New energy automobile power supply detection method and system

Abstract

The invention discloses a new energy automobile power supply detection method and system, and belongs to the technical field of power supply fault detection. The method comprises the following steps: acquiring three-phase voltage of a three-phase inverter, summing absolute values to obtain a three-phase composite signal, segmenting according to a period, dividing each frame into a left signal section and a right signal section, extracting a morphological feature vector to calculate a time domain morphological mismatch value, and constructing a time domain morphological mismatch matrix; fourier transform is carried out on the left signal segment and the right signal segment, a frequency domain form mismatch value is calculated based on the difference between the frequency gravity centers of the fundamental frequency and the non-fundamental frequency, and a frequency domain form mismatch matrix is formed; and further constructing a time domain and frequency domain form mismatch evolution matrix, processing the matrix by using a double-domain collaborative enhancement neural network, and finally outputting a power supply fault score. Through collaborative analysis of the time domain and frequency domain features and evolution matrix construction, the fault detection precision of the new energy automobile power supply is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power failure detection, and particularly relates to a new energy automobile power detection method and system. BACKGROUND

[0002] The three-phase inverter in the power supply system undertakes the key task of converting direct current into alternating current, and its performance directly affects the power output, endurance mileage and driving safety of the vehicle. Once the three-phase inverter fails, it may cause the vehicle power to be interrupted, the electronic system to malfunction, and even cause serious safety accidents such as fire, which not only threatens the safety of the driver and passengers, but also hinders the sustainable development of the new energy automobile industry.

[0003] At present, in the field of new energy automobile power detection, the traditional technology mainly monitors single electrical parameters such as voltage amplitude and current size. This kind of method usually only sets a simple threshold range to judge whether the power supply system is abnormal, and only evaluates the power supply running state from a limited angle. However, the monitoring method of single electrical parameter is difficult to capture the complex fault characteristics presented in the running process of the three-phase inverter, and it is even more difficult to detect the weak signal changes of early faults.

[0004] When facing early implicit faults, the traditional method often cannot timely perceive and give early warning due to the unobvious fault characteristics, resulting in the continuous development of the fault without being discovered, and finally causing serious consequences. Therefore, the traditional detection technology has the problem of low fault detection precision. SUMMARY

[0005] In view of the above problems in the prior art, the new energy automobile power detection method and system provided by the present application solve the problem of low fault detection precision in the prior art.

[0006] In order to achieve the above application purpose, the technical scheme adopted by the present application is as follows: a new energy automobile power detection method, comprising the following steps:

[0007] Collecting three-phase voltage of the three-phase inverter in the automobile power supply, taking absolute value of each voltage value in the three-phase voltage, and adding the absolute values of the three-phase voltage at the same time to obtain a three-phase synthesis signal;

[0008] The three-phase synthesis signal is divided according to the cycle length to obtain a plurality of synthesis signal frames;

[0009] Each synthesis signal frame is divided into a left signal segment and a right signal segment, and a shape feature vector is extracted, a time domain shape mismatch value is obtained based on the difference between the left and right shape feature vectors, and a synthesis signal time domain shape mismatch matrix is formed;

[0010] performing Fourier transform on the left signal segment and the right signal segment respectively, calculating the frequency domain form mismatch value according to the difference between the left side and the right side on the fundamental frequency and the difference on the non-fundamental frequency gravity center, and constructing the frequency domain form mismatch matrix of the synthesized signal;

[0011] constructing the time domain form mismatch evolution matrix according to the time domain form mismatch value, and constructing the frequency domain form mismatch evolution matrix according to the frequency domain form mismatch value;

[0012] adopting the dual-domain cooperative enhancement neural network to process the time domain form mismatch matrix of the synthesized signal, the frequency domain form mismatch matrix of the synthesized signal, the time domain form mismatch evolution matrix and the frequency domain form mismatch evolution matrix, and obtaining the power fault score.

[0013] Further, the process of constructing the time domain form mismatch matrix of the synthesized signal comprises:

[0014] performing segmentation on the synthesized signal frame to obtain the left signal segment and the right signal segment;

[0015] extracting the form feature vector from the left signal segment and the right signal segment respectively;

[0016] obtaining the time domain form mismatch value according to the difference between the left signal segment and the right signal segment on the form feature vector;

[0017] filling the matrix according to the Hankel rule with each time domain form mismatch value to obtain the time domain form mismatch matrix of the synthesized signal.

[0018] Further, the process of extracting the form feature vector comprises:

[0019] adopting the adjacent voltage value subtraction in the left signal segment and the right signal segment respectively, and taking the absolute value of the subtraction result to obtain the voltage increment;

[0020] calculating the voltage increment discrete value of each voltage increment in the left signal segment and the right signal segment respectively to obtain the voltage increment discrete value of the left signal segment and the voltage increment discrete value of the right signal segment;

[0021] calculating the energy of the left signal segment and the energy of the right signal segment respectively;

[0022] constructing the form feature vector with the energy and the voltage increment discrete value belonging to the same signal segment.

[0023] Further, the specific process of obtaining the time domain form mismatch value comprises: calculating the Euclidean distance between the form feature vector of the left signal segment and the form feature vector of the right signal segment, and performing normalization processing to obtain the time domain form mismatch value.

