A method for detecting the health status of power supply in new energy vehicles

Abstract

This invention discloses a method for detecting the health status of a new energy vehicle power supply, belonging to the field of power supply technology. The invention first extracts abnormal current signals from the phase line current signals of the three-phase inverter of the new energy vehicle power supply at different scale windows. It then extracts amplitude dispersion values ​​from the frequency domain, calculates the three-phase frequency domain cumulative values, generates frequency domain amplitude dispersion weights, and constructs a multi-scale frequency domain feature matrix and weight matrix. Next, it extracts effective abrupt change values ​​from the time domain, calculates the three-phase time domain cumulative values, generates time domain abrupt change weights, and constructs a multi-scale time domain feature matrix and weight matrix. Finally, it fuses the feature matrix and weight matrix of the same dimension, inputs them into a fault detection neural network for processing, and outputs the power supply health status. This invention improves the sensitivity and accuracy of early power supply fault identification, effectively reducing the probability of false positives and false negatives.

Description

Technical Field

[0001] This invention relates to the field of power supply technology, and more specifically to a method for detecting the health status of a power supply in a new energy vehicle. Background Technology

[0002] With the rapid development of the new energy vehicle industry, power batteries and supporting power systems, as core power units, directly affect vehicle driving safety, range, and service life through their operational stability and health status. Under the influence of long-term charge-discharge cycles, complex operating loads, and fluctuations in ambient temperature and humidity, power systems are prone to problems such as cell degradation, poor circuit contact, and inverter failure. Failure to detect these issues promptly and accurately can lead to a sudden drop in power performance, thermal runaway, or even safety accidents. Therefore, efficient detection technology for power system health status has become a key research focus in the industry.

[0003] In the prior art, there is a power supply health detection method based on single time-domain current signal analysis. This method collects the current time-domain signal in the power supply circuit, extracts characteristic parameters such as signal peak value, mean value, and variance, and combines them with preset thresholds to determine whether there is an abnormality in the power supply.

[0004] However, this existing technology has obvious technical problems: it only relies on the single-dimensional feature parameters in the time domain for analysis, which cannot capture the frequency domain anomaly information hidden in the current signal, and ignores the collaborative correlation characteristics of the current signals of each phase of the three-phase inverter. When faced with multi-scale operating condition fluctuations and early weak faults, the feature extraction is not comprehensive enough, and there is a problem of low sensitivity to early fault identification, which is prone to misjudgment or missed judgment. Summary of the Invention

[0005] In view of the above-mentioned shortcomings in the prior art, the present invention provides a method for detecting the power supply health status of new energy vehicles, which solves the problem that the prior art has low sensitivity to early fault identification and is prone to misjudgment or missed judgment.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for detecting the health status of a new energy vehicle's power supply, comprising the following steps:

[0007] S1. Extract abnormal current signals at different scale windows from the line current signal of each phase in the three-phase inverter of the new energy vehicle power supply.

[0008] S2. Convert the abnormal current signal to the frequency domain, extract the amplitude dispersion value at each moment, and add the corresponding amplitude dispersion values ​​of the three phases to obtain the cumulative value of the three-phase frequency domain, and construct a multi-scale frequency domain feature matrix.

[0009] S3. Generate frequency domain amplitude dispersion weights based on the difference between the maximum and minimum values ​​of the corresponding amplitude dispersion values ​​of the three phases, and construct a multi-scale frequency domain weight matrix;

[0010] S4. Extract effective abrupt change values ​​from the abnormal current signal, add the corresponding effective abrupt change values ​​of the three phases to obtain the three-phase time-domain cumulative value, and construct a multi-scale time-domain feature matrix;

[0011] S5. Based on the difference between the maximum and minimum values ​​of the three corresponding effective mutation values, generate the temporal mutation weights and construct a multi-scale temporal weight matrix.

[0012] S6. The multi-scale frequency domain feature matrix and the multi-scale frequency domain weight matrix are fused together, and the multi-scale time domain feature matrix and the multi-scale time domain weight matrix are fused together. The result is then processed through a fault detection neural network to obtain the power supply health status.

[0013] Furthermore, S1 includes the following sub-steps:

[0014] S11. Set multiple scale windows, take each current value in the line current signal of each phase as the center, use each scale window as the neighborhood range size, calculate the average current within the central neighborhood range, and obtain the average current of multiple scale windows.

[0015] S12. Subtract the average current value of the corresponding scale window from each current value to obtain the abnormal current component of the corresponding scale window.

[0016] S13. Arrange the abnormal current components in the same scale window according to their corresponding positions to obtain the abnormal current signal of the corresponding scale window.

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

[0018] S21. Perform a short-time Fourier transform on each abnormal current signal to obtain the frequency domain amplitude spectrum at each moment.

