A method for identifying differential anomalies in new energy vehicles
By collecting vibration signals from the differential housing and the drive motor connection end cover, a multi-scale relative fluctuation sequence and side-frequency feature matrix are constructed. Combined with a multi-source feature fusion convolutional neural network, the problem of low accuracy in differential anomaly identification is solved, and high-precision identification of early and subtle faults is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LUZHOU HAONENG DRIVETECH CO LTD
- Filing Date
- 2026-04-22
- Publication Date
- 2026-05-26
AI Technical Summary
Existing differential anomaly identification methods have low accuracy, especially in identifying early, minor faults.
Vibration signals from the differential housing and the drive motor connection end cover are collected. Positive and negative anomalies are screened through multi-scale relative fluctuation sequences, a side-frequency feature matrix is constructed, and these features are processed by a multi-source feature fusion convolutional neural network to identify anomalies.
It significantly improves the accuracy of differential anomaly identification, effectively identifies early and minor faults, and enhances the robustness and accuracy of identification.
Smart Images
Figure CN122087660A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of differential fault diagnosis technology, and specifically to a method for identifying differential anomalies in new energy vehicles. Background Technology
[0002] As a key component in the transmission system of new energy vehicles, the differential's operating status directly affects the vehicle's power, safety, and reliability. Under conditions of high-speed, high-torque output from the drive motor, the differential housing and its connection end cover to the drive motor are subjected to complex dynamic loads, which can easily induce faults such as abnormal gear meshing, bearing wear, and loosening of the fit. Therefore, accurate monitoring and anomaly identification of the differential's operating status are of great significance for ensuring safe vehicle operation and reducing maintenance costs.
[0003] Currently, vibration signal-based fault diagnosis methods for transmission systems are widely used. In existing technologies, a common method for differential anomaly identification involves placing vibration sensors on the differential housing and drive motor end cover to collect vibration signals during operation. By collecting the vibration signals from the differential housing, time-domain statistical features (such as root mean square value, kurtosis, and peak factor) or time-domain waveform features are extracted, and abnormal states are identified through threshold judgment or a classifier. However, this method extracts time-domain features from only a single dimension, which contains both fault components and interference components. Furthermore, its ability to identify early, subtle faults is limited, resulting in low identification accuracy. Summary of the Invention
[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides a method for identifying differential anomalies in new energy vehicles, which solves the problem of low identification accuracy in the prior art.
[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for identifying differential anomalies in new energy vehicles, comprising the following steps: S1. Collect vibration signals from the differential housing and the connection end cover between the drive motor and the differential assembly to obtain meshing vibration signals and drive vibration signals; S2. The meshing vibration signal and the driving vibration signal are synchronously divided into multiple non-overlapping sub-segments. The driving vibration interference component is removed from each sub-segment to obtain the interference-free signal sub-segment. S3. Obtain multi-scale relative fluctuation sequences from the interference-removed signal segments; S4. For multi-scale relative fluctuation sequences, screen for positive and negative outliers, and obtain positive and negative outlier sequences. S5. Perform envelope demodulation on the interference-removing signal segment, and obtain the modulation intensity, sideband attenuation rate and sideband peak factor based on the amplitude of each harmonic and its sideband, thus forming the sideband feature matrix. S6. A multi-source feature fusion convolutional neural network is used to process the side-frequency feature matrix, positive anomaly factor sequence, and negative anomaly factor sequence to obtain the differential anomaly result.
[0006] Furthermore, S2 includes the following sub-steps: S21. The meshing vibration signal and the driving vibration signal are synchronously divided into multiple non-overlapping sub-segments to obtain multiple meshing sub-segments and multiple driving sub-segments. S22. Obtain the mean amplitude of all sampling points for the meshing segment and the driving segment respectively, and obtain the mean value of the meshing segment and the mean value of the driving segment. S23. Based on the ratio of the average value of the meshing section to the average value of the driving section, perform amplitude matching on the corresponding driving sub-segment to obtain the amplitude-matched driving sub-segment. S24. Subtract the amplitude of the meshing segment from the corresponding amplitude-matched drive segment at the sampling point to obtain the interference-removed signal segment.
[0007] Furthermore, S3 includes the following sub-steps: S31. Taking each sampling point in the interference removal signal sub-segment as the center, set a scale window and calculate the average amplitude of each sampling point under the scale window. S32. Subtract the amplitude of each sampling point in the interference-removed signal segment from the mean amplitude under the corresponding scale window to obtain the fluctuation value under the corresponding scale. S33. Take the fluctuation values of the same scale in the same interference removal signal segment as elements to form a relative fluctuation sequence of the corresponding scale.
[0008] Furthermore, S4 includes the following sub-steps: S41. Set positive and negative thresholds. In the relative fluctuation sequence at each scale, when the fluctuation value of a sampling point is greater than the positive threshold, mark the corresponding sampling point as a positive outlier. S42. When the fluctuation value of a sampling point is less than the negative threshold, the corresponding sampling point is marked as a negative outlier. S43. Add up the fluctuation values of each positive anomaly point in the relative fluctuation sequence belonging to each scale to obtain the positive anomaly intensity of the corresponding scale. S44. Sum the fluctuation values of each negative anomaly point in the relative fluctuation sequence belonging to each scale to obtain the negative anomaly intensity of the corresponding scale. S45. Take the average value of the positive anomaly intensity at multiple scales and perform normalization mapping to obtain positive anomaly factors. Arrange the positive anomaly factors in chronological order to obtain a positive anomaly factor sequence. S46. Take the average of the negative anomaly intensities at multiple scales and perform normalization mapping to obtain negative anomaly factors. Arrange the negative anomaly factors in chronological order to obtain a negative anomaly factor sequence.