[0024] Further, the process of constructing the frequency domain form mismatch matrix of the synthesized signal comprises:

[0025] performing Fourier transform on the left signal segment and the right signal segment respectively to obtain a left spectrum and a right spectrum;

[0026] finding a frequency corresponding to a maximum amplitude in the left spectrum and the right spectrum respectively to obtain a fundamental frequency;

[0027] obtaining a left non-fundamental frequency gravity center according to other frequencies in the left spectrum except the fundamental frequency;

[0028] obtaining a right non-fundamental frequency gravity center according to other frequencies in the right spectrum except the fundamental frequency;

[0029] calculating a frequency domain form mismatch value according to a difference between the left and the right in the fundamental frequency and a difference between the left and the right in the non-fundamental frequency gravity center, filling a matrix according to a Hankel rule with each frequency domain form mismatch value to obtain a synthetic signal frequency domain form mismatch matrix.

[0030] Further, the process of calculating the frequency domain form mismatch value comprises: calculating a first difference between the fundamental frequency of the left and the fundamental frequency of the right, calculating a second difference between the left non-fundamental frequency gravity center and the right non-fundamental frequency gravity center, adding the first difference and the second difference, and normalizing to obtain the frequency domain form mismatch value.

[0031] Further, the process of constructing the time domain form mismatch evolution matrix and the frequency domain form mismatch evolution matrix comprises:

[0032] taking a ratio of a time domain form mismatch value at another time to a first time domain form mismatch value as a time domain form mismatch evolution ratio;

[0033] filling a matrix according to a Hankel rule with each time domain form mismatch evolution ratio to obtain a time domain form mismatch evolution matrix;

[0034] taking a ratio of a frequency domain form mismatch value at another time to a first frequency domain form mismatch value as a frequency domain form mismatch evolution ratio;

[0035] filling a matrix according to a Hankel rule with each frequency domain form mismatch evolution ratio to obtain a frequency domain form mismatch evolution matrix.

[0036] Further, the dual-domain collaborative enhancement neural network comprises: a time domain feature enhancement unit, a frequency domain feature enhancement unit, a first Concat layer, a dual-channel feature attention unit and a full connection layer.

[0037] a first input end of the time domain feature enhancement unit is configured to input the synthetic signal time domain form mismatch matrix, a second input end of the time domain feature enhancement unit is configured to input the time domain form mismatch evolution matrix, a first input end of the frequency domain feature enhancement unit is configured to input the synthetic signal frequency domain form mismatch matrix, and a second input end of the frequency domain feature enhancement unit is configured to input the frequency domain form mismatch evolution matrix.

[0038] The input end of the first Concat layer is connected with the output end of the time domain feature enhancement unit and the output end of the frequency domain feature enhancement unit respectively, and the output end thereof is connected with the input end of the dual-channel feature attention unit; the output end of the dual-channel feature attention unit is connected with the input end of the full connection layer; and the output end of the full connection layer is the output end of the dual-domain collaborative enhancement neural network.

[0039] Further, the time domain feature enhancement unit and the frequency domain feature enhancement unit have the same structure, and each includes a first convolution layer, a second convolution layer, a first average pooling layer, a second average pooling layer, a first maximum pooling layer, a second maximum pooling layer, a multiplier M1, a multiplier M2, a third convolution layer, a fourth convolution layer and an adder A1.

[0040] The input end of the first convolution layer is the first input end of the time domain feature enhancement unit and the frequency domain feature enhancement unit; and the input end of the second convolution layer is the second input end of the time domain feature enhancement unit and the frequency domain feature enhancement unit.

[0041] The output end of the first convolution layer is connected with the input end of the first average pooling layer and the input end of the first maximum pooling layer respectively; the output end of the second convolution layer is connected with the input end of the second average pooling layer and the input end of the second maximum pooling layer respectively; the input end of the multiplier M1 is connected with the output end of the first average pooling layer and the output end of the second average pooling layer respectively, and the output end thereof is connected with the input end of the third convolution layer; the input end of the multiplier M2 is connected with the output end of the first maximum pooling layer and the output end of the second maximum pooling layer respectively, and the output end thereof is connected with the input end of the fourth convolution layer; the input end of the adder A1 is connected with the output end of the third convolution layer and the output end of the fourth convolution layer respectively, and the output end thereof is the output end of the time domain feature enhancement unit and the frequency domain feature enhancement unit.

[0042] A new energy automobile power supply detection system, comprising: a signal synthesis subsystem, a segmentation subsystem, a time domain mode mismatch matrix construction subsystem, a frequency domain mode mismatch matrix construction subsystem, an evolution matrix construction subsystem and a fault classification subsystem;

[0043] The signal synthesis subsystem is used for acquiring three-phase voltages of a three-phase inverter in an automobile power supply, taking absolute values of each voltage value in the three-phase voltages, and adding the absolute values of the three-phase voltages at the same time to obtain a three-phase synthesis signal;

[0044] The segmentation subsystem is used for segmenting the three-phase synthesis signal according to a cycle length to obtain a plurality of synthesis signal frames;

[0045] The time-domain shape mismatch matrix construction subsystem is used for dividing each synthesized signal frame into a left signal segment and a right signal segment, extracting a shape feature vector, obtaining a time-domain shape mismatch value based on a difference between left and right shape feature vectors, and constructing a synthesized signal time-domain shape mismatch matrix;

[0046] The frequency-domain shape mismatch matrix construction subsystem is used for respectively performing Fourier transform on the left signal segment and the right signal segment, calculating a frequency-domain shape mismatch value based on a difference between the left and right in a fundamental frequency and a difference between the left and right in a non-fundamental frequency gravity center, and constructing a synthesized signal frequency-domain shape mismatch matrix;

[0047] The evolution matrix construction subsystem is used for constructing a time-domain shape mismatch evolution matrix according to the time-domain shape mismatch value, and constructing a frequency-domain shape mismatch evolution matrix according to the frequency-domain shape mismatch value;

[0048] The fault classification subsystem is used for processing the synthesized signal time-domain shape mismatch matrix, the synthesized signal frequency-domain shape mismatch matrix, the time-domain shape mismatch evolution matrix and the frequency-domain shape mismatch evolution matrix by using a dual-domain collaborative enhancement neural network to obtain a power supply fault score.

