High-speed rail bearing multi-scale feature fusion fault prediction method

Abstract

The invention discloses a multi-scale feature fusion fault prediction method for a high-speed rail bearing. The method comprises the following steps: acquiring a fault vibration signal of the high-speed rail bearing through a vibration signal comparison technology; noise and interference signals of the high-speed rail bearing fault original vibration signals are removed through an adaptive filter; performing multi-level decomposition on the vibration signal by using wavelet packet decomposition, extracting sub-band signals of different frequency bands, calculating instantaneous frequency and energy distribution of the sub-band signals by using Hilbert-Huang transform, and forming a time-frequency-energy three-dimensional feature map; carrying out weight distribution on each feature dimension in the three-dimensional feature map by adopting a self-adaptive weighting algorithm; constructing a training data set through the time-frequency-energy three-dimensional feature map and the fault type label; and training the long short-term memory network LSTM by using the training data set, constructing a fault prediction model, and performing fault prediction on the high-speed rail bearing. According to the invention, accurate prediction of the high-speed rail bearing fault is realized, and the accuracy of fault detection is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault prediction, and in particular to a high-speed rail bearing multi-scale feature fusion fault prediction method. Background Art

[0002] In modern high-speed railway systems, high-speed rail bearings are key components, and their operating status directly affects the safety and reliability of trains. Due to the complexity of the high-speed rail operating environment and high-load working conditions, bearings are easily affected by various factors, such as vibration, temperature changes, and mechanical wear, which lead to potential failure risks. Therefore, timely and accurate failure prediction of high-speed rail bearings has become crucial, which can not only improve the safety of train operation, but also reduce maintenance costs and extend equipment life.

[0003] Traditional high-speed rail bearing fault diagnosis methods mainly rely on the analysis of a single vibration fault signal source, which has obvious limitations. First, a single vibration fault signal source is difficult to fully and accurately reflect the operating status of the bearing, especially when faced with multiple complex fault modes. Secondly, due to the large noise interference in the high-speed rail operating environment, the collected signals often contain a large amount of irrelevant information, resulting in unreliable analysis results. Although these methods have improved the efficiency of fault diagnosis to a certain extent, they are still insufficient in dealing with changes in the working conditions of high-speed rail bearings and the identification of faults of different frequencies, solving multi-signal fusion and fault mode identification in complex environments.

[0004] In the context of today's industrial intelligence and digital transformation, big data, artificial intelligence and deep learning advanced technologies are bringing unprecedented changes to the industrial field. The integration of these technologies provides new ideas and solutions for the fault prediction of high-speed rail bearings. In particular, deep learning technology can better process large-scale unstructured data through neural network structures and mine deep fault characteristics from them. This application proposes a high-speed rail bearing fault prediction method based on multi-scale feature fusion. It captures the vibration signal of the high-speed rail bearing through an integrated sensor, combines adaptive filtering, wavelet packet decomposition, Hilbert-Huang transform and long short-term memory network (LSTM), and realizes efficient prediction of high-speed rail bearing faults. This method not only improves the accuracy of fault prediction, but also enhances the robustness and adaptability of the system, providing strong support for the intelligent maintenance of high-speed rail bearings. Summary of the invention

[0005] The present invention provides a high-speed rail bearing multi-scale feature fusion fault prediction method, comprising:

[0006] The vibration signal of the high-speed rail bearing is captured by an integrated sensor, and the vibration signal of the high-speed rail bearing fault is obtained by vibration signal comparison technology as the original vibration signal of the high-speed rail bearing fault;

[0007] The noise and interference signal of the original vibration signal of the high-speed rail bearing fault are removed by an adaptive filter to obtain a vibration signal;

[0008] The vibration signal is decomposed into multiple levels using wavelet packet decomposition to extract sub-band signals of different frequency bands. The instantaneous frequency and energy distribution of the sub-band signals are calculated using Hilbert-Huang transform to form a three-dimensional feature map of time-frequency-energy.

[0009] Adopt an adaptive weighting algorithm to assign weights to each feature dimension in the three-dimensional feature map to highlight the fault characteristics; construct a training data set through the time-frequency-energy three-dimensional feature map and fault type labels;

[0010] The long short-term memory network (LSTM) is trained using the training data set to build a fault prediction model, which is then used to predict faults in high-speed rail bearings.

[0011] Preferably, the integrated sensor includes but is not limited to an accelerometer and a displacement sensor. The integrated sensor is used to accurately measure the tiny displacements and acceleration changes in all directions of the bearing, and the original vibration signal generated during the operation of the high-speed rail bearing is captured in real time at a predetermined high sampling frequency. The physical quantity of the original vibration signal is converted into an electrical signal output to obtain the original vibration signal of the dynamic characteristics of the high-speed rail bearing.

[0012] Preferably, obtaining the bearing fault vibration signal by using the vibration signal comparison technology to compare the original vibration signal specifically includes:

[0013] Define the vibration signal of the high-speed rail bearing in a healthy state as x r (t), the real-time monitored vibration signal is x l (t), the amplitude difference is eliminated by normalization processing, and the vibration signal x of the high-speed rail bearing in a healthy state is calculated r (t) and the real-time monitored vibration signal is x l The cross-correlation function R of (t) x (t), the formula is: Where T is the time integration interval, τ is the delay time, and the formula Get the maximum correlation number ρ and set the fault threshold when When the vibration signal x is monitored, it is output in real time. l (t) is the fault vibration signal x(t).

