New energy automobile motor operation state diagnosis method

Abstract

The invention discloses a new energy automobile motor operation state diagnosis method, and belongs to the technical field of motor diagnosis. The method comprises the following steps: firstly, generating a vibration energy signal for a motor vibration signal, and obtaining short-scale, medium-scale and long-scale energy deviation signals through three windows with different lengths; segmenting the energy deviation signal of each scale, performing EMD decomposition, and separating out a long-term deviation component sub-signal and a local deviation fluctuation sub-signal; then respectively extracting a joint long-term deviation degree, a joint long-term fluctuation degree and a joint long-term skewness to form a long-term feature matrix, and extracting a joint local impact energy value, a joint local impact strength and a joint local impact form value to form a local feature matrix; and finally, processing the two types of normalized feature matrixes by using a double-branch fusion convolutional neural network, and outputting a motor operation state diagnosis result. According to the invention, the motor operation state characterization capability and the fault diagnosis precision are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motor diagnosis, and in particular to a new energy automobile motor operation state diagnosis method. BACKGROUND

[0002] As the power core, the new energy automobile driving motor works in complex conditions such as variable speed, variable load and strong impact for a long time. The vibration signal has typical nonlinear, non-stationary and strong interference characteristics. The state monitoring and fault diagnosis based on the vibration signal are crucial to ensure driving safety. Current motor operation state diagnosis methods generally use fixed time windows to extract vibration energy, combine empirical mode decomposition (EMD) to separate signal components, and use single features or conventional neural networks to complete fault recognition.

[0003] Such methods generally rely on single scale or fixed length analysis windows, and cannot consider the trend shift in long-term operation and the local impact characteristics caused by short-term faults. The multi-scale dynamic characteristics of the signal cannot be fully described. At the same time, when directly performing EMD decomposition on the original signal, the long-term trend component and the local fluctuation component will be mixed, and effective decoupling and accurate characterization cannot be achieved. The diagnosis model with single feature input cannot learn the essential differences of faults at different scales, and has insufficient sensitivity to early weak faults and compound faults, which may cause misjudgment and omission, and cannot meet the high reliability and high precision diagnosis requirements of new energy automobile driving motors. Therefore, the prior art has the problem of low motor operation state diagnosis accuracy. SUMMARY

[0004] In view of the above problems in the prior art, the new energy automobile motor operation state diagnosis method provided by the present application solves the problem of low motor operation state diagnosis accuracy in the prior art.

[0005] In order to achieve the above-mentioned purposes, the technical scheme adopted by the present application is as follows: a new energy automobile motor operation state diagnosis method, comprising the following steps:

[0006] S1, generating a vibration energy signal from the motor vibration signal, and obtaining short-scale, medium-scale and long-scale energy deviation signals based on three different length windows;

[0007] S2, segmenting the short-scale, medium-scale and long-scale energy deviation signals respectively, and then performing EMD decomposition on each segment to obtain long-term deviation component sub-signals and local deviation fluctuation sub-signals of each segment;

[0008] S3, extracting joint long-term deviation, joint long-term fluctuation and joint long-term skewness from the long-term deviation component sub-signals corresponding to the short-scale, medium-scale and long-scale to obtain a long-term feature matrix;

[0009] S4, extracting joint local impact energy value, joint local impact intensity and joint local impact morphology value from the local deviation fluctuation sub-signals corresponding to short scale, medium scale and long scale to obtain a local feature matrix;

[0010] S5, processing the normalized long-term feature matrix and the local feature matrix by using a double-branch fusion convolutional neural network to obtain a motor operation state diagnosis result.

[0011] Further, S1 includes the following sub-steps:

[0012] S11, calculating an instantaneous energy value for each sampling point in the motor vibration signal to obtain a vibration energy signal;

[0013] S12, setting three window lengths of short, medium and long;

[0014] S13, calculating energy mean values of the three window lengths in the center neighborhood range respectively with each sampling point in the vibration energy signal as the center;

[0015] S14, subtracting the three energy mean values respectively from the energy value of each sampling point to obtain three energy deviation values;

[0016] S15, constructing corresponding energy deviation signals from the energy deviation values of each sampling point belonging to the same window length to obtain short scale, medium scale and long scale energy deviation signals.

[0017] Further, S2 includes the following sub-steps:

[0018] S21, respectively segmenting the short scale, medium scale and long scale energy deviation signals to obtain multiple short scale energy deviation sub-signals, multiple medium scale energy deviation sub-signals and multiple long scale energy deviation sub-signals;

[0019] S22, respectively performing EMD decomposition on the short scale energy deviation sub-signals, the medium scale energy deviation sub-signals and the long scale energy deviation sub-signals to obtain multiple order intrinsic mode functions and residual terms;

[0020] S23, taking the residual term as a long-term deviation component sub-signal, and numerically adding 1, 2 and 3 orders of the multiple order intrinsic mode functions according to the sampling points to obtain a local deviation fluctuation sub-signal.

[0021] Further, S3 includes the following sub-steps:

[0022] S31, extracting joint long-term deviation degree, joint long-term fluctuation degree and joint long-term deviation degree from the long-term deviation component sub-signals corresponding to short scale, medium scale and long scale;

[0023] S32, splicing the joint long-term skewness, the joint long-term kurtosis and the joint long-term skewness of the same segment to form a long-term feature vector;

[0024] S33, forming a long-term feature matrix by using the long-term feature vectors corresponding to each segment.

