A bearing fault detection method based on wavelet transform and hybrid learning

Abstract

The application relates to the field of bearing fault detection, and specifically discloses a bearing fault detection method based on wavelet transformation and hybrid learning, which comprises the following steps: pre-processing bearing fault signal data, performing wavelet transformation on the pre-processed data according to a bearing fault principle, respectively calculating component energy of obtained detail components and approximate components and normalizing the component energy to obtain a six-dimensional feature set, randomly dividing the feature set into a training set and a test set, respectively inputting the training set and the test set into a support vector machine classifier for classification, adding classification results to feature spaces of the training set and the test set respectively, inputting the training set and the test set into a Gaussian naive Bayes classifier for training, respectively adding classification results to feature spaces of the training set and the test set, inputting the training set and the test set into a random forest classifier for training, and obtaining final classification results. The application can effectively diagnose bearing faults by analyzing and processing original vibration signals, extracting wavelet transformation features, and fusing and classifying by using a multi-machine learning method.

Description

Technical Field

[0001] This invention relates to the field of bearing fault detection, and specifically to a bearing fault detection method based on wavelet transform and hybrid learning. Background Technology

[0002] In an era of rapid technological advancement, the importance of engineering machinery is growing, finding applications in all aspects of people's production and daily life. Bearings are crucial components in mechanical equipment, primarily functioning to support rotating parts, reduce friction during movement, suppress friction generated during shaft rotation, reduce energy consumption and heat generation, and prevent component damage. They play a key role in mechanical equipment. According to Grand View Research, the global bearing market reached $118.7 billion in 2020 and is expected to maintain a compound annual growth rate of 8.5% until 2028. In 2020, the Chinese bearing industry's revenue reached 193 billion yuan, accounting for 25% of the global market. As one of the most frequently used components within engineering machinery, bearings often experience frequent failures. Therefore, developing effective equipment condition monitoring and fault diagnosis technologies has become an urgent need for equipment management and maintenance. Accurate and timely identification of faults arising and evolving during operation is of great significance for the safe operation of mechanical systems and the avoidance of major economic losses and catastrophic accidents.

[0003] However, existing bearing fault detection methods have certain shortcomings. First, traditional methods often struggle to accurately determine the type and location of the fault. Second, traditional methods often fail to effectively handle noise and useless information in the original signal, resulting in poor feature extraction. Furthermore, due to the insufficient sensitivity of the feature extraction methods, they often fail to capture the sensitive features of the vibration signal, thus affecting the accuracy of the classifier. Finally, traditional bearing fault detection methods typically use only a single machine learning algorithm for classification, failing to fully utilize the advantages of various algorithms, leading to limited classification results. Therefore, this invention proposes a bearing fault detection method based on wavelet transform and hybrid learning. This method extracts features by performing wavelet transform on the vibration signal and uses a composite machine learning method to achieve efficient and accurate classification. This technology overcomes the aforementioned shortcomings of traditional bearing fault detection methods and has broad application prospects. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a bearing fault detection method based on wavelet transform and hybrid learning.

[0005] The purpose of this invention is to provide a method for efficiently and accurately detecting bearing faults. This method utilizes wavelet transform to extract features from bearing vibration signals and employs machine learning algorithms such as support vector machines, Gaussian Naive Bayes, and random forests for classification and diagnosis.

[0006] A bearing fault detection method based on wavelet transform and hybrid learning includes:

[0007] Step S1 involves preprocessing the bearing fault signal data to achieve noise reduction and filtering of useful data.

[0008] Step S2: Based on the bearing fault principle, perform wavelet transform on the preprocessed bearing fault signal data, calculate the component energy of the obtained detail components and approximate components respectively and normalize them to obtain a six-dimensional feature set.

[0009] Step S3: Randomly divide the six-dimensional feature set into a training set and a test set, input the training set and the test set into the support vector machine classifier for classification, and add the classification results to the feature space of the training set and the test set respectively.

[0010] Step S4: Input the training set obtained in step S3 into the Gaussian Naive Bayes classifier for training, and then input the test set obtained in S3 into the Gaussian Naive Bayes classifier for classification. Add the classification results to the feature spaces of the training set and the test set obtained in step 3 respectively.