[0014] Preferably, the fault vibration signal is decomposed at multiple levels by wavelet packet decomposition to extract sub-band signals of different frequency bands, including the following contents:

[0015] Define the wavelet coefficients W j，k and the scale factor Vj，k , through the wavelet coefficient W j，k Capture the high-frequency vibration frequency of the vibration signal x(t) and calculate the wavelet coefficient W of the vibration signal x(t) j，k , the formula is: W j，k =∫x(t)ψ j，k (t)dt, where j represents the number of decomposition levels, k represents the index of position j, and ψ j，k (t) is the wavelet mother function, and the formula is By scaling factor V j，k Capture the low-frequency vibration frequency of the vibration signal x(t) and calculate the scale coefficient V of the vibration signal x(t) j，k , the formula is: V j，k =∫x(t)φ j，k (t)dt; where φ j，k (t) is the scale function, and the formula is: The vibration signal x(t) is decomposed into multi-layer frequency band signals by wavelet packet decomposition technology.

[0016] Preferably, extracting sub-band signals of different frequency segments from the multi-layer frequency band signals to identify the fault vibration signal type of the high-speed rail bearing specifically includes:

[0017] Through multi-level decomposition, a series of wavelet coefficients and scale coefficients are obtained, corresponding to different frequency band signals; for each layer of frequency band signals containing the nth node, each node can be further decomposed into two sub-nodes, including a high-frequency sub-node and a low-frequency sub-node, and applied to all nodes through a recursive mechanism until the required total decomposition layer number J is reached;

[0018] The wavelet coefficient W j，k and the scale factor V j，k Reconstruct the subband signal s j，n (t), the sub-band signal s j，n The formula for (t) is: j，n (t) = ∑ k W j，k ·V j，k , where n is the node position, j is the number of decomposition levels, and k represents the index of position j.

[0019] Preferably, based on the subband signal, the Hilbert-Huang transform is used to calculate the instantaneous frequency and energy distribution of each subband to form a time-frequency-energy three-dimensional feature map, including the following contents:

[0020] For the subband signal s at the nth node of the jth layer j，n (t), the analytical signal a is obtained by Hilbert transform j，n (t), the formula is: in represents the Hilbert transform function, i is the imaginary part of the subband signal;

[0021] By analyzing the signal a j，n (t) Calculate the instantaneous amplitude A j，n (t) and instantaneous phase θ j，n (t), the instantaneous amplitude A j，n The formula for (t) is: j，n (t)=|a j，n (t)|, the instantaneous phase θ j，n The formula for (t) is: j，n (t) = arg(a j，n (t));

[0022] According to the instantaneous phase θ j，n (t) Calculate the instantaneous frequency f j，n (t), the formula is:

[0023] The energy distribution of the subband signal is defined as E j，n (t, f), where t is time and f is frequency, the formula is: E j，n (t, f) = A 2 j，n (t)·δ(ff j，n (t)), where δ is the Dirac function;

[0024] The instantaneous frequency and energy distribution of the sub-band signal are combined to form a comprehensive time-frequency-energy three-dimensional feature map TFE (t, f, E), the formula is: Where j represents the number of the jth decomposition layer, J is the total number of decomposition layers, and n is the node of the jth layer.

[0025] Preferably, the adaptive weighting algorithm is used to assign weights to the feature dimensions in the three-dimensional feature map to highlight key fault features, specifically including:

[0026] The time-frequency-energy three-dimensional feature graph TFE(t, f, E) is discretized into multiple energy units according to the energy distribution, and the number of energy units is defined as N t , then each energy unit can be expressed as E n , where n = 1, 2, 3, ..., N t , each energy unit E n Mapped to the corresponding time-frequency domain, weights are assigned through an adaptive multi-scale transformation mechanism, and weighting functions w(f) are defined to adjust the weights of different frequency bands. The formula is: Convert the energy unit into a multi-scale feature matrix M along the time axis t，f , the formula is: Where δ is the Dirac function;

[0027] The adaptive weighting coefficient α(t, f) is used to dynamically adjust according to the changes in time and frequency. The adaptive weighting coefficient α(t, f) formula is: Where p is the adjustment weighted sensitivity coefficient;

[0028] The multi-scale feature matrix M is adjusted using the adaptive weighting coefficient α(t, f). t，f Weighted to form an optimized multi-scale feature map TFE opt (t, f, E), the formula is: TFE opt (t, f, E) = α(t, f)·M t，f .

[0029] Preferably, the fault type labels are constructed according to the fault types of multiple high-speed rail fault vibration signals, by defining a set of fault types of high-speed rail bearings, for each fault type of vibration signal, extracting the corresponding feature vector from the optimized multi-scale feature map, assigning a unique label to each fault type, and constructing the fault type label by using a unique hot encoding method;

[0030] Define the high-speed rail bearing fault type as l and the fault type label as y l , the y l =[y l，1 ,y l，2 , …, y l，L ] T , where L is the number of fault types corresponding to the label.