[0025] Further, the acquisition process of the joint long-term skewness comprises: adding the signal values of each sampling point in the long-term skewness component sub-signal of each segment to obtain the long-term skewness; under the same segment, adding the long-term skewness corresponding to the short scale, the medium scale and the long scale to obtain the joint long-term skewness;

[0026] The acquisition process of the joint long-term kurtosis comprises: extracting the standard deviation of the long-term skewness component sub-signal of each segment; under the same segment, adding the standard deviations corresponding to the short scale, the medium scale and the long scale to obtain the joint long-term kurtosis;

[0027] The acquisition process of the joint long-term skewness comprises: extracting the skewness of the long-term skewness component sub-signal of each segment; under the same segment, adding the skewness corresponding to the short scale, the medium scale and the long scale to obtain the joint long-term skewness.

[0028] Further, S4 comprises the following steps:

[0029] S41, extracting the joint local impact energy value, the joint local impact intensity and the joint local impact morphology value corresponding to the local skewness fluctuation sub-signal of the short scale, the medium scale and the long scale;

[0030] S42, splicing the joint local impact energy value, the joint local impact intensity and the joint local impact morphology value of the same segment to form a local feature vector;

[0031] S43, forming a local feature matrix by using the local feature vectors corresponding to each segment.

[0032] Further, the acquisition process of the joint local impact energy value comprises: setting a signal threshold, screening the signal values greater than the signal threshold in the local skewness fluctuation sub-signal of each segment to obtain the peak value; adding the peak values corresponding to each segment of the local skewness fluctuation sub-signal to obtain the impact energy value; under the same segment, adding the impact energy values corresponding to the short scale, the medium scale and the long scale to obtain the joint local impact energy value;

[0033] The acquisition process of the joint local impact intensity comprises: taking the absolute value of each signal value in the local skewness fluctuation sub-signal of each segment to obtain the signal absolute value, screening the maximum signal absolute value to obtain the peak amplitude; under the same segment, adding the peak amplitudes corresponding to the short scale, the medium scale and the long scale to obtain the joint local impact intensity;

[0034] The joint local impact morphology value acquisition process comprises: calculating the kurtosis of the local deviation fluctuation sub-signal of each segment; under the same segment, adding the kurtosis corresponding to the short scale, the medium scale and the long scale to obtain the joint local impact morphology value.

[0035] Further, the double-branch fusion convolutional neural network in S5 comprises: a first high-dimensional feature extraction unit, a second high-dimensional feature extraction unit, a channel attention feature fusion unit, a spatial attention feature fusion unit, a first Concat layer and a classification unit.

[0036] The input end of the first high-dimensional feature extraction unit is configured to input the normalized long-term feature matrix, and the output end thereof is connected with the first input end of the channel attention feature fusion unit and the first input end of the spatial attention feature fusion unit respectively.

[0037] The input end of the second high-dimensional feature extraction unit is configured to input the normalized local feature matrix, and the output end thereof is connected with the second input end of the channel attention feature fusion unit and the second input end of the spatial attention feature fusion unit respectively.

[0038] The input end of the first Concat layer is connected with the output end of the channel attention feature fusion unit and the output end of the spatial attention feature fusion unit respectively, and the output end thereof is connected with the input end of the classification unit.

[0039] The output end of the classification unit serves as the output end of the double-branch fusion convolutional neural network.

[0040] Further, S5 comprises the following steps:

[0041] S51, processing the normalized long-term feature matrix by using the first high-dimensional feature extraction unit to obtain long-term high-dimensional features;

[0042] S52, processing the normalized local feature matrix by using the second high-dimensional feature extraction unit to obtain local high-dimensional features;

[0043] S53, performing channel attention weighting on the long-term high-dimensional features and the local high-dimensional features by using the channel attention feature fusion unit to obtain high-dimensional channel fusion features;

[0044] S54, performing spatial attention weighting on the long-term high-dimensional features and the local high-dimensional features by using the spatial attention feature fusion unit to obtain high-dimensional spatial fusion features;

[0045] S55, splicing the high-dimensional channel fusion features and the high-dimensional spatial fusion features by using the first Concat layer to obtain spliced features;

[0046] S56, classifying the spliced features by using the classification unit to obtain the motor operation state diagnosis result.

[0047] Further, the channel attention feature fusion unit comprises a first channel attention generation module, a second channel attention generation module, a multiplier M1, a multiplier M2 and an adder A1;

[0048] The input end of the first channel attention generation module is connected with the first input end of the multiplier M1 and serves as the first input end of the channel attention feature fusion unit; the input end of the second channel attention generation module is connected with the first input end of the multiplier M2 and serves as the second input end of the channel attention feature fusion unit; the output end of the first channel attention generation module is connected with the second input end of the multiplier M1; the output end of the second channel attention generation module is connected with the second input end of the multiplier M2; the input ends of the adder A1 are respectively connected with the output end of the multiplier M1 and the output end of the multiplier M2, and the output end thereof serves as the output end of the channel attention feature fusion unit;

[0049] The spatial attention feature fusion unit comprises a first spatial attention generation module, a second spatial attention generation module, a multiplier M3, a multiplier M4 and an adder A2;

[0050] The input end of the first spatial attention generation module is connected with the first input end of the multiplier M3 and serves as the first input end of the spatial attention feature fusion unit; the input end of the second spatial attention generation module is connected with the first input end of the multiplier M4 and serves as the second input end of the spatial attention feature fusion unit; the output end of the first spatial attention generation module is connected with the second input end of the multiplier M3; the output end of the second spatial attention generation module is connected with the second input end of the multiplier M4; the input ends of the adder A2 are respectively connected with the output end of the multiplier M3 and the output end of the multiplier M4, and the output end thereof serves as the output end of the spatial attention feature fusion unit.

