A method for predicting rack life based on wear identification
By collecting sound signals on train gears and performing wavelet decomposition, we can identify wear abnormalities, and use the BP neural network model to predict the life of gear rails, the problem of insufficient life prediction of gear rails in the existing technology is solved, and high-precision wear recognition and predict.
Patent Information
- Application Number
- CN202411090073.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-08-09
AI Technical Summary
The prior art lacks effective gear rail life prediction methods, and cannot accurately predict the wear of gear rails, resulting in high maintenance costs and safety risks.
By installing sound sensors on train gears, acquiring sound signals and performing wavelet decomposition, high-frequency and low-frequency coefficient vectors are extracted, wear abnormalities are identified, difference vectors and amplitude coefficients are constructed, and life prediction is performed using the BP neural network model.
Improves the accuracy of gear rail life prediction, accurately identify wear conditions, and reduces maintenance costs and safety risks.
Smart Images

Figure CN118643298B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of rack life prediction, and in particular to a rack life prediction method based on wear identification. Background Art
[0002] A rack train is a specially designed rail transit vehicle used to climb steep hillsides. The running part of this train is equipped with a gear, which meshes with a rack installed in the middle of the track when the train is climbing. Through the bite force between the gear and the rack, the train can move forward steadily and travel safely even on very steep slopes. Rack trains can travel safely on slopes of up to 48%. Due to the harsh environmental conditions of mountain tracks, the racks are subjected to heavy loads and complex stresses while being affected by the natural environment (such as temperature changes, humidity, corrosion, etc.). Therefore, the wear of the racks is usually more serious than that of rail transit systems in flat areas. In order to ensure the safe operation of mountain tracks and reduce maintenance costs, it is particularly important to predict the life of the racks. In the existing technology, only the damage of the racks can be monitored, and there is a lack of a method for predicting the life of the racks. Summary of the invention
[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for predicting the life of a rack based on wear identification, which solves the problem of the lack of a method for predicting the life of a rack.
[0004] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for predicting the life of a rack based on wear identification, comprising the following steps:
[0005] S1, extracting multiple sound sub-signals through the sound sensor on the train gear;
[0006] S2, performing wavelet decomposition on each sound sub-signal to obtain a high-frequency coefficient vector and a low-frequency coefficient vector;
[0007] S3, constructing a high-frequency coefficient mean vector and a low-frequency coefficient mean vector according to the high-frequency coefficient vector and the low-frequency coefficient vector corresponding to each sound sub-signal;
[0008] S4, identifying the high-frequency coefficient vector and the low-frequency coefficient vector of wear abnormality according to the gap between the high-frequency coefficient vector and the high-frequency coefficient mean vector, and the gap between the low-frequency coefficient vector and the low-frequency coefficient mean vector;
[0009] S5. constructing a high-frequency difference vector and a low-frequency difference vector according to the high-frequency coefficient vector and the low-frequency coefficient vector of the wear abnormality;
[0010] S6, extracting a high-frequency amplitude coefficient and a low-frequency amplitude coefficient according to the high-frequency difference vector and the low-frequency difference vector;
[0011] S7. According to the high-frequency difference vector and the low-frequency difference vector, based on the compensation of the high-frequency amplitude coefficient and the low-frequency amplitude coefficient, a life prediction model is used to perform life prediction to obtain the life value of the rack at the abnormal position.
[0012] The beneficial effects of the present invention are as follows: the present invention collects multiple sound sub-signals through the sound sensor on the train gear, uses wavelet decomposition to analyze the sound signal, extracts high-frequency coefficient vectors and low-frequency coefficient vectors to characterize the characteristics of the signal components, constructs high-frequency coefficient mean vectors and low-frequency coefficient mean vectors to characterize the overall characteristic level of the signal, finds out the high-frequency coefficient vectors and low-frequency coefficient vectors of abnormal wear, obtains high-frequency difference vectors and low-frequency difference vectors, and uses the life prediction model to predict the life based on the compensation of the high-frequency amplitude coefficient and the low-frequency amplitude coefficient to obtain the life value of the rack at the abnormal location. The present invention reflects the degree of deviation of the abnormal sound sub-signal from the average level according to the high-frequency difference vector and the low-frequency difference vector, and uses the high-frequency amplitude coefficient and the low-frequency amplitude coefficient for compensation, thereby improving the accuracy of the life prediction model in predicting the life, and providing a rack life prediction method.