[0019] S22. Convert the frequency domain amplitude spectrum at each time step into the frequency domain energy probability;

[0020] S23. Based on the frequency domain energy probability, use Shannon entropy to calculate the amplitude dispersion value at each time step;

[0021] S24. Add the amplitude dispersion values ​​of the three-phase corresponding abnormal current signals at the same moment under the same scale window to obtain the three-phase frequency domain cumulative value of the corresponding scale window.

[0022] S25. Construct row vectors from the cumulative values ​​of each three-phase frequency domain corresponding to the same scale window to obtain the multi-scale frequency domain feature matrix.

[0023] Furthermore, S3 includes the following sub-steps:

[0024] S31. Under the same scale window, select the maximum and minimum amplitude dispersion values ​​of the three corresponding abnormal current signals at the same time to obtain the maximum amplitude dispersion value and the minimum amplitude dispersion value.

[0025] S32. Subtract the maximum amplitude dispersion value from the minimum amplitude dispersion value to obtain the amplitude dispersion difference;

[0026] S33. Perform nonlinear mapping on the amplitude dispersion difference to obtain the amplitude dispersion coefficient;

[0027] S34. Subtract the amplitude dispersion coefficient from 1 to obtain the frequency domain amplitude dispersion weight for the corresponding scale window;

[0028] S35. Distribute the weights of the frequency domain amplitudes corresponding to the same scale window into row vectors to obtain the multi-scale frequency domain weight matrix.

[0029] Furthermore, S4 includes the following sub-steps:

[0030] S41. Calculate the time-domain abrupt change value of the abnormal current signal at each moment;

[0031] S42. Add up the abnormal current components in the neighborhood range at each time step of the abnormal current signal and take the average value to obtain the average value of the abnormal current in the neighborhood.

[0032] S43. Multiply the mean of the abnormal current in the neighborhood by the component of the abnormal current at the center, and use the sign function to calculate the trend direction coefficient.

[0033] S44. Calculate the effective mutation value based on the time-domain mutation value and the trend direction coefficient;

[0034] S45. Add the effective abrupt change values ​​of the three-phase corresponding abnormal current signals at the same moment under the same scale window to obtain the three-phase time domain cumulative value of the corresponding scale window.

[0035] S46. Construct row vectors from the cumulative values ​​of each three-phase time domain corresponding to the same scale window to obtain the multi-scale time domain feature matrix.

[0036] Furthermore, the specific process of S41 includes: adopting the first The abnormal current component at time t and the first Subtracting the abnormal current components at time t from each other yields the difference between adjacent abnormal current components. The ratio of this difference to the standard abnormal current component difference is used as the enhancement coefficient. This enhancement coefficient is then compared with the... Multiplying the abnormal current components at time t, we get the first... The time-domain abrupt change value at time t. The time number;

[0037] The specific process of S43 includes: multiplying the mean of the neighborhood abnormal current with the component of the central abnormal current to obtain the multiplication result; when the multiplication result is greater than or equal to 0, the trend direction coefficient is assigned a value of 1 using a sign function; when the multiplication result is less than 0, the trend direction coefficient is assigned a value of -1 using a sign function.

[0038] The specific process of S44 includes: adding the trend direction coefficient to 1 to obtain the trend correction weight, multiplying the trend correction weight by the time-domain mutation value to obtain the effective mutation value.

[0039] Furthermore, S5 includes the following sub-steps:

[0040] S51. Under the same scale window, select the maximum and minimum effective mutation values ​​of the three corresponding abnormal current signals at the same time to obtain the maximum effective mutation value and the minimum effective mutation value.

[0041] S52. The mutation difference is obtained by subtracting the maximum effective mutation value from the minimum effective mutation value.

[0042] S53. Perform a nonlinear mapping on the mutation difference to obtain the mutation coefficient;

[0043] S54. Subtract the mutation coefficient from 1 to obtain the temporal mutation weight of the corresponding scale window;

[0044] S55. Construct row vectors from the time-domain abrupt change weights corresponding to the same scale window to obtain the multi-scale time-domain weight matrix.

[0045] Furthermore, S6 includes the following sub-steps:

[0046] S61. Multiply the multi-scale frequency domain feature matrix and the multi-scale frequency domain weight matrix element by element to obtain the frequency domain enhanced feature matrix;

[0047] S62. Multiply the multi-scale temporal feature matrix and the multi-scale temporal weight matrix element by element to obtain the temporal enhancement feature matrix;

[0048] S63. The time-domain enhanced feature matrix and the frequency-domain enhanced feature matrix are processed by a fault detection neural network to obtain the power supply health status.

[0049] Furthermore, S63 includes the following sub-steps:

[0050] S631. The frequency domain enhanced feature matrix is ​​processed through the frequency domain processing sub-network to obtain the frequency domain feature code;

[0051] S632. The temporal enhanced feature matrix is ​​processed through a temporal processing sub-network to obtain the temporal feature code;

[0052] S633. The frequency domain feature code and the time domain feature code are added together by adder A1 to obtain the fused feature;

[0053] S634. Attention-weighted fusion features are obtained by using dual-pooling enhancement units to perform attention-weighted fusion features.