[0051] The application has the beneficial effects that: the application can simultaneously consider the trend deviation of long-term operation of the motor and the local impact characteristics caused by short-time faults by generating energy deviation signals through three different scale analysis windows, and can comprehensively depict the multi-scale dynamic characteristics of the vibration signals; the long-term deviation components and the local fluctuation components are prevented from being mixed with each other by segmenting the multi-scale signals and then performing EMD decomposition, so that the two types of components are effectively decoupled and accurately characterized; the long-term deviation characteristics and the local impact characteristics are extracted, which can comprehensively reflect the essential differences of faults from multiple dimensions, significantly improve the sensitivity to early weak faults and compound faults, and reduce the misjudgment and omission probability; the long-term feature matrix and the local feature matrix are deeply fused and learned by using the double-branch fusion convolutional neural network, which can fully mine the associated information between different scales and different types of features, further improve the accuracy and reliability of fault identification, thereby improving the diagnosis accuracy of the operation state of the new energy automobile driving motor as a whole, and meeting the diagnosis requirements of high reliability and high precision under complex working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 It is a flowchart of a new energy automobile motor operation state diagnosis method.

[0053] Figure 2 It is a structure diagram of a double-branch fusion convolutional neural network.

[0054] Figure 3 It is a structure diagram of a first high-dimensional feature extraction unit and a second high-dimensional feature extraction unit.

[0055] Figure 4 It is a structure diagram of a channel attention feature fusion unit.

[0056] Figure 5 It is a structure diagram of a first channel attention generation module and a second channel attention generation module.

[0057] Figure 6 It is a structure diagram of a spatial attention feature fusion unit.

[0058] Figure 7 It is a structure diagram of a first spatial attention generation module and a second spatial attention generation module.

[0059] Figure 8 It is a structure diagram of a classification unit. DETAILED DESCRIPTION

[0060] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0061] As shown in Figure 1 A new energy vehicle motor running state diagnosis method, comprising the following steps:

[0062] S1, generating a vibration energy signal from the motor vibration signal, and obtaining short-scale, medium-scale and long-scale energy deviation signals based on three different window lengths;

[0063] S2, segmenting the short-scale, medium-scale and long-scale energy deviation signals respectively, and then performing EMD decomposition on each segment to obtain long-term deviation component sub-signals and local deviation fluctuation sub-signals of each segment;

[0064] S3, extracting joint long-term deviation, joint long-term fluctuation and joint long-term skewness from the long-term deviation component sub-signals corresponding to the short-scale, medium-scale and long-scale to obtain a long-term feature matrix;

[0065] S4, extracting joint local impact energy value, joint local impact intensity and joint local impact morphology value from the local deviation fluctuation sub-signals corresponding to the short-scale, medium-scale and long-scale to obtain a local feature matrix;

[0066] S5, processing the normalized long-term feature matrix and local feature matrix using a double-branch fusion convolutional neural network to obtain a motor running state diagnosis result.

[0067] In the present embodiment, S1 comprises the following sub-steps:

[0068] S11, calculating the instantaneous energy value of each sampling point in the motor vibration signal to obtain a vibration energy signal;

[0069] S12, setting three window lengths of short, medium and long;

[0070] S13, taking each sampling point of the vibration energy signal as the center, calculating the energy mean value under three window lengths in its neighborhood respectively to obtain the short window energy mean value, medium window energy mean value and long window energy mean value of the corresponding sampling point;

[0071] S14, subtracting the short window energy mean value of the corresponding sampling point from the energy value of each sampling point to obtain the short window energy deviation value of the corresponding sampling point; subtracting the medium window energy mean value of the corresponding sampling point from the energy value of each sampling point to obtain the medium window energy deviation value of the corresponding sampling point; subtracting the long window energy mean value of the corresponding sampling point from the energy value of each sampling point to obtain the long window energy deviation value of the corresponding sampling point,

[0072] S15, the short-scale energy deviation signal is composed of the short window energy deviation values of all sampling points, the medium-scale energy deviation signal is composed of the medium window energy deviation values of all sampling points, and the long-scale energy deviation signal is composed of the long window energy deviation values of all sampling points.

[0073] In the embodiment, the formula for calculating the instantaneous energy value is: , wherein, is the instantaneous energy value of the i th sampling point in the vibration energy signal, is the signal value of the i th sampling point in the motor vibration signal, is the signal value of the i th sampling point in the motor vibration signal, is the number of sampling points.

[0074] In the embodiment, the short window length is set to 10 sampling points, the medium window length is set to 20 sampling points, and the long window length is set to 30 sampling points, that is, the energy mean values in the neighborhood ranges with lengths of 10, 20 and 30 are respectively obtained for each sampling point.

[0075] The present application can convert the original vibration signal into an energy representation that can better reflect the degree of abnormal operation by calculating the instantaneous energy value of the motor vibration signal point by point, and highlight the energy mutation characteristics caused by faults. The energy mean values of the neighborhood are calculated with each sampling point as the center, and the energy deviation values are obtained by subtracting the corresponding window mean values from the energy values of the sampling points, which can effectively eliminate the energy fluctuations and random noise interference under normal working conditions and retain the abnormal deviation components caused by faults. The finally formed multi-scale energy deviation signal can comprehensively depict the running state changes of the motor under different time scales, and effectively improve the perception ability of weak faults and early faults.

[0076] In the embodiment, the motor vibration signal is a vibration acceleration signal collected in real time by a vibration acceleration sensor installed at the shell, end cover or bearing seat position of a new energy automobile driving motor.

