A hybrid model fault diagnosis method based on F1-score multi-level decision analysis

Abstract

The application discloses a kind of mixed model fault diagnosis methods based on F1-score multistage decision analysis, comprising the following steps: S1, obtains two groups of target rolling bearing vibration signals and is converted into time-frequency diagram;S2, establishes two kinds of deep learning models;S3, trains deep learning model and calculates the F1-score value of two kinds of models;S4, obtains another two groups of target rolling bearing vibration signals and extracts signal characteristics;S5, establishes two kinds of machine learning models;S6, trains machine learning model and calculates the F1-score value of two kinds of models;S7, by the F1-score value of deep learning model and machine learning model establishes F1-score database;S8, according to F1-score value formulates multistage decision analysis strategy and decides the best diagnostic result;The present application improves sparrow search algorithm by citing Logistc chaotic mapping method, and the multistage decision of F1-score quantifies the respective diagnostic capacity of two kinds of deep learning models and two kinds of machine learning models, and the most accurate diagnostic result can be obtained by multistage decision judgment analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical fault diagnosis, and particularly relates to a hybrid model fault diagnosis method based on F1-score multistage decision analysis. BACKGROUND

[0002] Mechanical equipment will inevitably fail during long-term operation, and accurate fault diagnosis can avoid accidents and safety problems caused by mechanical product failure. Among the existing mechanical product fault diagnosis methods, the data-driven method is the most widely used, which effectively captures fault information by monitoring the fault signals of mechanical equipment and then judges the health status of the mechanical equipment. The data-driven fault diagnosis method is mainly divided into two categories: one is the machine learning-based fault diagnosis method, and the other is the deep learning-based fault diagnosis method.

[0003] The effectiveness of traditional single intelligent diagnosis method is often based on specific conditions or specific occasions. For complex diagnosis problems, it is difficult to obtain a universal intelligent diagnosis model by relying on one or two methods, resulting in one-sidedness and uncertainty of the diagnosis decision result. SUMMARY

[0004] In order to solve the problem that the diagnosis decision structure of the existing diagnosis method has one-sidedness and uncertainty when facing complex diagnosis problems, the present application provides a hybrid model fault diagnosis method based on F1-score multistage decision analysis.

[0005] The technical solution adopted by the present application to achieve the above-mentioned purpose is: a hybrid model fault diagnosis method based on F1-score multistage decision analysis, comprising the following steps:

[0006] S1, obtaining two groups of target rolling bearing vibration signals, and converting the vibration signals into two groups of time-frequency graphs;

[0007] S2, establishing two deep learning models, including a residual attention network model and a multi-scale sliding convolutional neural network model;

[0008] S3, training the deep learning model: bringing the two groups of time-frequency graph data in S1 into the residual attention network model and the multi-scale sliding convolutional neural network model for training, and calculating the F1-score values of the two models through the training results;

[0009] S4, obtaining another two groups of target rolling bearing vibration signals, and extracting the signal features of the vibration signals;

[0010] S5, establishing two machine learning models, including a support vector machine model and an artificial neural network model;

[0011] S6. Training the machine learning model: The signal features extracted in S4 are respectively introduced into the support vector machine model and the artificial neural network model for training, and the F1-score values ​​of the two models are calculated based on the training results;

[0012] S7. Establish an F1-score database based on the F1-score values ​​of the deep learning model and the machine learning model;

[0013] S8. Develop a multi-level decision analysis strategy based on the F1-score value, and determine the best diagnostic result based on the multi-level decision analysis strategy.

[0014] Preferably, said S1 comprises the following steps:

[0015] S1-1: Obtain the vibration signal of the target rolling bearing through the acceleration sensor;

[0016] S1-2: Use the improved sparrow search algorithm to optimize the hyperparameters in the variational mode decomposition algorithm;

[0017] S1-3: The optimized variational mode decomposition algorithm decomposes the one-dimensional signal collected in S1-1 into multiple intrinsic mode components;

[0018] S1-4: remove high-frequency intrinsic modal components and some low-frequency intrinsic modal components from multiple intrinsic modal components and perform linear reconstruction to obtain two sets of filtered signals;

[0019] S1-5: The filtered signal is converted into a two-dimensional time-frequency diagram through the short-time Fourier transform method, and two sets of time-frequency diagrams focusing on high-frequency and low-frequency information are obtained.

[0020] Preferably, said S1-2 includes the following steps:

[0021] S1-2-1: Set the number of sparrow populations to 100, the search space dimension to 2, the modal component layer number K to the range [2, 10], the penalty factor α to the range [1, 10000], and the maximum number of iterations to 100;

[0022] S1-2-2: The Logistc chaotic mapping method is used to initialize the position of the sparrow population, and the chaotic system control parameter r is set to 3.85;

[0023] S1-2-3: The modal component layer number K and the penalty factor α are passed to the variational mode decomposition algorithm. After decomposition, multiple intrinsic modal components are obtained. The average walk entropy of each intrinsic modal component is used as the fitness value. The average walk entropy value is minimized by optimizing the modal component layer number K and the penalty factor α.

[0024] S1-2-4: Determine whether the convergence condition is met. If so, terminate the iteration and obtain the optimal hyperparameter combination [K, α]. If not, return to step S1-2-2 to continue iterating. The convergence condition is that the maximum number of iterations is reached or the fitness value fluctuation in S1-2-3 is less than 0.001.