[0011] Step S5: Input the training set obtained after training with the Gaussian Naive Bayes classifier in step S4 into the random forest classifier for training, and then input the test set obtained after training with the Gaussian Naive Bayes classifier in step S4 into the random forest classifier for classification to obtain the final classification result of the test set samples.

[0012] Furthermore, in step S1, invalid data is manually filtered, low-pass filtered, and the data is divided into multiple segments of 4096 points each.

[0013] Furthermore, the wavelet transform in step S2 uses the db8 wavelet basis and calculates the normalized energy of the six components.

[0014] Furthermore, the process of extracting features from the preprocessed data using wavelet transform in step S2 is as follows:

[0015] Step S21 involves performing a Discrete Wavelet Transform (DWT) on the original bearing fault signal data, including:

[0016] Step S211, assuming the original signal is x[n], the low-pass filter corresponding to the db8 wavelet is g[n], and the high-pass filter is h[n], then the result of the first layer DWT is:

[0017]

[0018]

[0019] Where, x 1，L The approximate component of the first layer, x 1，H The first layer is called the detail component, and K is the filter length; since the db series wavelets are difficult to represent analytically, g[n] and h[n] are given by the pywt library of Python;

[0020] Step S212, perform DWT on the approximate components of the (α-1)th layer to obtain:

[0021]

[0022]

[0023] Step S213: Perform DWT five times consecutively to obtain the detail components x of the first five layers. α，H (α = 1, 2, 3, 4, 5) and the approximate component x of the fifth layer 5，L ;

[0024] Step S22: Then, the energy of each detailed component is calculated, and the normalized energy is used as the feature of the bearing fault signal data of that segment.

[0025] Further, step S22 includes:

[0026] Step S221: Calculate the energy C of each detailed component. α (α = 1, 2, 3, 4, 5) and the energy A of the fifth-layer approximation component, where N is the signal length:

[0027]

[0028]

[0029] Step S222: Normalize the energy of each component to obtain C′ α (α=1, 2, 3, 4, 5) and A′:

[0030]

[0031]

[0032] The six normalized energies obtained are used as features of the bearing fault signal data for this segment, and are used for subsequent learning and classification tasks.

[0033] Furthermore, in step S3, the process of using a support vector machine classifier to classify the training set and the test set, and adding the classification results to the training set and the test set respectively, is as follows:

[0034] Step S31, solving for the support vectors refers to finding the maximum separating hyperplane of the form y = wx + b, where w and b are the parameters to be solved, x is the feature vector, and y is the sample class. For binary classification problems, this can be represented by -1 and 1, and can be transformed into the following mathematical programming problem:

[0035]

[0036]

[0037] Where x i Let y be the eigenvector. i The category to which the sample belongs can be represented by -1 and 1 for binary classification problems, where n is the number of samples and λ is the value of λ. i , λ j is the Lagrange multiplier, which is an intermediate variable for solving the problem. st means "satisfy", which introduces the condition that the subsequent parameters should satisfy. λ refers to the vector composed of all the Lagrange multipliers.

[0038] If we introduce soft margins to represent the severity of misclassification, then the problem to be solved is:

[0039]

[0040] stg i (w, b) = 1 - y i (w T x i +b)-ξ i ≤0, ξ i ≥0, i=1,2,…,n

[0041] Where C is the penalty coefficient, ξ i w is the slack variable used to measure the soft interval. T The transpose of vector w represents the dual problem of solving the above planning problem using the Lagrange method. The Lagrange function can be constructed as follows:

[0042]

[0043] stλ i ≥0μ i ≥0

[0044] Where λ i and μ i They are Lagrange multipliers, w, b, ξ i The main parameter.

[0045] By taking the derivative with respect to the principal parameters and setting the derivative to 0, the program can be transformed into the following program:

[0046]

[0047]

[0048] For linearly inseparable feature sets, the kernel function method is used, and the hyperplane is f(x) = w T The planning problem to be solved is φ(x)+b:

[0049]

[0050]

[0051] Common kernel functions include Gaussian kernel (rbf), etc. Where δ is a hyperparameter that controls the range of influence of the Gaussian kernel function. The larger its value, the greater the local influence range of the Gaussian kernel function.