[0013] Furthermore, the S1 comprises the following sub-steps:
[0014] S11, installing a sound sensor on the gear that contacts the train and the rack;
[0015] S12, collecting sound signals during the running of the train through a sound sensor;
[0016] S13: Segment the sound signal to obtain multiple sound sub-signals.
[0017] The beneficial effect of the above further scheme is: after the present invention collects the sound signal during the running of the train, the sound signal is segmented, so as to segment the sound signal corresponding to the train running a certain distance, thereby ensuring that the multiple sound sub-signals can reflect the working condition of the rack section.
[0018] Furthermore, S3 includes the following sub-steps:
[0019] S31, calculating the mean of the i-th high-frequency coefficient and the mean of the i-th low-frequency coefficient according to the high-frequency coefficient vector and the low-frequency coefficient vector corresponding to each sound sub-signal, wherein i is a positive integer, the value range of i is 1 to N, and N is the number of high-frequency coefficients in the high-frequency coefficient vector or the number of low-frequency coefficients in the low-frequency coefficient vector;
[0020] S32, constructing the means of the N high-frequency coefficients into a high-frequency coefficient mean vector;
[0021] S33. Construct the mean values of the N low-frequency coefficients into a low-frequency coefficient mean vector.
[0022] Furthermore, the formula for calculating the mean value of the i-th high frequency coefficient in S31 is: ,
[0023] The high frequency coefficient mean vector in S32 is: H={h1,…,h i ,…,h N};
[0024] The formula for calculating the mean value of the i-th low-frequency coefficient in S31 is: ,
[0025] The low-frequency coefficient mean vector in S32 is: L={l1,…,l i ,…,l N};
[0026] Among them, h i is the mean of the ith high frequency coefficient, h j,i is the i-th high-frequency coefficient in the high-frequency coefficient vector corresponding to the j-th sound sub-signal, M is the number of sound sub-signals, H is the high-frequency coefficient mean vector, h1 is the mean of the first high-frequency coefficient, h N is the mean value of the Nth high frequency coefficient, l i is the mean of the i-th low-frequency coefficient, l1 is the mean of the first low-frequency coefficient, l N is the mean of the Nth low-frequency coefficient, L is the mean vector of low-frequency coefficients, j is a positive integer, l j,i is the i-th low-frequency coefficient in the low-frequency coefficient vector corresponding to the j-th sound sub-signal.
[0027] The beneficial effect of the above further scheme is that the present invention calculates the mean of the high-frequency coefficients and the mean of the corresponding low-frequency coefficients corresponding to each segment of the sound sub-signal, reflecting the average level of the high-frequency coefficients and the low-frequency coefficients of the sound signal.
[0028] Furthermore, the S4 comprises the following sub-steps:
[0029] S41, calculating a high frequency gap value according to the distance between the high frequency coefficient vector corresponding to each sound sub-signal and the high frequency coefficient mean vector;
[0030] S42, calculating a low-frequency gap value according to the distance between the low-frequency coefficient vector corresponding to each sound sub-signal and the low-frequency coefficient mean vector;
[0031] S43, calculating the wear value according to the high-frequency gap value and the low-frequency gap value;
[0032] S44. When the wear value is higher than the threshold, the corresponding high-frequency coefficient vector and low-frequency coefficient vector are high-frequency coefficient vector and low-frequency coefficient vector of abnormal wear.
