A machine learning-based method for charging pile fault diagnosis

Abstract

This invention discloses a charging pile fault diagnosis method based on machine learning, belonging to the field of machine learning technology. The method includes: collecting charging pile operating status data and generating a segmented state sequence for the charging stage; constructing a charging pile component topology diagram and performing a Chirplet transform to generate topologically coupled Chirplet atoms; performing Chirplet time-frequency decomposition to generate charging ripple time-frequency evolution features; constructing a ripple dynamic degradation feature sequence based on the charging ripple time-frequency evolution features; constructing an improved liquid time constant network to generate a continuous degradation state vector of the charging pile; and performing state matching with a preset fault prototype library to generate a charging pile fault diagnosis report. This invention, by introducing the Chirplet transform and the improved liquid time constant network, achieves dynamic modeling and accurate fault diagnosis of the continuous degradation state of charging piles under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a machine learning-based method for diagnosing charging pile faults. Background Technology

[0002] With the continuous growth of new energy vehicle ownership, charging piles, as a crucial infrastructure for new energy vehicles, directly impact charging safety and equipment lifespan through their operational stability. During long-term operation, charging piles are susceptible to factors such as high load operation, ambient temperature changes, current surges, and frequent start-stop cycles, leading to problems such as abnormal temperature rise in power modules, contactor aging, unstable contact in the charging gun, and deterioration in insulation performance. Current charging pile maintenance methods typically monitor equipment status through output voltage, output current, temperature, and alarm information, and diagnose faults based on fixed thresholds. While this method can identify obvious faults, it is weak in recognizing slow degradation states in the early stages of fault formation, easily resulting in delayed fault detection and false alarms.

[0003] Currently, some charging pile fault diagnosis methods are beginning to use machine learning models to analyze operational data, extracting features such as voltage, current, temperature, and ripple to achieve state classification and fault identification. Some methods use time-frequency analysis to process ripple signals and combine them with recurrent neural networks, convolutional neural networks, or basic liquid time constant networks to complete state prediction. However, most existing methods only perform feature analysis on single signals, failing to adequately consider the correlation changes between different components and the continuous time state evolution process. Furthermore, under complex operating conditions and frequent switching of charging stages, problems such as unstable state features, inaccurate expression of degradation trends, and decreased fault identification accuracy are prone to occur.

[0004] Therefore, how to provide a machine learning-based method for diagnosing charging pile faults is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a machine learning-based method for fault diagnosis of charging piles. This invention utilizes charging pile operating status data, ripple signal time-frequency analysis technology, topology correlation modeling technology, and an improved liquid time constant network to jointly analyze ripple frequency drift, energy fluctuations, bandwidth spread, and continuous degradation states generated during charging pile operation. It details the continuous-time degradation state modeling, ripple propagation correlation analysis, and fault state identification process under complex operating conditions of the charging pile. This enables accurate identification and state diagnosis of faults such as power module aging, gun wire contact degradation, contactor jitter, insulation attenuation, and communication anomalies. This invention possesses advantages such as strong degradation state expression capability, high adaptability to complex operating conditions, strong continuous-time state tracking capability, and high fault identification accuracy, thereby improving the operational safety and fault early warning reliability of charging piles.

[0006] A machine learning-based fault diagnosis method for charging piles according to an embodiment of the present invention includes: Collect charging pile operation status data, perform preprocessing on the charging pile operation status data, and generate a segmented state sequence for the charging stage; A topology graph of charging pile components is constructed based on the segmented state sequence of the charging stage. Chirplet transformation is performed on the power ripple signal, current ripple signal and voltage ripple signal to establish the ripple propagation correlation between the nodes of the charging pile components. The center frequency parameter, tuning frequency parameter and window width parameter of the Chirplet atom are updated by graph convolution to generate topologically coupled Chirplet atoms. Chirplet time-frequency decomposition is performed on the power ripple signal, current ripple signal and voltage ripple signal based on topologically coupled Chirplet atoms, and temperature rise coupled weight modulation is performed on the Chirplet time-frequency energy according to the temperature change of the power module and the temperature change of the gun wire to generate charging ripple time-frequency evolution characteristics. Based on the time-frequency evolution characteristics of charging ripple, the ripple frequency drift characteristics, ripple energy fluctuation characteristics, and frequency band diffusion characteristics are calculated, and a dynamic degradation characteristic sequence of ripple is constructed. An improved liquid time constant network is constructed based on the ripple dynamic degradation feature sequence. An energy conservation-Hamilton liquid block is introduced to construct the Hamiltonian energy state and execute the liquid time constant constraint. Random diffusion intensity is generated based on sampling time difference, communication state and ripple mutation features, and continuous time liquid state update is performed to generate the continuous degradation state vector of the charging pile. The charging pile's continuous degradation state vector is matched with the execution state of a preset fault prototype library to generate a charging pile fault diagnosis report.