[0052] Step S32: Iterate using the sequence minimum optimization algorithm to find the optimal λ.

[0053] Step S33, when calculating the partial derivative, we have And for λ i Points with a value greater than 0 are all support vectors. A support vector (x) can be randomly selected. s y s If b = y, then b = y s -wx s ;

[0054] Step S34: The result of step S33 can be used to represent the maximum segmentation hyperplane: w T x+b=0 and the classification decision function: f(x)=sgn(w T x+b), where sgn(·) is the sign function, defined as follows:

[0055]

[0056] Step S35: Add the classification results of the training set to the original training set, and add the classification results of the test set to the original test set for subsequent processing.

[0057] Furthermore, step S32 specifically includes:

[0058] Step S321: Select two parameters λ that need to be updated. i and λ j With other parameters fixed, the constraint becomes:

[0059] λ i y i +λ j y j =c,λ i ≥0, λ j ≥0

[0060]

[0061] From this we can obtain This then transforms into a programming problem with only one variable constraint;

[0062] Step S322: For a programming problem with only one variable constraint, we can adjust λ... i Find the partial derivative, set the initial derivative to 0, and then find the variable λ. inew Then λ is calculated. jnew ;

[0063] Step S323: Repeat steps S321 and S322, iterating until convergence, and find the optimal λ.

[0064] Furthermore, step 4, which involves training the training set using a Gaussian Naive Bayes classifier and classifying the test set, is as follows:

[0065] Step S41, according to Bayes' theorem, the classification is as follows when feature x is x: y The probability of i is P(y i |x)=P(x|y i )P(y i ) / P(x), and then we can obtain the law of total probability from it.

[0066] Step S42, solve for the likelihood P(x|y) i Assume that for a continuous variable x, there is a Gaussian distribution assumption, that is, it is assumed that in y i Under the condition that x follows a Gaussian distribution The parameters μ and σ can be obtained by maximum likelihood estimation;

[0067] Step S43: For multidimensional features, there is the Naive Bayes assumption, which assumes that the features are independently distributed, i.e., P(x|y) i )=P(x1|y i )P(x2|y i ...P(x) n |y i );

[0068] Step S44: Based on the above assumptions and formulas, P(y) can be calculated. i |x) and select the category with the highest probability as the final category y. i .

[0069] Furthermore, the process of training the training set and classifying the test set using a random forest classifier in step 5 is as follows:

[0070] Step S51: Perform bootstrap sampling, that is, assuming there are a total of N samples in the feature space, N samples are drawn from the training set with replacement as the training set for one tree;

[0071] Step S52: If there are M features, take m<<M, keep m fixed throughout the whole process, randomly select m features, and use the best feature among these m feature dimensions as the split node to build a decision tree;

[0072] Step S53: Repeat steps S51 and S52 k times, that is, k decision trees are built, these k decision trees form a random forest, and the class that the data belongs to is determined by the voting result.

[0073] The beneficial effects of the present invention are as follows:

[0074] 1) The present invention adopts a filtering technology to preprocess the original data, which reduces the influence of noise on subsequent analysis and improves the signal-to-noise ratio of the signal;

[0075] 2) The multi-level wavelet transform used in the present invention can obtain multi-layer detail and approximation components. By calculating and normalizing the energy of each component, a six-dimensional feature set can be obtained. In addition, a cascaded classifier is used for classification and diagnosis, which can give full play to the advantages of multiple classifiers and improve the accuracy of fault detection;

[0076] 3) The classifier of the present invention is simple to train, occupies low machine resources during operation, has fast operation speed, and can perform fault detection in various working scenarios. Brief Description of the Drawings

[0077] Figure 1 is a flowchart of the bearing fault detection method based on the fusion of wavelet transform and multiple machine learning methods provided by this embodiment;

[0078] Figure 2 is a relative wavelet energy bar graph of samples under four working conditions provided by this embodiment;

[0079] Figure 3 is a specific flowchart of the bearing fault detection method based on the fusion of wavelet transform and multiple machine learning methods provided by this embodiment;

[0080] Figure 4 is a comparison diagram of the classification results between the bearing fault detection method based on the fusion of wavelet transform and multiple machine learning methods and the bearing fault detection method based on wavelet transform and a single machine learning method provided by this embodiment.

