diagnoses and outputs the specific fault category of the gearbox, thus providing crucial technical support for the safe and stable operation of the equipment. This process eliminates the need for complex hyperparameter selection and can obtain accurate diagnostic results even with relatively small datasets.

[0048] like Figure 1 As shown, the gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence in this embodiment includes the following steps one through four.

[0049] Step 1: Collect vibration signal data of the gearbox and preprocess it. Divide the data into multiple samples according to a uniform time length, and define the label of each sample according to the fault type.

[0050] The frequency response parameter of the vibration acceleration sensor used to collect vibration signal data of the gearbox should be no less than 1 kHz, and the sampling frequency of the vibration signal sequence should be no less than 2.56 kHz. When slicing the vibration signal data of the gearbox, the overlap between samples should not exceed 50%, resulting in multiple vibration signal samples with a time length of no less than 0.10 s.

[0051] Step 2: Perform variable step-size coarsening on each sample. At different step sizes for each coarsening scale, calculate the slope CvM distance divergence between the difference signal and the reference distribution. Then, average the slope CvM distance divergence at the same coarsening scale to compress the original high-dimensional vibration signal sequence. 3D eigenvectors.

[0052] Step two includes the following sub-steps:

[0053] (2-1) Given a series of values ​​for scale λ, perform variable step size at each λ. coarse-graining operation, Both and are positive integers, and This yields one coarse-grained subsequence at each scale λ. ;

[0054] For a given scale φ and step size At this time, variable step size coarsening can be expressed as:

[0055]

[0056] in, , ,scale and step length All are positive integers and satisfy the following conditions: .

[0057] (2-2) For each coarse-grained subsequence First, the reconstruction yields a series of embedding dimensions. subsequence of: , ,…, .

[0058] (2-3) For each element in each subsequence, compute its backward difference. The differential molecular sequence was obtained. ,…, .

[0059] (2-4) Choice A dimensional probability distribution function as a reference probability distribution ,in for A dimensional vector.

[0060] In this embodiment, considering the effect in practical applications, a Gaussian distribution is taken as the reference probability distribution. ,in Let the mean vector of the reference distribution be taken. The zero vector of dimension; The reference distribution covariance matrix is ​​specifically selected from the average covariance matrix of the normal samples (in the absence of faults) in the training set that have not been coarsened.

[0061] (2-5) For coarse-grained sequences A series of differential molecular sequences were obtained The empirical cumulative distribution function (ECDF) is used to estimate its distribution characteristics, and then the CvM distance is used to measure the empirical distribution. and reference probability distribution Differences between them:

[0062]

[0063] in This represents satisfaction in various dimensions. All less than of The number of.

[0064]

[0065] in, Represents the cumulative distribution of experience and The CvM distance between them, i.e., the signal The slope CvM distance divergence.

[0066] (2-6) After calculating the slope CvM divergence for each coarse-grained subsequence, and then averaging the slope CvM distance divergence at the same coarse-grained scale, the variable step-size slope CvM divergence of the vibration signal sequence at that scale can be obtained. For vibration signals at multiple scales ,…, By extracting the variable step-size slope CvM divergence, the variable step-size multi-scale slope CvM divergence can be obtained, thus constructing the variable step-size multi-scale slope CvM divergence of the signal. .

[0067] The formula for calculating the mean is as follows:

[0068]

[0069] like Figure 2 As shown, Figure 2 The left image shows In this case, a corresponding coarse-grained sequence (the yellow sequence in the gray area) and The two coarse-grained sequences corresponding to τ are shown (the yellow sequences within the two light blue boxes). For each τ subsequence under τ, the divergence is calculated separately, and finally, the average is obtained as a numerical value.

[0070] In this embodiment, considering the diagnostic effectiveness in practical applications and the timeliness of method implementation, the embedding dimension λ is set to 3, and the maximum scale of the extracted variable step size multi-scale slope CvM distance divergence is determined. Selected as 10.