[0049] The present application has the following advantages:

[0050] 1. The present application first takes absolute values of three-phase voltages output by a three-phase inverter, sums the absolute values, generates a three-phase synthesized signal, and then evaluates signal distortion based on shape features of each period of the synthesized signal. The synthesized signal effectively amplifies fault features by integrating the correlation information of the three-phase voltages, so that the fault features are more prominent and easier to capture in the face of early faults.

[0051] 2. The present application divides a synthesized signal frame into left and right segments in a period, and extracts time-domain and frequency-domain shape features. The time-domain shape mismatch value obtained based on the difference between left and right shape feature vectors can sensitively capture subtle changes in signal shape in the time dimension and effectively identify weak features of early hidden faults. In the frequency domain, the frequency-domain shape mismatch value is calculated based on the difference between the fundamental frequency and the non-fundamental frequency gravity center, realizing multi-dimensional evaluation of the signal in each period.

[0052] 3. The present application constructs time-domain and frequency-domain shape mismatch evolution matrices to further explore the signal change law from a dynamic perspective, so that the detection method can not only capture instantaneous fault features but also analyze fault development trends, thereby enhancing the ability to predict faults. Finally, a dual-domain collaborative enhancement neural network is used to fully integrate multi-source information of the time domain, the frequency domain and their evolution matrices, to realize accurate scoring of power supply faults and improve fault detection accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 It is a flowchart of a new energy automobile power supply detection method;

[0054] Figure 2 is the waveform of the ideal three-phase voltage;

[0055] Figure 3 is the waveform of the ideal three-phase synthesis signal;

[0056] Figure 4 is a schematic diagram of the three-phase synthesis signal being divided according to the cycle length;

[0057] Figure 5 is a structural schematic diagram of the dual-domain collaborative enhancement neural network;

[0058] Figure 6 is a structural schematic diagram of the time-domain feature enhancement unit and the frequency-domain feature enhancement unit;

[0059] Figure 7 is a structural schematic diagram of the dual-channel feature attention 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] Embodiment 1, as shown in a new energy vehicle power supply detection method, comprising the following steps: Figure 1

[0062] Collecting three-phase voltage of three-phase inverter in vehicle power supply, taking absolute value of each voltage value in three-phase voltage, and adding absolute values of three-phase voltage at the same time to obtain three-phase synthesis signal;

[0063] Dividing the three-phase synthesis signal according to the cycle length to obtain a plurality of synthesis signal frames;

[0064] Dividing each synthesis signal frame into left signal segment and right signal segment, and extracting morphological feature vector, obtaining time domain morphological mismatch value based on the gap between left and right morphological feature vectors, and constructing synthesis signal time domain morphological mismatch matrix;

[0065] Performing Fourier transform on the left signal segment and the right signal segment respectively, calculating the frequency domain morphological mismatch value according to the gap between the left and right at the fundamental frequency and the gap at the non-fundamental frequency gravity center, and constructing the synthesis signal frequency domain morphological mismatch matrix;

[0066] According to the time domain morphological mismatch value, constructing the time domain morphological mismatch evolution matrix, and according to the frequency domain morphological mismatch value, constructing the frequency domain morphological mismatch evolution matrix;​

[0067] The dual-domain collaborative enhancement neural network is used to process the time-domain morphology mismatch matrix of the synthetic signal, the frequency-domain morphology mismatch matrix of the synthetic signal, the time-domain morphology mismatch evolution matrix, and the frequency-domain morphology mismatch evolution matrix to obtain a power fault score.

[0068] Ideally, the waveform of the three-phase voltage is as shown in Figure 2 L1 is the waveform of the first-phase voltage, L2 is the waveform of the second-phase voltage, and L3 is the waveform of the third-phase voltage. In FIG. 2, the frequency of the three-phase voltage is 150 Hz.

[0069] In this embodiment, the expression of the three-phase synthetic signal is: where S(t) is the three-phase synthetic signal, A is the voltage effective value of the three-phase voltage, sin is a sine function, t is the time, and || is the absolute value.

[0070] The ideal waveform of the three-phase synthetic signal is as shown in Figure 3 .

[0071] The three-phase voltage (L1, L2, L3) of the three-phase inverter is interrelated and has a phase difference, and it is difficult to comprehensively reflect the running state of the three-phase as a whole by monitoring each phase voltage alone. By taking the absolute value of the three-phase voltage and adding it to obtain the three-phase synthetic signal, the information of the three-phase voltage can be integrated into one signal, and the running condition of the three-phase inverter can be evaluated as a whole. In an ideal case, the three-phase voltage is symmetrical, and the three-phase synthetic signal has a relatively regular and stable waveform (as shown in Figure 3 ). When the three-phase inverter fails, the symmetry of the three-phase voltage will be destroyed, and this destruction will be reflected as abnormal changes in the waveform of the three-phase synthetic signal. Compared with a single phase voltage, the synthetic signal can more significantly highlight the fault characteristics, making it easier to detect faults, especially early hidden faults, which helps to detect potential problems in advance and avoid further expansion of the fault.