[0081] In the drawings:

[0082] Normal - Normal bearing, IR007 - Inner ring fault, B007 - Rolling element fault, OR007@3 - Outer ring fault; "Support Vector Machine" - Classification result using only Support Vector Machine classifier; "Gaussian Naive Bayes" - Classification result using only Gaussian Naive Bayes classifier; "Random Forest" - Classification result using only Random Forest classifier; "Fusion Algorithm" - Classification result of bearing fault detection method based on wavelet transform and multiple machine learning methods fusion in this embodiment. Detailed Implementation

[0083] The invention will now be further described with reference to the accompanying drawings.

[0084] Please refer to Figures 1-4 The present invention provides a bearing fault detection method based on wavelet transform and hybrid learning, comprising the following steps:

[0085] Step S1 involves preprocessing the bearing fault signal data to remove noise and filter useful data. Specifically, this includes:

[0086] In this embodiment, the original dataset is manually screened to remove invalid data, low-pass filtered, and the data is divided into multiple slices of 4096 points each.

[0087] Step S2: Based on the bearing failure principle, perform wavelet transform on the preprocessed data, calculate the component energy of the obtained detail components and approximate components respectively and normalize them to obtain a six-dimensional feature set.

[0088] Based on the bearing fault principle, wavelet transform is performed on the preprocessed data. The energy of the obtained detail components and approximate components is calculated and normalized to obtain a six-dimensional feature set. In this embodiment, the essence of wavelet transform is to decompose the original signal using multiple wavelet basis functions. It has both frequency analysis properties and contains time-domain information. This method can effectively extract time-frequency domain information and effectively separate each component from the signal, and perform subscale refinement analysis on the function or signal.

[0089] The specific method for feature extraction from the preprocessed data using wavelet transform is as follows:

[0090] Step S21: Perform Discrete Wavelet Transform (DWT) on the original signal. The specific process is as follows:

[0091] Step S211, assuming the original signal is x[n], the low-pass filter corresponding to the db8 wavelet is g[n], and the high-pass filter is h[n], then the result of the first layer DWT is:

[0092]

[0093]

[0094] Where x 1，L The approximate component of the first layer, x 1，H The first layer of detail components, g[n] and h[n], are provided by the Python pywt library because the db series wavelets are difficult to represent analytically. Their specific parameters are shown in the table below:

[0095]

[0096] Step S212, perform DWT on the approximate components of the (α-1)th layer to obtain:

[0097]

[0098]

[0099] Step S213: Perform DWT five times consecutively to obtain the detail components x of the first five layers. α，H (α = 1, 2, 3, 4, 5) and the approximate component x of the fifth layer 5，L

[0100] Step S22, then calculate the energy for each detailed component, the specific process is as follows:

[0101] Step S221: Calculate the energy C of each of the above detailed components. α (α = 1, 2, 3, 4, 5) and the energy A of the fifth-layer approximate component:

[0102]

[0103]

[0104] step S 222, normalizing the energy of each component yields C′ α (α=1, 2, 3, 4, 5) and A′:

[0105]

[0106]

[0107] The six normalized energies obtained are used as features of this data segment for subsequent learning and classification tasks.

[0108] The characteristics obtained for bearings with different fault types vary significantly; please refer to [reference needed]. Figure 2 .

[0109] Step S3: Randomly divide the feature set into a training set and a test set, input the training set and the test set into a support vector machine classifier for classification, and add the classification results to the feature space of the training set and the test set respectively.

[0110] Support Vector Machine (SVM) is a supervised learning model and related learning algorithm used to analyze data in classification problems. The specific method for classifying a feature set using this method and adding the classification results to the training set is as follows:

[0111] Step S31, Support Vector Machine refers to finding a hyperplane w in the feature space. T The expression x + b = 0 maximizes the distance from the support vectors to the hyperplane, where support vectors refer to the feature points closest to the hyperplane. Solving for the support vectors can be transformed into the following mathematical programming problem:

[0112]

[0113]

[0114] Step S32 involves iteratively applying a sequence minimum optimization algorithm to find the optimal λ. Steps S321-S323 below describe the specific algorithm flow.