[0071] Step 3: Using the feature vector extracted from each signal as data features and the fault type as a label, train the LogitBoost model to obtain a multi-classification model for gearbox fault diagnosis.

[0072] The specific types of gearbox faults used as training labels should include, but are not limited to, missing teeth, tooth cracks, tooth surface peeling, and tooth tip peeling.

[0073] The hyperparameters of the LogitBoost model used are as follows: CART (Classification and Regression Tree) is selected as the weak learner, with the maximum depth of the CART regression tree set to no more than 4 (to avoid overfitting while ensuring feature fitting ability), the minimum number of sample splits no less than 5 (to ensure effective learning of fault features in small samples), and the minimum number of leaf node samples set to no less than 2; the number of weak learner iterations (i.e., the number of boosting rounds) is set to no more than 50; the log-likelihood loss function is selected, with the expression: ,in For sample labels, This is the output of a linear combination of classifiers.

[0074] The specific training process of the LogitBoost model is as follows:

[0075] (3-1) Model input vector

[0076] Let the maximum scale used in the variable step size multi-scale stage be . Let the extracted variable step size multi-scale slope CvM distance divergence MVCM be denoted as . ,but ,in , Number of samples; category labels , This represents the number of categories.

[0077] To simplify calculations, the labels are converted into one-hot vectors. :

[0078]

[0079] The model output is Cumulative score function Each corresponds to a different fault type, where Representative sample Belongs to the The score for the class.

[0080] Then, the softmax function is used to convert the score into a probability. :

[0081]

[0082] (3-2) For all samples and categories Initialize all cumulative score functions The initial probability is uniformly distributed. .

[0083] (3-3) Co-iteration Times. For each iteration Iterate through each category Train a CART regression tree for each category. The specific process is as follows:

[0084] (3-3-1) For each sample i, based on the current probability p ik Calculate weights and target value z ik :

[0085]

[0086]

[0087] (3-3-2) Using CART regression trees as weak learners, the input is samples The target value is The sample weights are CART regression trees minimize the weighted mean squared error. As an objective, we find the optimal split point and leaf node values ​​to obtain the... Categories in round iteration CART regression tree .

[0088] (3-3-3) in the first In the round of iteration, all categories are traversed to obtain After constructing the CART regression tree, decentralization is first performed to ensure that the sum of the cumulative score functions is zero:

[0089]

[0090] Then update the cumulative score function. ,in This is an additional learning rate added to prevent overfitting, and it is usually 0.1 or less.

[0091] (3-3-4) Using the updated cumulative score function Recalculate the probability of all samples. For use in the next iteration:

[0092]

[0093] Step 4: For gearboxes with unknown operating conditions, collect their vibration signals, calculate their variable step size multi-scale slope CvM divergence index, input it into the trained multi-classification model, and output the fault type.

[0094] The LogitBoost model used in the inference process is based on the results obtained after training. Group-strong classifier (each group contains) After accumulating the CART regression trees, the new feature vector... Make a prediction. The specific reasoning process is as follows:

[0095] (4-1) For each category The cumulative score function is obtained by summing the predicted values ​​of all historical CART regression trees. :

[0096]

[0097] (4-2) Select the category with the highest score as the prediction result:

[0098]

[0099] Corresponding to the aforementioned embodiment of the gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence, the present invention also provides an embodiment of a gearbox vibration fault diagnosis device based on variable step size multi-scale slope CvM divergence.

[0100] The gearbox vibration fault diagnosis device based on variable step size multi-scale slope CvM divergence provided in this embodiment of the invention includes one or more processors for implementing the gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence in the above embodiment.

[0101] The gearbox vibration fault diagnosis device based on variable step-size multi-scale slope CvM divergence of the present invention can be applied to any device with data processing capabilities, such as a computer. The device can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of the device with data processing capabilities reading the corresponding computer program instructions from non-volatile memory into memory and executing them. From a hardware perspective, in addition to the processor, memory, network interface, and non-volatile memory, the device with data processing capabilities in the embodiment may also include other hardware depending on its actual functions; these will not be elaborated further.

