A Fault Diagnosis Method and System for Leaf Spring Suspension Based on Multi-Sensor Data and Deep Learning

Abstract

This invention discloses a method and system for fault diagnosis of leaf spring suspension based on multi-sensor data and deep learning, relating to the field of vehicle fault diagnosis technology. The method includes: acquiring a fault diagnosis dataset for a leaf spring suspension system; processing each signal data in the dataset using a joint processing method of empirical mode decomposition and wavelet denoising; converting the two acceleration signal data and one displacement signal data corresponding to each fault diagnosis data into two-dimensional color image data; training a pre-constructed deep learning network model using the leaf spring suspension system fault diagnosis dataset and the two-dimensional color image data; and performing leaf spring suspension fault diagnosis using the trained deep learning network model. This invention enables real-time fault diagnosis of leaf spring suspension systems, achieving accurate identification and location of leaf spring suspension faults, and ensuring the accuracy and reliability of the diagnostic results.

Description

Technical Field

[0001] This invention relates to the field of vehicle fault diagnosis technology, and in particular to a method and system for diagnosing leaf spring suspension faults based on multi-sensor data and deep learning. Background Technology

[0002] Leaf spring suspension systems used in heavy vehicles are prone to malfunctions such as loosening of U-bolts, degradation of leaf spring performance, and degradation of shock absorber performance when operating under long-term complex conditions. Traditional methods for diagnosing vehicle leaf spring suspension faults combine manual inspection with periodic checks. However, in practical applications, these methods suffer from problems such as detection lag, difficulty in detecting early minor damage, heavy reliance on experience, and insufficient diagnostic accuracy.

[0003] Therefore, there is an urgent need for a leaf spring suspension fault diagnosis method and system that can detect faults in a timely manner and achieve accurate identification and location of leaf spring suspension faults. Summary of the Invention

[0004] To address some or all of the technical problems existing in the prior art, the present invention provides a method and system for diagnosing leaf spring suspension faults based on multi-sensor data and deep learning.

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

[0006] Firstly, a method for diagnosing leaf spring suspension faults based on multi-sensor data and deep learning is provided, including:

[0007] A fault diagnosis dataset for the leaf spring suspension system is obtained. The fault diagnosis dataset includes multiple fault diagnosis data, including axle acceleration signal data, frame acceleration signal data, suspension displacement signal data, and corresponding leaf spring suspension system status. The leaf spring suspension system status includes: normal status, loose leaf spring U-bolt status, leaf spring performance degradation status, and shock absorber performance degradation status.

[0008] Based on the joint processing method of empirical mode decomposition and wavelet denoising, data processing is performed on each signal data in the fault diagnosis dataset of the leaf spring suspension system to extract useful information from the signal data.

[0009] The two acceleration signal data and one displacement signal data corresponding to each of the fault diagnosis data are converted into two-dimensional color image data after data processing.

[0010] A pre-constructed deep learning network model is trained using the fault diagnosis dataset of the leaf spring suspension system and the two-dimensional color image data corresponding to the fault diagnosis data. The input of the deep learning network model is the two-dimensional color image data corresponding to the fault diagnosis data, and the output of the deep learning network model is the state of the leaf spring suspension system corresponding to the fault diagnosis data.

[0011] Real-time acquisition of axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed, and using the trained deep learning network model to perform leaf spring suspension fault diagnosis.

[0012] In some alternative implementations, fault diagnosis datasets for leaf spring suspension systems are obtained through real-vehicle road tests and simulation tests.

[0013] In some optional implementations, obtaining a fault diagnosis dataset for the leaf spring suspension system through real-vehicle road tests and simulation tests further includes the following steps:

[0014] Step 101: Install accelerometers at the axle and the frame respectively, and install a laser displacement sensor between the axle and the frame;

[0015] Step 102: Conduct a real-vehicle road test with the leaf spring suspension system in normal condition, and collect the acceleration signals from the two accelerometers and the displacement signals from the laser displacement sensor in real time to obtain multiple fault diagnosis data corresponding to the normal condition; and conduct a real-vehicle road test with the leaf spring suspension system in the leaf spring suspension U-bolt loose condition, and collect the acceleration signals from the two accelerometers and the displacement signals from the laser displacement sensor in real time to obtain multiple fault diagnosis data corresponding to the leaf spring suspension U-bolt loose condition.

[0016] Step 103: Based on the data obtained from real vehicle road tests conducted under normal conditions of the leaf spring suspension system, a high-precision rigid-flexible coupling simulation model is established using RecurDyn software. The stiffness parameters of the leaf spring are dynamically adjusted using the STEP function in RecurDyn software to simulate the performance degradation process of the leaf spring. Simulation tests are conducted based on the adjusted parameters to obtain fault diagnosis data corresponding to multiple leaf spring performance degradation states. Similarly, the damping parameters of the shock absorber are dynamically adjusted using the STEP function to simulate the performance degradation process of the shock absorber. Simulation tests are conducted based on the adjusted parameters to obtain fault diagnosis data corresponding to multiple shock absorber performance degradation states.

[0017] In some alternative implementations, a joint processing method based on empirical mode decomposition and wavelet denoising processes the signal data through the following steps:

[0018] Step 201: Identify all local maxima and local minima in the signal, and generate the upper and lower envelopes of the signal respectively using cubic spline interpolation.

[0019] Step 202: Calculate the average envelope based on the upper and lower envelope lines;

[0020] Step 203: Extract signal components based on the average envelope;

[0021] Step 204: Determine whether the signal component is an intrinsic mode function. If yes, the extracted signal component is used as an intrinsic mode function and output. Calculate the remaining signal and execute steps 201-204 based on the remaining signal. If no, output the remaining signal and proceed to the next step.