[0025] Preferably, the specific steps of iterating using the improved sparrow search algorithm in S1-2-1, S1-2-2 and S1-2-4 are as follows:

[0026] Assume that the number of sparrow population is N, the position of sparrow population is initialized by Logistc chaotic mapping method, the chaotic system control parameter r is set to 3.85, and the position of the i-th sparrow in the d-dimensional search space is Where i = 1, 2, ..., N, the execution process of the sparrow search algorithm is divided into the following three behaviors:

[0027] I. First, the "explorer" in the population will search for food for the entire population and provide directional guidance for the position update of the "follower" in the population. The position update of the "explorer" is described as follows:

[0028]

[0029] Where t represents the current iteration number, j = 1, 2, 3..., d; item max Indicates the maximum number of iterations; X ij represents the position information of the i-th sparrow in the j-th dimension; β is a random number between (0,1]; R2, ST represent the warning value and safety value respectively; Q is a random number that follows a normal distribution; L is a matrix with all elements being 1×d;

[0030] II. When it is safe, the Follower will follow the Explorer to find food and move closer to the current optimal food location. Its position update description is as follows:

[0031]

[0032] Where: X p This is the best position currently occupied by the "Explorer"; X worst represents the worst position of the current “explorer”; A represents a 1×d matrix, in which each element is randomly assigned to 1 or -1, and A + =A T (AA T ) -1 When i>n / 2, it indicates that the i-th "follower" with a poor location has no good food source and needs to fly to other places to forage (n represents the total number of sparrows);

[0033] III. When danger approaches, sparrows will quickly fly to other safe places to find food. The proportion of sparrows in this group is fixed, and their positions are updated as follows:

[0034]

[0035] Where: is the current global optimal position; η is the random step size control parameter, which is a random number that obeys the normal distribution with a mean of 0 and a variance of 1; θ∈[-1,1] is an arbitrary number, f i is the fitness value corresponding to the position of the i-th sparrow; f g and f w are the fitness values ​​corresponding to the best and worst positions currently explored; ψ is the smallest constant.

[0036] Preferably, the specific steps of the variational mode decomposition algorithm in S1-2-3 are as follows:

[0037]

[0038]

[0039] In the formula, {u k}={u1,u2,...,u K} are K modal components, {ω k}={ω1,ω2,...,ω K} is the corresponding center frequency, f is the original signal;

[0040] The Lagrangian operator is introduced to transform the constrained problem into an unconstrained problem. The corresponding expression is as follows:

[0041]

[0042] Where α is the penalty factor and λ is the Lagrange multiplier;

[0043] The modal components and center frequencies are updated as follows:

[0044]

[0045]

[0046] Where, and Represent f(t), u i (t), λ(t) and The Fourier transform of .

[0047] Preferably, the calculation formula of the average distribution entropy in S1-2-3 is as follows:

[0048]

[0049] Where X is a time series of length N, corresponding to the intrinsic modal component after decomposition by the strain modal decomposition algorithm, w is the window size, and p is the time series of length N. i is a subsequence X i The frequency of each value x in is calculated as where n i is x in X i The number of times it appears in .

[0050] Preferably, in S4, 23 signal features of the vibration signal are extracted, and the extracted signal features are respectively subjected to principal component analysis and Fisher feature evaluation criteria to select sensitive features, reduce the feature space dimension, and remove redundant and irrelevant features.

[0051] Preferably, in S5, the whale optimization algorithm is used to optimize the hyperparameters in the iterative support vector machine model, and the specific steps are as follows:

[0052] S5-1: Set the whale population size to 50, the search space dimension to 2, the penalty factor c to the range [0, 1000], the kernel function to the radial basis function, the parameter g to the range [0, 10], and the maximum number of iterations to 100;

[0053] S5-2: Update the whale's position and use the diagnostic accuracy as the fitness function value to train the support vector machine model;

[0054] S5-3: Determine whether the convergence condition is met. If so, terminate the iteration and obtain the optimal hyperparameter combination [c, g]. If not, return to step S5-2 to continue iterating. The convergence condition is that the maximum number of iterations is reached or the fitness value fluctuation in S5-2 is less than 0.001.

[0055] Preferably, in S3 and S6, the calculation formula of F1-score is as follows:

[0056]

[0057]

[0058]

[0059] Where, T P Indicates the number of correct positive class judgments; F P Indicates the number of positive class errors; F N It indicates the number of negative class misjudgments; Precision indicates the precision rate; Recall indicates the recall rate.

[0060] Preferably, the specific steps of S8 are as follows:

[0061] A, B, C, and D represent the four constructed models respectively. i 、B i 、C i 、D i They represent the F1-score values ​​of the four models on the i-th fault classification problem, a, b, c, and d represent the prediction results of the four models, and the hierarchical decision strategy is as follows:

[0062] First-level strategy: If there are three or more identical classification results in [a, b, c, d], the classification result is finally selected as the optimal value;

[0063] Second-level decision: If there are two identical elements in [a, b, c, d] (taking a = b as an example), perform the following operations:

[0064] I. Determine the following conditions:

[0065] ifA a +B b / 2>C c

[0066] ifA a +B b / 2>D d

[0067] If both conditions are met, the final classification result is a or b. If not, proceed to step II.