[0054] S635. Deep features are extracted by using a deep feature extraction unit to extract deep features from attention-weighted fusion features, resulting in deep fusion features;

[0055] S636. A fully connected layer is used to classify the deep fusion features to obtain the power supply health status.

[0056] Furthermore, the dual-pooling enhancement unit in S634 includes: a Maxpool layer, an Avgpool layer, a Concat layer, a first Conv layer, a Sigmoid layer, and a multiplier M1;

[0057] The input of the Maxpool layer is connected to the input of the Avgpool layer and used to input fused features;

[0058] The input of the Concat layer is connected to the output of the Maxpool layer and the output of the Avgpool layer, respectively, and its output is connected to the input of the first Conv layer and the first input of the multiplier M1, respectively.

[0059] The output of the first Conv layer is connected to the input of the Sigmoid layer;

[0060] The second input of multiplier M1 is connected to the output of the Sigmoid layer, and its output is used to output attention-weighted fusion features.

[0061] The beneficial effects of this invention are as follows:

[0062] 1. This invention targets the current signals of each phase line of a three-phase inverter, simultaneously covering both the time and frequency domains. It captures hidden frequency domain anomalies through frequency domain conversion and retains core time domain features through time domain mutation analysis. Combined with window processing of different scales, it fully explores weak features and related information related to the health status of the power supply, thus solving the problem of incomplete feature extraction in existing technologies.

[0063] 2. This invention relies on the construction of a targeted weight matrix to enhance the identification of key features and adapt to the synergistic correlation characteristics of three-phase currents. The frequency domain amplitude dispersion weight and the time domain abrupt change weight are generated based on the extreme differences of the corresponding parameters of the three phases, so that the weight allocation can accurately match the differences and synergistic relationships of each phase signal, avoid invalid feature interference, improve the effectiveness of feature data, and improve the accuracy of capturing early weak fault signals.

[0064] 3. This invention achieves the enhancement of key features and the weakening of redundant features by fusing dual-domain feature-weight matrix. Combined with the data analysis capabilities of fault detection neural network, it effectively solves the problems of low sensitivity and easy misjudgment and omission in the early fault identification of existing technologies. It can accurately identify early weak faults in the power supply, provide accurate assessment of the power supply health status, and ensure the safety and stability of the power supply system of new energy vehicles. Attached Figure Description

[0065] Figure 1 A flowchart of a method for detecting the power supply health status of a new energy vehicle;

[0066] Figure 2 This is a schematic diagram of the fault detection neural network structure;

[0067] Figure 3 This is a schematic diagram of the structure of the frequency domain processing subnetwork and the time domain processing subnetwork;

[0068] Figure 4 This is a schematic diagram of the dual-pooling enhancement unit.

[0069] Figure 5 This is a schematic diagram of the deep feature extraction unit. Detailed Implementation

[0070] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0071] like Figure 1 As shown, a method for detecting the health status of a new energy vehicle's power supply includes the following steps:

[0072] S1. Extract abnormal current signals at different scale windows from the line current signal of each phase in the three-phase inverter of the new energy vehicle power supply.

[0073] S2. Convert the abnormal current signal to the frequency domain, extract the amplitude dispersion value at each moment, and add the corresponding amplitude dispersion values ​​of the three phases to obtain the cumulative value of the three-phase frequency domain, and construct a multi-scale frequency domain feature matrix.

[0074] S3. Generate frequency domain amplitude dispersion weights based on the difference between the maximum and minimum values ​​of the corresponding amplitude dispersion values ​​of the three phases, and construct a multi-scale frequency domain weight matrix;

[0075] S4. Extract effective abrupt change values ​​from the abnormal current signal, add the corresponding effective abrupt change values ​​of the three phases to obtain the three-phase time-domain cumulative value, and construct a multi-scale time-domain feature matrix;

[0076] S5. Based on the difference between the maximum and minimum values ​​of the three corresponding effective mutation values, generate the temporal mutation weights and construct a multi-scale temporal weight matrix.

[0077] S6. The multi-scale frequency domain feature matrix and the multi-scale frequency domain weight matrix are fused together, and the multi-scale time domain feature matrix and the multi-scale time domain weight matrix are fused together. The result is then processed through a fault detection neural network to obtain the power supply health status.

[0078] In this embodiment, S1 includes the following sub-steps:

[0079] S11. Set multiple scale windows, take each current value in the line current signal of each phase as the center, use each scale window as the neighborhood range size, calculate the average current within the central neighborhood range, and obtain the average current of multiple scale windows.

[0080] S12. Subtract the average current value of the corresponding scale window from each current value to obtain the abnormal current component of the corresponding scale window.

[0081] S13. Arrange the abnormal current components in the same scale window according to their corresponding positions to obtain the abnormal current signal of the corresponding scale window.