[0031] Preferably, it is characterized in that the construction of a training data set using the time-frequency-energy three-dimensional feature map corresponding to the fault type and the corresponding fault type label specifically includes:

[0032] The optimized multi-scale feature map TFE opt (t, f, E) TFE is obtained based on the high-speed rail bearing fault type l l (t, f, E), from TFE l Extract the characteristic vector x of the fault type from (t, f, E) l , the formula is: l =[x l，1 , x l，2 , …, x l,M ] T , where M is the dimension of the fault feature vector, x l，1 , x l，2 , …, x l，M From TFE l (t, f, E) extracted eigenvalues;

[0033] The feature vector x of each fault type l With the fault type label y l Combined into a training sample (x l ,y l ), all training samples {(x1, y1), (x2, y2), ..., (X N ,y N )} form a training set D, where N is the number of training samples.

[0034] Preferably, the long short-term memory network LSTM is trained using the training data set D to obtain a fault prediction model for identifying and predicting high-speed rail bearing faults, specifically including:

[0035] The pre-built LSTM fault prediction model contains multiple hidden layers, each layer has H hidden units, and the weight matrix W and bias vector b of the LSTM fault prediction model are initialized;

[0036] Perform forward propagation through the long short-term memory network LSTM to calculate the hidden state h t , the formula is: t =o t ⊙tanh(c t ), where o t is the output gate state, ⊙ is the element-by-element multiplication, c t is the cell state; the output gate state o t The formula is: t =σ(W o [h t-1 , x t ]+b o ), where σ is the sigmoid activation function, W o is the output gate weight matrix, h t-1 is the hidden state at time t-1, x t is the eigenvector at time t, b o is the output gate bias vector; the cell state c t The formula is: Among them, f t is the forget gate state, i t is the input gate state, is the candidate cell state; the forget gate state f t The formula is: t =σ(W f [h t-1 , x t ]+b f ), where W f is the forget gate weight matrix, b fis the forget gate bias vector; the input gate state i t The formula is: t =σ(W i [h t-1 , x t ]+b i ), where W i is the input gate weight matrix, b i is the input gate bias vector; the candidate cell state The formula is: Where W c is the candidate cell state weight matrix, b c is the candidate cell state bias vector;

[0037] Through the hidden state h t Pass it to the fully connected layer to get the predicted fault type label The formula is The predicted fault results are output through the fault prediction model.

[0038] Compared with the prior art, the technical solution of this application has the following technical effects:

[0039] The present invention captures the vibration signal of the high-speed rail bearing through an integrated sensor, and combines it with an adaptive filter to remove noise and interference, thereby solving the problem in the traditional method that the fault signal caused by environmental noise and mechanical noise is difficult to accurately extract. It can more clearly identify the abnormal vibration of the high-speed rail bearing during operation, thereby improving the accuracy and reliability of fault detection, allowing maintenance personnel to take measures more promptly and effectively, and ensuring the safe and stable operation of the high-speed rail.

[0040] The present invention uses wavelet packet decomposition technology to perform multi-level decomposition of vibration signals to extract sub-band signals of different frequency bands, and uses Hilbert-Huang transform based on these sub-band signals to calculate the instantaneous frequency and energy distribution, forming a time-frequency-energy three-dimensional feature map, which solves the technical problem that a single-scale analysis cannot fully reflect the fault characteristics; such a multi-scale feature fusion method can comprehensively evaluate the working status of high-speed rail bearings from multiple angles, enhance the sensitivity of the fault diagnosis model to minor faults under complex working conditions, and thus improve the prediction accuracy and fault location capability.

[0041] The present invention adopts an adaptive weighting algorithm to assign weights to various dimensions in the time-frequency-energy three-dimensional feature graph, highlighting key fault features, and solving the problems of unreasonable feature selection and easy neglect of important information in the prior art; by dynamically adjusting the importance of each feature dimension, it ensures that the factors that really affect the health status of high-speed rail bearings can be focused on during the model training process, thereby optimizing the model performance and making it have better generalization ability and higher accuracy in practical applications.

[0042] The present invention uses the time-frequency-energy three-dimensional feature map corresponding to the fault type and the corresponding fault type label to construct a training data set, and trains the pre-constructed long short-term memory network LSTM, which solves the problems of limited memory length and insufficient modeling of long-term dependency relationships in traditional machine learning models when processing time series data. The powerful time series modeling capability of the LSTM network combined with high-quality feature input significantly enhances the model's learning and understanding of the failure modes of high-speed rail bearings throughout their life cycle, achieves effective early warning of possible future failures, and provides a scientific basis for preventive maintenance.

[0043] The present invention defines a set of fault types for high-speed rail bearings and assigns a unique label to each fault type, constructs fault type labels using one-hot encoding, and generates training samples in combination with optimized multi-scale feature maps, thereby solving the problems of fuzzy category division and insufficient label information in traditional fault classification methods. This method not only improves the distinction between different types of faults, but also simplifies the process of model understanding and interpretation, making it easier for technicians to quickly locate problems, further improving the practicality and user-friendliness of the entire fault prediction system.

[0044] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application so that it can be implemented in accordance with the contents of the specification, and to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following is a detailed description of the preferred embodiments of the present application in conjunction with the accompanying drawings as follows.

[0045] Based on the detailed description of the specific embodiments of the present application in combination with the accompanying drawings below, those skilled in the art will become more aware of the above and other objects, advantages and features of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings without creative work. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale.