[0068] II. Calculate the F1-score values ​​of the results predicted by models C and D respectively to see whether they are the highest values ​​among all models under the predicted result label, that is:

[0069] ifC c =max[A c ,B c ,C c ,D c ]

[0070] ifD d =max[A d ,B d ,C d ,D d ]

[0071] If only one condition is met, assuming that model C meets the condition, then c is selected as the optimal value; if both conditions are met, then c and d are selected as candidate results and proceed to step III; if neither condition is met, then proceed to step IV;

[0072] III. Compare Cc With D d The highest value is selected to determine the optimal value; assuming D d >C c , then the optimal result is d;

[0073] IV. In [A a , B b , C c , D d ] and select the highest value, and use the corresponding model judgment result as the optimal result;

[0074] Third-level decision: If all elements in [a, b, c, d] are different, perform the following operations:

[0075] Compare the F1-score values ​​of the four models A, B, C, and D in their respective current classification results, namely:

[0076] ifA a =max[A a ,B a ,C a ,D a ]

[0077] ifB b =max[A b ,B b ,C b ,D b ]

[0078] ifC c =max[A c ,B c ,C c ,D c ]

[0079] ifD d =max[A d ,B d ,C d ,D d ]

[0080] If only one condition is met, the optimal result is the result provided by the model that meets the condition; if multiple conditions are met, the operation of step III in the second-level decision is carried out; if none of the conditions are met, the operation of step IV in the second-level decision is carried out.

[0081] The present invention discloses a hybrid model fault diagnosis method based on F1-score multi-level decision analysis. The method improves the sparrow search algorithm by citing the Logistc chaotic mapping method, and optimizes the variational mode decomposition method with the improved sparrow search algorithm. The decomposed intrinsic modal components have a uniform frequency distribution and a single center frequency, eliminating the phenomenon of modal aliasing. The multi-scale sliding convolutional neural network in the deep learning model can achieve robustness to changes in the scale of target information and extract more feature information at different scales. The residual attention network and the multi-scale sliding convolutional neural network focus on the high-frequency and low-frequency features of the signal, respectively, and can play their respective diagnostic characteristics in different diagnostic tasks and complement each other's advantages. The multi-level decision analysis method based on F1-score can quantify the respective diagnostic capabilities of the two deep learning diagnostic models and the two machine learning diagnostic models, comprehensively consider the diagnostic effects of each model on different faults, and obtain the most accurate diagnostic results through multi-level decision judgment analysis. Compared with traditional decision tree (DT), K nearest neighbor (KNN), gradient boosting decision tree (Xgboost) and other methods, it has better results and better algorithm interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 The present invention is a flow chart of a hybrid model fault diagnosis method based on F1-score multi-level decision analysis.

[0083] Figure 2 It is a schematic diagram of signal processing and feature extraction of the present invention.

[0084] Figure 3 This is a flow chart of the present invention's improved sparrow search algorithm for optimizing variational mode decomposition.

[0085] Figure 4 It is a flow chart of the whale algorithm for optimizing the support vector machine model of the present invention.

[0086] Figure 5 It is a line graph of the average diagnostic accuracy of the four models of the present invention on the experimental data.

[0087] Figure 6 This is a flow chart of the F1-score multi-level decision analysis strategy of the present invention.

[0088] Figure 7 This is a comparison chart of the accuracy of the hybrid model based on F1-score multi-level decision analysis of the present invention, a single model, and other methods. DETAILED DESCRIPTION

[0089] The present invention is a hybrid model fault diagnosis method based on F1-score multi-level decision analysis, which consists of Figure 1 As shown, the following steps are included:

[0090] S1, acquire two groups of target rolling bearing vibration signals, and convert the vibration signals into time-frequency diagrams;

[0091] S2, establish two deep learning models, including a residual attention network model and a multi-scale sliding convolutional neural network model;

[0092] S3, train the deep learning model: the data of the two groups of time-frequency diagrams in S1 are respectively brought into the residual attention network model and the multi-scale sliding convolutional neural network model for training, and the F1-score values of the two models are calculated through the training results;

[0093] S4, acquire another two groups of target rolling bearing vibration signals, and extract signal features of the vibration signals;

[0094] S5, establish two machine learning models, including a support vector machine model and an artificial neural network model;

[0095] S6, train the machine learning model: the extracted signal features in S4 are respectively brought into the support vector machine model and the artificial neural network model for training, and the F1-score values of the two models are calculated through the training results;

[0096] S7, establish an F1-score database through the F1-score values of the deep learning model and the machine learning model;

[0097] S8, according to the F1-score values, a multi-level decision analysis strategy is formulated, and the best diagnosis result is decided according to the multi-level decision analysis strategy.

[0098] The application adopts the bearing experimental data set of Case Western Reserve University as experimental data for verification, and the data set includes vibration data of "Fan" fan end and "Drive" motor end under two rotating speeds of 1797r / min and 17307r / min. The four groups of vibration data are respectively allocated to the residual attention network model, the multi-scale sliding convolutional neural network model, the support vector machine model and the artificial neural network model for training and testing.