[0009] Furthermore, the formula for calculating the positive outlier factor in S45 is as follows: ,in, For the first One positive outlier, In scale window Next A positive anomaly intensity, The number of scale windows, For scale window, A set of scale windows; The formula for calculating the negative outlier factor in S46 is as follows: ,in, For the first One negative outlier, In scale window Next There are 1 negative anomaly intensity, where | is the absolute value.
[0010] Furthermore, S5 includes the following sub-steps: S51. Perform Hilbert envelope demodulation on each interference-removing signal segment to obtain the envelope signal; S52. Perform a Fourier transform on the envelope signal to obtain the envelope spectrum; S53. Calculate the differential engagement frequency based on the number of teeth on the differential gear and the speed of the drive motor; S54. Obtain multiple harmonics based on the differential engagement frequency; S55. Set the sideband frequency for each harmonic; S56. Based on the amplitudes corresponding to multiple harmonics and the amplitudes corresponding to sideband frequencies in the envelope spectrum, obtain the modulation intensity, sideband attenuation rate, and sideband peak factor to form a sideband feature matrix.
[0011] Furthermore, the process of obtaining the modulation intensity in S56 includes: extracting the amplitude corresponding to all sideband frequencies in the envelope spectrum, adding the amplitude corresponding to all sideband frequencies to obtain the total amplitude of the full-order sideband, extracting the amplitude of multiple harmonics from the envelope spectrum, adding the amplitude of all harmonics to obtain the total harmonic amplitude, and using the ratio of the total amplitude of the full-order sideband to the total harmonic amplitude as the modulation intensity; The process of obtaining the sideband attenuation rate in S56 includes: extracting the amplitude corresponding to each order sideband frequency of each harmonic in the envelope spectrum; adding the sideband amplitudes of all harmonics of the same order to obtain the total amplitude of each order sideband; adding the ratio of the total amplitude of the second order sideband to the total amplitude of the first order sideband to the ratio of the total amplitude of the third order sideband to the total amplitude of the second order sideband, and taking the average of the summed results to obtain the sideband attenuation rate; The process of obtaining the sideband peak factor in S56 includes: extracting the amplitude corresponding to each order sideband frequency of each harmonic in the envelope spectrum, taking the maximum value among all sideband amplitudes as the sideband peak; calculating the root mean square value of all sideband amplitudes; and taking the ratio of the sideband peak value to the root mean square value as the sideband peak factor.
[0012] Furthermore, the multi-source feature fusion convolutional neural network in S6 includes: The input of the first convolutional layer is used to input the sideband feature matrix; the input of the second convolutional layer is used to input the positive anomaly factor sequence; the input of the third convolutional layer is used to input the negative anomaly factor sequence. The input of adder A1 is connected to the output of the second convolutional layer and the output of the third convolutional layer, respectively; the first input of the spatial-channel dual-domain feature extraction module is connected to the output of the first convolutional layer, its second input is connected to the output of adder A1, and its output is connected to the input of the fourth convolutional layer; the input of the fully connected layer is connected to the output of the fourth convolutional layer, and its output serves as the output of the multi-source feature fusion convolutional neural network.
[0013] Furthermore, S6 includes the following sub-steps: S61. The sideband feature matrix is processed by the first convolutional layer to obtain the sideband temporal features; S62. The positive outlier sequence is processed by the second convolutional layer to obtain positive time series features; S63. The negative anomaly sequence is processed by the third convolutional layer to obtain negative temporal features; S64. Adder A1 is used to add the positive and negative time series features element by element to obtain the anomaly fusion time series features; S65. The spatial-channel dual-domain feature extraction module is used to extract multi-spatial features and multi-channel features from the side-frequency temporal features and the anomaly fusion temporal features respectively, and then fused and spliced to obtain spatial-channel dual-domain features. S66. The fourth convolutional layer is used to process the spatial-channel dual-domain features to obtain compressed features; S67. The compression characteristics are processed using a fully connected layer to obtain abnormal results of the differential.
[0014] Furthermore, the spatial-channel dual-domain feature extraction module includes: The input end of the first multi-spatial feature extraction unit is connected to the input end of the first multi-channel feature extraction unit and serves as the first input end of the spatial-channel dual-domain feature extraction module; the input end of the second multi-spatial feature extraction unit is connected to the input end of the second multi-channel feature extraction unit and serves as the second input end of the spatial-channel dual-domain feature extraction module. The input terminals of adder A2 are connected to the output terminals of the first multi-spatial feature extraction unit and the second multi-spatial feature extraction unit, respectively; the input terminals of adder A3 are connected to the output terminals of the first multi-channel feature extraction unit and the second multi-channel feature extraction unit, respectively. The input of the first Concat layer is connected to the output of adder A2 and adder A3 respectively, and its output serves as the output of the spatial-channel dual-domain feature extraction module.
