A data-driven method for detecting damage of primary steel springs in heavy-duty locomotives

Abstract

The present invention discloses a data-driven method for detecting damage to a series of steel springs of a heavy-loaded locomotive, and belongs to the technical field of detecting damage to a series of steel springs of a heavy-loaded locomotive. The detection method comprises the following steps: S1: obtaining the vertical vibration response signal of the frame of the heavy-loaded locomotive, and preprocessing it to obtain a sample set; S2: dividing the sample set into a training set, a test set and a validation set, and performing standardization processing; S3: constructing a series of spring damage detection model; S4: using the training set and the validation set to train and adjust the series of spring damage detection model, and obtain the final series of spring damage detection model; S5: using the test set to evaluate and verify the final series of spring damage detection model, and realize the detection of damage to the series of steel springs of the target heavy-loaded locomotive. The present invention can realize intelligent hierarchical detection of damage to the series of springs of a heavy-loaded locomotive, and has the characteristics of accuracy, high efficiency, strong generalization ability, etc., and is expected to develop into a long-term monitoring method for series of spring damage.

Description

Technical Field

[0001] The invention relates to the technical field of heavy-load locomotive primary spring damage detection, and in particular to a data-driven heavy-load locomotive primary steel spring damage detection method. Background Art

[0002] In recent years, the failure of primary steel springs of heavy-duty locomotives in my country has occurred frequently. As a key component of the primary suspension system of heavy-duty locomotives, the primary steel spring connects the axle box and the frame, and plays the role of connection, positioning, and vibration reduction and buffering during vehicle operation. It can effectively reduce the vertical and lateral vibrations caused by wheel-rail unevenness and improve the stability of vehicle operation. However, with the increase in the volume and speed of heavy-duty trains, the problem of irritation unevenness between wheels and rails has become more serious, which has aggravated the high-frequency and high-amplitude vibration of the primary steel springs, causing the springs to easily fail due to fatigue or even break. The failure of the primary steel spring will directly affect the stability and safety of heavy-duty train operation, and will also greatly increase the cost of operation and maintenance. Therefore, it is necessary to accurately detect the damage status of the primary steel springs in order to effectively guide the daily maintenance of locomotive steel springs and ensure the long-term healthy and stable operation of heavy-duty locomotives.

[0003] The existing methods for detecting damage to primary steel springs include model-based methods and data-driven methods. The model-based methods mainly use traditional methods such as Kalman filtering, principal component analysis, and least squares method to establish a fault detection and isolation system (FD / I) to monitor the condition of the suspension system of rail vehicles. However, the above methods are highly dependent on the accuracy of the dynamic model, and the calculation of the dynamic model is a time-consuming task. In recent years, with the development of artificial intelligence, data-driven methods have gradually attracted widespread attention. However, since the vibration response of the suspension system of heavy-duty locomotives is disturbed by severe random unevenness and complex operating conditions, the fault characteristics are often masked by strong background noise and exhibit complex non-stationary characteristics. At present, it is still a challenging task to achieve adaptive feature extraction and accurate fault identification of primary steel springs. Summary of the invention

[0004] In view of the above problems, the present invention aims to provide a data-driven method for detecting damage of a heavy-duty locomotive primary steel spring.

[0005] The technical solution of the present invention is as follows:

[0006] A data-driven method for detecting damage of a heavy-duty locomotive primary steel spring comprises the following steps:

[0007] S1: obtaining a vertical vibration response signal of a frame of a heavy-load locomotive, and preprocessing the vertical vibration response signal of the frame to obtain a sample set;

[0008] S2: Divide the sample set into a training set, a test set and a validation set, and standardize the data in each data set;

[0009] S3: constructing a series of spring damage detection model, wherein the series of spring damage detection model includes a shallow feature extraction module, a multi-scale feature extraction module, an adaptive feature fusion module and a fault classification module;

[0010] S4: using the training set to train the primary spring damage detection model, and using the validation set to adjust model hyperparameters and optimize model structure, to obtain a final primary spring damage detection model;

[0011] S5: The test set is used to evaluate and verify the final primary spring damage detection model to achieve primary spring damage detection for target heavy-load locomotives.

[0012] Preferably, in step S1, the vertical vibration response signal of the frame of the heavy-load locomotive is obtained by installing an acceleration sensor at a frame position above the first and second wheelsets of the heavy-load locomotive.

[0013] Preferably, in step S1, the preprocessing is data enhancement processing.

[0014] Preferably, sliding window sampling is used to expand and enhance the data to achieve data enhancement.