[0099] S1 includes the following steps:

[0100] S1-1: acquire two groups of target rolling bearing vibration signals through an acceleration sensor;

[0101] S1-2: The improved sparrow search algorithm (ISSA) is used to optimize the hyperparameters in the variational mode decomposition (VMD) algorithm. The sparrow search algorithm has the characteristics of fast convergence speed and strong search ability. However, it uses a random generation method when initializing the population. This method will greatly produce low-quality population individuals and have a negative impact on the search process. In addition, random initialization may produce similar individuals, resulting in a shrinkage of the search space, limiting the search ability of the algorithm and causing the algorithm to converge to a local optimal solution. Therefore, the Logistc chaos mapping method is introduced into the sparrow search algorithm technology to improve the randomness and distribution of the sparrow population initialization, while improving the global search performance of the algorithm. The specific steps include:

[0102] S1-2-1: Set the number of sparrow populations to 100, the search space dimension to 2, the modal component layer number K to the range [2, 10], the penalty factor α to the range [1, 10000], and the maximum number of iterations to 100;

[0103] S1-2-2: Use the Logistc chaotic mapping method to initialize the position of the sparrow population. The chaotic system control parameter r is set to 3.85. The specific expression of the Logistc chaotic mapping method is as follows:

[0104] x n+1 =rx n (1-x n )

[0105] Where x n represents the output value of the nth iteration step, and r is the system control parameter. In practical applications, r is usually between [3.57, 4.00]. At this time, the system exhibits typical chaotic phenomena, with high sensitivity and complexity.

[0106] S1-2-3: The modal component layer number K and the penalty factor α are passed to the variational mode decomposition algorithm. After decomposition, multiple intrinsic modal components are obtained. The average walk entropy of each intrinsic modal component is used as the fitness value. The average walk entropy value is minimized by optimizing the modal component layer number K and the penalty factor α.

[0107] S1-2-4: Determine whether the convergence condition is met. If so, terminate the iteration and obtain the optimal hyperparameter combination [K, α]. If not, return to step S1-2-2 to continue iterating. The convergence condition is that the maximum number of iterations is less than 10 or the fitness value fluctuation in S1-2-3 is less than 0.001.

[0108] S1-3: The optimized variational mode decomposition algorithm decomposes the one-dimensional signal collected in S1-1 into multiple intrinsic mode components;

[0109] S1-4: linearly reconstruct the multiple intrinsic modal components by removing some high-frequency intrinsic modal components and some low-frequency intrinsic modal components, and obtain two sets of filtered signals;

[0110] S1-5: The filtered signal is converted into a two-dimensional time-frequency diagram by the short-time Fourier transform method, thereby obtaining two sets of time-frequency diagrams that focus on high-frequency and low-frequency information respectively.

[0111] The sparrow search algorithm is a heuristic algorithm based on the life behavior of sparrows in nature. The specific steps are as follows:

[0112] Assume that the number of sparrows is N, and the position of the i-th sparrow in the d-dimensional search space is Where i = 1, 2, ..., N, the execution process of the sparrow search algorithm is divided into the following three behaviors:

[0113] I. First, the "explorer" in the population will search for food for the entire population and provide directional guidance for the position update of the "follower" in the population. The position update of the "explorer" is described as follows:

[0114]

[0115] Where t represents the current iteration number, j = 1, 2, 3..., d; item max Indicates the maximum number of iterations; X ij represents the position information of the i-th sparrow in the j-th dimension; β is a random number between (0,1]; R2, ST represent the warning value and safety value respectively; Q is a random number that follows a normal distribution; L is a matrix with all elements being 1×d;

[0116] II. When it is safe, the Follower will follow the Explorer to find food and move closer to the current optimal food location. Its position update description is as follows:

[0117]

[0118] Where: X p This is the best position currently occupied by the "Explorer"; X worst represents the worst position of the current “explorer”; A represents a 1×d matrix, in which each element is randomly assigned to 1 or -1, and A + =A T (AA T ) -1When i>n / 2, it indicates that the i-th "follower" with a poor location has no good food source and needs to fly to other places to forage (n represents the total number of sparrows);

[0119] III. When danger approaches, sparrows will quickly fly to other safe places to find food. The proportion of sparrows in this group is fixed, and their positions are updated as follows:

[0120]

[0121] Where: is the current global optimal position; η is the random step size control parameter, which is a random number that obeys the normal distribution with a mean of 0 and a variance of 1; θ∈[-1,1] is an arbitrary number, f i is the fitness value corresponding to the position of the i-th sparrow; f g and f w are the fitness values ​​corresponding to the best and worst positions currently explored; ψ is the smallest constant.

[0122] The variational mode decomposition algorithm is a signal processing method that can adjust the time-frequency scale. It can decompose the signal into multiple intrinsic modal components. These components are relatively sparse. Each intrinsic modal component has its own center frequency and limited bandwidth. The specific process of the variational mode decomposition algorithm in S1-2-3 is as follows:

[0123]

[0124]

[0125] In the formula, {u k}={u1,u2,...,u K} are K modal components, {ω k}={ω1,ω2,...,ω K} is the corresponding center frequency, f is the original signal;

[0126] The Lagrangian operator is introduced to transform the constrained problem into an unconstrained problem. The corresponding expression is as follows:

[0127]

[0128] Where α is the penalty factor and λ is the Lagrange multiplier;

[0129] The modal components and center frequencies are updated as follows:

[0130]

[0131]

[0132] Where, and Represent f(t), u i (t), λ(t) and The Fourier transform of .

[0133] The calculation formula for the average walk entropy in S1-2-3 is as follows:

[0134]

[0135] Where X is a time series of length N, corresponding to the intrinsic modal component after decomposition by the strain modal decomposition algorithm, w is the window size, and p is the time series of length N. i is a subsequence X i The frequency of each value x in is calculated as where n i is x in X i The number of times it appears in .