[0077] In the embodiment, S2 includes the following steps:

[0078] ​S21, respectively, the short scale, the middle scale and the long scale energy deviation signal is segmented and processed to obtain a plurality of short scale energy deviation sub-signals, a plurality of middle scale energy deviation sub-signals and a plurality of long scale energy deviation sub-signals, and the three scale energy deviation signals are consistent in segmentation mode;

[0079] S22, EMD decomposition is performed on the short scale energy deviation sub-signals, the middle scale energy deviation sub-signals and the long scale energy deviation sub-signals respectively to obtain a plurality of intrinsic mode functions and a residual term;

[0080] S23, the residual term is taken as a long-term deviation component sub-signal, and the first, second and third orders of the plurality of intrinsic mode functions are numerically added according to the sampling points to obtain a local deviation fluctuation sub-signal.

[0081] A segment of energy deviation sub-signal corresponds to a plurality of intrinsic mode functions and a residual term, and corresponds to a long-term deviation component sub-signal and a local deviation fluctuation sub-signal.

[0082] In this embodiment, the short scale, the middle scale and the long scale energy deviation signals are segmented and processed respectively, and the three scale energy deviation signals adopt the same segmentation rule: non-overlapping segmentation is performed with 64 sampling points as a segment, and the segmentation starting sampling points of each scale signal remain consistent, thereby obtaining a plurality of short scale energy deviation sub-signals, a plurality of middle scale energy deviation sub-signals and a plurality of long scale energy deviation sub-signals.

[0083] In this embodiment, the first, second and third intrinsic mode functions belonging to the same energy deviation sub-signal are numerically added according to the same sampling point to obtain a local deviation fluctuation sub-signal.

[0084] The present application reduces the influence of signal non-stationarity and working condition fluctuation on subsequent decomposition by segmenting and processing the multi-scale energy deviation signal, effectively suppresses the modal aliasing phenomenon; through EMD decomposition of each segmented sub-signal, the residual term is taken as a long-term deviation component sub-signal, which can accurately represent the slow deviation and long-term degradation trend in the motor running process, and the superposition of the first three intrinsic mode functions obtains a local deviation fluctuation sub-signal, which can highlight the local impact and short-time mutation characteristics caused by faults, thereby realizing effective decoupling and accurate representation of long-term trend components and local fluctuation components.

[0085] In this embodiment, S3 includes the following steps:

[0086] S31, extracting a joint long-term deviation degree, a joint long-term fluctuation degree and a joint long-term deviation degree from the long-term deviation component sub-signals corresponding to the short scale, the middle scale and the long scale;

[0087] S32, splicing the joint long-term deviation degree, the joint long-term fluctuation degree and the joint long-term deviation degree of the same segment to form a long-term feature vector;

[0088] S33. Take the long-term feature vector corresponding to each segment as a column vector, arrange the columns in chronological order, and form a long-term feature matrix.

[0089] In this embodiment, the size of the long-term feature matrix is ​​3×T, where T is the number of segments.

[0090] In this embodiment, the process of obtaining the joint long-term deviation includes: adding the signal values ​​of each sampling point in the long-term deviation of each segment into the sub-signal to obtain the long-term deviation; and adding the long-term deviations corresponding to the short-scale, medium-scale, and long-scale sub-segments within the same segment to obtain the joint long-term deviation, wherein the same segment refers to the same segment number, that is, adding the long-term deviation of the short-scale sub-signal of the m-th segment, the long-term deviation of the medium-scale sub-signal of the m-th segment, and the long-term deviation of the long-scale sub-signal of the m-th segment into the sub-signal.

[0091] The process of obtaining the joint long-term volatility includes: extracting the standard deviation of the long-term deviation of each segment into a molecular signal; and summing the standard deviations corresponding to the short-scale, medium-scale, and long-scale signals within the same segment to obtain the joint long-term volatility.

[0092] The process of obtaining the joint long-term skewness includes: extracting the skewness from the long-term deviation of each segment into a molecular signal; and summing the skewnesses corresponding to the short-scale, medium-scale, and long-scale signals within the same segment to obtain the joint long-term skewness.

[0093] The formula for calculating the joint long-term deviation is:

[0094] ,

[0095] in, To combine long-term deviation, For the first Long-term deviation of the scale in the molecular signal The signal value at each sampling point The number of sampling points, The numbers 1, 2, and 3 correspond to short-scale, medium-scale, and long-scale, respectively.

[0096] The formula for calculating the combined long-term volatility is:

[0097] ,

[0098] in, To combine long-term volatility, For the first The scale has long deviated from the standard deviation of the molecular signal.

[0099] The formula for calculating the joint long-term skewness is:

[0100] ,

[0101] wherein, is a joint long-term skewness, is a first scale long-term deviation component signal skewness.

[0102] The present application can comprehensively depict the slow drift, performance degradation and trend abnormality of the motor in the long-term running process from three dimensions of the amplitude accumulation deviation degree, the data discrete fluctuation degree and the signal distribution asymmetry by respectively extracting and fusing the accumulation, the standard deviation and the skewness of the long-term deviation component signals under the short scale, the medium scale and the long scale to construct the joint long-term skewness, the joint long-term kurtosis and the joint long-term skewness.

[0103] In the embodiment, S4 comprises the following steps:

[0104] S41, extracting a joint local impact energy value, a joint local impact intensity and a joint local impact morphology value for the local deviation fluctuation sub-signals corresponding to the short scale, the medium scale and the long scale;

[0105] S42, splicing the joint local impact energy value, the joint local impact intensity and the joint local impact morphology value of the same segment to form a local feature vector;

[0106] S43, taking the local feature vector corresponding to each segment as a column vector, arranging the columns in time sequence to form a local feature matrix.

[0107] In the embodiment, the size of the local feature matrix is 3xT, and T is the number of segments.