[0033] The beneficial effect of the above further scheme is: the present invention calculates the wear value according to the high-frequency gap value and the low-frequency gap value, and screens out the high-frequency coefficient vector and the low-frequency coefficient vector that seriously deviate from the average level as the high-frequency coefficient vector and the low-frequency coefficient vector of abnormal wear.
[0034] Furthermore, the formula for calculating the high frequency gap value in S41 is: , where d h is the high frequency gap value, h j,i is the i-th high-frequency coefficient in the high-frequency coefficient vector corresponding to the j-th sound sub-signal, h i is the mean of the i-th high-frequency coefficient, i and j are positive integers;
[0035] The formula for calculating the low-frequency gap value in S42 is: , where d l is the low frequency gap value, l j,i is the i-th low-frequency coefficient in the low-frequency coefficient vector corresponding to the j-th sound sub-signal, l i is the mean of the i-th low-frequency coefficient, N is the number of high-frequency coefficients in the high-frequency coefficient vector or the number of low-frequency coefficients in the low-frequency coefficient vector;
[0036] The formula for calculating the wear value in S43 is: , where γ is the wear value, θ h is the high frequency weighting coefficient, θ l is the low frequency weighting coefficient.
[0037] Furthermore, the S5 comprises the following sub-steps:
[0038] S51, subtracting the high-frequency coefficient vector of the wear anomaly from the high-frequency coefficient mean vector to obtain a high-frequency difference vector;
[0039] S52, subtracting the low-frequency coefficient vector of the wear anomaly from the low-frequency coefficient mean vector to obtain a low-frequency difference vector.
[0040] Furthermore, the S6 comprises the following sub-steps:
[0041] S61, reconstructing the high-frequency difference vector to obtain a high-frequency signal;
[0042] S62, reconstructing the low-frequency difference vector to obtain a low-frequency signal;
[0043] S63, calculating the average amplitude of the high-frequency signal to obtain the high-frequency amplitude;
[0044] S64, calculating the average amplitude of the low-frequency signal to obtain the low-frequency amplitude;
[0045] S65, normalizing the high frequency amplitude to obtain a high frequency amplitude coefficient, and normalizing the low frequency amplitude to obtain a low frequency amplitude coefficient.
[0046] The beneficial effect of the above further scheme is: the present invention reconstructs the high-frequency difference vector and the low-frequency difference vector to obtain a high-frequency signal and a low-frequency signal, obtains the corresponding high-frequency amplitude coefficient and the low-frequency amplitude coefficient to represent the amplitude in the time domain, and further represents the degree of deviation.
[0047] Further, the life prediction model in S7 includes: a first BP neural network, a second BP neural network, a first compensation layer, a second compensation layer and a life output layer;
[0048] The input end of the first BP neural network is used to input a high-frequency difference vector, and its output end is connected to the first input end of the first compensation layer; the input end of the second BP neural network is used to input a low-frequency difference vector, and its output end is connected to the first input end of the second compensation layer; the second input end of the first compensation layer is used to input a high-frequency amplitude coefficient; the second input end of the second compensation layer is used to input a low-frequency amplitude coefficient; the input end of the life output layer is respectively connected to the output end of the first compensation layer and the output end of the second compensation layer, and its output end serves as the output end of the life prediction model.
[0049] Furthermore, the expression of the first compensation layer is: , where r1 is the output of the first compensation layer, tanh is the hyperbolic tangent activation function, and α h is the high frequency amplitude coefficient, ω αh is α h The weight of b αh is α h The bias of , y1 is the output of the first BP neural network;
[0050] The expression of the second compensation layer is: , where r2 is the output of the second compensation layer, α l is the low frequency amplitude coefficient, ω αl is α l The weight of b αl is α l y2 is the output of the second BP neural network;
[0051] The expression of the lifetime output layer is: , where S is the lifespan value output by the lifespan output layer, ω r1 is the weight of r1, ω r2 is the weight of r2.