[0007] Optionally, the charging pile operation status data specifically includes output voltage data, output current data, charging power data, power ripple signal data, current ripple signal data, voltage ripple signal data, power module temperature data, gun wire temperature data, insulation resistance data, contactor operation status data, communication status data, charging stage identification data, and collection timestamp data.

[0008] Optionally, generating the segmented state sequence of the charging stage includes: The charging pile operation status data is sorted by time according to the collected timestamp data, linear interpolation is performed on missing data, bandpass filtering is performed on power ripple signal data, current ripple signal data and voltage ripple signal data, and status encoding is performed on contactor action status data and communication status data to generate a time-aligned status sequence. Read the charging stage identifier data in the time-aligned state sequence, divide the time-aligned state sequence into stages according to different charging stages, and arrange the charging pile operation status data and sampling time difference in the corresponding stage according to the order of the collected timestamp data to generate a segmented state sequence of charging stages.

[0009] Optionally, generating topologically coupled Chirplet atoms includes: Read the power ripple signal data, current ripple signal data, voltage ripple signal data, power module temperature data, gun wire temperature data, insulation resistance data, contactor operation status data, and communication status data from the segmented state sequence of the charging stage, and construct the topology diagram of the charging pile components; The power ripple signal data, current ripple signal data, and voltage ripple signal data are divided into corresponding node ripple signals according to the component nodes. Based on the node ripple signals, the potential distribution features corresponding to the center frequency parameter, frequency modulation parameter, and window width parameter are generated. Based on the potential distribution features, the center frequency parameter, frequency modulation parameter, and window width parameter are probabilistically reconstructed to generate adaptive Chirplet atoms. Based on adaptive Chirplet atoms, a set of phylogenetic evolutionary atoms is constructed. The nodal ripple signal reconstruction error is used as the fitness index. Iterative selection, crossover and mutation are performed on the center frequency parameter, frequency modulation parameter and window width parameter to generate evolutionary Chirplet atoms. Chirplet transform is performed on the node ripple signal based on the evolved Chirplet atom pair to generate time-frequency domain response results and instantaneous phase-frequency domain response results, respectively. A hypersphere rotation time-frequency space is constructed based on the time-frequency domain response results and instantaneous phase-frequency domain response results. In the hypersphere rotation time-frequency space, a joint rotation mapping is performed on the center frequency parameter, the tuning frequency parameter and the ripple energy distribution to generate the node time-frequency energy vector. Based on the connection edges in the charging pile component topology graph, the node time-frequency energy vector similarity between corresponding component nodes is calculated, and the node time-frequency energy vector similarity is written as the ripple propagation association weight into the corresponding connection edge. Based on the ripple propagation association weights in the charging pile component topology graph, graph convolution updates are performed on the center frequency parameter, tuning frequency parameter, and window width parameter of the Chirplet atom, and topologically coupled Chirplet atoms are generated based on the updated center frequency parameter, tuning frequency parameter, and window width parameter.

[0010] Optionally, the generation of charging ripple time-frequency evolution features includes: Read the topologically coupled Chirplet atom, power ripple signal data, current ripple signal data, voltage ripple signal data, power module temperature data, gun wire temperature data, and acquisition timestamp data. Divide the power ripple signal data, current ripple signal data, and voltage ripple signal data into ripple signal segments within the corresponding stages according to the segmented state sequence of the charging stage. Randomly mask part of the power module temperature data and gun wire temperature data to generate a temperature rise mask sequence. Chirplet time-frequency decomposition is performed on each ripple signal segment based on topologically coupled Chirplet atoms to obtain the power ripple time-frequency response, current ripple time-frequency response and voltage ripple time-frequency response. Temperature rise reconstruction is then performed on the temperature rise mask sequence based on the unmasked power module temperature data and gun wire temperature data to generate the temperature rise reconstruction sequence. The temperature rise reconstruction sequence is used to calculate the temperature change of the power module and the temperature change of the gun wire between adjacent sampling windows. The temperature rise distribution is constructed based on the temperature change of the power module and the temperature change of the gun wire. The Chirplet time-frequency energy distribution is constructed based on the power ripple time-frequency response, the current ripple time-frequency response and the voltage ripple time-frequency response. The optimal transmission mapping relationship is constructed based on the Chirplet time-frequency energy distribution and the temperature rise distribution. The optimal transmission cost from the Chirplet time-frequency energy distribution to the temperature rise distribution is calculated. The temperature rise coupling weight is generated based on the optimal transmission cost. The Chirplet time-frequency energy is modulated based on the temperature rise coupling weight to generate the temperature rise coupled Chirplet time-frequency energy. Based on temperature rise coupling, Chirplet time-frequency energy extraction is used to extract the ripple main frequency position, ripple energy peak value, frequency band spread range, and stage frequency offset, and these are combined in the order of the collected timestamp data to generate charging ripple time-frequency evolution characteristics.