[0022] Step 205: The extracted high-frequency intrinsic mode functions are decomposed using discrete wavelet transform to obtain the corresponding approximation coefficients and detail coefficients;

[0023] Step 206: Based on the signal length of the high-frequency intrinsic mode function, set a threshold to distinguish between signal and noise;

[0024] Step 207: Process the detail coefficients corresponding to the high-frequency intrinsic mode functions using the set threshold to obtain the threshold-processed detail coefficients;

[0025] Step 208: Based on the approximation coefficients corresponding to the high-frequency intrinsic mode functions and the detail coefficients after thresholding, the denoised high-frequency intrinsic mode functions are reconstructed by wavelet reconstruction.

[0026] Step 209: Add all the denoised high-frequency intrinsic mode functions, the unprocessed low-frequency intrinsic mode functions, and the final output residual signal to reconstruct the signal and obtain the reconstructed signal.

[0027] In some alternative implementations, based on the Gram angle field method, the two acceleration signal data and one displacement signal data corresponding to each fault diagnosis data are converted into two-dimensional color image data after data processing.

[0028] In some optional implementations, the two acceleration signal data and one displacement signal data corresponding to the fault diagnosis data are converted into two-dimensional color image data through the following steps:

[0029] Step 301: Based on the two acceleration signal data and one displacement signal data after data processing corresponding to the fault diagnosis data, obtain the corresponding axle acceleration sequence data, frame acceleration sequence data and suspension displacement sequence data, and normalize each sequence data to the [0,1] interval respectively;

[0030] Step 302: Encode each value in the normalized sequence data as an angle cosine, encode the timestamp corresponding to each value in the normalized sequence data as a radius, and map each normalized sequence data to the polar coordinate system according to the angle cosine and the radius.

[0031] Step 303: Calculate the cosine similarity between any two values ​​in each sequence data after normalization, and generate the GAF matrix corresponding to each sequence data based on the cosine similarity.

[0032] Step 304: Visualize the GAF matrix corresponding to the axle acceleration sequence data, the GAF matrix corresponding to the frame acceleration sequence data, and the GAF matrix corresponding to the suspension displacement sequence data as grayscale images, and then map the grayscale images to the R, G, and B channels and stack them to generate two-dimensional color image data.

[0033] In some optional implementations, the deep learning network model includes: a 7×7 convolutional layer, a first BN processing unit, a first ReLU processing unit, a max pooling layer, a first residual block, a second residual block, a third residual block, a fourth residual block, a global average pooling layer, and a fully connected layer connected in sequence.

[0034] The first residual block includes 3 BottleNeck structures, the second residual block includes 4 BottleNeck structures, the third residual block includes 6 BottleNeck structures, and the fourth residual block includes 3 BottleNeck structures.

[0035] The BottleNeck structure includes: a second BN processing unit, a second ReLU processing unit, a 3×3 convolutional layer, a third BN processing unit, a third ReLU processing unit, a first 1×1 convolutional layer, a fourth BN processing unit, an addition unit, and a fourth ReLU processing unit connected in sequence; and a second 1×1 convolutional layer and a fifth BN processing unit connected between the input of the BottleNeck structure and the input of the addition unit. The input of the second 1×1 convolutional layer is connected to the input of the BottleNeck structure, the output of the second 1×1 convolutional layer is connected to the input of the fifth BN processing unit, and the output of the fifth BN processing unit is connected to the input of the addition unit.

[0036] In some alternative implementations, the deep learning network model is trained in the following manner:

[0037] The deep learning network model is trained by using the two-dimensional color image data corresponding to the fault diagnosis data as the input and the leaf spring suspension system state in the fault diagnosis data as the output.

[0038] In some optional implementations, the deep learning network model is trained using the two-dimensional color image data corresponding to the fault diagnosis data as input and the leaf spring suspension system state in the fault diagnosis data as output, including:

[0039] Step 401: Input the two-dimensional color image data corresponding to the multiple fault diagnosis data into the deep learning network model in sequence to obtain the leaf spring suspension system state prediction result corresponding to each fault diagnosis data output by the deep learning network model.

[0040] Step 402: Calculate the loss function based on the leaf spring suspension system status in the fault diagnosis data and the leaf spring suspension system status prediction result corresponding to the fault diagnosis data;

[0041] Step 403: Determine whether the preset training stopping condition has been met. If yes, use the current deep learning network model as the deep learning network model that has completed training. If no, update the parameters of the deep learning network model using the loss function and return to step 401.

[0042] Secondly, a leaf spring suspension fault diagnosis system based on multi-sensor data and deep learning is also provided, including:

[0043] The fault diagnosis dataset acquisition unit is used to acquire the leaf spring suspension system fault diagnosis dataset. The leaf spring suspension system fault diagnosis dataset includes multiple fault diagnosis data, including axle acceleration signal data, frame acceleration signal data, suspension displacement signal data, and corresponding leaf spring suspension system status. The leaf spring suspension system status includes: normal status, leaf spring suspension U-bolt loosening status, leaf spring performance degradation status, and shock absorber performance degradation status.

[0044] The data processing unit is used to process each signal data in the fault diagnosis dataset of the leaf spring suspension system based on the joint processing method of empirical mode decomposition and wavelet denoising, and extract useful information from the signal data.

[0045] The data conversion unit is used to convert the two acceleration signal data and one displacement signal data corresponding to each fault diagnosis data into two-dimensional color image data;

[0046] A deep learning network model generation unit is used to train a pre-constructed deep learning network model using the fault diagnosis dataset of the leaf spring suspension system and the two-dimensional color image data corresponding to the fault diagnosis data. The input of the deep learning network model is the two-dimensional color image data corresponding to the fault diagnosis data, and the output of the deep learning network model is the leaf spring suspension system state corresponding to the fault diagnosis data.