[0052] The beneficial effects of the above further scheme are as follows: the present invention adopts the first BP neural network to process the high-frequency difference vector, adopts the second BP neural network to process the low-frequency difference vector, predicts the corresponding life ratio, and then adopts the first compensation layer and the second compensation layer respectively for compensation to improve the accuracy of life prediction, and the life output layer integrates the outputs of the first compensation layer and the second compensation layer to obtain the life value of the rack. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 It is a flowchart of a method for predicting the life of a rack based on wear identification;
[0054] Figure 2 This is a schematic diagram of the life prediction model. DETAILED DESCRIPTION
[0055] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0056] like Figure 1 As shown, a method for predicting the life of a rack rail based on wear identification comprises the following steps:
[0057] S1. Signal collection: extract multiple sound sub-signals through the sound sensor on the train gear;
[0058] S2, wavelet decomposition: perform wavelet decomposition on each sound sub-signal to obtain a high-frequency coefficient vector and a low-frequency coefficient vector;
[0059] In step S2, the high frequency coefficient vector is composed of a plurality of high frequency coefficients, and the low frequency coefficient vector is composed of a plurality of low frequency coefficients;
[0060] S3, constructing high-frequency and low-frequency coefficient mean vectors: constructing high-frequency coefficient mean vectors and low-frequency coefficient mean vectors according to the high-frequency coefficient vectors and low-frequency coefficient vectors corresponding to each sound sub-signal;
[0061] S4, identifying the high-frequency and low-frequency coefficient vectors of wear abnormality: identifying the high-frequency coefficient vector and the low-frequency coefficient vector of wear abnormality according to the gap between the high-frequency coefficient vector and the high-frequency coefficient mean vector, and the gap between the low-frequency coefficient vector and the low-frequency coefficient mean vector;
[0062] S5, constructing high-frequency and low-frequency difference vectors: constructing high-frequency difference vectors and low-frequency difference vectors according to the high-frequency coefficient vector and low-frequency coefficient vector of the wear abnormality;
[0063] S6, extracting high-frequency and low-frequency amplitude coefficients: extracting high-frequency amplitude coefficients and low-frequency amplitude coefficients according to the high-frequency difference vector and the low-frequency difference vector;
[0064] S7. Life prediction: Based on the high-frequency difference vector and the low-frequency difference vector, and based on the compensation of the high-frequency amplitude coefficient and the low-frequency amplitude coefficient, a life prediction model is used to perform life prediction to obtain the life value of the rack at the abnormal position.
[0065] The S1 comprises the following sub-steps:
[0066] S11, installing a sound sensor on the gear that contacts the train and the rack;
[0067] S12, collecting sound signals during the running of the train through a sound sensor;
[0068] S13: Segment the sound signal to obtain multiple sound sub-signals.
[0069] After collecting the sound signal during the running of the train, the present invention segments the sound signal, thereby segmenting the sound signal corresponding to a certain distance traveled by the train, and ensuring that the multiple sound sub-signals can reflect the working condition of the rack rail section.
[0070] In this embodiment, multi-layer wavelet decomposition is selected, for example, 5-layer wavelet decomposition is selected to obtain 5 high-frequency coefficients and 5 low-frequency coefficients. The wavelet decomposition method of each sound sub-signal is the same.
[0071] The S3 comprises the following sub-steps:
[0072] S31, calculating the mean of the i-th high-frequency coefficient and the mean of the i-th low-frequency coefficient according to the high-frequency coefficient vector and the low-frequency coefficient vector corresponding to each sound sub-signal, wherein i is a positive integer, the value range of i is 1 to N, and N is the number of high-frequency coefficients in the high-frequency coefficient vector or the number of low-frequency coefficients in the low-frequency coefficient vector;
[0073] S32, constructing the means of the N high-frequency coefficients into a high-frequency coefficient mean vector;
[0074] S33. Construct the mean values of the N low-frequency coefficients into a low-frequency coefficient mean vector.