[0082] In this embodiment, three scale windows are selected: the first scale window has 5 current sampling points, the second scale window has 9 current sampling points, and the third scale window has 15 current sampling points. The scale window includes a center value and multiple neighboring current values.

[0083] In this embodiment, the expression for the abnormal current component is obtained as follows:

[0084] ,

[0085] in, for The phase line current signal in The first obtained under the scale window abnormal current components at all times for Phase line current signal Current value at time, For Centered and in The average current obtained under various scale windows. Take 1, 2, and 3 in sequence, corresponding to phases A, B, and C. The numbers 1, 2, and 3 are taken sequentially to correspond to the first, second, and third scale windows, respectively.

[0086] Therefore, S13 will obtain the abnormal current signals of the three scale windows corresponding to A, the three scale windows corresponding to B, and the three scale windows corresponding to C.

[0087] This invention utilizes three window scales (5, 9, and 15) to capture current fluctuation characteristics across different time spans. Smaller windows identify instantaneous current surges, while larger windows capture long-term current drift, effectively covering the signal manifestations of early, weak faults at different time scales and avoiding the omission of subtle anomalies by a single scale. Furthermore, this invention directly quantifies the degree to which the current deviates from normal fluctuations at each moment by using the difference between the current value and the mean of its corresponding scale's neighborhood, enabling the precise extraction of minute anomalies hidden within stable currents.

[0088] In this embodiment, S2 includes the following sub-steps:

[0089] S21. Perform a short-time Fourier transform on each abnormal current signal to obtain the frequency domain amplitude spectrum at each moment.

[0090] S22. Convert the frequency domain amplitude spectrum at each time step into the frequency domain energy probability;

[0091] S23. Based on the frequency domain energy probability, use Shannon entropy to calculate the amplitude dispersion value at each time step;

[0092] S24. Add the amplitude dispersion values ​​of the three-phase corresponding abnormal current signals at the same moment under the same scale window to obtain the three-phase frequency domain cumulative value of the corresponding scale window.

[0093] S25. Construct row vectors from the cumulative values ​​of each three-phase frequency domain corresponding to the same scale window to obtain the multi-scale frequency domain feature matrix.

[0094] In this embodiment, the formula for calculating the frequency domain energy probability is:

[0095] ,

[0096] in, For the first Frequency domain energy probability at a given frequency point The first value in the frequency domain amplitude spectrum The amplitude at each frequency point The frequency point number, This represents the number of frequency points in the frequency domain amplitude spectrum.

[0097] In this embodiment, the formula for calculating the amplitude dispersion is:

[0098] ,

[0099] in, This represents the amplitude dispersion value.

[0100] This invention converts abnormal current signals to the frequency domain using short-time Fourier transform, enabling the capture of abnormal fluctuations that are difficult to detect in the time domain. The frequency domain energy probability formula transforms the frequency domain amplitude into a relative energy proportion, and combined with the amplitude dispersion value calculated by Shannon entropy, quantifies the uniformity of frequency domain energy distribution. When an early power supply failure occurs, it causes a significant change in the amplitude dispersion value.

[0101] This invention integrates the frequency domain anomaly characteristics of three-phase currents by accumulating the amplitude dispersion values ​​of the three phases at the same time.

[0102] The multi-scale frequency domain feature matrix is: ,in, 3 lines The multi-scale frequency domain feature matrix of the column, A row vector is constructed for each of the three-phase frequency domain cumulative values ​​corresponding to the first scale window. A row vector is constructed for each of the three-phase frequency domain cumulative values ​​corresponding to the second scale window. A row vector is constructed for each of the three-phase frequency domain cumulative values ​​corresponding to the third scale window. For transpose operation, It is a positive integer.

[0103] In this embodiment, S3 includes the following sub-steps:

[0104] S31. Under the same scale window, select the maximum and minimum amplitude dispersion values ​​of the three corresponding abnormal current signals at the same time to obtain the maximum amplitude dispersion value and the minimum amplitude dispersion value.

[0105] S32. Subtract the maximum amplitude dispersion value from the minimum amplitude dispersion value to obtain the amplitude dispersion difference;

[0106] S33. Perform nonlinear mapping on the amplitude dispersion difference to obtain the amplitude dispersion coefficient in the range of [0~1].

[0107] S34. Subtract the amplitude dispersion coefficient from 1 to obtain the frequency domain amplitude dispersion weight for the corresponding scale window;

[0108] S35. Distribute the weights of the frequency domain amplitudes corresponding to the same scale window into row vectors to obtain the multi-scale frequency domain weight matrix.

[0109] In this embodiment, in S31, , ,in, for The abnormal current signal of the phase in the first The first scale window Amplitude dispersion at time 10:00 For the first The first scale window Maximum amplitude dispersion at time 10:00 For the first The first scale window Minimum amplitude dispersion at time 1 / 2 In various Take the maximum value below the given value. In various Take the minimum value below the given value.

[0110] The formula for calculating the frequency domain amplitude dispersion weight is: ,in, For the first Type of scale window Dispersion weights of frequency domain amplitude at time step.