[0072] In this embodiment, the process of constructing the time-domain morphology mismatch matrix of the synthetic signal includes:

[0073] The synthetic signal frame is divided into a left signal segment and a right signal segment;

[0074] The morphology feature vectors of the left signal segment and the right signal segment are extracted, respectively;

[0075] The time-domain morphology mismatch value is obtained according to the difference between the left signal segment and the right signal segment in the morphology feature vector;

[0076] The time-domain shape mismatch values are filled into the matrix according to the Hankel rule (in the filling process, the time-domain shape mismatch values are arranged in time sequence first, and then the matrix is filled according to the Hankel rule), to obtain a synthesized signal time-domain shape mismatch matrix.

[0077] In the embodiment, the peak value is found in each synthesized signal frame, the synthesized signal frame is cut according to the peak value, to obtain a left signal segment and a right signal segment, or the synthesized signal frame is cut into the left signal segment and the right signal segment according to the position of half of the time of the synthesized signal frame, i.e. half of the time of the period.

[0078] In the embodiment, as shown in Figure 3 , the ideal waveform of the three-phase synthesized signal fluctuates periodically. As shown in Figure 4 , the three-phase synthesized signal is cut according to the period length, Figure 4 In the embodiment, a total of 6 periods are shown, but the waveform of the first half period and the last half period is not complete, so the first half period and the last half period are removed, and the synthesized signal frame is extracted from the position with a complete period waveform, as shown in Figure 4 , a total of 5 synthesized signal frames are obtained.

[0079] The left signal segment and the right signal segment are cut according to the peak value, the time-domain shape mismatch value is obtained by calculating the difference between the shape feature vectors of the left signal segment and the right signal segment, the shape variation before and after the signal in the period is captured, and the abnormal features caused by the fault are amplified.

[0080] In the embodiment, the process of extracting the shape feature vector includes:

[0081] In the left signal segment and the right signal segment, the adjacent voltage values are subtracted, and the absolute value of the subtraction result is taken, to obtain a voltage increment;

[0082] The voltage increment discrete values are calculated for each voltage increment in the left signal segment and the right signal segment, to obtain the voltage increment discrete values of the left signal segment and the voltage increment discrete values of the right signal segment;

[0083] The energy is calculated for the left signal segment and the right signal segment, to obtain the energy of the left signal segment and the energy of the right signal segment;

[0084] The energy and the voltage increment discrete values belonging to the same signal segment are constructed into a shape feature vector.

[0085] In the embodiment, the adjacent voltage values are subtracted, i.e. the voltage value at the t+1 moment is subtracted from the voltage value at the t moment. The voltage increment discrete value is the variance or the standard deviation of each voltage increment.

[0086] The process of calculating the energy is as follows: each voltage value in the same signal segment is squared, and the squared values are added, to obtain the energy.

[0087] In an ideal case, each voltage increment discrete degree is low, and the voltage increment discrete value is calculated to quantify the discrete degree of the signal in the left and right signal segments. When a fault exists, the stability of the signal is affected, and the discrete degree changes. The energy of the signal segment is calculated in the application, and the energy distribution of the signal in different segments can be reflected. Abnormal distribution of the signal energy is caused by the fault, and the energy value can further assist in judging the type and severity of the fault, thereby enriching the dimension of the fault feature.

[0088] In the embodiment, the specific process of obtaining the time-domain shape mismatch value includes: calculating the Euclidean distance between the shape feature vector of the left signal segment and the shape feature vector of the right signal segment, and performing normalization processing to obtain the time-domain shape mismatch value.

[0089] The formula of the time-domain shape mismatch value is: wherein, ε is the time-domain shape mismatch value, E L is the energy of the shape feature vector of the left signal segment, V L is the voltage increment discrete value of the shape feature vector of the left signal segment, E R is the energy of the shape feature vector of the right signal segment, V R is the voltage increment discrete value of the shape feature vector of the right signal segment, and D is a voltage normalization parameter.

[0090] The Euclidean distance between the shape feature vectors of the left and right signal segments is calculated, the difference degree of the shape of the two is quantified, and the distortion condition of each cycle is evaluated.

[0091] In the embodiment, the process of constructing the frequency-domain shape mismatch matrix of the synthesized signal includes:

[0092] Performing Fourier transform on the left signal segment and the right signal segment respectively to obtain the left spectrum and the right spectrum;

[0093] Finding the frequency corresponding to the maximum amplitude in the left spectrum and the right spectrum respectively to obtain the fundamental frequency;

[0094] According to the other frequencies except the fundamental frequency in the left spectrum, the left non-fundamental frequency gravity center is obtained;

[0095] According to the other frequencies except the fundamental frequency in the right spectrum, the right non-fundamental frequency gravity center is obtained;

[0096] According to the difference between the left and right in the fundamental frequency and the difference between the non-fundamental frequency gravity centers, the frequency-domain shape mismatch value is calculated, the frequency-domain shape mismatch values are arranged in time sequence, and the matrix is filled according to the Hankel rule to obtain the frequency-domain shape mismatch matrix of the synthesized signal.

[0097] In the embodiment, the process of obtaining the non-fundamental frequency center of gravity comprises: removing the fundamental frequency from the spectrum, and then calculating the weighted average of all non-fundamental frequencies by taking the amplitude of each frequency as the weight, and the value is the non-fundamental frequency center of gravity.