[0108] In the embodiment, the acquisition process of the joint local impact energy value comprises: setting a signal threshold, screening the signal values greater than the signal threshold in the local deviation fluctuation sub-signals of each segment to obtain a peak value; adding the peak values corresponding to each segment of the local deviation fluctuation sub-signals to obtain an impact energy value; under the same segment, adding the impact energy values corresponding to the short scale, the medium scale and the long scale to obtain the joint local impact energy value;

[0109] The acquisition process of the joint local impact intensity comprises: taking the absolute value of each signal value in the local deviation fluctuation sub-signals of each segment to obtain a signal absolute value, screening the maximum signal absolute value to obtain a peak amplitude; under the same segment, adding the peak amplitudes corresponding to the short scale, the medium scale and the long scale to obtain the joint local impact intensity;

[0110] The process of obtaining the joint local impact morphology value includes: calculating the kurtosis of the local deviation wavelet signal for each segment; and adding the kurtosis corresponding to the short-scale, medium-scale, and long-scale signals within the same segment to obtain the joint local impact morphology value.

[0111] In this embodiment, the signal threshold is 0.5 to 2 times the standard deviation of the local deviation of the wavelet signal in the current segment.

[0112] The formula for calculating the combined local impact energy value is:

[0113] ,

[0114] in, To combine local impact energy values, For the first The first type of local deviation wave sub-signal with a scale greater than the signal threshold A peak value, This represents the number of peak values ​​exceeding the signal threshold in the local deviation wave sub-signal at the corresponding scale.

[0115] The formula for calculating the combined local impact strength is:

[0116] ,

[0117] in, To combine local impact intensity, In the first The maximum absolute value of the signal is taken from the local deviation wave sub-signal of the various scales.

[0118] The formula for calculating the combined local impact shape value is:

[0119] ,

[0120] in, To combine local impact morphological values, For the first The kurtosis of the local deviation from the wavelet signal at this scale.

[0121] This invention constructs a joint local impact energy value, a joint local impact intensity value, and a joint local impact morphology value by fusing the threshold peak sum, maximum absolute amplitude, and kurtosis of local deviation wave sub-signals at short, medium, and long scales. This allows for the capture of instantaneous impact and local abrupt change characteristics caused by motor faults from three dimensions: impact energy magnitude, impact amplitude strength, and impact distribution morphology. The fusion of multi-scale local features can significantly improve the sensitivity to early weak faults, complex faults, and short-term impact faults, effectively enhancing the identification of fault features.

[0122] In the embodiment, the long-term feature matrix and the local feature matrix are normalized by row (each row element is divided by the maximum value of the row). The maximum value is found on each row, and each row element is divided by the maximum value of the corresponding row to achieve normalization.

[0123] As shown in Figure 2 The double-branch fusion convolutional neural network in S5 includes a first high-dimensional feature extraction unit, a second high-dimensional feature extraction unit, a channel attention feature fusion unit, a spatial attention feature fusion unit, a first Concat layer, and a classification unit.

[0124] The input end of the first high-dimensional feature extraction unit is configured to input the normalized long-term feature matrix, and the output end thereof is connected with the first input end of the channel attention feature fusion unit and the first input end of the spatial attention feature fusion unit respectively.

[0125] The input end of the second high-dimensional feature extraction unit is configured to input the normalized local feature matrix, and the output end thereof is connected with the second input end of the channel attention feature fusion unit and the second input end of the spatial attention feature fusion unit respectively.

[0126] The input end of the first Concat layer is connected with the output end of the channel attention feature fusion unit and the output end of the spatial attention feature fusion unit respectively, and the output end thereof is connected with the input end of the classification unit.

[0127] The output end of the classification unit serves as the output end of the double-branch fusion convolutional neural network.

[0128] In the embodiment, S5 includes the following steps:

[0129] S51, the normalized long-term feature matrix is processed by the first high-dimensional feature extraction unit to obtain long-term high-dimensional features.

[0130] S52, the normalized local feature matrix is processed by the second high-dimensional feature extraction unit to obtain local high-dimensional features.

[0131] S53, the channel attention feature fusion unit is adopted to perform channel attention weighting on the long-term high-dimensional features and the local high-dimensional features to obtain high-dimensional channel fusion features.

[0132] S54, the spatial attention feature fusion unit is adopted to perform spatial attention weighting on the long-term high-dimensional features and the local high-dimensional features to obtain high-dimensional spatial fusion features.

[0133] S55, the first Concat layer is adopted to splice the high-dimensional channel fusion features and the high-dimensional spatial fusion features to obtain spliced features.

[0134] S56, the classification unit is used to classify the spliced features to obtain a motor operation state diagnosis result.

[0135] As shown in Figure 3 The first high-dimensional feature extraction unit and the second high-dimensional feature extraction unit are the same in structure and each include a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a first pooling layer, a second pooling layer, a third pooling layer, and a second Concat layer.

[0136] The input end of the first convolutional layer is connected with the input end of the second convolutional layer and the input end of the third convolutional layer respectively and serves as the input end of the first high-dimensional feature extraction unit and the second high-dimensional feature extraction unit.

[0137] The first convolutional layer, the fourth convolutional layer, and the first pooling layer are connected in sequence; the second convolutional layer, the fifth convolutional layer, and the second pooling layer are connected in sequence; and the third convolutional layer, the sixth convolutional layer, and the third pooling layer are connected in sequence.

[0138] The input end of the second Concat layer is connected with the output end of the first pooling layer, the output end of the second pooling layer, and the output end of the third pooling layer respectively, and the output end thereof serves as the output end of the first high-dimensional feature extraction unit and the second high-dimensional feature extraction unit.