[0011] Optionally, constructing the ripple dynamic degradation feature sequence includes: Read the ripple dominance frequency position, ripple energy peak, frequency band spread range, stage frequency offset and acquisition timestamp data from the charging ripple time-frequency evolution characteristics. Calculate the ratio between the ripple dominance frequency position difference and the sampling time difference between adjacent sampling windows according to the acquisition timestamp data order to generate ripple frequency drift change characteristics. The ratio between the peak ripple energy difference and the sampling time difference between adjacent sampling windows is calculated according to the order of the collected timestamp data to generate ripple energy fluctuation characteristics. The ratio between the frequency band spread range difference and the sampling time difference between adjacent sampling windows is also calculated to generate frequency band spread change characteristics. The ripple frequency drift characteristics, ripple energy fluctuation characteristics, frequency band spread characteristics, and stage frequency offset corresponding to the same sampling window are combined in the order of the acquisition timestamp data to generate a ripple dynamic degradation characteristic sequence.

[0012] Optionally, generating the continuous degradation state vector of the charging pile includes: An improved liquid time constant network is constructed, comprising a ripple dynamic degradation input layer, a liquid state encoding layer, a continuous-time liquid update layer, and a multi-state fusion layer; The ripple dynamic degradation input layer performs time jitter enhancement and amplitude occlusion enhancement on the ripple dynamic degradation feature sequence within the same sampling window to generate a dual-path temporal enhancement sequence. Based on the dual-path temporal enhancement sequence, it performs temporal comparison self-supervised enhancement to generate a liquid input state sequence. The liquid state coding layer reads the liquid input state sequence, extracts the ripple dynamic degradation features of adjacent sampling windows, calculates the time decay coefficient corresponding to the sampling time difference, and performs state coupling mapping on the ripple dynamic degradation features, insulation resistance change features and contactor action state features to generate the initial liquid state sequence and ripple abrupt change features. The continuous-time liquid update layer is based on output voltage data, output current data, charging power data and ripple energy fluctuation characteristics. It introduces energy conservation-Hamilton liquid block to construct Hamiltonian energy state and executes liquid time constant constraint to generate Hamiltonian constrained liquid state sequence. The initial liquid state sequence is mapped to the observable function space, and spectral operator decomposition and exponential mapping time advancement are performed based on the Koopman spectral decomposition mechanism to generate the spectral decomposed liquid state sequence. Based on the topology of charging pile components, node diffusion connection relationships are constructed. Through graph-ODE collaborative diffusion coupling, a state differential equation containing liquid state evolution, node diffusion propagation and random disturbance diffusion is constructed. Random diffusion intensity is generated based on sampling time difference, communication state data and ripple mutation characteristics. The state differential equation is solved by continuous-time diffusion integral to generate a diffusion-coupled liquid state sequence. A symptotic integral time update block is introduced to perform split integral updates, generating a continuously degenerate liquid state sequence and a randomly diffused liquid state sequence. The multi-state fusion layer performs state splicing and weight fusion on Hamilton-constrained liquid state sequences, spectral decomposition liquid state sequences, diffusion-coupled liquid state sequences, continuously degraded liquid state sequences, and random diffusion liquid state sequences to generate a continuously degraded state vector for the charging pile. The improved liquid time constant network was trained, and the prediction error of the continuous degenerate state vector, the reconstruction error of the ripple dynamic degenerate feature sequence, and the continuous time liquid state evolution error were used as joint optimization objectives. The parameters of the ripple dynamic degenerate input layer, the liquid state encoding layer, the continuous time liquid update layer, and the multi-state fusion layer were continuously optimized. Training was stopped when the change value of the joint loss was less than the convergence threshold for five consecutive training cycles.