[0047] Figure 1 This is an overall template diagram of a high-speed rail bearing multi-scale feature fusion fault prediction method of the present invention;

[0048] Figure 2 This is an overall flow chart of a high-speed rail bearing multi-scale feature fusion fault prediction method of the present invention;

[0049] Figure 3 It is a time-frequency-energy three-dimensional characteristic graph of a high-speed rail bearing multi-scale feature fusion fault prediction method of the present invention;

[0050] Figure 4 It is an optimized time-frequency-energy three-dimensional feature graph of a high-speed rail bearing multi-scale feature fusion fault prediction method of the present invention. DETAILED DESCRIPTION

[0051] To make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiment of the present application. Obviously, the described embodiment is a part of the embodiment of the present application, rather than all of the embodiments. In the following description, specific details such as specific configuration and components are provided only to help fully understand the embodiments of the present application. Therefore, it should be clear to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, for clarity and brevity, the description of known functions and structures is omitted in the embodiment.

[0052] It should be understood that the references to "one embodiment" or "this embodiment" throughout the specification mean that the specific features, structures, or characteristics associated with the embodiment are included in at least one embodiment of the present application. Therefore, the references to "one embodiment" or "this embodiment" appearing throughout the specification do not necessarily refer to the same embodiment. In addition, these specific features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0053] In addition, the present application may repeat reference numerals and / or letters in different examples. This repetition is for the purpose of simplicity and clarity, and does not in itself indicate the relationship between the various embodiments and / or settings discussed.

[0054] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist at the same time. The term " / and" in this article describes another type of association object relationship, indicating that there can be two relationships. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the previous and next associated objects are in an "or" relationship.

[0055] The term "at least one" in this article is merely a description of the association relationship of associated objects, indicating that there may be three relationships. For example, at least one of A and B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0056] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusions.

[0057] Example 1

[0058] This embodiment mainly describes a high-speed rail bearing multi-scale feature fusion fault prediction method. Figure 2 As shown, including:

[0059] S 1. Capture the vibration signal of the high-speed rail bearing through the integrated sensor, and obtain the vibration signal of the high-speed rail bearing fault through the vibration signal comparison technology as the original vibration signal of the high-speed rail bearing fault, such as Figure 1 As shown;

[0060] Furthermore, the integrated sensor includes but is not limited to an accelerometer and a displacement sensor. The integrated sensor is used to accurately measure the small displacement and acceleration changes in all directions of the bearing. The accelerometer is mainly used to capture the acceleration changes. It can very sensitively sense the high-frequency vibration signal caused by the internal defects of the bearing; and the displacement sensor can accurately measure the small displacement caused by the relative position change of the mechanical parts. The combination of the two can fully reflect the working state of the bearing.

[0061] The original vibration signal generated during the operation of the high-speed rail bearing is captured in real time through a predetermined high sampling frequency. The high sampling frequency collects a large number of data points per second, so that even transient or periodic anomalies can be captured in time, and the physical quantity of the original vibration signal is converted into an electrical signal output to obtain the original vibration signal of the dynamic characteristics of the high-speed rail bearing;

[0062] Further, the method of obtaining the bearing fault vibration signal by using the vibration signal comparison technology to obtain the bearing fault vibration signal specifically includes:

[0063] S2, removing the noise and interference signal of the original vibration signal of the high-speed rail bearing fault by an adaptive filter to obtain a vibration signal;

[0064] Define the vibration signal of the high-speed rail bearing in a healthy state as x r (t), the real-time monitored vibration signal is x l (t), the amplitude difference is eliminated by normalization processing, and the vibration signal x of the high-speed rail bearing in a healthy state is calculated r (t) and the real-time monitored vibration signal is x l The cross-correlation function R of (t) x (t), the formula is: Where T is the time integration interval, τ is the delay time, and the formula Get the maximum correlation number ρ and set the fault threshold when When the vibration signal x is monitored, it is output in real time. l (t) is the fault vibration signal x(t);

[0065] The real-time monitored vibration signal is compared with the vibration signal in a healthy state, the amplitude difference between the two is eliminated by normalization processing, and the cross-correlation function is calculated to evaluate the similarity between the two. If the maximum correlation number between the real-time monitoring signal and the healthy signal is lower than the preset fault threshold, it indicates that the currently monitored vibration signal is abnormal, that is, it is judged as a fault vibration signal; this method based on vibration signal comparison can effectively identify abnormal vibration patterns caused by wear, cracks or other mechanical problems.

[0066] S3. Use wavelet packet decomposition to decompose the vibration signal at multiple levels, extract sub-band signals of different frequency bands, and use Hilbert-Huang transform to calculate the instantaneous frequency and energy distribution of the sub-band signals to form a three-dimensional feature map of time-frequency-energy, such as Figure 3 As shown;

[0067] Furthermore, the fault vibration signal is decomposed at multiple levels using wavelet packet decomposition to extract sub-band signals of different frequency bands, including the following:

[0068] Define the wavelet coefficients W j，k and the scale factor V j，k , through the wavelet coefficient W j，k Capture the high-frequency vibration frequency of the vibration signal x(t) and calculate the wavelet coefficient W of the vibration signal x(t) j，k , the formula is: W j，k =∫x(t)ψj，k (t)dt, where j represents the number of decomposition levels, k represents the index of position j, and ψ j，k (t) is the wavelet mother function, and the formula is By scaling factor V j，k Capture the low-frequency vibration frequency of the vibration signal x(t) and calculate the scale coefficient V of the vibration signal x(t) j，k , the formula is: V j，k =∫x(t)φ j，k (t)dt; where φ j，k (t) is the scale function, and the formula is: The vibration signal x(t) is decomposed into multi-layer frequency band signals by wavelet packet decomposition technology.