[0098] In the actual output of the three-phase inverter, there are other frequency components in addition to the fundamental frequency. The reason is that: the switching devices (such as IGBT) of the inverter will generate switching loss and harmonic during high-frequency switching; the nonlinear characteristics of the load (such as motors, power electronic loads) will induce harmonic currents, which will in turn reflect as harmonic components of the voltage; in addition, electromagnetic interference in the circuit, device aging and other factors will also cause the generation of stray frequency components of non-fundamental frequency.

[0099] In the embodiment, the process of calculating the frequency domain shape mismatch value comprises: calculating the difference between the left fundamental frequency and the right fundamental frequency to obtain a first difference, calculating the difference between the left non-fundamental frequency center of gravity and the right non-fundamental frequency center of gravity to obtain a second difference, adding the first difference and the second difference, and normalizing to obtain the frequency domain shape mismatch value.

[0100] The formula of the frequency domain shape mismatch value is: Wherein, γ is the frequency domain shape mismatch value, f L is the left fundamental frequency, f R is the right fundamental frequency, F L is the left non-fundamental frequency center of gravity, F R is the right non-fundamental frequency center of gravity, and T is the frequency normalization parameter.

[0101] The application focuses on the fundamental frequency and the non-fundamental frequency center of gravity (focusing on harmonics, stray frequencies, etc.), and evaluates the distortion of the signal in the same period from two dimensions of main frequency stability and secondary frequency abnormal distribution.

[0102] In the embodiment, the process of constructing the time domain shape mismatch evolution matrix and the frequency domain shape mismatch evolution matrix comprises:

[0103] The ratio of the time domain shape mismatch value at other time to the first time domain shape mismatch value is taken as the time domain shape mismatch evolution ratio;

[0104] The time domain shape mismatch evolution ratios are filled in the matrix according to the Hankel rule (when filling, the time domain shape mismatch evolution ratios are arranged in chronological order first, and then the matrix is filled according to the Hankel rule), to obtain the time domain shape mismatch evolution matrix;

[0105] The ratio of the frequency domain shape mismatch value at other time to the first frequency domain shape mismatch value is taken as the frequency domain shape mismatch evolution ratio;

[0106] The frequency domain morphological mismatch evolution ratios are filled into the matrix according to the Hankel rule (when filling, the frequency domain morphological mismatch evolution ratios are arranged in time sequence first, and then the matrix is filled according to the Hankel rule), to obtain a frequency domain morphological mismatch evolution matrix.

[0107] In the embodiment, N synthesized signal frames are extracted, and then one time domain morphological mismatch value and one frequency domain morphological mismatch value are obtained for each synthesized signal frame, to obtain N-1 time domain morphological mismatch evolution ratios and N-1 frequency domain morphological mismatch values.

[0108] The application represents the relative change amplitude and development trend of the fault feature over time by calculating the ratio (evolution ratio) of the mismatch value at other time to the first mismatch value. Compared with only focusing on the instantaneous mismatch value, the evolution ratio can amplify the subtle dynamic difference caused by the early and implicit fault, and even if the absolute change of the mismatch value is weak in the early stage of the fault, the development trend can be clearly reflected through the ratio, and the perception ability for the early stage of the fault is improved.

[0109] In the embodiment, the matrix filling according to the Hankel rule is exemplified by taking a group of data values arranged in time sequence first, x1, x2, x3, x4 and x5 as an example. After the matrix is filled according to the Hankel rule, the obtained matrix is: .

[0110] In the embodiment, the size of the synthesized signal time domain morphological mismatch matrix, the synthesized signal frequency domain morphological mismatch matrix, the time domain morphological mismatch evolution matrix and the frequency domain morphological mismatch evolution matrix is consistent. When the sizes are inconsistent, 0 is used to fill the end data, so that the sizes are consistent. For example, if the matrix size is set to 12x12, the number of original time domain morphological mismatch values and frequency domain morphological mismatch values to be filled is 23, and the number of time domain morphological mismatch evolution ratios and frequency domain morphological mismatch values is 22. Therefore, after filling, one element is less at the end of the time domain morphological mismatch evolution matrix and the frequency domain morphological mismatch evolution matrix, and 0 can be used for filling.

[0111] As shown in Figure 5 , the dual-domain cooperative enhancement neural network comprises a time domain feature enhancement unit, a frequency domain feature enhancement unit, a first Concat layer, a dual-channel feature attention unit and a full connection layer.

[0112] The first input end of the time domain feature enhancement unit is used for inputting the synthesized signal time domain morphological mismatch matrix, and the second input end is used for inputting the time domain morphological mismatch evolution matrix. The first input end of the frequency domain feature enhancement unit is used for inputting the synthesized signal frequency domain morphological mismatch matrix, and the second input end is used for inputting the frequency domain morphological mismatch evolution matrix.

[0113] The input end of the first Concat layer is connected with the output end of the time domain feature enhancement unit and the output end of the frequency domain feature enhancement unit respectively, and the output end thereof is connected with the input end of the dual-channel feature attention unit; the output end of the dual-channel feature attention unit is connected with the input end of the full connection layer; and the output end of the full connection layer is the output end of the dual-domain collaborative enhancement neural network.

[0114] The time domain feature enhancement unit is adopted to fuse the time domain morphology mismatch matrix and the time domain morphology mismatch evolution matrix of the synthesized signal, so that the time domain fault feature is enhanced; the frequency domain feature enhancement unit is adopted to fuse the frequency domain morphology mismatch matrix and the frequency domain morphology mismatch evolution matrix of the synthesized signal, so that the frequency domain fault feature is enhanced; the first Concat layer is adopted for feature splicing, the dual-channel feature attention unit is adopted for dual-channel feature enhancement, and finally the full connection layer is adopted for output fault scoring.