[0047] The data acquisition unit is used to collect axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed in real time.

[0048] The fault diagnosis unit is used to perform leaf spring suspension fault diagnosis based on the axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed in real time, using the trained deep learning network model.

[0049] The main advantages of the technical solution of this invention are as follows:

[0050] The present invention relates to a leaf spring suspension fault diagnosis method and system based on multi-sensor data and deep learning. By constructing and training a deep learning network model to fit the mapping relationship between the acceleration signal data and suspension displacement signal data of the leaf spring suspension system and the state of the leaf spring suspension system, the deep learning network model is then used to diagnose leaf spring suspension faults. This enables real-time fault diagnosis of the leaf spring suspension system, accurate identification and location of leaf spring suspension faults, and ensures the accuracy and reliability of the diagnostic results. Attached Figure Description

[0051] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and constitute a part of this invention, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0052] Figure 1 A flowchart illustrating a leaf spring suspension fault diagnosis method based on multi-sensor data and deep learning, provided as an embodiment of the present invention;

[0053] Figure 2 A flowchart illustrating the data processing of signal data provided in this embodiment of the invention;

[0054] Figure 3 This is a schematic diagram of the structure of a deep learning network model provided in an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram of a leaf spring suspension fault diagnosis system based on multi-sensor data and deep learning, provided as an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0057] The technical solutions provided by the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0058] refer to Figure 1 In a first aspect, embodiments of the present invention provide a method for diagnosing leaf spring suspension faults based on multi-sensor data and deep learning, the method comprising the following steps:

[0059] Step 1: Obtain the fault diagnosis dataset for the leaf spring suspension system;

[0060] In this embodiment of the invention, the leaf spring suspension system fault diagnosis dataset includes multiple fault diagnosis data, which include axle acceleration signal data, frame acceleration signal data, suspension displacement signal data, and the corresponding leaf spring suspension system status.

[0061] In this embodiment of the invention, the states of the leaf spring suspension system include: normal state, loose leaf spring suspension U-bolt state, leaf spring performance degradation state, and shock absorber performance degradation state.

[0062] Step 2: Based on the joint processing method of empirical mode decomposition and wavelet denoising, data processing is performed on each signal data in the fault diagnosis dataset of the leaf spring suspension system to extract useful information from the signal data;

[0063] In this embodiment of the invention, considering that the signal data in the fault diagnosis dataset of the leaf spring suspension system may contain noise information, a combined processing method based on empirical mode decomposition and wavelet denoising is used to process each signal data in the fault diagnosis dataset of the leaf spring suspension system, extract useful information from the signal data, and reduce noise interference.

[0064] Step 3: Convert the two acceleration signal data and one displacement signal data corresponding to each fault diagnosis data into two-dimensional color image data;

[0065] In this embodiment of the invention, based on the fault diagnosis dataset of the leaf spring suspension system after data processing, for each fault diagnosis data, the two acceleration signal data and one displacement signal data corresponding to the fault diagnosis data after data processing are converted into two-dimensional color image data, thereby obtaining the two-dimensional color image data corresponding to each fault diagnosis data.

[0066] Step 4: Train a pre-built deep learning network model using the leaf spring suspension system fault diagnosis dataset and the corresponding two-dimensional color image data of the fault diagnosis data;

[0067] In this embodiment of the invention, the input to the deep learning network model is the two-dimensional color image data corresponding to the fault diagnosis data, and the output of the deep learning network model is the state of the leaf spring suspension system corresponding to the fault diagnosis data.

[0068] In this embodiment of the invention, a deep learning network model is used to predict the corresponding state type of the leaf spring suspension system based on the input information.

[0069] Step 5: Collect axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed in real time, and use the trained deep learning network model to diagnose leaf spring suspension faults.

[0070] In this embodiment of the invention, when diagnosing a leaf spring suspension fault, the axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed are first collected. The acquired acceleration signal data and suspension displacement signal data are processed and converted to obtain corresponding two-dimensional color image data. The two-dimensional color image data is then input into a trained deep learning network model to obtain the leaf spring suspension system state output by the deep learning network model. The leaf spring suspension system state output by the deep learning network model is the current state of the leaf spring suspension system.

[0071] In this embodiment of the invention, if the leaf spring suspension system status output by the deep learning network model is normal, it indicates that the leaf spring suspension system is normal and without fault; if the leaf spring suspension system status output by the deep learning network model is loose leaf spring U-bolts, it indicates that the leaf spring suspension system has a loose leaf spring U-bolt fault; if the leaf spring suspension system status output by the deep learning network model is degraded leaf spring performance, it indicates that the leaf spring suspension system has a degraded leaf spring performance fault; if the leaf spring suspension system status output by the deep learning network model is degraded shock absorber performance, it indicates that the leaf spring suspension system has a degraded shock absorber performance fault.

[0072] The leaf spring suspension fault diagnosis method based on multi-sensor data and deep learning provided in this invention constructs and trains a deep learning network model to fit the mapping relationship between the acceleration signal data and suspension displacement signal data of the leaf spring suspension system and the state of the leaf spring suspension system. Then, the deep learning network model is used to diagnose leaf spring suspension faults, which can realize real-time fault diagnosis of the leaf spring suspension system, accurately identify and locate leaf spring suspension faults, and ensure the accuracy and reliability of the diagnostic results.

[0073] Furthermore, in this embodiment of the invention, a fault diagnosis dataset for the leaf spring suspension system is obtained through real vehicle road tests and simulation tests.