[0075] The formula for calculating the mean value of the i-th high frequency coefficient in S31 is: ,
[0076] The high frequency coefficient mean vector in S32 is: H={h1,…,h i ,…,h N};
[0077] The formula for calculating the mean value of the i-th low-frequency coefficient in S31 is: ,
[0078] The low-frequency coefficient mean vector in S32 is: L={l1,…,l i ,…,l N};
[0079] Among them, h i is the mean of the ith high frequency coefficient, h j,i is the i-th high-frequency coefficient in the high-frequency coefficient vector corresponding to the j-th sound sub-signal, M is the number of sound sub-signals, H is the high-frequency coefficient mean vector, h1 is the mean of the first high-frequency coefficient, h N is the mean value of the Nth high frequency coefficient, l i is the mean of the i-th low-frequency coefficient, l1 is the mean of the first low-frequency coefficient, l N is the mean of the Nth low-frequency coefficient, L is the mean vector of low-frequency coefficients, j is a positive integer, l j,i is the i-th low-frequency coefficient in the low-frequency coefficient vector corresponding to the j-th sound sub-signal.
[0080] The present invention calculates the mean value of the high-frequency coefficients and the mean value of the low-frequency coefficients corresponding to each segment of the sound sub-signal, reflecting the average level of the high-frequency coefficients and the low-frequency coefficients of the sound signal.
[0081] The S4 comprises the following sub-steps:
[0082] S41, calculating a high frequency gap value according to the distance between the high frequency coefficient vector corresponding to each sound sub-signal and the high frequency coefficient mean vector;
[0083] S42, calculating a low-frequency gap value according to the distance between the low-frequency coefficient vector corresponding to each sound sub-signal and the low-frequency coefficient mean vector;
[0084] S43, calculating the wear value according to the high-frequency gap value and the low-frequency gap value;
[0085] S44. When the wear value is higher than the threshold, the corresponding high-frequency coefficient vector and low-frequency coefficient vector are high-frequency coefficient vector and low-frequency coefficient vector of abnormal wear.
[0086] In this embodiment, the threshold is set based on experiments or experience.
[0087] The present invention calculates the wear value according to the high-frequency gap value and the low-frequency gap value, and selects the high-frequency coefficient vector and the low-frequency coefficient vector that seriously deviate from the average level as the high-frequency coefficient vector and the low-frequency coefficient vector of abnormal wear.
[0088] The formula for calculating the high frequency gap value in S41 is: , where dh is the high frequency gap value, h j,i is the i-th high-frequency coefficient in the high-frequency coefficient vector corresponding to the j-th sound sub-signal, h i is the mean of the i-th high-frequency coefficient, i and j are positive integers;
[0089] The formula for calculating the low-frequency gap value in S42 is: , where d l is the low frequency gap value, l j,i is the i-th low-frequency coefficient in the low-frequency coefficient vector corresponding to the j-th sound sub-signal, l i is the mean of the i-th low-frequency coefficient, N is the number of high-frequency coefficients in the high-frequency coefficient vector or the number of low-frequency coefficients in the low-frequency coefficient vector;
[0090] The formula for calculating the wear value in S43 is: , where γ is the wear value, θ h is the high frequency weighting coefficient, θ l is the low frequency weighting coefficient, θ h +θ l =1.
[0091] In this embodiment, the specific value of the high-frequency weighting coefficient is greater than the specific value of the low-frequency weighting coefficient, which reflects the detailed information of the sound and highlights the edge and transient state of the sound.
[0092] The S5 comprises the following sub-steps:
[0093] S51, subtracting the high-frequency coefficient vector of the wear anomaly from the high-frequency coefficient mean vector to obtain a high-frequency difference vector;
[0094] S52, subtracting the low-frequency coefficient vector of the wear anomaly from the low-frequency coefficient mean vector to obtain a low-frequency difference vector.