[0015] The beneficial effects of this invention are as follows: 1. This invention simultaneously acquires vibration signals from the differential housing and the drive motor connection end cover, and eliminates drive vibration interference components at the sub-segment level, effectively removing signal fluctuations caused by non-fault factors. Compared to existing methods that rely solely on single-dimensional time-domain features, this invention fundamentally reduces the impact of operating condition fluctuations and noise interference on the identification results, significantly improving the accuracy of differential anomaly identification.
[0016] 2. This invention, on the one hand, constructs a multi-scale relative fluctuation sequence from the interference-removing signal segment and filters positive and negative outliers to capture subtle dynamic changes in the time domain; on the other hand, it extracts features such as modulation intensity, sideband attenuation rate, and sideband peak factor through envelope demodulation to construct a sideband feature matrix, characterizing fault modulation characteristics in the frequency domain. This multi-dimensional, multi-scale feature extraction method enables this invention to effectively identify early, subtle faults, overcoming the shortcomings of existing technologies in identifying early faults.
[0017] 3. This invention employs a multi-source feature fusion convolutional neural network to collaboratively process the side-frequency feature matrix, positive anomaly sequence, and negative anomaly sequence. Through automatic fusion and high-order abstraction of heterogeneous features, it fully leverages the complementary advantages between time-domain and frequency-domain multi-source information, further enhancing the robustness and accuracy of classification and recognition. Attached Figure Description
[0018] Figure 1 A flowchart of a method for identifying differential anomalies in new energy vehicles; Figure 2 This is a schematic diagram of the structure of a multi-source feature fusion convolutional neural network; Figure 3 This is a schematic diagram of the spatial-channel dual-domain feature extraction module; Figure 4 This is a schematic diagram of the structure of the first multi-space feature extraction unit and the second multi-space feature extraction unit; Figure 5 This is a schematic diagram of the structure of the first multi-channel feature extraction unit and the second multi-channel feature extraction unit. Detailed Implementation
[0019] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0020] like Figure 1 As shown, a method for identifying differential anomalies in new energy vehicles includes the following steps: S1. Collect vibration signals from the differential housing and the connection end cover between the drive motor and the differential assembly to obtain meshing vibration signals and drive vibration signals; S2. The meshing vibration signal and the driving vibration signal are synchronously divided into multiple non-overlapping sub-segments. The driving vibration interference component is removed from each sub-segment to obtain the interference-free signal sub-segment. S3. Obtain multi-scale relative fluctuation sequences from the interference-removed signal segments; S4. For multi-scale relative fluctuation sequences, screen for positive and negative outliers, and obtain positive and negative outlier sequences. S5. Perform envelope demodulation on the interference-removing signal segment, and obtain the modulation intensity, sideband attenuation rate and sideband peak factor based on the amplitude of each harmonic and its sideband, thus forming the sideband feature matrix. S6. A multi-source feature fusion convolutional neural network is used to process the side-frequency feature matrix, positive anomaly factor sequence, and negative anomaly factor sequence to obtain the differential anomaly result.
[0021] Vibration signals from the differential housing are collected to obtain meshing vibration signals, and vibration signals from the connection end cover between the drive motor and the differential assembly are collected to obtain drive vibration signals.
[0022] In this embodiment, the vibration signal sampling rate is set to 12800Hz. The meshing vibration signal and the driving vibration signal are synchronously divided into K non-overlapping segments, each containing 2048 sampling points, corresponding to a duration of approximately 0.16 seconds. When the total acquisition time is 30 seconds, K is set to 187.
[0023] In this embodiment, S2 includes the following sub-steps: S21. The meshing vibration signal and the driving vibration signal are synchronously divided into multiple non-overlapping sub-segments to obtain multiple meshing sub-segments and multiple driving sub-segments. S22. Obtain the mean amplitude of all sampling points for both the meshing segment and the driving segment, to obtain the mean amplitude of the meshing segment and the mean amplitude of the driving segment respectively: , ,in, For the first The average value of the meshing segments of each meshing segment For the first The first meshing segment The amplitude of each sampling point The number of sampling points in the sub-segment. For the first The average value of each driver segment. For the first The first driver segment The amplitude of each sampling point; S23. Based on the ratio of the average value of the meshing section to the average value of the driving section, perform amplitude matching on the corresponding driving sub-segment to obtain the amplitude-matched driving sub-segment: ,in, For the first The first amplitude-matched driving segment The amplitude of each sampling point; S24. Subtract the amplitude of the meshing segment from that of the corresponding amplitude-matched drive segment at each sampling point to obtain the interference-removed signal segment: ,in, For the first The first segment of the interference signal sub-segment The amplitude of each sampling point.
[0024] This invention obtains a de-interference signal segment by synchronous segmentation, segment mean matching, and subtracting the amplitude of the meshing segment from the amplitude of the drive segment after amplitude matching at sampling points. This eliminates interference components related to the vibration of the drive end cover, such as motor electromagnetic excitation and sudden speed changes, while retaining the fault signal components of the differential itself.