[0015] Preferably, in step S3, the shallow feature extraction module is composed of a convolutional neural network and an effective attention mechanism;

[0016] The multi-scale feature extraction module uses a multi-scale parallel residual network to deeply mine the multi-scale potential features of the signal , the relationship between the number of branches and the branch scale in the multi-scale potential feature is:

[0017] (1)

[0018] Where: For the The feature extraction scale of each branch; is the number of branches;

[0019] The adaptive feature fusion module completes the fusion of parallel branch features through the following sub-steps:

[0020] First, the multi-scale potential features extracted by the multi-scale feature extraction module are fused through feature cascade:

[0021] (2)

[0022] Then, the weight relationship of each branch feature is adaptively learned through one-dimensional convolution to obtain the feature :

[0023] (3)

[0024] Where: is global average pooling; It is a one-dimensional convolution; is the size of the one-dimensional convolution kernel;

[0025] Finally, the Softmax function is used to obtain the weight of each branch and assign it to each branch to obtain the fusion feature. :

[0026] (4)

[0027] Where: is the attention weight for each branch, and .

[0028] Preferably, the multi-scale parallel residual network adopts a parallel structure with three branches, and the scales of each branch are 3×1, 5×1 and 7×1 respectively.

[0029] Preferably, the size of the one-dimensional convolution kernel is By channel dimension Decide:

[0030] (5)

[0031] Where: is the odd number closest to the calculated value; and are all constants.

[0032] Preferably, the constant The value is 2, constant The value is 1.

[0033] Preferably, in step S3, the classification rule of the fault classification module is:

[0034] When the vertical stiffness coefficient decay is equal to 0%, the spring is normal;

[0035] When 0%<vertical stiffness coefficient attenuation≤50%, the spring has a minor fault;

[0036] When 50%<vertical stiffness coefficient attenuation≤75%, the spring is moderately faulty;

[0037] When the vertical stiffness coefficient attenuation is 75%<≤100%, the spring is seriously faulty.

[0038] The beneficial effects of the present invention are:

[0039] 1. The present invention uses the vertical vibration response signal of the frame as the monitoring signal, does not need to rely on manual signal processing, can adaptively realize the damage detection of a series of steel springs, and is expected to develop into a long-term monitoring method.

[0040] 2. The first series of steel spring damage detection models of the present invention include a multi-scale feature extraction module using a multi-scale parallel residual structure, which can fully mine multi-scale potential features containing spring fault information. Compared with the deep residual structure, the multi-scale parallel residual structure can effectively enhance the feature learning ability of the method by adjusting the number of branches and the corresponding scale, effectively overcoming the network degradation problem.

[0041] 3. The first-class steel spring damage detection model of the present invention includes an adaptive feature fusion module, which introduces the attention mechanism to adaptively learn the importance difference of each branch, ensure the accurate fusion of branch features, and then realize the accurate detection of spring damage degree. The present invention has the characteristics of accuracy, high efficiency, strong generalization ability, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0043] Figure 1 It is a flow chart of a data-driven heavy-duty locomotive primary steel spring damage detection method according to the present invention;

[0044] Figure 2 This is a schematic diagram of the installation position of the acceleration sensor of the framework of the present invention;

[0045] Figure 3 It is a structural schematic diagram of a spring damage detection model of the present invention;

[0046] Figure 4 It is a structural schematic diagram of a multi-scale feature extraction module of a series spring damage detection model in a specific embodiment;

[0047] Figure 5 It is a structural schematic diagram of an adaptive feature fusion module of a series spring damage detection model in a specific embodiment;

[0048] Figure 6Schematic diagrams of acceleration time domain and frequency domain waveforms of a simulation framework in a specific embodiment; wherein (a) is a schematic diagram of a time domain waveform of a weak fault acceleration, (b) is a schematic diagram of a frequency domain waveform of a weak fault acceleration, (c) is a schematic diagram of a time domain waveform of a moderate fault acceleration, (d) is a schematic diagram of a frequency domain waveform of a moderate fault acceleration, (e) is a schematic diagram of a time domain waveform of a severe fault acceleration, and (f) is a schematic diagram of a frequency domain waveform of a severe fault acceleration;

[0049] Figure 7 A visualization diagram of a primary steel spring damage detection process obtained according to a simulated frame acceleration in a specific embodiment;

[0050] Figure 8 It is a schematic diagram of a primary steel spring damage detection result obtained according to a simulated frame acceleration in a specific embodiment;

[0051] Fig. 9 Schematic diagram of the time domain and frequency domain waveforms of the measured frame acceleration in another specific embodiment; wherein (a) is a schematic diagram of the acceleration time domain waveform, and (b) is a schematic diagram of the acceleration frequency domain waveform;

[0052] Fig.10 is a visualization diagram of a primary steel spring damage detection process obtained according to the measured frame acceleration in another specific embodiment;

[0053] Fig.11 It is a schematic diagram of the damage detection result of a series of steel springs obtained according to the measured frame acceleration in another specific embodiment. DETAILED DESCRIPTION

[0054] The present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, in the absence of conflict, the embodiments in this application and the technical features in the embodiments can be combined with each other. It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meanings as those generally understood by those of ordinary skill in the art to which this application belongs. The words "including" or "comprising" and the like used in the disclosure of the present invention mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0055] like Figure 1 As shown, the present invention provides a data-driven method for detecting damage of a heavy-duty locomotive primary steel spring, comprising the following steps:

[0056] S1: Obtain a vertical vibration response signal of a frame of a heavy-load locomotive, and preprocess the vertical vibration response signal of the frame to obtain a sample set.