[0139] In the embodiment, the size of the long-term feature matrix and the local feature matrix is 3xT, the convolution kernel of the first convolutional layer, the second convolutional layer, and the third convolutional layer is 3x1, the step is 1, the padding is 1, the output channel is 1, and the output size is 1xT. The convolution kernel of the fourth convolutional layer is 1x3, the step is 1, the padding is 1, the output channel is 8, and the output size is 8xT. The convolution kernel of the fifth convolutional layer is 1x5, the step is 1, the padding is 2, the output channel is 8, and the output size is 8xT. The convolution kernel of the sixth convolutional layer is 1x7, the step is 1, the padding is 3, the output channel is 8, and the output size is 8xT. The pooling window size of the first pooling layer, the second pooling layer, and the third pooling layer is 1x2, the step is 2, the output size is 8xT / 2. The second Concat layer performs channel splicing to obtain 24xT / 2.

[0140] The application can precisely capture local patterns, short-term fluctuations, and long-range dependencies in the input matrix at different time resolutions by using three-branch parallel convolution structures and different sizes of time convolution kernels (1x3, 1x5, 1x7) to extract multi-scale time features from the input features, thereby effectively improving the richness and representation ability of the features.

[0141] As shown in Figure 4 The channel attention feature fusion unit includes a first channel attention generation module, a second channel attention generation module, a multiplier M1, a multiplier M2, and an adder A1.

[0142] The input end of the first channel attention generation module is connected with the first input end of the multiplier M1 and serves as the first input end of the channel attention feature fusion unit; the input end of the second channel attention generation module is connected with the first input end of the multiplier M2 and serves as the second input end of the channel attention feature fusion unit; the output end of the first channel attention generation module is connected with the second input end of the multiplier M1; the output end of the second channel attention generation module is connected with the second input end of the multiplier M2; the input end of the adder A1 is connected with the output end of the multiplier M1 and the output end of the multiplier M2 respectively, and the output end thereof serves as the output end of the channel attention feature fusion unit.

[0143] As shown in Figure 5 The first channel attention generation module and the second channel attention generation module are of the same structure and both include, in sequence, a first global average pooling layer, a first full connection layer and a first Sigmoid layer.

[0144] The first global average pooling layer: performs global average pooling on the input 24xT / 2 feature in the time sequence dimension, and outputs a feature with a size of 24x1 (the channel number is kept as 24 and the time sequence length is compressed to 1). The multiplier M1: performs element-by-element multiplication in the channel dimension between the first input feature (24xT / 2) and the channel weight M1 (24x1), and outputs a feature with a size of 24xT / 2. The multiplier M2: performs element-by-element multiplication in the channel dimension between the second input feature (24xT / 2) and the channel weight M2 (24x1), and outputs a feature with a size of 24xT / 2.

[0145] As shown in Figure 6 The spatial attention feature fusion unit includes: a first spatial attention generation module, a second spatial attention generation module, a multiplier M3, a multiplier M4 and an adder A2.

[0146] The input end of the first spatial attention generation module is connected with the first input end of the multiplier M3 and serves as the first input end of the spatial attention feature fusion unit; the input end of the second spatial attention generation module is connected with the first input end of the multiplier M4 and serves as the second input end of the spatial attention feature fusion unit; the output end of the first spatial attention generation module is connected with the second input end of the multiplier M3; the output end of the second spatial attention generation module is connected with the second input end of the multiplier M4; the input end of the adder A2 is connected with the output end of the multiplier M3 and the output end of the multiplier M4 respectively, and the output end thereof serves as the output end of the spatial attention feature fusion unit.

[0147] As shown in Figure 7As shown, the first spatial attention generation module and the second spatial attention generation module comprise a second global average pooling layer, a global maximum pooling layer, a third Concat layer, a seventh convolutional layer, and a second Sigmoid layer.

[0148] The second global average pooling layer: global average pooling is performed in the channel dimension, and the output size is 1xT / 2. The global maximum pooling layer: global maximum pooling is performed in the channel dimension, and the output size is 1xT / 2. The third Concat layer: two 1xT / 2 features are spliced in the channel dimension, and the output size is 2xT / 2. The seventh convolutional layer: the convolution kernel size is 1x1, the output channel is 1, and the output size is 1xT / 2. The second Sigmoid layer: the output spatial attention weight, the size is still 1xT / 2.

[0149] The multiplier M3: the first input feature (24xT / 2) is multiplied with the spatial weight M3 (1xT / 2) in the spatial dimension, and the output size is 24xT / 2. The multiplier M4: the second input feature (24xT / 2) is multiplied with the spatial weight M4 (1xT / 2) in the spatial dimension, and the output size is 24xT / 2.

[0150] The first high-dimensional feature extraction unit is adopted to process the long-term feature matrix, the second high-dimensional feature extraction unit is adopted to process the local feature matrix, high-dimensional features are extracted, and then the channel attention feature fusion unit is adopted to perform channel attention weighting on the long-term high-dimensional features and the local high-dimensional features. In the channel attention feature fusion unit, the first and second channel attention generation modules are respectively used to perform adaptive weighting in the channel dimension (M1, M2) on the long-term high-dimensional features and the local high-dimensional features, and then the element-level fusion (A1) is used to obtain high-dimensional channel fusion features, so that the key feature channels with higher distinguishability for state recognition can be automatically focused, and redundant channel information can be suppressed.

[0151] In the spatial attention feature fusion unit, the first and second spatial attention generation modules are respectively used to perform adaptive weighting in the spatial dimension (M3, M4) on the long-term high-dimensional features and the local high-dimensional features, and then the element-level fusion (A2) is used to obtain high-dimensional spatial fusion features, so that the spatial positions more important for fault representation can be accurately positioned, and the feature response of the local fault area can be strengthened.