[0111] The multi-scale frequency domain weight matrix is: ,in, 3 lines The multi-scale frequency domain weight matrix of the column, A row vector is constructed for the frequency domain amplitude dispersion weights corresponding to the first scale window. The row vector is constructed for the frequency domain amplitude dispersion weights corresponding to the second scale window. The row vector is formed by dispersing the frequency domain amplitude weights corresponding to the third scale window.

[0112] This invention quantifies the degree of coordination among the three phases by using frequency domain amplitude dispersion weights. When the frequency domain amplitude dispersion weights are closer to 1, the changes in the three phases are consistent.

[0113] In this embodiment, S4 includes the following sub-steps:

[0114] S41. Calculate the time-domain abrupt change value of the abnormal current signal at each moment;

[0115] S42. Add up the abnormal current components in the neighborhood range at each time step of the abnormal current signal and take the average value to obtain the average value of the abnormal current in the neighborhood.

[0116] S43. Multiply the mean of the abnormal current in the neighborhood by the component of the abnormal current at the center, and use the sign function to calculate the trend direction coefficient.

[0117] S44. Calculate the effective mutation value based on the time-domain mutation value and the trend direction coefficient;

[0118] S45. Add the effective abrupt change values ​​of the three-phase corresponding abnormal current signals at the same moment under the same scale window to obtain the three-phase time domain cumulative value of the corresponding scale window.

[0119] S46. Construct row vectors from the cumulative values ​​of each three-phase time domain corresponding to the same scale window to obtain the multi-scale time domain feature matrix.

[0120] In this embodiment, the specific process of S41 includes: using the first The abnormal current component at time t and the first Subtracting the abnormal current components at time t from each other yields the difference between adjacent abnormal current components. The ratio of this difference to the standard abnormal current component difference is used as the enhancement coefficient. This enhancement coefficient is then compared with the... Multiplying the abnormal current components at time t, we get the first... The time-domain abrupt change value at time t. The time number: ,in, For the first The time-domain abrupt change value at time t. For the first Abnormal current components at time, For the first Abnormal current components at time, The difference in standard abnormal current components. This is the time number.

[0121] In this embodiment, the standard abnormal current component difference is the statistical mean of the differences between adjacent abnormal current components calculated from the abnormal current component sample set within a preset time period when the three-phase inverter of the new energy vehicle power supply is in normal and healthy operating conditions.

[0122] The formula for calculating the average neighborhood abnormal current in S42 is: ,in, For the first The mean of the neighborhood abnormal current at time 1. For the first Abnormal current component at time Within the neighborhood of One abnormal current component This represents the number of abnormal current components within the neighborhood. This is the number of the abnormal current component. The time number, The value is 5.

[0123] The specific process of S43 includes: multiplying the mean of the neighborhood abnormal current with the component of the central abnormal current to obtain the multiplication result; when the multiplication result is greater than or equal to 0, assigning a value of 1 to the trend direction coefficient using a sign function; when the multiplication result is less than 0, assigning a value of -1 to the trend direction coefficient using a sign function. ,in, For the first The trend direction coefficient at any given time. The value is 1 when the value inside the parentheses is greater than or equal to 0, and -1 when the value inside the parentheses is less than 0.

[0124] The specific process of S44 includes: adding the trend direction coefficient to 1 to obtain the trend correction weight, and multiplying the trend correction weight by the time-domain abrupt change value to obtain the effective abrupt change value. ,in, For the first The effective mutation value at time t, For the first The time-domain abrupt change value at time t.

[0125] The multi-scale time-domain feature matrix is ​​as follows: ,in, 3 lines The multi-scale time-domain feature matrix of the column, A row vector is constructed for each of the three-phase time-domain cumulative values ​​corresponding to the first scale window. A row vector is constructed for each of the three-phase time-domain cumulative values ​​corresponding to the second scale window. The row vector is formed for each of the three-phase time-domain cumulative values ​​corresponding to the third scale window.

[0126] This invention extracts multi-dimensional features from abnormal current signals in the time domain. First, it calculates the time-domain abrupt change value by combining the difference between adjacent time points with the enhancement coefficient, accurately capturing the instantaneous abrupt change characteristics of the current signal. Then, it obtains the trend direction coefficient by using the mean of the abnormal current in the neighborhood and the sign function, achieving effective discrimination of the current abrupt change trend. Furthermore, it combines the two to calculate the effective abrupt change value, screening out effective abrupt change features with actual trend significance and eliminating invalid abrupt change interference without trend. At the same time, it achieves the synergistic integration of three-phase time-domain features by calculating the three-phase time-domain cumulative value.

[0127] In this embodiment, S5 includes the following sub-steps:

[0128] S51. Under the same scale window, select the maximum and minimum effective mutation values ​​of the three corresponding abnormal current signals at the same time to obtain the maximum effective mutation value and the minimum effective mutation value.