[0136] In the short-time Fourier transform in S1-5, the non-stationary signal is first truncated using an analysis window that slides over time, and then the truncated signal is Fourier transformed to obtain the time-frequency characteristics of the response signal. The short-time Fourier transform expression is as follows:

[0137]

[0138] Where x(τ) is the time course of a response signal, t is the time to be analyzed, f is the frequency, i is an imaginary number, g(τ-t) is the window function, and τ is the time corresponding to the center coordinate of the window function.

[0139] In S2, two sets of deep learning models are established to sample and train the two sets of time-frequency graphs obtained by the short-time Fourier transform method.

[0140] The first model is the Residual Attention Network (ResNet-Attention), which is a residual network with an attention mechanism added at the end. The residual network can directly pass the information of the previous layer to the next layer through cross-layer connections, which helps to speed up the flow of information and the transmission of gradients. The attention mechanism can help the model focus on learning important features to enhance the model's discriminative ability and stability, thereby achieving better results in tasks such as classification and detection. In addition, the attention mechanism can enhance the robustness of the model, thereby resisting small perturbations of information.

[0141] The second model is a multi-scale sliding convolutional neural network (M-SSCNN). This network model uses multi-scale convolutions in parallel to achieve robustness to changes in the scale of the target information, while extracting more feature information at different scales. Using sliding convolutions instead of the pooling operation of the original convolutional neural network avoids the feature loss caused by maximum pooling and the feature blurring caused by average pooling.

[0142] The specific structures and parameters of the two models are shown in Table 1 below.

[0143] Table 1 Model specific parameters

[0144]

[0145]

[0146] In S4, a total of 23 signal features are extracted from the time domain and frequency domain of the one-dimensional signal, including peak-to-peak value, root mean square, center frequency, etc. The extracted signal features are then subjected to principal component analysis (PCA) and Fisher feature evaluation criteria to select sensitive features, reduce the dimension of the feature space, remove redundant and irrelevant features, enhance the model's ability to process relevant feature information, and improve the performance and recognition ability of the classifier.

[0147] In S5, the whale optimization algorithm is used to optimize the hyperparameters in the support vector machine model. The specific steps are as follows:

[0148] S5-1: Set the whale population size to 50, the search space dimension to 2, the penalty factor c to the range [0, 1000], the kernel function to the radial basis function, the parameter g to the range [0, 10], and the maximum number of iterations to 100;

[0149] S5-2: Update the whale's position and use the diagnostic accuracy as the fitness function value to train the support vector machine model;

[0150] S5-3: Determine whether the convergence condition is met. If so, terminate the iteration and obtain the optimal hyperparameter combination [c, g]. If not, return to step S5-2 to continue iterating. The convergence condition is that the maximum number of iterations is reached or the fitness value fluctuation in S5-2 is less than 0.001.

[0151] The Whale Optimization Algorithm is an intelligent optimization algorithm developed by simulating the hunting behavior of humpback whales. It has the advantages of strong stability, few adjustment parameters, and strong optimization ability. The specific steps of the Whale Optimization Algorithm are as follows:

[0152] Assume that the number of whale populations is N, and the position of the i-th whale in the d-dimensional search space is Where i = 1, 2, ..., N, the optimal solution is the whale position corresponding to the optimal objective function; the whale optimization algorithm consists of the following three stages:

[0153] I. Surround the prey

[0154] The current optimal position is set to the location of the target prey, and the position coordinates of the remaining whales are updated according to the following two formulas:

[0155] X(t+1)=X * (t)-AD

[0156] D=|CX * (t)-X(t)|

[0157] Where: t is the current iteration number; X * (t) is the optimal whale position at the current iteration number; X(t) is the whale position at the current iteration number; D is the distance parameter between the whale position and the current optimal position. The expressions of A and C are as follows:

[0158] A=2ar-a

[0159] C=2r

[0160]

[0161] Where r is a random number in the range [0,1]; a is the convergence factor, which decreases linearly from 2 to 0 during the iteration process; M is the maximum number of iterations;

[0162] II. Spiral Bubble Net Predation

[0163] At this stage, the distance between the whale and the target prey is first calculated, which is expressed as follows:

[0164] D=|CX rand -X|

[0165] X(t-1)=X rand -AD

[0166] Then the predation is carried out in a spiral upward manner. The mathematical model is as follows:

[0167] X(t+1)=D′e bl cos(2πl)+X * (t)

[0168] D′=|X * (t)-X(t)|

[0169] Where: b is a constant that defines the shape of the logarithmic spiral; l is a random number in the range [-1, 1]; D′ is the distance between the whale's position and the current optimal position;

[0170] The probability of the above two predation behaviors, encircling the prey and spiral bubble net, is 50%. The mathematical model is as follows:

[0171]

[0172] Where: p is a random number in the range [0,1];

[0173] III. Searching for Prey

[0174] In the whale optimization algorithm, when the convergence factor A>1, individual whales will perform a large-scale random search based on the locations of other whales in the group. This can avoid falling into the local optimal situation. The position update formula is as follows:

[0175] D=|CX rand -X|

[0176] X(t-1)=X rand -AD

[0177] Where: X rand The position of is the position of a whale individual randomly selected from the current population.