[0102] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0103] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0104] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence described in the above embodiments.

[0105] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0106] The following specific implementation example demonstrates the beneficial effects of the present invention.

[0107] In this embodiment, a multi-classification model is trained on gearbox vibration signal data containing nine different fault types, and small-sample fault diagnosis is achieved.

[0108] In this embodiment, a vibration acceleration sensor is arranged on the gearbox housing to extract vibration signal data, and the sensor sampling frequency is 20000Hz. Figure 3 On the gearbox failure simulation test bench shown, different gear failure components (such as...) were replaced and adjusted. Figure 4As shown, the vibration signal data simulates different typical gearbox faults, including missing teeth, tooth cracks, tooth surface spalling, and tooth tip spalling. When dividing the gearbox vibration signal data, a time length of 0.18s was used, resulting in a dataset containing 9 state types, each type including 100 signals of length 3600. 20% of the dataset was randomly selected as the training set (20 data points for each fault type), and the remainder was used as the test set. The gearbox states include healthy, missing teeth, tooth cracks, tooth surface spalling, tooth tip spalling 5, tooth tip spalling 4, tooth tip spalling 3, tooth tip spalling 2, and tooth tip spalling 1 (where the degree of tooth tip spalling ranges from 5 to 1, representing progressively increasing severity). The time-domain plots of the vibration signal sequences under different gearbox fault types are shown below. Figure 5 As shown.

[0109] In this embodiment, the error bar plot of the slope CvM divergence of vibration signals for different gearbox fault types at 10 scales is shown below. Figure 6 As shown.

[0110] In this embodiment, the hyperparameters of the LogitBoost multi-class classification model are as follows: a CART regression tree is selected as the weak learner, with the maximum depth of the CART regression tree set to 3, the minimum number of sample splits set to 8, and the minimum number of leaf node samples set to 2; the number of iterations for the weak learner is selected to be 30; and the learning rate is selected to be 0.05. The curve showing the change in the LogitBoost multi-class classification model loss with the number of iterations for the extracted features is shown below. Figure 7 As shown. From Figure 7 It can be seen that, due to the small sample size, the model quickly fits the training set; the loss on the test set also tends to stabilize after a rapid decrease.

[0111] After extracting features from the gearbox vibration signal sequences of each state in the test set, the features are input into a LogitBoost multi-classification model trained based on source domain data, and the final diagnostic output is the specific fault category of the gearbox. The confusion matrix results on the test set are as follows: Figure 8 As shown, the LogitBoost multi-class classification model achieves an average gearbox fault identification accuracy of 99.58% on the test dataset. Out of a total of 720 test samples, one "healthy" sample and two "missing tooth" samples were incorrectly identified as "tooth tip peeling 5" and "healthy," respectively. Two "missing tooth" samples, two "tooth tip peeling 4" samples, and one "tooth tip peeling 2" sample were incorrectly identified as "healthy," "tooth tip peeling 2," and "tooth tip peeling 4," respectively. The remaining 717 samples were correctly identified, demonstrating the excellent application effect of the method of this invention in the diagnosis and identification of gearbox vibration faults.

[0112] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for diagnosing gearbox vibration faults based on variable step-size multi-scale slope CvM divergence, characterized in that, include: S1: Collect vibration signal data of the gearbox and preprocess it. Divide it into multiple samples according to a uniform time length and define the label of each sample according to the fault type. S2: Perform variable-step coarsening on each sample, at each coarsening scale. At different step sizes, the slope CvM distance divergence between the differential signal and the reference distribution is calculated, and the average slope CvM distance divergence at the same coarse-grained scale is taken to compress the original high-dimensional vibration signal sequence into... 3D feature vector; S3: Using the feature vector extracted from each signal as data features and the fault type as a label, train the LogitBoost model to obtain a multi-classification model for gearbox fault diagnosis; S4: For gearboxes with unknown operating conditions, collect their vibration signals, calculate their variable step size multi-scale slope CvM divergence index, input it into the trained multi-classification model, and output the fault type.