[0129] S52. The mutation difference is obtained by subtracting the maximum effective mutation value from the minimum effective mutation value.

[0130] S53. Perform nonlinear mapping on the mutation difference to obtain the mutation coefficient in the range of [0~1];

[0131] S54. Subtract the mutation coefficient from 1 to obtain the temporal mutation weight of the corresponding scale window;

[0132] S55. Construct row vectors from the time-domain abrupt change weights corresponding to the same scale window to obtain the multi-scale time-domain weight matrix.

[0133] In this embodiment, in S51, , ,in, for The abnormal current signal of the phase in the first The first scale window The effective mutation value at time t, For the first The first scale window The maximum effective mutation value at time t. For the first The first scale window The minimum effective mutation value at time 1. In various Take the maximum value below the given value. In various Take the minimum value below the given value.

[0134] The formula for calculating the temporal abrupt change weight is: ,in, For the first Type of scale window The temporal change weight at time step.

[0135] The multi-scale time-domain weight matrix is ​​as follows: ,in, 3 lines The multi-scale time-domain weight matrix of the column, A row vector is constructed for each temporal abrupt change weight corresponding to the first scale window. A row vector is constructed for each temporal abrupt change weight corresponding to the second scale window. The row vector is formed for each temporal mutation weight corresponding to the third scale window.

[0136] This invention achieves a quantitative assessment of the degree of coordination of the three-phase time-domain features by calculating the extreme difference of the effective mutation values ​​of the three phases at the same scale and time. Then, the difference is transformed into a time-domain mutation weight in the [0,1] interval through nonlinear mapping. The weight can accurately reflect the coordination state of the three-phase time-domain features.

[0137] In this embodiment, S6 includes the following sub-steps:

[0138] S61. Multiply the multi-scale frequency domain feature matrix and the multi-scale frequency domain weight matrix element by element to obtain the frequency domain enhanced feature matrix;

[0139] S62. Multiply the multi-scale temporal feature matrix and the multi-scale temporal weight matrix element by element to obtain the temporal enhancement feature matrix;

[0140] S63. The time-domain enhanced feature matrix and the frequency-domain enhanced feature matrix are processed by a fault detection neural network to obtain the power supply health status.

[0141] like Figure 2 As shown, the fault detection neural network includes: a frequency domain processing subnetwork, a time domain processing subnetwork, an adder A1, a double pooling enhancement unit, a deep feature extraction unit, and a fully connected layer.

[0142] In this embodiment, S63 includes the following sub-steps:

[0143] S631. The frequency domain enhanced feature matrix is ​​processed through the frequency domain processing sub-network to obtain the frequency domain feature code;

[0144] S632. The temporal enhanced feature matrix is ​​processed through a temporal processing sub-network to obtain the temporal feature code;

[0145] S633. The frequency domain feature code and the time domain feature code are added together by adder A1 to obtain the fused feature;

[0146] S634. Attention-weighted fusion features are obtained by using dual-pooling enhancement units to perform attention-weighted fusion features.

[0147] S635. Deep features are extracted by using a deep feature extraction unit to extract deep features from attention-weighted fusion features, resulting in deep fusion features;

[0148] S636. A fully connected layer is used to classify the deep fusion features to obtain the power supply health status.

[0149] This invention achieves refined extraction and representation of frequency and time domain features by separately encoding the corresponding enhanced feature matrices through frequency and time domain processing sub-networks, preserving the unique information of features in different dimensions. Then, an adder is used to fuse the frequency and time domain feature codes, achieving effective integration of basic features in both domains. Combined with the attention weighting of the dual-pooling enhancement unit, key fused features strongly correlated with power supply health status can be adaptively highlighted, redundant information is weakened, and the effectiveness of features is improved. Subsequently, a deep feature extraction unit is used to mine the deep correlation information of the fused features, further enhancing the representational ability of the features. Finally, classification is completed through a fully connected layer.

[0150] like Figure 3As shown, the frequency domain processing subnetwork and the time domain processing subnetwork have the same structure, both including: a first Conv block, a second Conv block, and a third Conv block. The convolution kernel size of the first Conv block is 1×1, and the convolution kernel size of the second Conv block and the third Conv block is 1×3.

[0151] like Figure 4 As shown, the dual-pooling enhancement unit in S634 includes: a Maxpool layer, an Avgpool layer, a Concat layer, a first Conv layer, a Sigmoid layer, and a multiplier M1;

[0152] The input of the Maxpool layer is connected to the input of the Avgpool layer and used to input fused features;

[0153] The input of the Concat layer is connected to the output of the Maxpool layer and the output of the Avgpool layer, respectively, and its output is connected to the input of the first Conv layer and the first input of the multiplier M1, respectively.

[0154] The output of the first Conv layer is connected to the input of the Sigmoid layer;

[0155] The second input of multiplier M1 is connected to the output of the Sigmoid layer, and its output is used to output attention-weighted fusion features.