[0115] As shown in Figure 6 The time domain feature enhancement unit and the frequency domain feature enhancement unit have the same structure, and both include a first convolution layer, a second convolution layer, a first average pooling layer, a second average pooling layer, a first maximum pooling layer, a second maximum pooling layer, a multiplier M1, a multiplier M2, a third convolution layer, a fourth convolution layer and an adder A1.

[0116] The input end of the first convolution layer is the first input end of the time domain feature enhancement unit and the frequency domain feature enhancement unit; and the input end of the second convolution layer is the second input end of the time domain feature enhancement unit and the frequency domain feature enhancement unit.

[0117] The output end of the first convolution layer is connected with the input end of the first average pooling layer and the input end of the first maximum pooling layer respectively; the output end of the second convolution layer is connected with the input end of the second average pooling layer and the input end of the second maximum pooling layer respectively; the input end of the multiplier M1 is connected with the output end of the first average pooling layer and the output end of the second average pooling layer respectively, and the output end thereof is connected with the input end of the third convolution layer; the input end of the multiplier M2 is connected with the output end of the first maximum pooling layer and the output end of the second maximum pooling layer respectively, and the output end thereof is connected with the input end of the fourth convolution layer; the input end of the adder A1 is connected with the output end of the third convolution layer and the output end of the fourth convolution layer respectively, and the output end thereof is the output end of the time domain feature enhancement unit and the frequency domain feature enhancement unit.

[0118] The application can simultaneously extract basic features of different input data by respectively processing two types of input matrices through the first and second convolution layers, respectively extract mean value features and maximum value features through the average pooling layer and the maximum pooling layer, respectively, fuse two types of mean value features through the multiplier M1 to obtain first mean value enhanced features, fuse two types of maximum value features through the multiplier M2 to obtain second maximum value enhanced features, further extract the enhanced features through the third and fourth convolution layers, and fuse them through the adder A1.

[0119] As shown in Figure 7 The double-channel feature attention unit includes a fifth convolution layer, a sixth convolution layer, a first Softmax layer, a second Softmax layer, a multiplier M3, a multiplier M4, and a second Concat layer.

[0120] The input end of the fifth convolution layer is connected with the input end of the sixth convolution layer and serves as the input end of the double-channel feature attention unit; the output end of the fifth convolution layer is connected with the input end of the first Softmax layer and the first input end of the multiplier M3; the output end of the sixth convolution layer is connected with the input end of the second Softmax layer and the first input end of the multiplier M4.

[0121] The second input end of the multiplier M3 is connected with the output end of the first Softmax layer; the second input end of the multiplier M4 is connected with the output end of the second Softmax layer.

[0122] The input end of the second Concat layer is 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 double-channel feature attention unit.

[0123] In the application, the fifth convolution layer and the sixth convolution layer process input features in parallel, then generate weights of respective features according to the first Softmax layer and the second Softmax layer, the multiplier M3 and the multiplier M4 weight the features of the corresponding convolution layer through the attention weights, realize the enhancement of the key fault features, and at the same time, suppress the redundant or interference features, and the second Concat layer fuses the two weighted features.

[0124] In the embodiment, the convolution kernel size of the first convolution layer and the second convolution layer is 1x1, the convolution kernel size of the third convolution layer and the fourth convolution layer is 3x3, the convolution kernel size of the fifth convolution layer is 3x3, and the convolution kernel size of the sixth convolution layer is 5x5.

[0125] Embodiment 2: A new energy automobile power supply detection system, comprising: a signal synthesis subsystem, a segmentation subsystem, a time domain pattern mismatch matrix construction subsystem, a frequency domain pattern mismatch matrix construction subsystem, an evolution matrix construction subsystem, and a fault classification subsystem.

[0126] The signal synthesis subsystem is used for collecting three-phase voltages of a three-phase inverter in an automobile power supply, taking absolute values of each voltage value in the three-phase voltages, and adding the absolute values of the three-phase voltages at the same time to obtain a three-phase synthesis signal;

[0127] The cutting subsystem is used for cutting the three-phase synthesis signal according to a cycle length to obtain a plurality of synthesis signal frames;

[0128] The time-domain form mismatch matrix construction subsystem is used for dividing each synthesis signal frame into a left signal segment and a right signal segment, extracting form feature vectors, obtaining time-domain form mismatch values based on differences between the left and right form feature vectors, and constructing a synthesis signal time-domain form mismatch matrix;

[0129] The frequency-domain form mismatch matrix construction subsystem is used for respectively performing Fourier transform on the left signal segment and the right signal segment, calculating frequency-domain form mismatch values according to differences between the left and right in the fundamental frequency and differences between the left and right in the non-fundamental frequency gravity center, and constructing a synthesis signal frequency-domain form mismatch matrix;

[0130] The evolution matrix construction subsystem is used for constructing a time-domain form mismatch evolution matrix according to the time-domain form mismatch values, and constructing a frequency-domain form mismatch evolution matrix according to the frequency-domain form mismatch values;

[0131] The fault classification subsystem is used for processing the synthesis signal time-domain form mismatch matrix, the synthesis signal frequency-domain form mismatch matrix, the time-domain form mismatch evolution matrix and the frequency-domain form mismatch evolution matrix by using a dual-domain cooperative enhancement neural network to obtain a power supply fault score.

[0132] The specific implementation mode of embodiment 2 is the same as that of embodiment 1.