[0025] In this embodiment, S3 includes the following sub-steps: S31. Taking each sampling point in the interference-removed signal segment as the center, set a scale window and calculate the average amplitude of each sampling point under the scale window: ,in, For the first The first segment of the interference signal sub-segment Each sampling point in the scale window The mean amplitude below, For the first The first segment of the interference signal sub-segment The amplitude of each sampling point It is a positive integer; S32. Subtract the amplitude of each sampling point in the interference-removed signal segment from the mean amplitude under the corresponding scale window to obtain the fluctuation value under the corresponding scale: ,in, For the first The first segment of the interference signal sub-segment Each sampling point in the scale window The fluctuation value below, For the first The first segment of the interference signal sub-segment The amplitude of each sampling point For the first The first segment of the interference signal sub-segment Each sampling point in the scale window The mean amplitude below; S33. Take the fluctuation values of the same scale in the same interference removal signal segment as elements to form a relative fluctuation sequence of the corresponding scale.
[0026] In this embodiment, the scale window set is as follows: In S31, for the endpoints and ends of the interference-removing signal segment, when the scale window cannot satisfy the requirement of a symmetrical neighborhood when the sampling point is the center, the sample is directly taken from the neighborhood of that point. sampling points (i.e., continuous) Calculate the mean of (points).
[0027] This invention first sets up multiple scale windows centered on each sampling point in the interference-removing signal segment, calculating the mean amplitude of each sampling point within its neighborhood. This operation adaptively acquires the local background level of the signal. Second, the original amplitude is subtracted from the local mean to obtain the fluctuation value at that scale. This difference effectively eliminates the slow-changing trend and local mean components of the signal, highlighting the instantaneous changes of the signal within a local range. By setting multiple scale windows, this method can simultaneously capture signal fluctuation characteristics at different time scales: small-scale windows are sensitive to local impacts and transient spikes, which is beneficial for identifying high-frequency impact components caused by early faults such as pitting and cracks; large-scale windows reflect the slow-changing trend and envelope fluctuations of the signal, which helps identify low-frequency modulation characteristics caused by gear wear and loose fit.
[0028] In this embodiment, S4 includes the following sub-steps: S41. Set positive and negative thresholds. In the relative fluctuation sequence at each scale, when the fluctuation value of a sampling point is greater than the positive threshold, mark the corresponding sampling point as a positive outlier. S42. When the fluctuation value of a sampling point is less than the negative threshold, the corresponding sampling point is marked as a negative outlier. S43. Sum the fluctuation values of each positive anomaly point in the relative fluctuation sequence belonging to each scale to obtain the positive anomaly intensity at the corresponding scale: ,in, In scale window Next A positive anomaly intensity, In scale window Next In the relative fluctuation sequence, the th The fluctuation value of each positive outlier. This represents the number of positive outliers. S44. Sum the fluctuation values of each negative anomaly point in the relative fluctuation sequence belonging to each scale to obtain the negative anomaly intensity at the corresponding scale: ,in, In scale window Next One negative anomaly intensity, In scale window Next In the relative fluctuation sequence, the th The fluctuation value of each negative outlier. The number of negative outliers; S45. Take the average value of the positive anomaly intensity at multiple scales and perform normalization mapping to obtain positive anomaly factors. Arrange the positive anomaly factors in chronological order to obtain a positive anomaly factor sequence. S46. Take the average of the negative anomaly intensities at multiple scales and perform normalization mapping to obtain negative anomaly factors. Arrange the negative anomaly factors in chronological order to obtain a negative anomaly factor sequence.
[0029] In this embodiment, the formula for calculating the positive anomaly factor in S45 is: ,in, For the first One positive outlier, In scale window Next A positive anomaly intensity, The number of scale windows, For scale windows; The formula for calculating the negative outlier factor in S46 is as follows: ,in, For the first One negative outlier, In scale window Next There are 1 negative anomaly intensity, where | is the absolute value.
[0030] This invention filters out outliers by setting positive and negative thresholds, identifying excessively large and small fluctuation values. The fluctuation values of all positive outliers within the same relative fluctuation sequence are summed to obtain the positive anomaly intensity, reflecting the accumulated energy of positive impacts in the sequence. Similarly, the fluctuation values of all negative outliers within the same relative fluctuation sequence are summed to obtain the negative anomaly intensity, reflecting the accumulated energy of negative impacts in the sequence. The positive anomaly intensities across multiple scales are summed and averaged to represent the overall level of positive impact energy of the sub-segment at different time scales. The negative anomaly intensities across multiple scales are summed and averaged to represent the overall level of negative impact energy of the sub-segment at different time scales. Finally, a nonlinear normalization mapping is performed to obtain positive and negative anomaly factors in the range of 0 to 1.
[0031] In this embodiment, under fault-free (normal) conditions, a segment of the de-interference signal is acquired, and the standard deviation of its relative fluctuation sequence at each scale is calculated. The positive threshold for each scale is set to 3 times the standard deviation at the corresponding scale, and the negative threshold for each scale is set to -3 times the standard deviation at the corresponding scale. The specific values of the positive and negative thresholds can be adjusted according to experiments and are not limited to the values specified in this embodiment.
[0032] Each sub-segment corresponds to one positive anomaly factor and one negative anomaly factor. Therefore, the positive anomaly factors are arranged in chronological order to obtain a positive anomaly factor sequence, and the negative anomaly factors are arranged in chronological order to obtain a negative anomaly factor sequence.