[0115] Step S321: Select two parameters λ that need to be updated. i and λ j With other parameters fixed, the constraint now becomes:

[0116] λ i y i +λ j y j =c,λ i ≥0, λ j ≥0

[0117]

[0118] From this we can obtain This then transforms into a programming problem with only one variable constraint;

[0119] Step S322, for problems with only one variable constraint, we can adjust λ i Find the partial derivative, set the initial derivative to 0, and then find the variable λ. inew Then λ is calculated. jnew ;

[0120] Step S323: Repeat steps S321 and S322 iteratively until convergence, and find the optimal λ.

[0121] Step S33, when calculating the partial derivative, we have And for λi Points with a value greater than 0 are all support vectors. A support vector (x) can be randomly selected. s y s If b = y, then b = y s -wx s ;

[0122] Step S34: The result of step S33 can be used to represent the maximum segmentation hyperplane: w T x+b=0 and the classification decision function: f(x)=sgn(w T x+b);

[0123] Step S35: Simultaneously, considering the possibility of sample misclassification, the present invention introduces a soft margin:

[0124]

[0125] stg i (w, b) = 1 - y i (w T x i +b)-ξ i ≤0, ξ i ≥0, i=1,2,...,n

[0126] Where C is the penalty coefficient, ξ i Let be the slack variable used to measure the soft margin. The dual problem of the above planning problem, solved using the Lagrange method, can be constructed using the following Lagrange function:

[0127]

[0128] stλ i ≥0μ i ≥0

[0129] Where λ i and μ i They are Lagrange multipliers, w, b, ξ i The main parameter.

[0130] By taking the derivative with respect to the principal parameters and setting the derivative to 0, the program can be transformed into the following program:

[0131]

[0132]

[0133] For linearly inseparable feature sets, the kernel function method is used, and the hyperplane is f(x) = w T The planning problem to be solved is φ(x)+b:

[0134]

[0135]

[0136] Common kernel functions include Gaussian kernel (rbf), etc.

[0137] After merging the feature sets, proceed to step S4.

[0138] Step S4: Input the training set obtained in step S3 into the Gaussian Naive Bayes classifier for training, and then input the test set obtained in step S3 into it for classification. Add the classification results to the feature spaces of the training set and the test set obtained in step 3 respectively.

[0139] Among them, Gaussian Naive Bayes is a variant of the Naive Bayes classifier. It is a probabilistic classification method based on Bayes' theorem, assuming that the probability distribution of each feature follows a Gaussian distribution. The specific method for training and classifying the feature training set using this approach is as follows:

[0140] Step S41, according to Bayes' theorem, we can obtain the probability of classifying as yi when feature x is P(y i |x)=P(x|y i )P(y i ) / P(x), and then we can obtain the law of total probability from it.

[0141] Step S42, solve for the likelihood P(x|y) i Assume that for a continuous variable x, there is a Gaussian distribution assumption, that is, it is assumed that in y... i Under the condition that x follows a Gaussian distribution The parameters can be obtained by maximum likelihood estimation;

[0142] Step S43: For multidimensional features, there is the Naive Bayes assumption, which assumes that the features are independently distributed, i.e., P(x|y) i )=P(x1|y i )P(x2|y i ...P(x) n |y i );

[0143] Step S44: Based on the above assumptions and formulas, P(y) can be calculated. i |x) and select the category with the highest probability as the final category y. i .

[0144] Step S5: Input the training set obtained in step S4 into the random forest classifier for training, and then input the test set obtained in S4 into it for classification to obtain the final classification result of the test set samples.