[0133] In this embodiment, when the power supply fault score is 0-4, the power supply is fault-free, the three-phase voltages are symmetrical and stable, and the synthesis signal waveform is regular. When the power supply fault score is 4-7, it corresponds to early slight faults (such as slight aging of a phase device), and the local waveform of the synthesis signal appears slight distortion. When the power supply fault score is 7-9, it corresponds to moderate faults (such as poor contact of a switching device), and the form difference in the synthesis signal cycle is obvious. When the power supply fault score is 9-10, it corresponds to serious faults (such as short circuit of a phase), and the synthesis signal waveform is severely distorted.

[0134] The three-phase voltages output by the three-phase inverter are first taken as absolute values and then summed to generate a three-phase synthesis signal; then, based on the form features of each cycle of the synthesis signal, the signal distortion condition is evaluated, the synthesis signal integrates the related information of the three-phase voltages, can effectively amplify the fault features, makes the fault features more prominent, and is easier to capture in the face of early faults.

[0135] In the period, the left and right segments of the synthetic signal frame are segmented, and the time domain and frequency domain morphological features are extracted respectively. In the time domain, the time domain morphological mismatch value obtained based on the left and right morphological feature vector gap can sensitively capture the subtle changes of signal morphology in the time dimension, and effectively identify the weak features of early hidden faults. In the frequency domain, the frequency domain morphological mismatch value is calculated according to the fundamental frequency gap and the non-fundamental frequency frequency center gap, and the multi-dimensional evaluation of the signal of each cycle is realized.

[0136] The application constructs the time domain and frequency domain morphological mismatch evolution matrix, further excavates the signal change law from the dynamic angle, so that the detection method can not only capture the instantaneous fault features, but also analyze the fault development trend, and enhance the prediction ability of the fault. Finally, the dual-domain cooperative enhancement neural network is adopted, the multi-source information of the time domain, the frequency domain and the evolution matrix is fully fused, the accurate score of the power supply fault is realized, and the fault detection precision is improved.

[0137] The above is only the preferred embodiment of the application and is not used to limit the application. For those skilled in the art, the application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A new energy vehicle power supply detection method, characterized in that, The method comprises the following steps: Collecting three-phase voltage of three-phase inverter in automobile power supply, taking absolute value of each voltage value in three-phase voltage, and adding absolute values of three-phase voltage at the same time to obtain three-phase synthesis signal; Dividing three-phase synthesis signal according to cycle length to obtain multiple synthesis signal frames; Dividing each synthesis signal frame into left signal segment and right signal segment, and extracting shape feature vector, obtaining time domain shape mismatch value based on difference between left and right shape feature vectors, and constructing time domain shape mismatch matrix of synthesis signal; Respectively performing Fourier transform on left signal segment and right signal segment, calculating frequency domain shape mismatch value based on difference between left and right at fundamental frequency and difference between non-fundamental frequency gravity center, and constructing frequency domain shape mismatch matrix of synthesis signal; Constructing time domain shape mismatch evolution matrix according to time domain shape mismatch value, and constructing frequency domain shape mismatch evolution matrix according to frequency domain shape mismatch value; Processing time domain shape mismatch matrix of synthesis signal, frequency domain shape mismatch matrix of synthesis signal, time domain shape mismatch evolution matrix and frequency domain shape mismatch evolution matrix by using dual-domain collaborative enhancement neural network to obtain power supply fault score.

2. The new energy vehicle power supply detection method according to claim 1, characterized in that, The process of constructing time domain shape mismatch matrix of synthesis signal comprises: Dividing synthesis signal frame to obtain left signal segment and right signal segment; Respectively extracting shape feature vector of left signal segment and right signal segment; Obtaining time domain shape mismatch value based on difference between left signal segment and right signal segment in shape feature vector; Filling matrix according to Hankel rule for each time domain shape mismatch value to obtain time domain shape mismatch matrix of synthesis signal.

3. The new energy vehicle power supply detection method according to claim 1 or 2, characterized in that, The process of extracting shape feature vector comprises: Respectively subtracting adjacent voltage values in left signal segment and right signal segment, and taking absolute value of subtraction result to obtain voltage increment; Respectively calculating voltage increment discrete value of each voltage increment in left signal segment and right signal segment to obtain voltage increment discrete value of left signal segment and voltage increment discrete value of right signal segment; Respectively calculating energy of left signal segment and right signal segment; Constructing shape feature vector by using energy and voltage increment discrete value belonging to same signal segment.

4. The new energy vehicle power supply detection method according to claim 1 or 2, characterized in that, The specific process of obtaining time domain shape mismatch value comprises: calculating Euclidean distance between shape feature vector of left signal segment and shape feature vector of right signal segment, and performing normalization processing to obtain time domain shape mismatch value.

5. The new energy vehicle power supply detection method according to claim 1, characterized in that, The process of constructing frequency domain shape mismatch matrix of synthesis signal comprises: Respectively performing Fourier transform on left signal segment and right signal segment to obtain left spectrum and right spectrum; Respectively finding maximum amplitude corresponding frequency in left spectrum and right spectrum to obtain fundamental frequency; Obtaining left non-fundamental frequency gravity center based on other frequencies except fundamental frequency in left spectrum; Obtaining right non-fundamental frequency gravity center based on other frequencies except fundamental frequency in right spectrum; Calculating frequency domain shape mismatch value based on difference between left and right at fundamental frequency and difference between non-fundamental frequency gravity center, and filling matrix according to Hankel rule for each frequency domain shape mismatch value to obtain frequency domain shape mismatch matrix of synthesis signal.