[0156] The kernel size of the first Conv layer is 1×1.

[0157] The dual-pooling enhancement unit processes the fused features in parallel through the Maxpool and Avgpool layers, capturing peak and overall average information from the features respectively. After integration by the Concat layer, the multi-dimensional statistical characteristics of the fused features are preserved. Attention weights are then generated through the first Conv layer and the Sigmoid layer, which can adaptively identify and highlight key features strongly correlated with the power supply health status. Finally, attention weighting is completed through a multiplier, which effectively enhances the representational ability of key features, weakens the interference of redundant information, and improves the efficiency and accuracy of subsequent deep feature extraction. This provides more discriminative fused feature inputs for accurate judgment of power supply health status.

[0158] like Figure 5 As shown, the deep feature extraction unit includes: a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, and a Concat layer; the convolutional kernels of the second and fourth convolutional layers are 1×1, and the convolutional kernels of the third, fifth, and sixth convolutional layers are 3×3.

[0159] In this embodiment, the power supply health status includes: healthy, slightly abnormal, moderately abnormal, and severely abnormal.

[0160] This invention extracts input features differentially through two parallel convolutional branches, and then combines the features of the two branches through a Concat layer, which not only preserves the detailed information of shallow basic features, but also incorporates deep complex features.

[0161] This invention overcomes the shortcomings of existing technologies that rely solely on single time-domain features and neglect frequency-domain information and three-phase collaborative characteristics by fusing multi-scale time-domain and frequency-domain dual-dimensional features and combining them with the synergistic correlation analysis of three-phase currents. It highlights key anomaly features through weight enhancement and significantly improves the sensitivity of identifying early, subtle faults by deeply fusing dual-domain, multi-scale features through neural networks, effectively avoiding false positives and false negatives.

[0162] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A new energy vehicle power supply health state detection method, characterized in that, Includes the following steps: S1. Extract abnormal current signals at different scale windows from the line current signal of each phase in the three-phase inverter of the new energy vehicle power supply. S2. Convert the abnormal current signal to the frequency domain, extract the amplitude dispersion value at each moment, and add the corresponding amplitude dispersion values ​​of the three phases to obtain the cumulative value of the three-phase frequency domain, and construct a multi-scale frequency domain feature matrix. S3. Generate frequency domain amplitude dispersion weights based on the difference between the maximum and minimum values ​​of the corresponding amplitude dispersion values ​​of the three phases, and construct a multi-scale frequency domain weight matrix; S4. Extract effective abrupt change values ​​from the abnormal current signal, add the corresponding effective abrupt change values ​​of the three phases to obtain the three-phase time-domain cumulative value, and construct a multi-scale time-domain feature matrix; S5. Based on the difference between the maximum and minimum values ​​of the three corresponding effective mutation values, generate the temporal mutation weights and construct a multi-scale temporal weight matrix. S6. The multi-scale frequency domain feature matrix is ​​fused with the multi-scale frequency domain weight matrix, and the multi-scale time domain feature matrix is ​​fused with the multi-scale time domain weight matrix. The result is then processed by a fault detection neural network to obtain the power supply health status. S4 includes the following steps: S41. Calculate the time-domain abrupt change value of the abnormal current signal at each moment; S42. Add up the abnormal current components in the neighborhood range at each time step of the abnormal current signal and take the average value to obtain the average value of the abnormal current in the neighborhood. S43. Multiply the mean of the abnormal current in the neighborhood by the component of the abnormal current at the center, and use the sign function to calculate the trend direction coefficient. S44. Calculate the effective mutation value based on the time-domain mutation value and the trend direction coefficient; S45. Add the effective abrupt change values ​​of the three-phase corresponding abnormal current signals at the same moment under the same scale window to obtain the three-phase time domain cumulative value of the corresponding scale window. S46. Construct a row vector from the cumulative values ​​of each three-phase time domain corresponding to the same scale window to obtain the multi-scale time domain feature matrix. The specific process of S41 comprises: The abnormal current component at the first time point is subtracted from the abnormal current component at the second time point to obtain a difference between adjacent abnormal current components. The difference between adjacent abnormal current components is divided by a standard difference between abnormal current components to obtain an enhancement coefficient. The enhancement coefficient is multiplied by the abnormal current component at the first time point to obtain a time-domain mutation value at the second time point. t is the number of the time point; and t is the number of the time point. The specific process of S43 includes: multiplying the mean of the neighborhood abnormal current with the component of the central abnormal current to obtain the multiplication result; when the multiplication result is greater than or equal to 0, the trend direction coefficient is assigned a value of 1 using a sign function; when the multiplication result is less than 0, the trend direction coefficient is assigned a value of -1 using a sign function. The specific process of S44 includes: adding the trend direction coefficient to 1 to obtain the trend correction weight, multiplying the trend correction weight by the time-domain mutation value to obtain the effective mutation value.