[0033] In this embodiment, S5 includes the following sub-steps: S51. Perform Hilbert envelope demodulation on each interference-removing signal segment to obtain the envelope signal; S52. Perform a Fourier transform on the envelope signal to obtain the envelope spectrum; S53. Calculate the differential engagement frequency based on the number of teeth on the differential gears and the speed of the drive motor: ,in, The differential engagement frequency, The number of teeth on the differential gear. This refers to the speed of the drive motor; S54. Obtain multiple harmonics based on the differential engagement frequency: , ,in, For the first Second harmonic The integer is a positive integer, taking the values 1, 2, or 3. S55. Set the sideband frequency for each harmonic: and , This refers to the numbering of the sideband order. Take 1, 2, 3 in sequence, corresponding to and , and , and , The differential output shaft rotational frequency; S56. Based on the amplitudes corresponding to multiple harmonics and the amplitudes corresponding to sideband frequencies in the envelope spectrum, obtain the modulation intensity, sideband attenuation rate, and sideband peak factor to form a sideband feature matrix.
[0034] In this embodiment, the process of obtaining the modulation intensity in S56 includes: extracting the amplitude corresponding to all sideband frequencies in the envelope spectrum to obtain the sideband amplitude; summing all sideband amplitudes to obtain the total amplitude of the full-order sidebands; extracting the amplitudes of multiple harmonics from the envelope spectrum; summing all harmonic amplitudes to obtain the total harmonic amplitude; and using the ratio of the total amplitude of the full-order sidebands to the total harmonic amplitude as the modulation intensity. ,in, For modulation intensity, For the first The amplitude of the second harmonic. For the denominator parameter, For the first Second harmonic, the first Amplitude of the right sideband of the first order. For the first Second harmonic, the first Amplitude of the left sideband of the order; The process of obtaining the sideband attenuation rate in S56 includes: extracting the amplitude corresponding to each order sideband frequency of each harmonic from the envelope spectrum; summing the sideband amplitudes of all harmonics of the same order to obtain the total amplitude of each order sideband; adding the ratio of the total amplitude of the second order sideband to the total amplitude of the first order sideband to the ratio of the total amplitude of the third order sideband to the total amplitude of the second order sideband, and averaging the sums to obtain the sideband attenuation rate. , ,in, This refers to the sideband attenuation rate. For the first Total amplitude of the edge band, For the first Total amplitude of the edge band; The process of obtaining the sideband peak factor in S56 includes: extracting the amplitude corresponding to each order sideband frequency of each harmonic from the envelope spectrum, taking the maximum value among all sideband amplitudes as the sideband peak; calculating the root mean square value of all sideband amplitudes; and taking the ratio of the sideband peak value to the root mean square value as the sideband peak factor. , ,in, For sideband peak factor, To take all The maximum value in the middle.
[0035] This invention first performs Hilbert envelope demodulation and Fourier transform on each interference-removing signal segment to obtain the envelope spectrum. This processing can effectively extract the envelope component in the modulated signal, highlighting the periodic impact characteristics caused by the fault. Then, the meshing frequency is calculated based on the number of teeth of the differential gear and the speed of the drive motor, and the fundamental frequency, second harmonic, and third harmonic are obtained as harmonics for analysis. At the same time, the third-order sideband amplitude of each harmonic is extracted using the differential output shaft rotation frequency as the sideband interval. Based on this, three frequency domain characteristics are calculated: modulation intensity, i.e., the ratio of the total amplitude of all sidebands to the total amplitude of the harmonics, reflecting the overall strength of the modulation effect; sideband attenuation rate, i.e., the average of the ratios of the total amplitudes of adjacent order sidebands, reflecting the attenuation rate of sideband energy with increasing order; and sideband peak factor, i.e., the ratio of the maximum amplitude of the sideband to the root mean square value, reflecting whether there are prominent peak components in the sideband.
[0036] like Figure 2 As shown, the multi-source feature fusion convolutional neural network in S6 includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, an adder A1, a spatial-channel dual-domain feature extraction module, a fourth convolutional layer, and a fully connected layer; The input of the first convolutional layer is used to input the sideband feature matrix; the input of the second convolutional layer is used to input the positive anomaly factor sequence; the input of the third convolutional layer is used to input the negative anomaly factor sequence. The input of adder A1 is connected to the output of the second convolutional layer and the output of the third convolutional layer, respectively; the first input of the spatial-channel dual-domain feature extraction module is connected to the output of the first convolutional layer, its second input is connected to the output of adder A1, and its output is connected to the input of the fourth convolutional layer; the input of the fully connected layer is connected to the output of the fourth convolutional layer, and its output serves as the output of the multi-source feature fusion convolutional neural network.
[0037] S6 includes the following steps: S61. The sideband feature matrix is processed by the first convolutional layer to obtain the sideband temporal features; S62. The positive outlier sequence is processed by the second convolutional layer to obtain positive time series features; S63. The negative anomaly sequence is processed by the third convolutional layer to obtain negative temporal features; S64. Adder A1 is used to add the positive and negative time series features element by element to obtain the anomaly fusion time series features; S65. The spatial-channel dual-domain feature extraction module is used to extract multi-spatial features and multi-channel features from the side-frequency temporal features and the anomaly fusion temporal features respectively, and then fused and spliced to obtain spatial-channel dual-domain features. S66. The fourth convolutional layer is used to process the spatial-channel dual-domain features to obtain compressed features; S67. The compression characteristics are processed using a fully connected layer to obtain abnormal results of the differential.