[0074] In this embodiment of the invention, obtaining a fault diagnosis dataset for the leaf spring suspension system through real vehicle road tests and simulation tests further includes the following steps:

[0075] Step 101: Install accelerometers at the axle and the frame respectively, and install a laser displacement sensor between the axle and the frame;

[0076] Step 102: Conduct a real-vehicle road test with the leaf spring suspension system in normal condition, and collect the acceleration signals from the two accelerometers and the displacement signals from the laser displacement sensor in real time to obtain multiple fault diagnosis data corresponding to the normal condition; and conduct a real-vehicle road test with the leaf spring suspension system in the leaf spring suspension U-bolt loose condition, and collect the acceleration signals from the two accelerometers and the displacement signals from the laser displacement sensor in real time to obtain multiple fault diagnosis data corresponding to the leaf spring suspension U-bolt loose condition.

[0077] Step 103: Based on the data obtained from real vehicle road tests conducted under normal conditions of the leaf spring suspension system, a high-precision rigid-flexible coupling simulation model is established using RecurDyn (Recursive Dynamic) software. The STEP function in RecurDyn software is used to dynamically adjust the stiffness parameters of the leaf spring to simulate the performance degradation process of the leaf spring. Simulation tests are conducted based on the adjusted parameters to obtain fault diagnosis data corresponding to multiple leaf spring performance degradation states. Similarly, the STEP function is used to dynamically adjust the damping parameters of the shock absorber to simulate the performance degradation process of the shock absorber. Simulation tests are conducted based on the adjusted parameters to obtain fault diagnosis data corresponding to multiple shock absorber performance degradation states.

[0078] In this embodiment of the invention, the fault diagnosis data corresponding to multiple normal states, the fault diagnosis data corresponding to multiple leaf spring suspension U-bolt loose states, the fault diagnosis data corresponding to multiple leaf spring performance degradation states, and the fault diagnosis data corresponding to multiple shock absorber performance degradation states, together constitute the leaf spring suspension system fault diagnosis dataset.

[0079] In this embodiment of the invention, the fault diagnosis dataset of the leaf spring suspension system is obtained in a safe and fast manner by means of the above method, and the accuracy of the obtained dataset is guaranteed.

[0080] Furthermore, in this embodiment of the invention, when the leaf spring suspension system is in normal condition and a real vehicle road test is conducted, the tire pressure is controlled at 900 kPa, and the vehicle is driven at speeds of 5 km / h, 10 km / h, 15 km / h, and 20 km / h on four different road surfaces: concrete road, asphalt road, gravel road, and washboard road. Acceleration signals from the accelerometer and displacement signals from the laser displacement sensor are collected at different vehicle speeds and on different road surfaces.

[0081] When conducting real-vehicle road tests with the leaf spring suspension system in a loose leaf spring suspension U-bolt state, the vehicle was driven at speeds of 5km / h, 10km / h, 15km / h, and 20km / h on four different road surfaces: concrete road, asphalt road, gravel road, and corrugated road. Acceleration signals from the accelerometer and displacement signals from the laser displacement sensor were collected at different vehicle speeds and on different road surfaces.

[0082] Furthermore, in this embodiment of the invention, during the simulation test, road surface models corresponding to concrete road, asphalt road, gravel road, and corrugated road are established respectively. Simulations are conducted at vehicle speeds of 5km / h, 10km / h, 15km / h, and 20km / h under the four different road surface models. Accelerometer acceleration signals and laser displacement sensor displacement signals are collected at different vehicle speeds and under different road surface models.

[0083] In this embodiment of the invention, the dataset obtained in the above manner can cover multi-dimensional working conditions, approximate real complex scenarios, provide an unbiased, complete, and physically interpretable training foundation for subsequent deep learning network models, and significantly improve the generalization ability of the models.

[0084] Furthermore, in this embodiment of the invention, in step 2, when processing the signal data based on the combined processing method of empirical mode decomposition and wavelet denoising, the signal data is first processed by the empirical mode decomposition method, and then processed by the wavelet denoising method based on the result of the first processing step.

[0085] refer to Figure 2 Furthermore, in this embodiment of the invention, in step 2, the signal data is processed through the following steps:

[0086] Step 201: Identify all local maxima and local minima in the signal, and generate the upper and lower envelopes of the signal respectively using cubic spline interpolation.

[0087] Step 202: Calculate the average envelope based on the upper and lower envelope lines;

[0088] Step 203: Extract signal components based on the average envelope;

[0089] Step 204: Determine whether the signal component is an intrinsic mode function (IMF). If yes, the extracted signal component is used as an IMF and output. Calculate the remaining signal and execute steps 201-204 based on the remaining signal. If no, output the remaining signal and proceed to the next step.

[0090] Step 205: The extracted high-frequency intrinsic mode functions are decomposed using discrete wavelet transform to obtain the corresponding approximation coefficients and detail coefficients;

[0091] Step 206: Based on the signal length of the high-frequency intrinsic mode function, set a threshold to distinguish between signal and noise;

[0092] Step 207: Process the detail coefficients corresponding to the high-frequency intrinsic mode functions using the set threshold to obtain the threshold-processed detail coefficients;

[0093] Step 208: Based on the approximation coefficients corresponding to the high-frequency intrinsic mode functions and the detail coefficients after thresholding, the denoised high-frequency intrinsic mode functions are reconstructed by wavelet reconstruction.

[0094] Step 209: Add all the denoised high-frequency intrinsic mode functions, the unprocessed low-frequency intrinsic mode functions, and the final output residual signal to reconstruct the signal and obtain the reconstructed signal.

[0095] In this embodiment of the invention, in step 202, the average envelope is calculated using the following formula:

[0096]

[0097] Where m(t) represents the average envelope, e max (t) represents the upper envelope, e min (t) represents the lower envelope.