[0145] Wherein, the random forest is a classifier composed of a plurality of decision trees, and its output is comprehensively determined by the outputs of all decision trees. The specific manner of using this method to train the feature training set and perform classification is as follows:

[0146] Step S51: Perform bootstrap sampling, that is, assuming there are a total of N samples in the feature space, N samples are drawn from the training set with replacement as the training set for one tree;

[0147] Step S52: If there are M features, take m<<M and fix m throughout the whole process, randomly select m features, and use the optimal feature among these m feature dimensions (with maximum entropy gain as the splitting principle) as the splitting node to construct a decision tree

[0148] Step S53: Repeat the steps S51 and S52 k times, that is, construct k decision trees, and these k decision trees form a random forest, and the category to which the data belongs is determined by the voting result.

[0149] The finally obtained classification result is taken as the final classification result of the test sample. Please refer to the result Figure 4 .

[0150] In summary, the present invention provides a bearing fault detection method based on wavelet transform and hybrid learning. By preprocessing the original vibration signal, extracting wavelet transform features, and then fusing and classifying using multiple machine learning methods, bearing faults can be effectively diagnosed. Compared with previous feature extraction methods, the present invention can better extract features indicating fault characteristics, and fully utilizes the advantages of each machine learning algorithm, improving the accuracy of fault diagnosis.

[0151] In addition, it should be recognized that the above description is only of the preferred embodiments of the present invention, and is not intended to limit the present invention. For those skilled in the art, various modifications and variations may be made to the present invention. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present invention shall all be included within the protection scope of the present invention.

Claims

1. A bearing fault detection method based on wavelet transform and hybrid learning, characterized in that, include: Step S1 involves preprocessing the bearing fault signal data to achieve noise reduction and filtering of useful data. Step S2: Based on the bearing fault principle, perform wavelet transform on the preprocessed bearing fault signal data, calculate the component energy of the obtained detail components and approximate components respectively and normalize them to obtain a six-dimensional feature set. Step S3: Randomly divide the six-dimensional feature set into a training set and a test set, input the training set and the test set into the support vector machine classifier for classification, and add the classification results to the feature space of the training set and the test set respectively. The process of using a support vector machine classifier to classify the training and test sets, and then adding the classification results to the training and test sets respectively, is as follows: Step S31, solving for the support vectors refers to finding the maximum separating hyperplane of the form y = wx + b, where w and b are the parameters to be solved, x is the feature vector, and y is the sample class. For binary classification problems, this is represented by -1 and 1, which transforms into the following mathematical programming problem: ， Where, x i Let y be the eigenvector. i The category to which the sample belongs is represented by -1 and 1 for binary classification problems, where n is the number of samples and λ is the number of samples. i ,λ j Let be the Lagrange multiplier, which is an intermediate variable for solving the problem. st represents the condition that the parameter should satisfy, which leads to the condition that the subsequent parameter should satisfy. λ refers to the vector composed of all the Lagrange multipliers. If we introduce soft margins to represent the severity of misclassification, then the problem to be solved is: ， ， Where i = 1, 2, ..., n, C is the penalty coefficient, ξ i w is a slack variable used to measure the soft interval. T Let w represent the transpose of vector w. The dual problem of the above planning problem is solved using the Lagrange method. The Lagrange function is constructed as follows: ， Where λ i and μ i They are Lagrange multipliers, w, b, ξ i Principal parameters; By taking the derivative with respect to the principal parameters and setting the derivative to 0, the program is transformed into the following program: ， ， For linearly inseparable feature sets, the kernel function method is used, and the hyperplane is then... The planning problem that needs to be solved is: ， Step S32 involves iteratively employing a sequence minimum optimization algorithm to find the optimal λ, specifically including: Step S321: Select two parameters λ that need to be updated. i and λ j With other parameters fixed, the constraint becomes: ， ， Therefore Then it transforms into a programming problem with only one variable constraint; Step S322, for a programming problem with only one variable constraint, for λ i Find the partial derivative, set the partial derivative to 0, and then find the variable λ. inew Then λ is calculated. jnew ; Step S323: Repeat steps S321 and S322, iterating until convergence, and find the optimal λ. Step S33, when calculating the partial derivative, we have , λ i Points with a value greater than 0 are all support vectors. A support vector (x) is randomly selected. s ,y s If b = y, then b = y s -wx s ; Step S34, the result of step S33 represents the maximum segmentation hyperplane: w T x+b=0 and the classification decision function: f(x)=sgn(w T x+b), where sgn(·) is the sign function, defined as follows: ， Step S35: Add the classification results of the training set to the original training set, and add the classification results of the test set to the original test set. Step S4: Input the training set obtained in step S3 into the Gaussian Naive Bayes classifier for training, and then input the test set obtained in S3 into the Gaussian Naive Bayes classifier for classification. Add the classification results to the feature spaces of the training set and the test set obtained in step S3 respectively. Step S5: Input the training set obtained after training with the Gaussian Naive Bayes classifier in step S4 into the random forest classifier for training, and then input the test set obtained after training with the Gaussian Naive Bayes classifier in step S4 into the random forest classifier for classification to obtain the final classification result of the test set samples.