[0152] Finally, the high-dimensional channel fusion features and the high-dimensional spatial fusion features are spliced through the first Concat layer, so that double attention weighting fusion in the channel dimension and the spatial dimension is realized, the representation ability and the recognition degree of the features are significantly improved, and thus the accuracy of motor operation state diagnosis is effectively improved.

[0153] As Figure 8As shown, the classification unit includes, in sequence, an eighth convolutional layer, a fourth pooling layer, a flattening layer, a second full connection layer and a classification layer; the convolution kernel of the eighth convolutional layer is 1x3, the step is 1, the padding is 1, the output channel is 48, and the output size is 48xT / 2; the pooling window of the fourth pooling layer is 1x2, the step is 2, the padding is 0, and the output size is 48xT / 4. The activation function of the classification layer is Softmax.

[0154] In the embodiment, the motor operation state diagnosis result includes: a normal state, a rotor fault state, a stator fault state and a bearing fault state.

[0155] In the embodiment, the dual-branch fusion convolutional neural network is trained by using an existing gradient descent method.

[0156] The present application can simultaneously consider the trend deviation of long-term operation of the motor and the local impact characteristics caused by short-term faults by using three different scale analysis windows, i.e., short, medium and long, to comprehensively depict the multi-scale dynamic characteristics of the vibration signal; the multi-scale signal is segmented and then decomposed by EMD to avoid the mutual aliasing of the long-term deviation component and the local fluctuation component, to effectively decouple and accurately represent the two types of components; the long-term deviation features and the local impact features are extracted, which can comprehensively reflect the essential differences of faults from multiple dimensions, significantly improve the sensitivity to early weak faults and compound faults, and reduce the probability of misjudgment and omission; the long-term feature matrix and the local feature matrix are deeply fused and learned by using the dual-branch fusion convolutional neural network, which can fully mine the associated information between different scales and different types of features, further improve the accuracy and reliability of fault identification, and thus improve the diagnosis accuracy of the operation state of the new energy automobile driving motor as a whole, and meet the diagnosis requirements of high reliability and high precision under complex working conditions.

[0157] The above is only a preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for diagnosing the operating status of a new energy vehicle motor, characterized in that, Includes the following steps: S1. Generate vibration energy signals from motor vibration signals, and obtain short-scale, medium-scale, and long-scale energy deviation signals based on three different window lengths; S2. Segment the energy deviation signals at short, medium and long scales respectively, and then perform EMD decomposition on each segment to obtain the long-term deviation component signal and the local deviation wave component signal of each segment. S3. Extract the joint long-term deviation, joint long-term volatility, and joint long-term skewness from the long-term deviation component signals corresponding to short-scale, medium-scale, and long-scale to obtain the long-term feature matrix. S4. Extract the joint local impact energy value, joint local impact intensity, and joint local impact morphology value from the local deviation wave sub-signals corresponding to short-scale, medium-scale, and long-scale to obtain the local feature matrix; S5. A dual-branch fusion convolutional neural network is used to process the normalized long-term feature matrix and local feature matrix to obtain the motor operating status diagnosis results.

2. The method for diagnosing the operating status of a new energy vehicle motor according to claim 1, characterized in that, S1 includes the following steps: S11. Calculate the instantaneous energy value for each sampling point in the motor vibration signal to obtain the vibration energy signal; S12. Set the window length to short, medium, and long. S13. Taking each sampling point in the vibration energy signal as the center, calculate the average energy of the three window lengths in the neighborhood of the center respectively; S14. Subtract the three energy averages from the energy value of each sampling point to obtain the three energy deviation values; S15. The energy deviation values ​​of each sampling point belonging to the same window length are used to form the corresponding energy deviation signal, thus obtaining the short-scale, medium-scale, and long-scale energy deviation signals.

3. The method for diagnosing the operating status of a new energy vehicle motor according to claim 1, characterized in that, S2 includes the following steps: S21. The energy deviation signals of short scale, medium scale and long scale are segmented and processed respectively to obtain multiple short-scale energy deviation sub-signals, multiple medium-scale energy deviation sub-signals and multiple long-scale energy deviation sub-signals. S22. Perform EMD decomposition on the short-scale energy deviation sub-signal, the medium-scale energy deviation sub-signal, and the long-scale energy deviation sub-signal respectively to obtain the multi-order intrinsic mode functions and residual terms; S23. The residual term is taken as the long-term deviation component of the sub-signal. The values ​​of the 1st, 2nd and 3rd orders in the multi-order intrinsic mode functions are added together according to the sampling points to obtain the local deviation wave sub-signal.

4. The method for diagnosing the operating status of a new energy vehicle motor according to claim 1, characterized in that, S3 includes the following steps: S31. Extract the joint long-term deviation, joint long-term volatility, and joint long-term skewness from the long-term deviations of the molecular signals corresponding to short-scale, medium-scale, and long-scale. S32. Combine the joint long-term deviation, joint long-term volatility, and joint long-term skewness of the same segment to form a long-term feature vector. S33. The long-term feature vectors corresponding to each segment are used to form a long-term feature matrix.