[0178] In S5-2, the support vector machine model in the machine learning model is a new type of learning machine based on unified learning theory. Through some nonlinear mapping, the input variable x is transformed into a high-dimensional space z, so that the input variable that is difficult to be linearly separated in the low space can be easily constructed into an optimal hyperplane in the high-dimensional space, and the sample data is divided into two categories: positive examples and negative examples, and the distance between them and the optimal hyperplane is maximized.

[0179] The process of finding the optimal hyperplane can be expressed as a quadratic programming problem, as follows:

[0180]

[0181] Where: w is the weight coefficient vector of the classification surface; b is the threshold of the classification surface; relaxation term ζ i ≥0, used to correct the deviation of data from similar points due to noise; c is the penalty parameter, used to adjust the confidence range and empirical risk ratio of the learning machine;

[0182] The present invention adopts a one-to-one multi-classification support vector machine for fault diagnosis and incorporates the radial kernel function RBF to map to a high-dimensional space. The expression is as follows:

[0183] K(x i ,x j)=exp(-g||x i -x j || 2 )

[0184] Where g is a positive real number. The support vector machine with this kernel function has two parameters: penalty parameter C and kernel function parameter g.

[0185] Another type of machine learning model is the artificial neural network (ANN). ANN can adaptively learn and extract features from historical fault data and perform fault diagnosis based on the extracted features. Its numerous hidden layers approximate nonlinear functions, enhancing the model's expressive power and nonlinear fitting capabilities. During training, the ANN model can automatically select the correct features and weights and is less susceptible to random perturbations or data noise during prediction. At the same time, its training model is highly robust to noise. The neural network uses the cross-entropy function as the loss function, and the "Adam" algorithm is used for optimization.

[0186] The specific structure of the artificial neural network model is shown in Table 2

[0187] Table 2 Artificial neural network model

[0188]

[0189] The expression of the cross entropy function is as follows:

[0190] H(p,q)=-(p*log(q)+(1-p)*log(1-q))

[0191] Where p is the actual value and q is the predicted probability value.

[0192] F1-score is a commonly used indicator for evaluating the effectiveness of classification tasks. It is used to evaluate the efficiency of two-class or multi-class models. Taking into account both precision and recall, the present invention uses F1-score to quantify the diagnostic capabilities of each model under each fault type. In S7, the calculation formula of F1-score is as follows:

[0193]

[0194]

[0195]

[0196] Where, T P Indicates the number of correct positive class judgments; F P Indicates the number of positive class errors; F NFalse Negative represents the number of negative class misjudgments; Precision represents the precision rate; Recall represents the recall rate.

[0197] After establishing two deep learning models and two machine learning models, training and testing are respectively performed on the respective data sets, and the training set and the test set are divided in a ratio of 7:3, and the F1-score values of the four models on each type of fault are calculated and the F1-score database of the four models is constructed, and the specific steps of S8 are as follows:

[0198] A, B, C, and D represent the four models constructed, A i , B i , C i , and D i represent the F1-score values of the four models on the i-th fault classification problem, and a, b, c, and d represent the prediction results of the four models, and the hierarchical decision strategy is as follows:

[0199] First-level strategy: If there are three or more identical classification results in [a, b, c, d], the final optimal value is selected as the classification result;

[0200] Second-level decision: If there are two identical elements in [a, b, c, d] (for example, a = b), the following operations are performed:

[0201] I. Determine the following conditions:

[0202] if A a +B b / 2>C c

[0203] if A a +B b / 2>D d

[0204] If both conditions are met, the final classification result is a or b. If the conditions are not met, proceed to step II.

[0205] II. Calculate whether the F1-score values of the prediction results of models C and D are the highest values among all models under the predicted result label, that is:

[0206] ifC c =max[A c ,B c ,C c ,D c ]

[0207] ifD d =max[A d ,B d ,Cd ,D d ]

[0208] If only one satisfies the condition, assume that the C model satisfies the condition, then select c as the optimal value; if both satisfy, then c, d are candidate results into step III; if both do not satisfy the condition, then enter step IV;

[0209] III, compare the values of C c and D d respectively, select the highest one as the optimal value; assume that D d >C c , the optimal result is d;

[0210] IV, select the highest value in [A a , B b , C c , D d ], and take the model judgment result corresponding to it as the optimal result;

[0211] Third level decision: if [a, b, c, d] exists and is not the same, then the following operations are performed:

[0212] Compare the F1-score values of A, B, C, D four models in their current classification results respectively, that is:

[0213] ifA a =max[A a ,B a ,C a ,D a ]

[0214] ifB b =max[A b ,B b ,C b ,D b ]

[0215] ifC c =max[A c ,B c ,C c ,D c ]

[0216] ifD d =max[A d ,B d ,C d ,D d ]

[0217] If only one condition is met, the optimal result is the result provided by the model that meets the condition; if multiple conditions are met, the operation of step III in the second-level decision is carried out; if none of the conditions are met, the operation of step IV in the second-level decision is carried out.