[0057] In a specific embodiment, Figure 2As shown, the vertical vibration response signal of the frame of the heavy-load locomotive is obtained by installing an acceleration sensor at the frame position above the first and second wheel pairs of the heavy-load locomotive.

[0058] In a specific embodiment, the preprocessing is data enhancement processing. Optionally, data enhancement is achieved by expanding and enhancing the data using sliding window sampling.

[0059] S2: Divide the sample set into a training set, a test set and a validation set, and standardize the data in each data set.

[0060] S3: construct a series of spring damage detection model, which includes a shallow feature extraction module, a multi-scale feature extraction module, an adaptive feature fusion module and a fault classification module.

[0061] In a specific embodiment, the shallow feature extraction module is composed of a convolutional neural network and an effective attention mechanism;

[0062] The multi-scale feature extraction module uses a multi-scale parallel residual network to deeply mine the multi-scale potential features of the signal , the relationship between the number of branches and the branch scale in the multi-scale potential feature is:

[0063] (1)

[0064] Where: For the The feature extraction scale of each branch; is the number of branches;

[0065] The adaptive feature fusion module completes the fusion of parallel branch features through the following sub-steps:

[0066] First, the multi-scale potential features extracted by the multi-scale feature extraction module are fused through feature cascade:

[0067] (2)

[0068] Then, the weight relationship of each branch feature is adaptively learned through one-dimensional convolution to obtain the feature :

[0069] (3)

[0070] Where: is global average pooling; It is a one-dimensional convolution; is the size of the one-dimensional convolution kernel;

[0071] Finally, the Softmax function is used to obtain the weight of each branch and assign it to each branch to obtain the fusion feature. :

[0072] (4)

[0073] Where: is the attention weight for each branch, and .

[0074] In a specific embodiment, the multi-scale parallel residual network adopts a parallel structure of three branches, and the scales of each branch (i.e., the convolution kernel size) are 3×1, 5×1, and 7×1. It should be noted that the number of parallel structure branches and the scale of each branch will affect the detection time and accuracy. In addition to the preferred solution of this embodiment, a multi-scale parallel residual network with other parameters can also be used according to the requirements of detection time and accuracy.

[0075] In a specific embodiment, the size of the one-dimensional convolution kernel is By channel dimension Decide:

[0076] (5)

[0077] Where: is the odd number closest to the calculated value; and are all constants.

[0078] In a specific embodiment, the constant The value is 2, constant The value is 1. It should be noted that the constant is determined by experiments (different constants are selected for experiments, and the final constant is determined according to the accuracy of the results). In addition to the constants used in this embodiment, other constants may also be used as needed.

[0079] In a specific embodiment, the primary spring damage detection model is as follows: Figure 3-5As shown, in this embodiment, the shallow feature extraction module is mainly composed of a convolutional neural network (CNN) and an effective attention mechanism (ECA), wherein the convolution basic unit structure includes a convolution layer Conv, a batch normalization layer BN, an activation function ReLU and a maximum pooling layer Maxpool; the multi-scale feature extraction module is composed of three single-scale parallel residual blocks with different sizes, each single-scale parallel residual block contains two series-connected residual structures, and each residual structure is composed of the above-mentioned convolution basic unit structure and residual connection; the adaptive feature fusion module is mainly composed of global average pooling GAP, Conv, activation function Softmax, etc., to realize the adaptive feature fusion process of branch feature fusion, branch weight learning and branch weight allocation.

[0080] In a specific embodiment, the classification rule of the fault classification module is:

[0081] When the vertical stiffness coefficient decay is equal to 0%, the spring is normal;

[0082] When 0%<vertical stiffness coefficient attenuation≤50%, the spring has a minor fault;

[0083] When 50%<vertical stiffness coefficient attenuation≤75%, the spring is moderately faulty;

[0084] When the vertical stiffness coefficient attenuation is 75%<≤100%, the spring is seriously faulty.

[0085] S4: using the training set to train the primary spring damage detection model, and using the validation set to adjust model hyperparameters and optimize model structure, to obtain a final primary spring damage detection model.

[0086] S5: The test set is used to evaluate and verify the final primary spring damage detection model to achieve primary spring damage detection for target heavy-load locomotives.