[0069] Furthermore, the sub-band signals of different frequency bands are extracted from the multi-layer frequency band signals to identify the fault vibration signal type of the high-speed rail bearing, specifically including:

[0070] Through multi-level decomposition, a series of wavelet coefficients and scale coefficients are obtained, corresponding to different frequency band signals; for each layer of frequency band signals containing the nth node, each node can be further decomposed into two sub-nodes, including a high-frequency sub-node and a low-frequency sub-node, and applied to all nodes through a recursive mechanism until the required total decomposition layer number J is reached;

[0071] The wavelet coefficient W j，k and the scale factor V j，k Reconstruct the subband signal s j，n (t), the sub-band signal s j，n The formula for (t) is: j，n (t) = ∑ k W j，k ·V j，k , where n is the node position, j is the number of decomposition levels, and k represents the index of position j.

[0072] Further, based on the sub-band signal, the Hilbert-Huang transform is used to calculate the instantaneous frequency and energy distribution of each sub-band to form a time-frequency-energy three-dimensional feature map, including the following contents:

[0073] For the subband signal s at the nth node of the jth layer j，n (t), the analytical signal a is obtained by Hilbert transform j，n (t), the formula is: in represents the Hilbert transform function, i is the imaginary part of the subband signal;

[0074] By analyzing the signal a j，n (t) Calculate the instantaneous amplitude A j，n (t) and instantaneous phase θj，n (t), the instantaneous amplitude A j，n The formula for (t) is: j，n (t)=|a j，n (t)|, the instantaneous phase θ j，n The formula for (t) is: j，n (t) = arg(a j，n (t));

[0075] According to the instantaneous phase θ j，n (t) Calculate the instantaneous frequency f j，n (t), the formula is:

[0076] The energy distribution of the subband signal is defined as E j，n (t, f), where t is time and f is frequency, the formula is: E j，n (t, f) = A 2 j，n (t)·δ(ff j，n (t)), where δ is the Dirac function;

[0077] The instantaneous frequency and energy distribution of the sub-band signal are combined to form a comprehensive time-frequency-energy three-dimensional feature map TFE (t, f, E), the formula is: Where j represents the number of the jth decomposition layer, J is the total number of decomposition layers, and n is the node of the jth layer.

[0078] S4. Use an adaptive weighting algorithm to assign weights to each feature dimension in the three-dimensional feature map to highlight the fault characteristics; construct a training data set through the time-frequency-energy three-dimensional feature map and the fault type label;

[0079] Furthermore, the adaptive weighting algorithm is used to assign weights to each feature dimension in the three-dimensional feature map to highlight key fault features, specifically including:

[0080] The time-frequency-energy three-dimensional feature graph TFE(t, f, E) is discretized into multiple energy units according to the energy distribution, and the number of energy units is defined as N t , then each energy unit can be expressed as E n , where n = 1, 2, 3, ..., N t , each energy unit E n Mapped to the corresponding time-frequency domain, weights are assigned through an adaptive multi-scale transformation mechanism, and weighting functions w(f) are defined to adjust the weights of different frequency bands. The formula is: Convert the energy unit into a multi-scale feature matrix M along the time axis t，f , the formula is: Where δ is the Dirac function;

[0081] The adaptive weighting coefficient α(t, f) is used to perform dynamic adjustment according to the change of time and frequency. The formula of the adaptive weighting coefficient α(t, f) is: Where p is the adjustment weighted sensitivity coefficient;

[0082] The multi-scale feature matrix M is adjusted using the adaptive weighting coefficient α(t, f). t，f Weighted to form an optimized multi-scale feature map TFE opt (t, f, E), the formula is: TFE opt (t, f, E) = α(t, f)·M t，f ,like Figure 4 shown.

[0083] Furthermore, the fault type labels are constructed according to the fault types of the various high-speed rail fault vibration signals. By defining a set of fault types of high-speed rail bearings, for each fault type of vibration signal, a corresponding feature vector is extracted from the optimized multi-scale feature map, a unique label is assigned to each fault type, and a unique hot encoding method is used to construct the fault type label;

[0084] Define the high-speed rail bearing fault type as l and the fault type label as y l , the y l =[y l，1 ,y l，2 , …, y l，L ] T , where L is the number of fault types corresponding to the label.

[0085] Preferably, it is characterized in that the construction of a training data set using the time-frequency-energy three-dimensional feature map corresponding to the fault type and the corresponding fault type label specifically includes:

[0086] The optimized multi-scale feature map TFE opt (t, f, E) TFE is obtained based on the high-speed rail bearing fault type l l (t, f, E), from TFE l Extract the characteristic vector x of the fault type from (t, f, E) l , the formula is: l =[x l，1 , x l，2 , …, x l,M ] T , where M is the dimension of the fault feature vector, x l，1 , x l，2 , …, x l，M From TFE l (t, f, E) extracted eigenvalues;

[0087] The feature vector x of each fault type l With the fault type label y l Combined into a training sample (x l ,y l ), all training samples {(x1, y1), (x2, y2), ..., (x N ,y N )} form a training set D, where N is the number of training samples.