[0098] In step 203, the signal components are extracted using the following formula:

[0099] h(t) = x(t) - m(t);

[0100] Where h(t) represents the extracted signal component and x(t) represents the current signal.

[0101] In step 204, the remaining signal is calculated using the following formula:

[0102] r(t) = x(t) - IMF(t);

[0103] Where r(t) represents the residual signal and IMF(t) represents the intrinsic mode function extracted from the current signal x(t).

[0104] In step 205, the intrinsic mode functions are decomposed using the following formula:

[0105]

[0106] Among them, IMF k (t) represents the k-th eigenmode function, a k,j d represents the approximation coefficient of the j-th layer corresponding to the k-th intrinsic mode function. k,j represents the detail coefficients of the j-th layer corresponding to the k-th eigenmode function, and N represents the number of layers in the wavelet decomposition.

[0107] In step 206, the threshold is determined using the following formula:

[0108]

[0109] Where λ represents the threshold, σ represents the noise standard deviation, and n represents the signal length of the intrinsic mode function.

[0110] In step 207, the detail coefficients corresponding to the intrinsic mode functions are processed using the following formula:

[0111]

[0112] in, This represents the detail coefficients after thresholding, and sign represents the sign function.

[0113] In this embodiment of the invention, by processing the detail coefficients corresponding to the intrinsic mode functions using a set threshold, small-amplitude vibrations caused by noise in the detail coefficients can be suppressed.

[0114] In step 208, the denoised intrinsic mode functions are reconstructed using the following formula:

[0115]

[0116] Among them, WTD (IMF) k (t) represents the k-th intrinsic mode function after denoising.

[0117] In step 209, the reconstruction signal is obtained using the following formula:

[0118]

[0119] Where, x clean(t) represents the reconstructed signal, m represents the number of high-frequency intrinsic mode functions (IMFs) among the extracted IMFs, n represents the number of extracted IMFs, and r n (t) represents the remaining signal after extracting n intrinsic mode functions from the initial signal.

[0120] In this embodiment of the invention, m and n are determined according to the actual situation. Among the extracted n intrinsic mode functions, the first m intrinsic mode functions are high-frequency intrinsic mode functions, and the last nm intrinsic mode functions are low-frequency intrinsic mode functions.

[0121] In this embodiment of the invention, the processing steps 201-204 correspond to the empirical mode decomposition process, and the processing steps 205-209 correspond to the wavelet denoising process.

[0122] Furthermore, in this embodiment of the invention, in step 3, based on the Gram angle field method, the two acceleration signal data and one displacement signal data corresponding to each fault diagnosis data are converted into two-dimensional color image data after data processing.

[0123] In this embodiment of the invention, the two acceleration signal data and one displacement signal data corresponding to the fault diagnosis data are converted into two-dimensional color image data through the following steps:

[0124] Step 301: Based on the two acceleration signal data and one displacement signal data after data processing corresponding to the fault diagnosis data, obtain the corresponding axle acceleration sequence data, frame acceleration sequence data and suspension displacement sequence data, and normalize each sequence data to the [0,1] interval respectively;

[0125] Step 302: Encode each value in the normalized sequence data as an angle cosine, encode the timestamp corresponding to each value in the normalized sequence data as a radius, and map each normalized sequence data to the polar coordinate system according to the angle cosine and the radius.

[0126] Step 303: Calculate the cosine similarity between any two values ​​in each sequence data after normalization, and generate the GAF matrix (Gramian Angular Field Matrix) corresponding to each sequence data based on the cosine similarity.

[0127] Step 304: Visualize the GAF matrix corresponding to the axle acceleration sequence data, the GAF matrix corresponding to the frame acceleration sequence data, and the GAF matrix corresponding to the suspension displacement sequence data as grayscale images, and then map the grayscale images to the R, G, and B channels and stack them to generate two-dimensional color image data.

[0128] In this embodiment of the invention, for each fault diagnosis data, corresponding two-dimensional color image data is generated in the manner described above.

[0129] In this embodiment of the invention, in step 301, the sequence data is normalized using the following formula:

[0130]

[0131] Where, x i Let x represent the i-th value in the sequence data, max(x) represent the maximum value in the sequence data, and min(x) represent the minimum value in the sequence data. x represents i The corresponding normalized value is the i-th value in the normalized sequence data.

[0132] In step 302, each value in the normalized sequence data is encoded as an angle cosine using the following formula:

[0133]

[0134] The timestamp corresponding to each value in the normalized sequence data is encoded as a radius using the following formula:

[0135]

[0136] Where, φ i Representing numerical values The corresponding angle cosine, arccos represents the calculation of the inverse cosine function, r i t represents the radius corresponding to the i-th value in the normalized sequence data. i The timestamp represents the i-th value in the sequence data, and N represents the amount of data contained in the sequence data.

[0137] In step 303, the matrix elements in the GAF matrix are determined using the following formula:

[0138]

[0139] Where GAF(i,j) represents the matrix element in the i-th row and j-th column of the GAF matrix, φ j Representing numerical values The corresponding angle cosine, This represents the j-th value in the normalized sequence data.

[0140] In this embodiment of the invention, the two acceleration signal data and one displacement signal data corresponding to the fault diagnosis data are converted into two-dimensional color image data by the above method, so that they can be input into the deep learning network model for subsequent processing, thereby improving the generalization ability and prediction performance of the deep learning network model.

[0141] refer to Figure 3 Furthermore, in this embodiment of the invention, the deep learning network model includes: a 7×7 convolutional layer, a first BN processing unit, a first ReLU processing unit, a max pooling layer, a first residual block, a second residual block, a third residual block, a fourth residual block, a global average pooling layer, and a fully connected layer connected in sequence.