[0087] In a specific embodiment, the data-driven primary steel spring damage detection method for heavy-duty locomotives of the present invention is used to detect primary steel spring damage on simulation data. In this embodiment, a frame acceleration sensor installed above the first and second wheel sets of a heavy-duty locomotive is used to obtain a vertical vibration response signal, with a sampling frequency of 2000 Hz. The obtained simulation frame acceleration signal is as follows: Figure 6 As shown in the figure, the damage detection results of the first series steel spring are as follows: Figure 7 and Figure 8 shown.

[0088] In another specific embodiment, measured data uses springs A and B with two different stiffnesses to simulate different degrees of spring damage, wherein the stiffnesses of the springs A and B are 2.522 MN / m and 2.1234 MN / m, respectively. Fig. 9 This is a schematic diagram of the measured vertical vibration signal waveform of the frame. Fig.10 This is a visualization diagram of the damage detection process of a series of steel springs. Fig.11 This is a schematic diagram of the damage detection results of a series of steel springs of the present invention. Fig.11 It can be seen that the present invention can accurately identify different degrees of damage to the actual primary spring.

[0089] In summary, the present invention can utilize the bogie frame acceleration to accurately and efficiently perform intelligent graded detection of primary spring damage in heavy-loaded locomotives. Compared with the prior art, the present invention is a significant improvement.

[0090] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A data-driven method for detecting damage to a heavy-duty locomotive primary steel spring, characterized in that: The following steps are involved: S1: obtaining a vertical vibration response signal of a frame of a heavy-load locomotive, and preprocessing the vertical vibration response signal of the frame to obtain a sample set; S2: Divide the sample set into a training set, a test set and a validation set, and standardize the data in each data set; S3: constructing a series of spring damage detection model, wherein the series of spring damage detection model includes a shallow feature extraction module, a multi-scale feature extraction module, an adaptive feature fusion module and a fault classification module; The shallow feature extraction module consists of a convolutional neural network and an effective attention mechanism; The multi-scale feature extraction module uses a multi-scale parallel residual network to deeply mine the multi-scale potential features of the signal , the relationship between the number of branches and the branch scale in the multi-scale potential feature is: （1） Where: For the The feature extraction scale of each branch; is the number of branches; The adaptive feature fusion module completes the fusion of parallel branch features through the following sub-steps: First, the multi-scale potential features extracted by the multi-scale feature extraction module are fused through feature cascade: （2） Then, the weight relationship of each branch feature is adaptively learned through one-dimensional convolution to obtain the feature : （3） Where: is global average pooling; It is a one-dimensional convolution; is the size of the one-dimensional convolution kernel; Finally, the Softmax function is used to obtain the weight of each branch and assign it to each branch to obtain the fusion feature. : （4） Where: is the attention weight for each branch, and ; S4: using the training set to train the primary spring damage detection model, and using the validation set to adjust model hyperparameters and optimize model structure, to obtain a final primary spring damage detection model; S5: The test set is used to evaluate and verify the final primary spring damage detection model to achieve primary spring damage detection for target heavy-load locomotives.

2. The data-driven heavy-duty locomotive primary steel spring damage detection method according to claim 1 is characterized in that: In step S1, the vertical vibration response signal of the frame of the heavy-load locomotive is obtained by installing an acceleration sensor at the frame position above the first and second wheelsets of the heavy-load locomotive.

3. The data-driven heavy-duty locomotive primary steel spring damage detection method according to claim 1 is characterized in that: In step S1, the preprocessing is data enhancement processing.

4. The data-driven heavy-duty locomotive primary steel spring damage detection method according to claim 3 is characterized in that: Sliding window sampling is used to expand and enhance the data to achieve data enhancement.

5. The data-driven heavy-duty locomotive primary steel spring damage detection method according to claim 1 is characterized in that: The multi-scale parallel residual network adopts a parallel structure of three branches, and the scales of each branch are 3×1, 5×1 and 7×1 respectively.

6. The data-driven heavy-duty locomotive primary steel spring damage detection method according to claim 1 is characterized in that: The size of the one-dimensional convolution kernel By channel dimension Decide: （5） Where: is the odd number closest to the calculated value; and are all constants.

7. The data-driven heavy-duty locomotive primary steel spring damage detection method according to claim 6 is characterized in that: constant The value is 2, constant The value is 1.

8. The data-driven heavy-duty locomotive primary steel spring damage detection method according to any one of claims 1 to 7, characterized in that: In step S3, the classification rule of the fault classification module is: When the vertical stiffness coefficient decay is equal to 0%, the spring is normal; When 0%<vertical stiffness coefficient attenuation≤50%, the spring has a minor fault; When 50%<vertical stiffness coefficient attenuation≤75%, the spring is moderately faulty; When the vertical stiffness coefficient attenuation is 75%<≤100%, the spring is seriously faulty.

Images