[0095] The S6 comprises the following sub-steps:
[0096] S61, reconstructing the high-frequency difference vector to obtain a high-frequency signal;
[0097] S62, reconstructing the low-frequency difference vector to obtain a low-frequency signal;
[0098] S63, calculating the average amplitude of the high-frequency signal to obtain the high-frequency amplitude;
[0099] S64, calculating the average amplitude of the low-frequency signal to obtain the low-frequency amplitude;
[0100] S65, normalizing the high frequency amplitude to obtain a high frequency amplitude coefficient, and normalizing the low frequency amplitude to obtain a low frequency amplitude coefficient.
[0101] The present invention reconstructs the high-frequency difference vector and the low-frequency difference vector to obtain a high-frequency signal and a low-frequency signal, obtains the corresponding high-frequency amplitude coefficient and the low-frequency amplitude coefficient to represent the amplitude in the time domain, and further represents the degree of deviation.
[0102] like Figure 2 As shown, the life prediction model in S7 includes: a first BP neural network, a second BP neural network, a first compensation layer, a second compensation layer and a life output layer;
[0103] The input end of the first BP neural network is used to input a high-frequency difference vector, and its output end is connected to the first input end of the first compensation layer; the input end of the second BP neural network is used to input a low-frequency difference vector, and its output end is connected to the first input end of the second compensation layer; the second input end of the first compensation layer is used to input a high-frequency amplitude coefficient; the second input end of the second compensation layer is used to input a low-frequency amplitude coefficient; the input end of the life output layer is respectively connected to the output end of the first compensation layer and the output end of the second compensation layer, and its output end serves as the output end of the life prediction model.
[0104] In this embodiment, the English full name of BP in the first BP neural network and the second BP neural network is backpropagation. The first BP neural network and the second BP neural network both include an input layer, a hidden layer and an output layer. The first BP neural network inputs a high-frequency difference vector, and the second BP neural network inputs a low-frequency difference vector. The input nodes of the input layer of the first BP neural network are the same as the elements in the high-frequency difference vector, and the input nodes of the second BP neural network are the same as the elements in the low-frequency difference vector. The output layer of the first BP neural network has 1 output node, and the output layer of the second BP neural network has 1 output node. The activation functions of the hidden layer and the output layer can be selected from the S-type activation function and the tanh hyperbolic tangent activation function.
[0105] The expression of the first compensation layer is: , where r1 is the output of the first compensation layer, tanh is the hyperbolic tangent activation function, and α h is the high frequency amplitude coefficient, ω αh is α h The weight of b αh is α h The bias of , y1 is the output of the first BP neural network;
[0106] The expression of the second compensation layer is: , where r2 is the output of the second compensation layer, α l is the low frequency amplitude coefficient, ω αl is α l The weight of b αl is α ly2 is the output of the second BP neural network;
[0107] The expression of the lifetime output layer is: , where S is the lifespan value output by the lifespan output layer, ω r1 is the weight of r1, ω r2 is the weight of r2.
[0108] The present invention uses a first BP neural network to process high-frequency difference vectors, and uses a second BP neural network to process low-frequency difference vectors, predicts the corresponding life ratio, and then uses a first compensation layer and a second compensation layer to compensate respectively to improve the accuracy of life prediction. The life output layer combines the outputs of the first compensation layer and the second compensation layer to obtain the life value of the rack.
[0109] In this embodiment, the weights and biases in the life prediction model may be trained using a loss function, and the loss function may be calculated using the difference between the predicted life value and the marked life value.