[0088] S5. Use the training data set to train the long short-term memory network LSTM, build a fault prediction model, and use the fault prediction model to predict faults of high-speed rail bearings.

[0089] Furthermore, the long short-term memory network LSTM is trained using the training data set D to obtain a fault prediction model for identifying and predicting high-speed rail bearing faults, specifically including:

[0090] The pre-built LSTM fault prediction model contains multiple hidden layers, each layer has H hidden units, and the weight matrix W and bias vector b of the LSTM fault prediction model are initialized;

[0091] Perform forward propagation through the long short-term memory network LSTM to calculate the hidden state h t , the formula is: t =o t ⊙tanh(c t ), where o t is the output gate state, ⊙ is the element-by-element multiplication, c t is the cell state; the output gate state o t The formula is: t =σ(W o [h t-1 , x t ]+b o ), where σ is the sigmoid activation function, W o is the output gate weight matrix, h t-1 is the hidden state at time t-1, x t is the eigenvector at time t, b o is the output gate bias vector; the cell state c t The formula is: Among them, f t is the forget gate state, i t is the input gate state, is the candidate cell state; the forget gate state f t The formula is: t =σ(W f [ht-1 , x t ]+b f ), where W f is the forget gate weight matrix, b f is the forget gate bias vector; the input gate state i t The formula is: t =σ(W i [h t-1 , x t ]+b i ), where W i is the input gate weight matrix, b i is the input gate bias vector; the candidate cell state The formula is: Where W c is the candidate cell state weight matrix, b c is the candidate cell state bias vector;

[0092] Through the hidden state h t Pass it to the fully connected layer to get the predicted fault type label The formula is The predicted fault results are output through the fault prediction model.

[0093] This embodiment uses an integrated sensor to capture the vibration signal of the high-speed rail bearing, uses adaptive filtering to remove noise interference, combines wavelet packet decomposition and Hilbert-Huang transform to extract multi-scale features, forms a time-frequency-energy three-dimensional feature map, and then optimizes the feature importance through an adaptive weighting algorithm. Finally, a long short-term memory network is used for fault prediction, thereby improving the accuracy and reliability of fault detection, enhancing the sensitivity to minor faults under complex working conditions, and providing a scientific basis for the preventive maintenance of high-speed rail bearings. This series of methods not only optimizes the performance of the model, giving it better generalization ability and accuracy, but also simplifies the process of fault classification and location, and improves the practicality and user-friendliness of the system.

[0094] Example 2

[0095] This embodiment is based on Embodiment 1 and describes in detail the updating mechanism of the fault prediction model of high-speed rail bearing faults;

[0096] The long short-term memory network (LSTM) is trained using the training data set to build a fault prediction model, which is then used to predict faults in high-speed rail bearings.

[0097] Furthermore, the long short-term memory network LSTM is trained using the training data set D to obtain a fault prediction model for identifying and predicting high-speed rail bearing faults, specifically including:

[0098] The pre-built LSTM fault prediction model contains multiple hidden layers, each layer has H hidden units, and the weight matrix W and bias vector b of the LSTM fault prediction model are initialized;

[0099] Perform forward propagation through the long short-term memory network LSTM to calculate the hidden state h t , the formula is: t =o t ⊙tanh(c t ), where o t is the output gate state, ⊙ is the element-by-element multiplication, c t is the cell state; the output gate state o t The formula is: t =σ(W o [h t-1 , x t ]+b o ), where σ is the sigmoid activation function, W o is the output gate weight matrix, h t-1 is the hidden state at time t-1, x t is the eigenvector at time t, b o is the output gate bias vector; the cell state c t The formula is: Among them, f t is the forget gate state, i t is the input gate state, is the candidate cell state; the forget gate state f t The formula is: t =σ(W f [h t-1 , x t ]+b f ), where W f is the forget gate weight matrix, b f is the forget gate bias vector; the input gate state i t The formula is: t =σ(W i [h t-1 , x t ]+b i ), where W i is the input gate weight matrix, b i is the input gate bias vector; the candidate cell state The formula is: Where W c is the candidate cell state weight matrix, b c is the candidate cell state bias vector;

[0100] Through the hidden state h t Pass it to the fully connected layer to get the predicted fault type label The formula is Output predicted fault results through the fault prediction model;

[0101] Define the cross entropy loss function L to measure the predicted fault type label and the true label y l The difference between the two is: Use the gradient descent method to optimize the loss function L, and obtain the minimized loss function L, so as to update the weight matrix W and bias vector b in the fault prediction model;

[0102] The update formula of the weight matrix W is: μ1 is the learning rate of the weight matrix W;

[0103] The update formula of the bias vector b is: μ2 is the learning rate of bias vector b;

[0104] Repeat the above content until the cross entropy loss function L reaches the predetermined maximum number of iterations, output the optimized fault prediction model, and predict the high-speed rail bearing fault.