[0038] The sideband feature matrix has a size of 3×K, where K is the number of sub-segments. The first convolutional layer uses a 3×1 kernel with 1 output channel, performing convolutional fusion on the three features of each sub-segment along the column direction, outputting a sideband temporal feature of size 1×K. The second and third convolutional layers have 1×1 kernels with 1 output channel, and the positive and negative temporal features have a size of 1×K.
[0039] The first convolutional layer of this invention performs convolutional fusion on the side-frequency feature matrix to extract frequency-domain side-frequency time-series features. The second and third convolutional layers perform feature transformations on the positive and negative anomaly factor sequences, respectively, to obtain positive and negative time-series features. Adder A1 adds the positive and negative time-series features element-wise to obtain anomaly fusion time-series features, achieving complementary enhancement of positive and negative time-domain anomaly information. The spatial-channel dual-domain feature extraction module performs parallel extraction and fusion of multi-spatial and multi-channel features on the side-frequency time-series features and anomaly fusion time-series features, generating spatial-channel dual-domain features, enabling the network to simultaneously learn the deep correlation between frequency-domain modulation information and time-domain impact information. The fourth convolutional layer compresses and reduces the dimensionality of the fused dual-domain features to reduce the number of parameters. The fully connected layer outputs the final differential anomaly identification result. This invention fully utilizes the complementarity between frequency-domain modulation features and time-domain positive and negative impact features, significantly improving the accuracy and robustness of differential anomaly identification compared to single-feature or simple splicing identification methods.
[0040] like Figure 3 As shown, the spatial-channel dual-domain feature extraction module includes: a first multi-spatial feature extraction unit, a first multi-channel feature extraction unit, a second multi-spatial feature extraction unit, a second multi-channel feature extraction unit, an adder A2, an adder A3, and a first Concat layer; The input end of the first multi-spatial feature extraction unit is connected to the input end of the first multi-channel feature extraction unit and serves as the first input end of the spatial-channel dual-domain feature extraction module; the input end of the second multi-spatial feature extraction unit is connected to the input end of the second multi-channel feature extraction unit and serves as the second input end of the spatial-channel dual-domain feature extraction module. The input terminals of adder A2 are connected to the output terminals of the first multi-spatial feature extraction unit and the second multi-spatial feature extraction unit, respectively; the input terminals of adder A3 are connected to the output terminals of the first multi-channel feature extraction unit and the second multi-channel feature extraction unit, respectively. The input of the first Concat layer is connected to the output of adder A2 and adder A3 respectively, and its output serves as the output of the spatial-channel dual-domain feature extraction module.
[0041] like Figure 4 As shown, the first multi-spatial feature extraction unit and the second multi-spatial feature extraction unit have the same structure, both including: a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, and a second concat layer. The input of the fifth convolutional layer is connected to the input of the sixth convolutional layer and the input of the seventh convolutional layer, respectively. The input of the second concat layer is connected to the output of the fifth convolutional layer, the output of the sixth convolutional layer, and the output of the seventh convolutional layer, respectively. Its output serves as the output of the first multi-spatial feature extraction unit and the second multi-spatial feature extraction unit.
[0042] The fifth convolutional layer has a 1×3 kernel and 32 output channels, the sixth convolutional layer has a 1×5 kernel and 32 output channels, and the seventh convolutional layer has a 1×7 kernel and 32 output channels. The second concat layer is used to concatenate the outputs of the fifth, sixth, and seventh convolutional layers to obtain 96×K.
[0043] This invention employs three parallel one-dimensional convolutional layers. Input features are simultaneously fed into these three layers for independent convolution operations, and the outputs of the three layers are concatenated along the channel dimension to obtain 96×K multi-spatial features. Convolutional kernels of different sizes can capture spatial features within different receptive fields: small kernels are sensitive to local details, facilitating the extraction of fine features from transient impacts; medium kernels balance local and neighborhood information, capturing medium-scale fluctuation patterns; and large kernels have a larger receptive field, reflecting the overall trend and gradual change components of the signal. Through parallel extraction of multi-scale convolutional kernels and channel concatenation, efficient extraction of multi-scale spatial features of the signal is achieved, enhancing the network's ability to express differential fault features.
[0044] like Figure 5As shown, the first multi-channel feature extraction unit and the second multi-channel feature extraction unit have the same structure, both including: an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, and a third concat layer; the input of the eighth convolutional layer serves as the input of the first multi-channel feature extraction unit and the second multi-channel feature extraction unit, and its output is connected to the input of the ninth convolutional layer and the input of the tenth convolutional layer, respectively. The input of the third concat layer is connected to the output of the ninth convolutional layer and the output of the tenth convolutional layer, respectively, and its output serves as the output of the first multi-channel feature extraction unit and the second multi-channel feature extraction unit.
[0045] The eighth convolutional layer has a 1×1 kernel and 64 output channels. The ninth convolutional layer has a 1×3 kernel and 32 output channels. The tenth convolutional layer has a 1×5 kernel and 32 output channels. The third concat layer is used to concatenate the outputs of the ninth and tenth convolutional layers to obtain 64×K.