[0110] The present invention collects multiple sound sub-signals through sound sensors on train gears, analyzes the sound signals using wavelet decomposition, extracts high-frequency coefficient vectors and low-frequency coefficient vectors for characterizing the characteristics of signal components, constructs high-frequency coefficient mean vectors and low-frequency coefficient mean vectors for characterizing the overall characteristic level of the signal, finds out the high-frequency coefficient vectors and low-frequency coefficient vectors of abnormal wear, obtains high-frequency difference vectors and low-frequency difference vectors, and uses a life prediction model to perform life prediction based on compensation of high-frequency amplitude coefficients and low-frequency amplitude coefficients to obtain the life value of the rack at the abnormal location. The present invention reflects the degree of deviation of the abnormal sound sub-signal from the average level according to the high-frequency difference vector and the low-frequency difference vector, and uses high-frequency amplitude coefficients and low-frequency amplitude coefficients for compensation, thereby improving the accuracy of the life prediction model in predicting the life, and providing a rack life prediction method.
[0111] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for predicting the life of a rack based on wear identification, characterized in that: The following steps are involved: S1, extracting multiple sound sub-signals through the sound sensor on the train gear; S2, performing wavelet decomposition on each sound sub-signal to obtain a high-frequency coefficient vector and a low-frequency coefficient vector; S3, constructing a high-frequency coefficient mean vector and a low-frequency coefficient mean vector according to the high-frequency coefficient vector and the low-frequency coefficient vector corresponding to each sound sub-signal; S4, identifying the high-frequency coefficient vector and the low-frequency coefficient vector of wear abnormality according to the gap between the high-frequency coefficient vector and the high-frequency coefficient mean vector, and the gap between the low-frequency coefficient vector and the low-frequency coefficient mean vector; S5. constructing a high-frequency difference vector and a low-frequency difference vector according to the high-frequency coefficient vector and the low-frequency coefficient vector of the wear abnormality; S6, extracting a high-frequency amplitude coefficient and a low-frequency amplitude coefficient according to the high-frequency difference vector and the low-frequency difference vector; S7, according to the high-frequency difference vector and the low-frequency difference vector, based on the compensation of the high-frequency amplitude coefficient and the low-frequency amplitude coefficient, using the life prediction model to perform life prediction, and obtain the life value of the rack at the abnormal position; The S6 comprises the following sub-steps: S61, reconstructing the high-frequency difference vector to obtain a high-frequency signal; S62, reconstructing the low-frequency difference vector to obtain a low-frequency signal; S63, calculating the average amplitude of the high-frequency signal to obtain the high-frequency amplitude; S64, calculating the average amplitude of the low-frequency signal to obtain the low-frequency amplitude; S65, normalizing the high frequency amplitude to obtain a high frequency amplitude coefficient, and normalizing the low frequency amplitude to obtain a low frequency amplitude coefficient; The life prediction model in S7 includes: a first BP neural network, a second BP neural network, a first compensation layer, a second compensation layer and a life output layer; The input end of the first BP neural network is used to input a high-frequency difference vector, and its output end is connected to the first input end of the first compensation layer; the input end of the second BP neural network is used to input a low-frequency difference vector, and its output end is connected to the first input end of the second compensation layer; the second input end of the first compensation layer is used to input a high-frequency amplitude coefficient; the second input end of the second compensation layer is used to input a low-frequency amplitude coefficient; the input end of the life output layer is respectively connected to the output end of the first compensation layer and the output end of the second compensation layer, and its output end serves as the output end of the life prediction model; The expression of the first compensation layer is: , where r1 is the output of the first compensation layer, tanh is the hyperbolic tangent activation function, and α h is the high frequency amplitude coefficient, ω αh is α h The weight of b αh is α h The bias of , y1 is the output of the first BP neural network; The expression of the second compensation layer is: , where r2 is the output of the second compensation layer, α l is the low frequency amplitude coefficient, ω αl is α l The weight of b αl is α l y2 is the output of the second BP neural network; The expression of the lifetime output layer is: , where S is the lifespan value output by the lifespan output layer, ω r1 is the weight of r1, ω r2 is the weight of r2.
2. The rack life prediction method based on wear identification according to claim 1 is characterized in that: The S1 comprises the following sub-steps: S11, installing a sound sensor on the gear that contacts the train and the rack; S12, collecting sound signals during the running of the train through a sound sensor; S13: Segment the sound signal to obtain multiple sound sub-signals.