[0105] This embodiment describes in detail that the optimized fault prediction model uses the cross entropy loss function L to measure the prediction, updates the weight matrix W and the bias vector b in the fault prediction model, enhances the fault prediction model's ability to learn and understand the high-speed rail bearing failure mode, and achieves effective early warning of possible future failures of high-speed rail bearings, providing a solid foundation for preventive maintenance, while also improving the generalization ability and accuracy of the model.

[0106] The above are only preferred embodiments of the present invention, which do not limit the scope of protection of the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any changes, modifications, replacements, integrations and parameter changes to these embodiments within the spirit and principles of the present invention through conventional substitutions or without departing from the principles and spirit of the present invention fall within the scope of protection of the present invention.

Claims

1. A high-speed rail bearing multi-scale feature fusion fault prediction method, characterized in that: include: The vibration signal of the high-speed rail bearing is captured by an integrated sensor, and the vibration signal of the high-speed rail bearing fault is obtained by vibration signal comparison technology as the original vibration signal of the high-speed rail bearing fault; The noise and interference signal of the original vibration signal of the high-speed rail bearing fault are removed by an adaptive filter to obtain a vibration signal; The vibration signal is decomposed into multiple levels using wavelet packet decomposition to extract sub-band signals of different frequency bands. The instantaneous frequency and energy distribution of the sub-band signals are calculated using Hilbert-Huang transform to form a three-dimensional feature map of time-frequency-energy. Adopt an adaptive weighting algorithm to assign weights to each feature dimension in the three-dimensional feature map to highlight the fault characteristics; construct a training data set through the time-frequency-energy three-dimensional feature map and fault type labels; The long short-term memory network (LSTM) is trained using the training data set to build a fault prediction model, which is then used to predict faults in high-speed rail bearings.

2. A high-speed rail bearing multi-scale feature fusion fault prediction method according to claim 1, characterized in that: The integrated sensors include but are not limited to accelerometers and displacement sensors. The integrated sensors are used to accurately measure the tiny displacements and acceleration changes in all directions of the bearing, and the original vibration signals generated during the operation of the high-speed rail bearing are captured in real time at a predetermined high sampling frequency. The physical quantity of the original vibration signal is converted into an electrical signal output to obtain the original vibration signal of the dynamic characteristics of the high-speed rail bearing.

3. The high-speed rail bearing multi-scale feature fusion fault prediction method according to claim 2 is characterized in that: The method of obtaining the bearing fault vibration signal by using the vibration signal comparison technology to obtain the bearing fault vibration signal specifically includes: Define the vibration signal of the high-speed rail bearing in a healthy state as x r (t), the real-time monitored vibration signal is x l (t), the amplitude difference is eliminated by normalization processing, and the vibration signal x of the high-speed rail bearing in a healthy state is calculated r (t) and the real-time monitored vibration signal is x l The cross-correlation function R of (t) x (t), the formula is: Where T is the time integration interval, τ is the delay time, and the formula Get the maximum correlation number ρ, set the fault threshold θ, when ρ≤θ, then output the monitored vibration signal x in real time l (t) is the fault vibration signal x(t).

4. The high-speed rail bearing multi-scale feature fusion fault prediction method according to claim 1 is characterized in that: The fault vibration signal is decomposed at multiple levels using wavelet packet decomposition to extract sub-band signals of different frequency bands, including the following: Define the wavelet coefficients W j,k and the scale factor V j,k , through the wavelet coefficient W j,k Capture the high-frequency vibration frequency of the vibration signal x(t) and calculate the wavelet coefficient W of the vibration signal x(t) j,k , the formula is: W j,k =∫x(t)ψ j,k (t)dt, where j represents the number of decomposition levels, k represents the index of position j, and ψ j,k (t) is the wavelet mother function, and the formula is By scaling factor V j,k Capture the low-frequency vibration frequency of the vibration signal x(t) and calculate the scale coefficient V of the vibration signal x(t) j,k , the formula is: V j,k =∫x(t)φ j,k (t)dt; where φ j,k (t) is the scale function, and the formula is: The vibration signal x(t) is decomposed into multi-layer frequency band signals by wavelet packet decomposition technology.

5. The high-speed rail bearing multi-scale feature fusion fault prediction method according to claim 1 is characterized in that: Extract sub-band signals of different frequency bands from multi-layer frequency band signals to identify the fault vibration signal type of high-speed rail bearings, including: Through multi-level decomposition, a series of wavelet coefficients and scale coefficients are obtained, corresponding to different frequency band signals; for each layer of frequency band signals containing the nth node, each node can be further decomposed into two sub-nodes, including a high-frequency sub-node and a low-frequency sub-node, and applied to all nodes through a recursive mechanism until the required total decomposition layer number J is reached; The wavelet coefficient W j,k and the scale factor V j,k Reconstruct the subband signal s j,n (t), the sub-band signal s j,n The formula for (t) is: j,n (t)=∑ k W j,k ·V j,k , where n is the node position, j is the number of decomposition levels, and k represents the index of position j.