6. The new energy vehicle power supply detection method according to claim 1 or 5, characterized in that, The process of calculating the frequency domain shape mismatch value comprises: calculating the difference between the fundamental frequency of the left side and the fundamental frequency of the right side to obtain a first difference, calculating the difference between the non-fundamental frequency gravity center of the left side and the non-fundamental frequency gravity center of the right side to obtain a second difference, adding the first difference and the second difference, and normalizing to obtain the frequency domain shape mismatch value.

7. The new energy vehicle power supply detection method according to claim 1, characterized in that, The process of constructing the time domain shape mismatch evolution matrix and the frequency domain shape mismatch evolution matrix comprises: The ratio of the time domain shape mismatch value at other time to the first time domain shape mismatch value is taken as the time domain shape mismatch evolution ratio; The time domain shape mismatch evolution matrix is obtained by filling the matrix according to the Hankel rule. The ratio of the frequency domain shape mismatch value at other time to the first frequency domain shape mismatch value is taken as the frequency domain shape mismatch evolution ratio; The frequency domain shape mismatch evolution matrix is obtained by filling the matrix according to the Hankel rule.

8. The new energy vehicle power supply detection method according to claim 1, characterized in that, The dual-domain collaborative enhancement neural network comprises: a time domain feature enhancement unit, a frequency domain feature enhancement unit, a first Concat layer, a dual-channel feature attention unit and a full connection layer; The first input end of the time domain feature enhancement unit is used for inputting the synthesized signal time domain shape mismatch matrix, and the second input end thereof is used for inputting the time domain shape mismatch evolution matrix; the first input end of the frequency domain feature enhancement unit is used for inputting the synthesized signal frequency domain shape mismatch matrix, and the second input end thereof is used for inputting the frequency domain shape mismatch evolution matrix; The input end of the first Concat layer is connected with the output end of the time domain feature enhancement unit and the output end of the frequency domain feature enhancement unit respectively, the output end thereof is connected with the input end of the dual-channel feature attention unit, the output end of the dual-channel feature attention unit is connected with the input end of the full connection layer, and the output end of the full connection layer is taken as the output end of the dual-domain collaborative enhancement neural network.

9. The new energy vehicle power supply detection method according to claim 8, characterized in that, The time domain feature enhancement unit and the frequency domain feature enhancement unit have the same structure, and each comprises: a first convolutional layer, a second convolutional layer, a first average pooling layer, a second average pooling layer, a first maximum pooling layer, a second maximum pooling layer, a multiplier M1, a multiplier M2, a third convolutional layer, a fourth convolutional layer and an adder A1; The input end of the first convolutional layer is taken as the first input end of the time domain feature enhancement unit and the frequency domain feature enhancement unit; and the input end of the second convolutional layer is taken as the second input end of the time domain feature enhancement unit and the frequency domain feature enhancement unit. The output end of the first convolutional layer is connected with the input end of the first average pooling layer and the input end of the first maximum pooling layer respectively; the output end of the second convolutional layer is connected with the input end of the second average pooling layer and the input end of the second maximum pooling layer respectively; the input end of the multiplier M1 is connected with the output end of the first average pooling layer and the output end of the second average pooling layer respectively, the output end thereof is connected with the input end of the third convolutional layer; the input end of the multiplier M2 is connected with the output end of the first maximum pooling layer and the output end of the second maximum pooling layer respectively, the output end thereof is connected with the input end of the fourth convolutional layer; the input end of the adder A1 is connected with the output end of the third convolutional layer and the output end of the fourth convolutional layer respectively, and the output end thereof is taken as the output end of the time domain feature enhancement unit and the frequency domain feature enhancement unit.

10. A new energy vehicle power supply detection system, based on the new energy vehicle power supply detection method of any one of claims 1-9, characterized in that, The process comprises: The signal synthesis subsystem, the segmentation subsystem, the time-domain pattern mismatch matrix construction subsystem, the frequency-domain pattern mismatch matrix construction subsystem, the evolution matrix construction subsystem and the fault classification subsystem; The signal synthesis subsystem is used for collecting three-phase voltages of a three-phase inverter in an automobile power supply, taking absolute values of each voltage value in the three-phase voltages, and adding the absolute values of the three-phase voltages at the same time to obtain a three-phase synthesis signal; The segmentation subsystem is used for segmenting the three-phase synthesis signal according to a cycle length to obtain a plurality of synthesis signal frames; The time-domain pattern mismatch matrix construction subsystem is used for dividing each synthesis signal frame into a left signal segment and a right signal segment, extracting a pattern feature vector, obtaining a time-domain pattern mismatch value based on a difference between the left and right pattern feature vectors, and constructing a synthesis signal time-domain pattern mismatch matrix; The frequency-domain pattern mismatch matrix construction subsystem is used for respectively performing Fourier transform on the left signal segment and the right signal segment, calculating a frequency-domain pattern mismatch value based on a difference between the left and right signal segments at a fundamental frequency and a difference between the left and right signal segments at a non-fundamental frequency gravity center, and constructing a synthesis signal frequency-domain pattern mismatch matrix; The evolution matrix construction subsystem is used for constructing a time-domain pattern mismatch evolution matrix according to the time-domain pattern mismatch value, and constructing a frequency-domain pattern mismatch evolution matrix according to the frequency-domain pattern mismatch value; The fault classification subsystem is used for processing the synthesis signal time-domain pattern mismatch matrix, the synthesis signal frequency-domain pattern mismatch matrix, the time-domain pattern mismatch evolution matrix and the frequency-domain pattern mismatch evolution matrix by using a dual-domain cooperative enhancement neural network to obtain a power supply fault score.

Images