[0218] The present invention discloses a hybrid model fault diagnosis method based on F1-score multi-level decision analysis. The method improves the sparrow search algorithm by citing the Logistc chaotic mapping method, and optimizes the variational mode decomposition method with the improved sparrow search algorithm. The decomposed intrinsic modal components have a uniform frequency distribution and a single center frequency, eliminating the phenomenon of modal aliasing. The multi-scale sliding convolutional neural network in the deep learning model can achieve robustness to changes in the scale of target information and extract more feature information at different scales. The residual attention network and the multi-scale sliding convolutional neural network focus on the high-frequency and low-frequency features of the signal, respectively, and can play their respective diagnostic characteristics in different diagnostic tasks and complement each other's advantages. The multi-level decision analysis method based on F1-score can quantify the respective diagnostic capabilities of the two deep learning diagnostic models and the two machine learning diagnostic models, comprehensively consider the diagnostic effects of each model on different faults, and obtain the most accurate diagnostic results through multi-level decision judgment analysis. Compared with traditional decision tree (DT), K nearest neighbor (KNN), gradient boosting decision tree (Xgboost) and other methods, it has better results and better algorithm interpretability.

[0219] The present invention is described by way of example, and those skilled in the art will appreciate that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments may be modified to suit specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be within the scope of the present invention.

Claims

1. A hybrid model fault diagnosis method based on F1-score multi-level decision analysis, characterized in that: The following steps are involved: S1. Obtain two sets of target rolling bearing vibration signals and convert the vibration signals into two sets of time-frequency graphs; S2. Establish two deep learning models, including a residual attention network model and a multi-scale sliding convolutional neural network model; S3, training deep learning model: the two sets of time-frequency graph data in S1 are respectively brought into the residual attention network model and the multi-scale sliding convolutional neural network model for training, and the F1-score values ​​of the two models are calculated based on the training results; S4. Obtaining another two sets of target rolling bearing vibration signals and extracting signal features of the vibration signals; S5. Establish two machine learning models, including a support vector machine model and an artificial neural network model; S6. Training the machine learning model: The signal features extracted in S4 are respectively introduced into the support vector machine model and the artificial neural network model for training, and the F1-score values ​​of the two models are calculated based on the training results; S7. Establish an F1-score database based on the F1-score values ​​of the deep learning model and the machine learning model; S8. Develop a multi-level decision analysis strategy based on the F1-score value, and determine the best diagnostic result based on the multi-level decision analysis strategy.

2. A hybrid model fault diagnosis method based on F1-score multi-level decision analysis according to claim 1, characterized in that: Said S1 comprises the following steps: S1-1: Obtain the vibration signal of the target rolling bearing through the acceleration sensor; S1-2: Use the improved sparrow search algorithm to optimize the hyperparameters in the variational mode decomposition algorithm; S1-3: The optimized variational mode decomposition algorithm decomposes the one-dimensional signal collected in S1-1 into multiple intrinsic mode components; S1-4: remove high-frequency intrinsic modal components and some low-frequency intrinsic modal components from multiple intrinsic modal components and perform linear reconstruction to obtain two sets of filtered signals; S1-5: The filtered signal is converted into a two-dimensional time-frequency diagram through the short-time Fourier transform method, and two sets of time-frequency diagrams focusing on high-frequency and low-frequency information are obtained.

3. A hybrid model fault diagnosis method based on F1-score multi-level decision analysis according to claim 2, characterized in that: The S1-2 includes the following steps: S1-2-1: Set the number of sparrow populations to 100, the search space dimension to 2, the modal component layer number K to the range [2, 10], the penalty factor α to the range [1, 10000], and the maximum number of iterations to 100; S1-2-2: The Logistc chaotic mapping method is used to initialize the position of the sparrow population, and the chaotic system control parameter r is set to 3.85; S1-2-3: The modal component layer number K and the penalty factor α are passed to the variational mode decomposition algorithm. After decomposition, multiple intrinsic modal components are obtained. The average walk entropy of each intrinsic modal component is used as the fitness value. The average walk entropy value is minimized by optimizing the modal component layer number K and the penalty factor α. S1-2-4: Determine whether the convergence condition is met. If so, terminate the iteration and obtain the optimal hyperparameter combination [K, α]. If not, return to step S1-2-2 to continue iterating. The convergence condition is that the maximum number of iterations is reached or the fitness value fluctuation in S1-2-3 is less than 0.

001.

4. A hybrid model fault diagnosis method based on F1-score multi-level decision analysis according to claim 3, characterized in that: The specific steps of iterating using the improved sparrow search algorithm in S1-2-1, S1-2-2 and S1-2-4 are as follows: Assume that the number of sparrow population is N, the position of sparrow population is initialized by Logistc chaotic mapping method, the chaotic system control parameter r is set to 3.85, and the position of the i-th sparrow in the d-dimensional search space is Where i = 1, 2, ..., N, the execution process of the sparrow search algorithm is divided into the following three behaviors: I. First, the "explorer" in the population will search for food for the entire population and provide directional guidance for the position update of the "follower" in the population. The position update of the "explorer" is described as follows: Where t represents the current iteration number, j = 1, 2, 3..., d; item max Indicates the maximum number of iterations; X ij represents the position information of the i-th sparrow in the j-th dimension; β is a random number between (0,1]; R2, ST represent the warning value and safety value respectively; Q is a random number that follows a normal distribution; L is a matrix with all elements being 1×d; II. When it is safe, the Follower will follow the Explorer to find food and move closer to the current optimal food location. Its position update description is as follows: Where: X p This is the best position currently occupied by "Explorers"; X worst Represents the worst position of the current "explorer"; A represents a 1×d matrix, in which each element is randomly assigned to 1 or -1, and A + =A T (AA T ) -1 When i>n / 2, it indicates that the i-th "follower" with a poor location has no good food source and needs to fly to other places to forage (n represents the total number of sparrows); III. When danger approaches, sparrows will quickly fly to other safe places to find food. The proportion of sparrows in this group is fixed, and their positions are updated as follows: Where: is the current global optimal position; η is the random step size control parameter, which is a random number that obeys the normal distribution with a mean of 0 and a variance of 1; θ∈[-1,1] is an arbitrary number, f i is the fitness value corresponding to the position of the i-th sparrow; f g and f w are the fitness values ​​corresponding to the best and worst positions currently explored; ψ is the smallest constant.