[0046] This invention first employs an eighth convolutional layer to perform channel transformation and dimensionality upscaling on the input features, expanding the number of input channels to 64 and enhancing the expressive power of the features. Subsequently, the dimensionality-upgraded features are input into the ninth and tenth convolutional layers for parallel convolution operations, with each convolutional layer outputting 32 channels. Finally, the outputs of both layers are concatenated along the channel dimension to obtain multi-channel features. The 1×1 convolution, through linear combination and non-linear activation, achieves cross-channel information interaction and feature recombination. Convolutional kernels of different sizes (1×3 and 1×5) can capture channel-related features under different receptive fields; small kernels extract local channel relationships, while large kernels extract a wider range of channel dependencies.
[0047] The first Concat layer is used to concatenate the outputs of adder A2 (96×K) and adder A3 (64×K) to obtain 160×K. The fourth convolutional layer has a kernel size of 1×1 and an output channel of 1, resulting in a 1×K compression feature. The fully connected layer obtains the differential anomaly result based on the compression feature.
[0048] Abnormal differential results include: normal, gear wear, bearing pitting, tooth surface scuffing, and loose fit.
[0049] This invention first synchronously acquires vibration signals from the differential housing and the drive motor connection end cover, and removes drive vibration interference components at the sub-segment level, effectively stripping interference components mixed in with the features from the signal source, thus solving the problem of fault components and interference components mixed in the time-domain features. Second, features are extracted from the de-interference signal sub-segments from both the time and frequency domains—in the time domain, a multi-scale relative fluctuation sequence is constructed and positive and negative anomalies are screened to capture early weak fluctuation features; in the frequency domain, modulation intensity, sideband attenuation rate, and sideband peak factor are extracted through envelope demodulation to characterize fault modulation characteristics—breaking through the limitations of single-dimensional feature extraction and enhancing the ability to perceive early weak faults. Finally, a multi-source feature fusion convolutional neural network is used to fuse the sideband feature matrix, positive anomaly factor sequence, and negative anomaly factor sequence, achieving automatic fusion and high-order abstraction of multi-source features, further improving the robustness and accuracy of the recognition results. In summary, this invention effectively overcomes the problem of low recognition accuracy in existing technologies, achieving high-precision and high-reliability recognition of differential abnormal states.
[0050] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for identifying differential anomalies in new energy vehicles, characterized in that, Includes the following steps: S1. Collect vibration signals from the differential housing and the connection end cover between the drive motor and the differential assembly to obtain meshing vibration signals and drive vibration signals; S2. The meshing vibration signal and the driving vibration signal are synchronously divided into multiple non-overlapping sub-segments. The driving vibration interference component is removed from each sub-segment to obtain the interference-free signal sub-segment. S3. Obtain multi-scale relative fluctuation sequences from the interference-removed signal segments; S4. For multi-scale relative fluctuation sequences, screen for positive and negative outliers, and obtain positive and negative outlier sequences. S5. Perform envelope demodulation on the interference-removing signal segment, and obtain the modulation intensity, sideband attenuation rate and sideband peak factor based on the amplitude of each harmonic and its sideband, thus forming the sideband feature matrix. S6. A multi-source feature fusion convolutional neural network is used to process the side-frequency feature matrix, positive anomaly factor sequence, and negative anomaly factor sequence to obtain the differential anomaly result.
2. The method for identifying differential anomalies in new energy vehicles according to claim 1, characterized in that, S2 includes the following steps: S21. The meshing vibration signal and the driving vibration signal are synchronously divided into multiple non-overlapping sub-segments to obtain multiple meshing sub-segments and multiple driving sub-segments. S22. Obtain the mean amplitude of all sampling points for the meshing segment and the driving segment respectively, and obtain the mean value of the meshing segment and the mean value of the driving segment. S23. Based on the ratio of the average value of the meshing section to the average value of the driving section, perform amplitude matching on the corresponding driving sub-segment to obtain the amplitude-matched driving sub-segment. S24. Subtract the amplitude of the meshing segment from the corresponding amplitude-matched drive segment at the sampling point to obtain the interference-removed signal segment.
3. The method for identifying differential anomalies in new energy vehicles according to claim 1, characterized in that, S3 includes the following steps: S31. Taking each sampling point in the interference removal signal sub-segment as the center, set a scale window and calculate the average amplitude of each sampling point under the scale window. S32. Subtract the amplitude of each sampling point in the interference-removed signal segment from the mean amplitude under the corresponding scale window to obtain the fluctuation value under the corresponding scale. S33. Take the fluctuation values of the same scale in the same interference removal signal segment as elements to form a relative fluctuation sequence of the corresponding scale.
4. The method for identifying differential anomalies in new energy vehicles according to claim 1, characterized in that, S4 includes the following steps: S41. Set positive and negative thresholds. In the relative fluctuation sequence at each scale, when the fluctuation value of a sampling point is greater than the positive threshold, mark the corresponding sampling point as a positive outlier. S42. When the fluctuation value of a sampling point is less than the negative threshold, the corresponding sampling point is marked as a negative outlier. S43. Add up the fluctuation values of each positive anomaly point in the relative fluctuation sequence belonging to each scale to obtain the positive anomaly intensity of the corresponding scale. S44. Sum the fluctuation values of each negative anomaly point in the relative fluctuation sequence belonging to each scale to obtain the negative anomaly intensity of the corresponding scale. S45. Take the average value of the positive anomaly intensity at multiple scales and perform normalization mapping to obtain positive anomaly factors. Arrange the positive anomaly factors in chronological order to obtain a positive anomaly factor sequence. S46. Take the average of the negative anomaly intensities at multiple scales and perform normalization mapping to obtain negative anomaly factors. Arrange the negative anomaly factors in chronological order to obtain a negative anomaly factor sequence.