2. The bearing fault detection method based on wavelet transform and hybrid learning according to claim 1, characterized in that, Step S1 involves manually filtering invalid data, performing low-pass filtering, and dividing the data into multiple segments of 4096 points each.

3. The bearing fault detection method based on wavelet transform and hybrid learning according to claim 1, characterized in that, The wavelet transform in step S2 uses the db8 wavelet basis and calculates the normalized energy of the six components.

4. The bearing fault detection method based on wavelet transform and hybrid learning according to claim 3, characterized in that, The process of extracting features from the preprocessed data using wavelet transform in step S2 is as follows: Step S21, perform discrete wavelet transform on the original bearing fault signal data, including: Step S211, assuming the original signal is x[n], the low-pass filter corresponding to the db8 wavelet is g[n], and the high-pass filter is h[n], then the result of the first-level discrete wavelet transform is: ， ， Where x1,L are called the approximate components of the first layer, x1,H are called the detail components of the first layer, K is the filter length; g[n] and h[n] are given by the pywt library of Python; Step S212, for the first The approximate components of the layer are then obtained by performing DWT: ， ， Step S213: Perform DWT five times consecutively to obtain the detail components x of the first five layers. α，H And the approximate component x of the fifth layer 5,L , where α = 1, 2, 3, 4, 5; Step S22: Then, the energy of each detailed component is calculated, and the normalized energy is used as the feature of the bearing fault signal data of that segment.

5. The bearing fault detection method based on wavelet transform and hybrid learning according to claim 4, characterized in that, Step S22 includes: Step S221, calculate the energy C of each detailed component. α And the energy A of the fifth approximate component, where N is the signal length, α = 1, 2, 3, 4, 5: ， ， Step S222: Normalize the energy of each component to obtain C′ α And A′, where α=1,2,3,4,5: ， ， The normalized energy of the obtained six components is taken as the feature of the bearing fault signal data of this segment, for subsequent learning and classification tasks.

6. The bearing fault detection method based on wavelet transform and hybrid learning according to claim 1, characterized in that, The process of using a Gaussian Naive Bayes classifier to train the training set and classify the test set in the step S4 is as follows: Step S41, according to Bayes' theorem, the probability of classifying as yi when feature x is obtained as follows: Then, using the law of total probability, we obtain... ; Step S42, solve for the likelihood. Assume that for a continuous variable x, there is a Gaussian distribution assumption, that is, assume that in y i Under the condition that x follows a Gaussian distribution The parameters μ and σ are obtained by maximum likelihood estimation; Step S43: For multidimensional features, there is the Naive Bayes assumption, which assumes that the features are independently distributed. ; Step S44, calculate based on the above assumptions and formulas. And select the category with the highest probability as the final category y. i .

7. The bearing fault detection method based on wavelet transform and hybrid learning according to claim 1, characterized in that, The process of using a random forest classifier to train the training set and classify the test set in the step S5 is as follows: Step S51, perform bootstrap sampling, that is, assuming there are a total of N samples in the feature space, N samples are drawn from the training set with replacement as the training set for one tree; Step S52, if there are M features, take m<<M and fix m throughout the whole process, randomly select m features, and use the best feature among these m feature dimensions as the split node to build a decision tree; Step S53, repeat the steps of step S51 and step S52 k times, that is, k decision trees are built, and the k decision trees form a random forest, and the category of the data is determined through voting results.

Images