2. The method for detecting the power supply health status of new energy vehicles according to claim 1, characterized in that, S1 includes the following steps: S11. Set multiple scale windows, take each current value in the line current signal of each phase as the center, use each scale window as the neighborhood range size, calculate the average current within the central neighborhood range, and obtain the average current of multiple scale windows. S12. Subtract the average current value of the corresponding scale window from each current value to obtain the abnormal current component of the corresponding scale window. S13. Arrange the abnormal current components in the same scale window according to their corresponding positions to obtain the abnormal current signal of the corresponding scale window.

3. The method for detecting the power supply health status of new energy vehicles according to claim 1, characterized in that, S2 includes the following steps: S21. Perform a short-time Fourier transform on each abnormal current signal to obtain the frequency domain amplitude spectrum at each moment. S22. Convert the frequency domain amplitude spectrum at each time step into the frequency domain energy probability; S23. Based on the frequency domain energy probability, use Shannon entropy to calculate the amplitude dispersion value at each time step; S24. Add the amplitude dispersion values ​​of the three-phase corresponding abnormal current signals at the same moment under the same scale window to obtain the three-phase frequency domain cumulative value of the corresponding scale window. S25. Construct row vectors from the cumulative values ​​of each three-phase frequency domain corresponding to the same scale window to obtain the multi-scale frequency domain feature matrix.

4. The method for detecting the power supply health status of new energy vehicles according to claim 1, characterized in that, S3 includes the following steps: S31. Under the same scale window, select the maximum and minimum amplitude dispersion values ​​of the three corresponding abnormal current signals at the same time to obtain the maximum amplitude dispersion value and the minimum amplitude dispersion value. S32. Subtract the maximum amplitude dispersion value from the minimum amplitude dispersion value to obtain the amplitude dispersion difference; S33. Perform nonlinear mapping on the amplitude dispersion difference to obtain the amplitude dispersion coefficient; S34. Subtract the amplitude dispersion coefficient from 1 to obtain the frequency domain amplitude dispersion weight for the corresponding scale window; S35. Distribute the weights of the frequency domain amplitudes corresponding to the same scale window into row vectors to obtain the multi-scale frequency domain weight matrix.

5. The method for detecting the power supply health status of new energy vehicles according to claim 1, characterized in that, S5 includes the following steps: S51. Under the same scale window, select the maximum and minimum effective mutation values ​​of the three corresponding abnormal current signals at the same time to obtain the maximum effective mutation value and the minimum effective mutation value. S52. The mutation difference is obtained by subtracting the maximum effective mutation value from the minimum effective mutation value. S53. Perform a nonlinear mapping on the mutation difference to obtain the mutation coefficient; S54. Subtract the mutation coefficient from 1 to obtain the temporal mutation weight of the corresponding scale window; S55. Construct row vectors from the time-domain abrupt change weights corresponding to the same scale window to obtain the multi-scale time-domain weight matrix.

6. The method for detecting the power supply health status of new energy vehicles according to claim 1, characterized in that, S6 includes the following sub-steps: S61. Multiply the multi-scale frequency domain feature matrix and the multi-scale frequency domain weight matrix element by element to obtain the frequency domain enhanced feature matrix; S62. Multiply the multi-scale temporal feature matrix and the multi-scale temporal weight matrix element by element to obtain the temporal enhancement feature matrix; S63. The time-domain enhanced feature matrix and the frequency-domain enhanced feature matrix are processed by a fault detection neural network to obtain the power supply health status.

7. The method for detecting the power supply health status of new energy vehicles according to claim 6, characterized in that, S63 includes the following sub-steps: S631. The frequency domain enhanced feature matrix is ​​processed through the frequency domain processing sub-network to obtain the frequency domain feature code; S632. The temporal enhanced feature matrix is ​​processed through a temporal processing sub-network to obtain the temporal feature code; S633. The frequency domain feature code and the time domain feature code are added together by adder A1 to obtain the fused feature; S634. Attention-weighted fusion features are obtained by using dual-pooling enhancement units to perform attention-weighted fusion features. S635. Deep features are extracted by using a deep feature extraction unit to extract deep features from attention-weighted fusion features, resulting in deep fusion features; S636. A fully connected layer is used to classify the deep fusion features to obtain the power supply health status.

8. The method for detecting the power supply health status of new energy vehicles according to claim 7, characterized in that, The dual-pooling enhancement unit in S634 includes: a Maxpool layer, an Avgpool layer, a Concat layer, a first Conv layer, a Sigmoid layer, and a multiplier M1; The input of the Maxpool layer is connected to the input of the Avgpool layer and is used to input fused features; The input of the Concat layer is connected to the output of the Maxpool layer and the output of the Avgpool layer, respectively, and its output is connected to the input of the first Conv layer and the first input of the multiplier M1, respectively. The output of the first Conv layer is connected to the input of the Sigmoid layer; The second input of the multiplier M1 is connected to the output of the Sigmoid layer, and its output is used to output attention-weighted fusion features.

Images