5. The hybrid model fault diagnosis method based on F1-score multi-level decision analysis according to claim 3 is characterized in that: The specific steps of the variational mode decomposition algorithm in S1-2-3 are as follows: In the formula, {u k }={u1,u2,...,u K } are K modal components, {ω k }={ω1,ω2,...,ω K } is the corresponding center frequency, f is the original signal; The Lagrangian operator is introduced to transform the constrained problem into an unconstrained problem. The corresponding expression is as follows: Where α is the penalty factor and λ is the Lagrange multiplier; The modal components and center frequencies are updated as follows: Where, and Represent f(t), u i (t), λ(t) and The Fourier transform of .

6. The hybrid model fault diagnosis method based on F1-score multi-level decision analysis according to claim 3, characterized in that: The calculation formula of the average distribution entropy in S1-2-3 is as follows: Where X is a time series of length N, corresponding to the intrinsic modal component after decomposition by the strain modal decomposition algorithm, w is the window size, and p is the time series of length N. i is a subsequence X i The frequency of each value x in is calculated as where n i is x in X i The number of times it appears in .

7. The hybrid model fault diagnosis method based on F1-score multi-level decision analysis according to claim 1, characterized in that: In S4, 23 signal features of the vibration signal are extracted, and the extracted signal features are respectively subjected to principal component analysis and Fisher feature evaluation criteria to select sensitive features, reduce the feature space dimension, and remove redundant and irrelevant features.

8. The hybrid model fault diagnosis method based on F1-score multi-level decision analysis according to claim 1 is characterized in that: In S5, the whale optimization algorithm is used to optimize the hyperparameters in the iterative support vector machine model. The specific steps are as follows: S5-1: Set the whale population size to 50, the search space dimension to 2, the penalty factor c to the range [0, 1000], the kernel function to the radial basis function, the parameter g to the range [0, 10], and the maximum number of iterations to 100; S5-2: Update the whale's position and use the diagnostic accuracy as the fitness function value to train the support vector machine model; S5-3: Determine whether the convergence condition is met. If so, terminate the iteration and obtain the optimal hyperparameter combination [c, g]. If not, return to step S5-2 to continue iterating. The convergence condition is that the maximum number of iterations is reached or the fitness value fluctuation in S5-2 is less than 0.

001.

9. The hybrid model fault diagnosis method based on F1-score multi-level decision analysis according to claim 1, characterized in that: In S3 and S6, the calculation formula of F1-score is as follows: Where, T P Indicates the number of correct positive class judgments; F P Indicates the number of positive class judgment errors; F N Indicates the number of negative class judgment errors; Precision means precision; Recall means recall.

10. The hybrid model fault diagnosis method based on F1-score multi-level decision analysis according to claim 1, characterized in that: The specific steps of S8 are as follows: A, B, C, and D represent the four constructed models respectively. i 、B i 、C i 、D i They represent the F1-score values ​​of the four models on the i-th fault classification problem, a, b, c, and d represent the prediction results of the four models, and the hierarchical decision strategy is as follows: First-level strategy: If there are three or more identical classification results in [a, b, c, d], the classification result is finally selected as the optimal value; Second-level decision: If there are two identical elements in [a, b, c, d] (taking a = b as an example), perform the following operations: I. Determine the following conditions: if A a +B b / 2>C c if A a +B b / 2>D d If both conditions are met, the final classification result is a or b. If the conditions are not met, proceed to step II; II. Calculate the F1-score values ​​of the results predicted by models C and D respectively to see whether they are the highest values ​​among all models under the predicted result label, that is: if C c =max[A c ,B c ,C c ,D c ] if D d =max[A d ,B d ,C d ,D d ] If only one condition is met, assuming that model C meets the condition, then c is selected as the optimal value; if both conditions are met, then c and d are selected as candidate results and enter step III; if neither condition is met, then enter step IV; III. Compare C c With D d The highest value is selected to determine the optimal value; assuming D d >C c , then the optimal result is d; IV. In [A a , B b , C c , D d ] and select the highest value, and use the corresponding model judgment result as the optimal result; Third-level decision: If all elements in [a, b, c, d] are different, perform the following operations: Compare the F1-score values ​​of the four models A, B, C, and D in their respective current classification results, namely: if A a =max[A a ,B a ,C a ,D a ] ifB b =max[A b ,B b ,C b ,D b ] ifC c =max[A c ,B c ,C c ,D c ] ifD d =max[A d ,B d ,C d ,D d ] If only one condition is met, the optimal result is the result provided by the model that meets the condition; if multiple conditions are met, the operation of step III in the second-level decision is carried out; if none of the conditions are met, the operation of step IV in the second-level decision is carried out.

Images