[0142] The first residual block includes 3 BottleNeck structures, the second residual block includes 4 BottleNeck structures, the third residual block includes 6 BottleNeck structures, and the fourth residual block includes 3 BottleNeck structures.

[0143] The BottleNeck structure includes: a second BN processing unit, a second ReLU processing unit, a 3×3 convolutional layer, a third BN processing unit, a third ReLU processing unit, a first 1×1 convolutional layer, a fourth BN processing unit, an addition unit, and a fourth ReLU processing unit connected in sequence; and a second 1×1 convolutional layer and a fifth BN processing unit connected between the input of the BottleNeck structure and the input of the addition unit. The input of the second 1×1 convolutional layer is connected to the input of the BottleNeck structure, the output of the second 1×1 convolutional layer is connected to the input of the fifth BN processing unit, and the output of the fifth BN processing unit is connected to the input of the addition unit.

[0144] A 7×7 convolutional layer performs a 7×7 convolution operation on the input and outputs the resulting features; a Batch Normalization (BN) unit performs batch normalization on the input features; a ReLU unit performs a non-linear transformation on the input features using the ReLU activation function; a max pooling layer performs max pooling on the input features; a 3×3 convolutional layer performs a 3×3 convolution operation on the input features; a 1×1 convolutional layer performs a 1×1 convolution operation on the input features; an addition unit performs addition on the inputs; a global average pooling layer averages the feature maps of each channel of the input and performs dimensionality reduction; and a fully connected layer maps the output of the global average pooling layer to the final classification space.

[0145] In this embodiment of the invention, the 7×7 convolutional layer uses a 7×7 convolutional kernel with a stride of 2 and 64 channels; the max pooling layer uses a 3×3 pooling kernel with a stride of 2; the 3×3 convolutional layer uses a 3×3 convolutional kernel with a stride of 2 and 64 channels; the first 1×1 convolutional layer uses a 1×1 convolutional kernel with a stride of 2 and 256 channels; and the second 1×1 convolutional layer uses a 1×1 convolutional kernel with a stride of 2.

[0146] It should be noted that, in the appendix Figure 3 In this context, Conv7×7 represents a 7×7 convolutional layer, BN represents a BN processing unit, ReLU represents a ReLU processing unit, MaxPool represents a max pooling layer, AvgPool represents a global average pooling layer, FC represents a fully connected layer, Conv3×3 represents a 3×3 convolutional layer, Conv1×1 represents a 1×1 convolutional layer, and ⊕ represents an addition unit.

[0147] Furthermore, in this embodiment of the invention, in step 4, the deep learning network model is trained in the following manner:

[0148] The deep learning network model is trained by using the two-dimensional color image data corresponding to the fault diagnosis data as input and the leaf spring suspension system state in the fault diagnosis data as output.

[0149] In this embodiment of the invention, the deep learning network model is trained by using two-dimensional color image data corresponding to the fault diagnosis data as input and the leaf spring suspension system state in the fault diagnosis data as output. The training of the deep learning network model further includes the following steps:

[0150] Step 401: Input the two-dimensional color image data corresponding to multiple fault diagnosis data into the deep learning network model in sequence to obtain the leaf spring suspension system state prediction result corresponding to each fault diagnosis data output by the deep learning network model.

[0151] Step 402: Calculate the loss function based on the leaf spring suspension system status in the fault diagnosis data and the leaf spring suspension system status prediction results corresponding to the fault diagnosis data;

[0152] Step 403: Determine whether the preset training stopping condition has been met. If yes, use the current deep learning network model as the deep learning network model that has completed training. If no, update the parameters of the deep learning network model using the loss function and return to step 401.

[0153] In this embodiment of the invention, before training begins, the parameters of each layer of the deep learning network model are initialized parameters, and during training, the parameters of each layer of the deep learning network model are continuously updated.

[0154] In this embodiment of the invention, the loss function is the cross-entropy loss function.

[0155] In this embodiment of the invention, the training stopping condition is set according to the actual situation, such as the number of training iterations reaching a set number or the optimization index reaching a set threshold. The optimization index can be the loss function value mentioned above.

[0156] In this embodiment of the invention, gradient descent is used to iteratively update the parameters of the deep learning network model.

[0157] Specifically, the parameters of the deep learning network model are updated using the following formula:

[0158]

[0159] Where, θ t+1 Let θ represent the parameters of the deep learning network model at the (t+1)th iteration. t Let η represent the parameters of the deep learning network model at the t-th iteration, η represent the learning rate, L represent the loss function, and θ represent the parameters of the deep learning network model. The learning rate is preset and used to control the speed at which the parameters of the deep learning network model are updated.

[0160] refer to Figure 4 Secondly, embodiments of the present invention provide a leaf spring suspension fault diagnosis system based on multi-sensor data and deep learning, the system comprising:

[0161] The fault diagnosis dataset acquisition unit is used to acquire the fault diagnosis dataset of the leaf spring suspension system. The fault diagnosis dataset of the leaf spring suspension system includes multiple fault diagnosis data, including axle acceleration signal data, frame acceleration signal data, suspension displacement signal data, and corresponding leaf spring suspension system status. The leaf spring suspension system status includes: normal status, leaf spring suspension U-bolt loose status, leaf spring performance degradation status, and shock absorber performance degradation status.

[0162] The data processing unit is used to process each signal data in the fault diagnosis dataset of the leaf spring suspension system based on the joint processing method of empirical mode decomposition and wavelet denoising, and extract useful information from the signal data.