5. The method for identifying differential anomalies in new energy vehicles according to claim 4, characterized in that, The formula for calculating the positive outlier factor in S45 is as follows: ,in, For the first One positive outlier, In scale window Next A positive anomaly intensity, The number of scale windows, For scale window, A set of scale windows; The formula for calculating the negative outlier factor in S46 is as follows: ,in, For the first One negative outlier, In scale window Next There are 1 negative anomaly intensity, where | is the absolute value.
6. The method for identifying differential anomalies in new energy vehicles according to claim 1, characterized in that, S5 includes the following steps: S51. Perform Hilbert envelope demodulation on each interference-removing signal segment to obtain the envelope signal; S52. Perform a Fourier transform on the envelope signal to obtain the envelope spectrum; S53. Calculate the differential engagement frequency based on the number of teeth on the differential gear and the speed of the drive motor; S54. Obtain multiple harmonics based on the differential engagement frequency; S55. Set the sideband frequency for each harmonic; S56. Based on the amplitudes corresponding to multiple harmonics and the amplitudes corresponding to sideband frequencies in the envelope spectrum, obtain the modulation intensity, sideband attenuation rate, and sideband peak factor to form a sideband feature matrix.
7. The method for identifying differential anomalies in new energy vehicles according to claim 6, characterized in that, The process of obtaining the modulation intensity in S56 includes: extracting the amplitude corresponding to all sideband frequencies in the envelope spectrum to obtain the sideband amplitude; adding all sideband amplitudes to obtain the total amplitude of the full-order sideband; extracting the amplitude of multiple harmonics from the envelope spectrum; adding all harmonic amplitudes to obtain the total harmonic amplitude; and using the ratio of the total amplitude of the full-order sideband to the total harmonic amplitude as the modulation intensity. The process of obtaining the sideband attenuation rate in S56 includes: adding the sideband amplitudes of all harmonics of the same order to obtain the total amplitude of each sideband; adding the ratio of the total amplitude of the second-order sideband to the total amplitude of the first-order sideband to the ratio of the total amplitude of the third-order sideband to the total amplitude of the second-order sideband, and taking the average of the summed results to obtain the sideband attenuation rate; The process of obtaining the sideband peak factor in S56 includes: taking the maximum value among all sideband amplitudes as the sideband peak value; calculating the root mean square value of all sideband amplitudes; and taking the ratio of the sideband peak value to the root mean square value as the sideband peak factor.
8. The method for identifying differential anomalies in new energy vehicles according to claim 1, characterized in that, The multi-source feature fusion convolutional neural network in S6 includes: The input of the first convolutional layer is used to input the sideband feature matrix; the input of the second convolutional layer is used to input the positive anomaly factor sequence; the input of the third convolutional layer is used to input the negative anomaly factor sequence. The input of adder A1 is connected to the output of the second convolutional layer and the output of the third convolutional layer, respectively; the first input of the spatial-channel dual-domain feature extraction module is connected to the output of the first convolutional layer, its second input is connected to the output of adder A1, and its output is connected to the input of the fourth convolutional layer; the input of the fully connected layer is connected to the output of the fourth convolutional layer, and its output serves as the output of the multi-source feature fusion convolutional neural network.
9. The method for identifying differential anomalies in new energy vehicles according to claim 8, characterized in that, S6 includes the following steps: S61. The sideband feature matrix is processed by the first convolutional layer to obtain the sideband temporal features; S62. The positive outlier sequence is processed by the second convolutional layer to obtain positive time series features; S63. The negative anomaly sequence is processed by the third convolutional layer to obtain negative temporal features; S64. Adder A1 is used to add the positive and negative time series features element by element to obtain the anomaly fusion time series features; S65. The spatial-channel dual-domain feature extraction module is used to extract multi-spatial features and multi-channel features from the side-frequency temporal features and the anomaly fusion temporal features respectively, and then fused and spliced to obtain spatial-channel dual-domain features. S66. The fourth convolutional layer is used to process the spatial-channel dual-domain features to obtain compressed features; S67. The compression characteristics are processed using a fully connected layer to obtain abnormal results of the differential.
10. The method for identifying differential anomalies in new energy vehicles according to claim 8 or 9, characterized in that, The spatial-channel dual-domain feature extraction module includes: The input end of the first multi-spatial feature extraction unit is connected to the input end of the first multi-channel feature extraction unit and serves as the first input end of the spatial-channel dual-domain feature extraction module; the input end of the second multi-spatial feature extraction unit is connected to the input end of the second multi-channel feature extraction unit and serves as the second input end of the spatial-channel dual-domain feature extraction module. The input terminals of adder A2 are connected to the output terminals of the first multi-spatial feature extraction unit and the second multi-spatial feature extraction unit, respectively; the input terminals of adder A3 are connected to the output terminals of the first multi-channel feature extraction unit and the second multi-channel feature extraction unit, respectively. The input of the first Concat layer is connected to the output of adder A2 and adder A3 respectively, and its output serves as the output of the spatial-channel dual-domain feature extraction module.