5. The method for diagnosing the operating status of a new energy vehicle motor according to claim 1 or 4, characterized in that, The process of obtaining the joint long-term deviation includes: summing the signal values ​​of each sampling point in the molecular signal for the long-term deviation of each segment to obtain the long-term deviation; and summing the long-term deviations corresponding to the short-scale, medium-scale, and long-scale signals in the same segment to obtain the joint long-term deviation. The process of obtaining the joint long-term volatility includes: extracting the standard deviation of the long-term deviation of each segment into a molecular signal; and summing the standard deviations corresponding to the short-scale, medium-scale, and long-scale signals within the same segment to obtain the joint long-term volatility. The process of obtaining the joint long-term skewness includes: extracting the skewness from the long-term deviation of each segment into a molecular signal; and summing the skewnesses corresponding to the short-scale, medium-scale, and long-scale signals within the same segment to obtain the joint long-term skewness.

6. The method for diagnosing the operating status of a new energy vehicle motor according to claim 1, characterized in that, S4 includes the following steps: S41. Extract the joint local impact energy value, joint local impact intensity value, and joint local impact morphology value from the local deviation wave sub-signals corresponding to short-scale, medium-scale, and long-scale. S42. Combine the joint local impact energy value, joint local impact intensity, and joint local impact morphology value of the same segment to form a local feature vector; S43. The local feature vectors corresponding to each segment are used to form a local feature matrix.

7. The method for diagnosing the operating status of a new energy vehicle motor according to claim 1 or 6, characterized in that, The process of obtaining the joint local impact energy value includes: setting a signal threshold, filtering out signal values ​​greater than the signal threshold from the local deviation wave sub-signals of each segment to obtain peak values; adding up the peak values ​​corresponding to the local deviation wave sub-signals of each segment to obtain the impact energy value; and adding up the impact energy values ​​corresponding to the short-scale, medium-scale, and long-scale segments within the same segment to obtain the joint local impact energy value. The process of obtaining the joint local impact intensity includes: taking the absolute value of each signal value in the local deviation wave sub-signal of each segment to obtain the signal absolute value, filtering the largest signal absolute value to obtain the peak amplitude; and adding the peak amplitudes corresponding to the short-scale, medium-scale and long-scale in the same segment to obtain the joint local impact intensity. The process of obtaining the joint local impact morphology value includes: calculating the kurtosis of the local deviation wavelet signal for each segment; and adding the kurtosis corresponding to the short-scale, medium-scale, and long-scale signals within the same segment to obtain the joint local impact morphology value.

8. The method for diagnosing the operating status of a new energy vehicle motor according to claim 1, characterized in that, The dual-branch fusion convolutional neural network in S5 includes: a first high-dimensional feature extraction unit, a second high-dimensional feature extraction unit, a channel attention feature fusion unit, a spatial attention feature fusion unit, a first concat layer, and a classification unit; The input of the first high-dimensional feature extraction unit is used to input the normalized long-term feature matrix, and its output is connected to the first input of the channel attention feature fusion unit and the first input of the spatial attention feature fusion unit, respectively. The input of the second high-dimensional feature extraction unit is used to input the normalized local feature matrix, and its output is connected to the second input of the channel attention feature fusion unit and the second input of the spatial attention feature fusion unit, respectively. The input of the first Concat layer is connected to the output of the channel attention feature fusion unit and the spatial attention feature fusion unit, respectively, and its output is connected to the input of the classification unit. The output of the classification unit serves as the output of the dual-branch fusion convolutional neural network.

9. The method for diagnosing the operating status of a new energy vehicle motor according to claim 8, characterized in that, S5 includes the following steps: S51. The normalized long-term feature matrix is ​​processed by the first high-dimensional feature extraction unit to obtain long-term high-dimensional features. S52. The normalized local feature matrix is ​​processed by the second high-dimensional feature extraction unit to obtain local high-dimensional features; S53. Channel attention feature fusion unit is used to perform channel attention weighting on long-term high-dimensional features and local high-dimensional features to obtain high-dimensional channel fusion features; S54. Spatial attention feature fusion unit is used to perform spatial attention weighting on long-term high-dimensional features and local high-dimensional features to obtain high-dimensional spatial fusion features; S55. The first Concat layer is used to splice the high-dimensional channel fusion feature and the high-dimensional space fusion feature to obtain the spliced ​​feature; S56. Classify the spliced ​​features using classification units to obtain the motor operating status diagnosis results.

10. The method for diagnosing the operating status of a new energy vehicle motor according to claim 8 or 9, characterized in that, The channel attention feature fusion unit includes: a first channel attention generation module, a second channel attention generation module, a multiplier M1, a multiplier M2, and an adder A1; The input of the first channel attention generation module is connected to the first input of multiplier M1 and serves as the first input of the channel attention feature fusion unit; the input of the second channel attention generation module is connected to the first input of multiplier M2 and serves as the second input of the channel attention feature fusion unit; the output of the first channel attention generation module is connected to the second input of multiplier M1; the output of the second channel attention generation module is connected to the second input of multiplier M2; the input of adder A1 is connected to the outputs of multiplier M1 and multiplier M2 respectively, and its output serves as the output of the channel attention feature fusion unit. The spatial attention feature fusion unit includes: a first spatial attention generation module, a second spatial attention generation module, a multiplier M3, a multiplier M4, and an adder A2; The input terminal of the first spatial attention generation module is connected to the first input terminal of the multiplier M3 and serves as the first input terminal of the spatial attention feature fusion unit; the input terminal of the second spatial attention generation module is connected to the first input terminal of the multiplier M4 and serves as the second input terminal of the spatial attention feature fusion unit; the output terminal of the first spatial attention generation module is connected to the second input terminal of the multiplier M3; the output terminal of the second spatial attention generation module is connected to the second input terminal of the multiplier M4; the input terminal of the adder A2 is connected to the output terminals of the multipliers M3 and M4 respectively, and its output terminal serves as the output terminal of the spatial attention feature fusion unit.

Images