3. The method for predicting rack life based on wear identification according to claim 1, characterized in that: The S3 comprises the following sub-steps: S31, calculating the mean of the i-th high-frequency coefficient and the mean of the i-th low-frequency coefficient according to the high-frequency coefficient vector and the low-frequency coefficient vector corresponding to each sound sub-signal, wherein i is a positive integer, the value range of i is 1 to N, and N is the number of high-frequency coefficients in the high-frequency coefficient vector or the number of low-frequency coefficients in the low-frequency coefficient vector; S32, constructing the means of the N high-frequency coefficients into a high-frequency coefficient mean vector; S33. Construct the mean values of the N low-frequency coefficients into a low-frequency coefficient mean vector.
4. The method for predicting rack life based on wear identification according to claim 3 is characterized in that: The formula for calculating the mean value of the i-th high frequency coefficient in S31 is: , The high frequency coefficient mean vector in S32 is: H={h1,…,h i ,…,h N }; The formula for calculating the mean value of the i-th low-frequency coefficient in S31 is: , The low-frequency coefficient mean vector in S32 is: L={l1,…,l i ,…,l N }; Among them, h i is the mean of the i-th high frequency coefficient, h j,i is the i-th high-frequency coefficient in the high-frequency coefficient vector corresponding to the j-th sound sub-signal, M is the number of sound sub-signals, H is the high-frequency coefficient mean vector, h1 is the mean of the first high-frequency coefficient, h N is the mean value of the Nth high frequency coefficient, l i is the mean of the i-th low-frequency coefficient, l1 is the mean of the first low-frequency coefficient, l N is the mean of the Nth low-frequency coefficient, L is the mean vector of low-frequency coefficients, j is a positive integer, l j,i is the i-th low-frequency coefficient in the low-frequency coefficient vector corresponding to the j-th sound sub-signal.
5. The rack life prediction method based on wear identification according to claim 1 is characterized in that: The S4 comprises the following sub-steps: S41, calculating a high frequency gap value according to the distance between the high frequency coefficient vector corresponding to each sound sub-signal and the high frequency coefficient mean vector; S42, calculating a low-frequency gap value according to the distance between the low-frequency coefficient vector corresponding to each sound sub-signal and the low-frequency coefficient mean vector; S43, calculating the wear value according to the high-frequency gap value and the low-frequency gap value; S44. When the wear value is higher than the threshold, the corresponding high-frequency coefficient vector and low-frequency coefficient vector are high-frequency coefficient vector and low-frequency coefficient vector of abnormal wear.
6. The method for predicting rack life based on wear identification according to claim 5 is characterized in that: The formula for calculating the high frequency gap value in S41 is: , where d h is the high frequency gap value, h j,i is the i-th high-frequency coefficient in the high-frequency coefficient vector corresponding to the j-th sound sub-signal, h i is the mean of the i-th high-frequency coefficient, i and j are positive integers; The formula for calculating the low-frequency gap value in S42 is: , where d l is the low frequency gap value, l j,i is the i-th low-frequency coefficient in the low-frequency coefficient vector corresponding to the j-th sound sub-signal, l i is the mean of the i-th low-frequency coefficient, N is the number of high-frequency coefficients in the high-frequency coefficient vector or the number of low-frequency coefficients in the low-frequency coefficient vector; The formula for calculating the wear value in S43 is: , where γ is the wear value, θ h is the high frequency weighting coefficient, θ l is the low frequency weighting coefficient.
7. The method for predicting rack life based on wear identification according to claim 1, characterized in that: The S5 comprises the following sub-steps: S51, subtracting the high-frequency coefficient vector of the wear anomaly from the high-frequency coefficient mean vector to obtain a high-frequency difference vector; S52, subtracting the low-frequency coefficient vector of the wear anomaly from the low-frequency coefficient mean vector to obtain a low-frequency difference vector.