[0163] The data conversion unit is used to convert the two acceleration signal data and one displacement signal data corresponding to each fault diagnosis data into two-dimensional color image data after data processing.

[0164] The deep learning network model generation unit is used to train a pre-built deep learning network model using the fault diagnosis dataset of the leaf spring suspension system and the two-dimensional color image data corresponding to the fault diagnosis data. The input of the deep learning network model is the two-dimensional color image data corresponding to the fault diagnosis data, and the output of the deep learning network model is the state of the leaf spring suspension system corresponding to the fault diagnosis data.

[0165] The data acquisition unit is used to collect axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed in real time.

[0166] The fault diagnosis unit is used to perform leaf spring suspension fault diagnosis based on the real-time collected axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed, using a trained deep learning network model.

[0167] The leaf spring suspension fault diagnosis system based on multi-sensor data and deep learning provided in this embodiment of the invention consists of devices corresponding to the steps of the above method. It can realize all the processes of the leaf spring suspension fault diagnosis method based on multi-sensor data and deep learning described in any of the above embodiments. The specific working principle, function and technical effect of each unit are the same as those of the leaf spring suspension fault diagnosis method based on multi-sensor data and deep learning described in the above embodiments, and will not be repeated here.

[0168] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Additionally, the terms "front," "back," "left," "right," "upper," and "lower" in this document refer to the placement shown in the accompanying drawings.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fault diagnosis of leaf spring suspension based on multi-sensor data and deep learning, characterized in that, include: A fault diagnosis dataset for the leaf spring suspension system is obtained. The fault diagnosis dataset includes multiple fault diagnosis data, including axle acceleration signal data, frame acceleration signal data, suspension displacement signal data, and corresponding leaf spring suspension system status. The leaf spring suspension system status includes: normal status, loose leaf spring U-bolt status, leaf spring performance degradation status, and shock absorber performance degradation status. Based on the joint processing method of empirical mode decomposition and wavelet denoising, data processing is performed on each signal data in the fault diagnosis dataset of the leaf spring suspension system to extract useful information from the signal data. The two acceleration signal data and one displacement signal data corresponding to each of the fault diagnosis data are converted into two-dimensional color image data after data processing. A pre-constructed deep learning network model is trained using the fault diagnosis dataset of the leaf spring suspension system and the two-dimensional color image data corresponding to the fault diagnosis data. The input of the deep learning network model is the two-dimensional color image data corresponding to the fault diagnosis data, and the output of the deep learning network model is the state of the leaf spring suspension system corresponding to the fault diagnosis data. Real-time acquisition of axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed, and using the trained deep learning network model to perform leaf spring suspension fault diagnosis.

2. The leaf spring suspension fault diagnosis method based on multi-sensor data and deep learning according to claim 1, characterized in that, A fault diagnosis dataset for the leaf spring suspension system was obtained through real vehicle road tests and simulation tests.

3. The leaf spring suspension fault diagnosis method based on multi-sensor data and deep learning according to claim 2, characterized in that, The fault diagnosis dataset for the leaf spring suspension system is obtained through real vehicle road tests and simulation tests, and further includes the following steps: Step 101: Install accelerometers at the axle and the frame respectively, and install a laser displacement sensor between the axle and the frame; Step 102: Conduct a real-vehicle road test with the leaf spring suspension system in normal condition, and collect the acceleration signals from the two accelerometers and the displacement signals from the laser displacement sensor in real time to obtain multiple fault diagnosis data corresponding to the normal condition; and conduct a real-vehicle road test with the leaf spring suspension system in the leaf spring suspension U-bolt loose condition, and collect the acceleration signals from the two accelerometers and the displacement signals from the laser displacement sensor in real time to obtain multiple fault diagnosis data corresponding to the leaf spring suspension U-bolt loose condition. Step 103: Based on the data obtained from real vehicle road tests conducted under normal conditions of the leaf spring suspension system, a high-precision rigid-flexible coupling simulation model is established using RecurDyn software. The stiffness parameters of the leaf spring are dynamically adjusted using the STEP function in RecurDyn software to simulate the performance degradation process of the leaf spring. Simulation tests are conducted based on the adjusted parameters to obtain fault diagnosis data corresponding to multiple leaf spring performance degradation states. Similarly, the damping parameters of the shock absorber are dynamically adjusted using the STEP function to simulate the performance degradation process of the shock absorber. Simulation tests are conducted based on the adjusted parameters to obtain fault diagnosis data corresponding to multiple shock absorber performance degradation states.

4. The leaf spring suspension fault diagnosis method based on multi-sensor data and deep learning according to claim 1, characterized in that, Based on the joint processing method of empirical mode decomposition and wavelet denoising, the signal data is processed through the following steps: Step 201: Identify all local maxima and local minima in the signal, and generate the upper and lower envelopes of the signal respectively using cubic spline interpolation. Step 202: Calculate the average envelope based on the upper and lower envelope lines; Step 203: Extract signal components based on the average envelope; Step 204: Determine whether the signal component is an intrinsic mode function. If yes, the extracted signal component is used as an intrinsic mode function and output. Calculate the remaining signal and execute steps 201-204 based on the remaining signal. If no, output the remaining signal and proceed to the next step. Step 205: The extracted high-frequency intrinsic mode functions are decomposed using discrete wavelet transform to obtain the corresponding approximation coefficients and detail coefficients; Step 206: Based on the signal length of the high-frequency intrinsic mode function, set a threshold to distinguish between signal and noise; Step 207: Process the detail coefficients corresponding to the high-frequency intrinsic mode functions using the set threshold to obtain the threshold-processed detail coefficients; Step 208: Based on the approximation coefficients corresponding to the high-frequency intrinsic mode functions and the detail coefficients after thresholding, the denoised high-frequency intrinsic mode functions are reconstructed by wavelet reconstruction. Step 209: Add all the denoised high-frequency intrinsic mode functions, the unprocessed low-frequency intrinsic mode functions, and the final output residual signal to reconstruct the signal and obtain the reconstructed signal.