2. The gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence according to claim 1, characterized in that, The frequency response parameter of the vibration acceleration sensor that collects vibration signal data of the gearbox should be no less than 1kHz, and the sampling frequency of the vibration signal sequence should be no less than 2.56kHz. When dividing a single sample under a uniform time length, the overlap between samples should not be higher than 50%, and multiple vibration signal samples with a time length of no less than 0.10s should be obtained.

3. The gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence according to claim 1, characterized in that, S2 includes the following sub-steps: S2.1: Given a series of values ​​for the coarsening scale, perform a coarsening operation with a variable step size at each coarsening scale; S2.2: For each coarse-grained subsequence, a series of subsequences with an embedding dimension of m are first reconstructed; S2.3: For each element in each subsequence of dimension m, calculate its backward difference to obtain the difference numerator sequence; S2.4: Select an m-1 dimensional probability distribution function as a reference probability distribution; S2.5: For each difference numerator sequence, the distribution characteristics are estimated using the empirical cumulative distribution function to obtain the corresponding empirical distribution. Then, the CvM distance is used to measure the difference between the empirical distribution and the reference probability distribution to obtain the corresponding slope CvM distance divergence. S2.6: Take the average of the slope CvM distance divergence at the same coarse-grained scale to obtain the variable step-size slope CvM divergence of the vibration signal sequence at that scale. This allows us to construct the variable step-size multi-scale slope CvM divergence of the vibration signal sequence at all scales, which serves as the feature vector extracted for each signal.

4. The gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence according to claim 3, characterized in that, The reference probability distribution is a Gaussian distribution, where the mean of the Gaussian distribution is an m-1 dimensional zero vector; the covariance matrix of the Gaussian distribution is selected from the average covariance matrix of the normal samples in the training set that have not been coarsened.

5. The gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence according to claim 1, characterized in that, The types of gearbox failures tagged as such include missing teeth, tooth cracks, tooth surface spalling, and tooth tip spalling.

6. The gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence according to claim 1, characterized in that, The hyperparameters of the LogitBoost model are as follows: The CART regression tree is selected as the weak learner, with the maximum depth of the CART regression tree set to no more than 4, the minimum number of sample splits no less than 5, and the minimum number of leaf node samples no less than 2; the number of iterations of the weak learner is set to no more than 50; and the log-likelihood loss function is selected as the loss function.

7. The gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence according to claim 1, characterized in that, The training process of the LogitBoost model is as follows: Initialize the cumulative score function for all samples and classes, and initialize the probabilities after the scores are transformed using the Softmax function. Iterate for a total of R times; for each iteration... r The process involves iterating through each category and training a CART regression tree for each category. The specific steps are as follows: (1) For each sample i Based on the current probability p ik Calculate weights and target value z ik ; (2) Using a CART regression tree as a weak learner, the inputs are samples, target values, and sample weights; the CART regression tree aims to minimize the weighted mean squared error, finding the optimal split point and leaf node values ​​to obtain the first... r Categories in round iteration k Regression tree model; In the r After traversing all categories and obtaining K CART regression trees in each iteration, the trees are first decentralized to ensure that the sum of the score functions is zero, then the cumulative score function is updated; finally, the updated cumulative score function is used. F k ( x Recalculate the probability of all samples. p ik This is used for the next iteration.

8. The gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence according to claim 7, characterized in that, The specific reasoning process of the multi-class classification model that has completed training in S4 is as follows: For each category k The predicted values ​​of all historical CART regression trees are summed up, and the category with the highest score is selected as the prediction result.

9. A gearbox vibration fault diagnosis device based on variable step size multi-scale slope CvM divergence, characterized in that, It includes one or more processors for implementing the gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the gearbox vibration fault diagnosis method based on variable step size multi-scale slope CvM divergence as described in any one of claims 1 to 8.

Images