6. A high-speed rail bearing multi-scale feature fusion fault prediction method according to claim 1 or 5, characterized in that: Based on the sub-band signals, the Hilbert-Huang transform is used to calculate the instantaneous frequency and energy distribution of each sub-band to form a three-dimensional feature map of time-frequency-energy, including the following contents: For the subband signal s at the nth node of the jth layer j,n (t), the analytical signal a is obtained by Hilbert transform j,n (t), the formula is: in represents the Hilbert transform function, i is the imaginary part of the subband signal; By analyzing the signal a j,n (t) Calculate the instantaneous amplitude A j,n (t) and instantaneous phase θ j,n (t), the instantaneous amplitude A j,n The formula for (t) is: j,n (t)=|a j,n (t)|, the instantaneous phase θ j,n The formula for (t) is: j,n (t) = arg(a j,n (t)); According to the instantaneous phase θ j,n (t) Calculate the instantaneous frequency f j,n (t), the formula is: The energy distribution of the subband signal is defined as E j,n (t,f), where t is time and f is frequency, the formula is: E j,n (t,f)=A 2 j,n (t)·δ(ff j,n (t)), where δ is the Dirac function; The instantaneous frequency and energy distribution of the sub-band signal are combined to form a comprehensive time-frequency-energy three-dimensional feature map TFE (t, f, E), the formula is: Where j represents the number of the jth decomposition layer, J is the total number of decomposition layers, and n is the node of the jth layer.

7. The high-speed rail bearing multi-scale feature fusion fault prediction method according to claim 1 is characterized in that: The adaptive weighting algorithm is used to assign weights to each feature dimension in the three-dimensional feature map to highlight key fault features, specifically including: The time-frequency-energy three-dimensional feature map TFE(t,f,E) is discretized into multiple energy units according to the energy distribution, and the number of energy units is defined as N t , then each energy unit can be expressed as E n , where n = 1, 2, 3..., N t , each energy unit E n Mapped to the corresponding time-frequency domain, weights are assigned through an adaptive multi-scale transformation mechanism, and weighting functions w(f) are defined to adjust the weights of different frequency bands. The formula is: Convert the energy unit into a multi-scale feature matrix M along the time axis t,f , the formula is: Where δ is the Dirac function; The adaptive weighting coefficient α(t, f) is used to dynamically adjust according to the changes in time and frequency. The adaptive weighting coefficient α(t, f) formula is: Where p is the adjustment weighted sensitivity coefficient; The multi-scale feature matrix M is adjusted using the adaptive weighting coefficient α(t,f). t,f Weighted to form an optimized multi-scale feature map TFE opt (t,f,E), the formula is: TFE opt (t,f,E)=α(t,f)·M t,f .

8. The high-speed rail bearing multi-scale feature fusion fault prediction method according to claim 1 is characterized in that: The fault type labels are constructed according to the fault types of multiple high-speed rail fault vibration signals. By defining a set of fault types of high-speed rail bearings, for each fault type of vibration signal, a corresponding feature vector is extracted from the optimized multi-scale feature map, a unique label is assigned to each fault type, and a unique hot encoding method is used to construct the fault type label; Define the high-speed rail bearing fault type as l and the fault type label as y l , the y l =[y l,1 ,y l,2 ,…,y l,L ] T , where L is the number of fault types corresponding to the label.

9. A high-speed rail bearing multi-scale feature fusion fault prediction method according to claim 1 or 8, characterized in that: The method of constructing a training data set using the time-frequency-energy three-dimensional feature map corresponding to the fault type and the corresponding fault type label specifically includes: The optimized multi-scale feature map TFE opt (t,f,E) TFE is obtained based on the high-speed rail bearing fault type l l (t,f,E), from TFE l Extract the characteristic vector x of the fault type from (t,f,E) l , the formula is: l =[x l,1 ,x l,2 ,…,x l,M ] T , where M is the dimension of the fault feature vector, x l,1 ,x l,2 ,…,x l,M From TFE l (t,f,E) extracted eigenvalues; The feature vector x of each fault type l With the fault type label y l Combined into a training sample (x l ,y l ), all training samples {(x1,y1),(x2,y2),…,(x N ,y N )} form a training set D, where N is the number of training samples.

10. A high-speed rail bearing multi-scale feature fusion fault prediction method according to claim 9, characterized in that: The long short-term memory network LSTM is trained using the training data set D to obtain a fault prediction model for identifying and predicting high-speed rail bearing faults, specifically including: The pre-built LSTM fault prediction model contains multiple hidden layers, each layer has H hidden units, and the weight matrix W and bias vector b of the LSTM fault prediction model are initialized; Perform forward propagation through the long short-term memory network LSTM to calculate the hidden state h t , the formula is: t =o t ⊙tanh(c t ), where o t is the output gate state, ⊙ is the element-by-element multiplication, c t is the cell state; the output gate state o t The formula is: t =σ(W o [h t-1 ,x t ]+b o ), where σ is the sigmoid activation function, W o is the output gate weight matrix, h t-1 is the hidden state at time t-1, x t is the eigenvector at time t, b o is the output gate bias vector; the cell state c t The formula is: Among them, f t is the forget gate state, i t is the input gate state, is the candidate cell state; the forget gate state f t The formula is: t =σ(W f [h t-1 ,x t ]+b f ), where W f is the forget gate weight matrix, b f is the forget gate bias vector; the input gate state i t The formula is: t =σ(W i [h t-1 ,x t ]+b i ), where W i is the input gate weight matrix, b i is the input gate bias vector; the candidate cell state The formula is: Where W c is the candidate cell state weight matrix, b c is the candidate cell state bias vector; Through the hidden state h t Pass it to the fully connected layer to get the predicted fault type label The formula is The predicted fault results are output through the fault prediction model.

Images