5. The leaf spring suspension fault diagnosis method based on multi-sensor data and deep learning according to claim 1, characterized in that, Based on the Gram angle field method, the two acceleration signal data and one displacement signal data corresponding to each fault diagnosis data are converted into two-dimensional color image data after data processing.

6. The high ride comfort design method for a truck crane based on deep learning according to claim 5, characterized in that, The following steps convert the two acceleration signal data and one displacement signal data corresponding to the fault diagnosis data into two-dimensional color image data: Step 301: Based on the two acceleration signal data and one displacement signal data after data processing corresponding to the fault diagnosis data, obtain the corresponding axle acceleration sequence data, frame acceleration sequence data and suspension displacement sequence data, and normalize each sequence data to the [0,1] interval respectively; Step 302: Encode each value in the normalized sequence data as an angle cosine, encode the timestamp corresponding to each value in the normalized sequence data as a radius, and map each normalized sequence data to the polar coordinate system according to the angle cosine and the radius. Step 303: Calculate the cosine similarity between any two values ​​in each sequence data after normalization, and generate the GAF matrix corresponding to each sequence data based on the cosine similarity. Step 304: Visualize the GAF matrix corresponding to the axle acceleration sequence data, the GAF matrix corresponding to the frame acceleration sequence data, and the GAF matrix corresponding to the suspension displacement sequence data as grayscale images, and then map the grayscale images to the R, G, and B channels and stack them to generate two-dimensional color image data.

7. The high ride comfort design method for a truck crane based on deep learning according to claim 1, characterized in that, The deep learning network model includes: a 7×7 convolutional layer, a first BN processing unit, a first ReLU processing unit, a max pooling layer, a first residual block, a second residual block, a third residual block, a fourth residual block, a global average pooling layer, and a fully connected layer connected in sequence. The first residual block includes 3 BottleNeck structures, the second residual block includes 4 BottleNeck structures, the third residual block includes 6 BottleNeck structures, and the fourth residual block includes 3 BottleNeck structures. The BottleNeck structure includes: a second BN processing unit, a second ReLU processing unit, a 3×3 convolutional layer, a third BN processing unit, a third ReLU processing unit, a first 1×1 convolutional layer, a fourth BN processing unit, an addition unit, and a fourth ReLU processing unit connected in sequence; and a second 1×1 convolutional layer and a fifth BN processing unit connected between the input of the BottleNeck structure and the input of the addition unit. The input of the second 1×1 convolutional layer is connected to the input of the BottleNeck structure, the output of the second 1×1 convolutional layer is connected to the input of the fifth BN processing unit, and the output of the fifth BN processing unit is connected to the input of the addition unit.

8. The high ride comfort design method for a truck crane based on deep learning according to claim 1 or 7, characterized in that, The deep learning network model is trained using the following method: The deep learning network model is trained by using the two-dimensional color image data corresponding to the fault diagnosis data as the input and the leaf spring suspension system state in the fault diagnosis data as the output.

9. The high ride comfort design method for a truck crane based on deep learning according to claim 8, characterized in that, The deep learning network model is trained by using the two-dimensional color image data corresponding to the fault diagnosis data as input and the leaf spring suspension system state in the fault diagnosis data as output. The training process includes: Step 401: Input the two-dimensional color image data corresponding to the multiple fault diagnosis data into the deep learning network model in sequence to obtain the leaf spring suspension system state prediction result corresponding to each fault diagnosis data output by the deep learning network model. Step 402: Calculate the loss function based on the leaf spring suspension system status in the fault diagnosis data and the leaf spring suspension system status prediction result corresponding to the fault diagnosis data; Step 403: Determine whether the preset training stopping condition has been met. If yes, use the current deep learning network model as the deep learning network model that has completed training. If no, update the parameters of the deep learning network model using the loss function and return to step 401.

10. A leaf spring suspension fault diagnosis system based on multi-sensor data and deep learning, characterized in that, include: The fault diagnosis dataset acquisition unit is used to acquire the leaf spring suspension system fault diagnosis dataset. The leaf spring suspension system fault diagnosis dataset includes multiple fault diagnosis data, including axle acceleration signal data, frame acceleration signal data, suspension displacement signal data, and corresponding leaf spring suspension system status. The leaf spring suspension system status includes: normal status, leaf spring suspension U-bolt loosening status, leaf spring performance degradation status, and shock absorber performance degradation status. The data processing unit is used to process each signal data in the fault diagnosis dataset of the leaf spring suspension system based on the joint processing method of empirical mode decomposition and wavelet denoising, and extract useful information from the signal data. The data conversion unit is used to convert the two acceleration signal data and one displacement signal data corresponding to each fault diagnosis data into two-dimensional color image data after data processing. A deep learning network model generation unit is used to train a pre-constructed deep learning network model using the fault diagnosis dataset of the leaf spring suspension system and the two-dimensional color image data corresponding to the fault diagnosis data. The input of the deep learning network model is the two-dimensional color image data corresponding to the fault diagnosis data, and the output of the deep learning network model is the leaf spring suspension system state corresponding to the fault diagnosis data. The data acquisition unit is used to collect axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed in real time. The fault diagnosis unit is used to perform leaf spring suspension fault diagnosis based on the axle acceleration signal data, frame acceleration signal data, and suspension displacement signal data of the leaf spring suspension system to be diagnosed in real time, using the trained deep learning network model.

Images