[0013] Optionally, generating the charging pile fault diagnosis report includes: Read the charging pile continuous degradation state vector, collect timestamp data and charging stage identification data, and read the power module aging prototype, gun line contact degradation prototype, contactor jitter prototype, insulation attenuation prototype and communication abnormality prototype in the preset fault prototype library. Each fault prototype includes a fault prototype state vector, fault type label, fault location label and fault cause label. Calculate the state matching distance between the continuous degradation state vector of the charging pile and the state vector of each fault prototype. Determine the fault prototype with the smallest state matching distance as the target fault prototype. Calculate the fault confidence based on the state matching distance. When the fault confidence is greater than or equal to 0.75, generate the corresponding fault identification result. When the fault confidence is less than 0.75, generate the fault identification result to be reviewed. A charging pile fault diagnosis report is generated based on the fault identification results, the fault identification results to be reviewed, the target fault prototype, the collected timestamp data, and the charging stage identification data. The charging pile fault diagnosis report includes the fault occurrence stage, fault status, fault type, fault location, fault cause label, and fault confidence level.

[0014] The beneficial effects of this invention are: This invention proposes a machine learning-based fault diagnosis method for charging piles. By jointly analyzing operational status data such as output voltage, output current, charging power, power ripple signal, current ripple signal, voltage ripple signal, power module temperature, and charging gun wire temperature, it achieves continuous modeling of the fault evolution process under complex operating conditions of charging piles. Compared with traditional diagnostic methods that rely on fixed thresholds or single state parameters, this invention can utilize the frequency drift, energy fluctuation, and bandwidth spread characteristics in the ripple signal to reflect the early degradation trend of internal components of the charging pile, thereby improving the identification capability of latent faults and slow degradation faults, and reducing the occurrence of missed fault detections and false alarms.

[0015] This invention constructs a topology diagram of charging pile components and establishes ripple propagation correlations, enabling state coupling changes between different components to participate in the fault analysis process. This enhances the ability to represent states under complex operating conditions. Simultaneously, by utilizing an improved liquid time constant network to dynamically update continuously degraded states, the state change process under charging phase switching, communication disturbances, and random fluctuations remains continuous and stable, improving fault state tracking capabilities and diagnostic accuracy. This invention can classify and identify faults such as power module aging, gun wire contact degradation, contactor jitter, insulation attenuation, and communication anomalies, and generate corresponding fault diagnosis results, thereby improving the operational safety of charging piles, equipment maintenance efficiency, and the reliability of fault early warning. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a machine learning-based fault diagnosis method for charging piles proposed in this invention. Figure 2 This is a functional flowchart of the Chirplet transform in a machine learning-based charging pile fault diagnosis method proposed in this invention. Figure 3 This is a schematic diagram of the structure of the improved liquid time constant network for a machine learning-based charging pile fault diagnosis method proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 , Figure 2 and Figure 3 A machine learning-based method for fault diagnosis of charging piles includes: Collect charging pile operation status data, perform preprocessing on the charging pile operation status data, and generate a segmented state sequence for the charging stage; A topology graph of charging pile components is constructed based on the segmented state sequence of the charging stage. Chirplet transformation is performed on the power ripple signal, current ripple signal and voltage ripple signal to establish the ripple propagation correlation between the nodes of the charging pile components. The center frequency parameter, tuning frequency parameter and window width parameter of the Chirplet atom are updated by graph convolution to generate topologically coupled Chirplet atoms. Chirplet time-frequency decomposition is performed on the power ripple signal, current ripple signal and voltage ripple signal based on topologically coupled Chirplet atoms, and temperature rise coupled weight modulation is performed on the Chirplet time-frequency energy according to the temperature change of the power module and the temperature change of the gun wire to generate charging ripple time-frequency evolution characteristics. Based on the time-frequency evolution characteristics of charging ripple, the ripple frequency drift characteristics, ripple energy fluctuation characteristics, and frequency band diffusion characteristics are calculated, and a dynamic degradation characteristic sequence of ripple is constructed. An improved liquid time constant network is constructed based on the ripple dynamic degradation feature sequence. An energy conservation-Hamilton liquid block is introduced to construct the Hamiltonian energy state and execute the liquid time constant constraint. Random diffusion intensity is generated based on sampling time difference, communication state and ripple mutation features, and continuous time liquid state update is performed to generate the continuous degradation state vector of the charging pile. The charging pile's continuous degradation state vector is matched with the execution state of a preset fault prototype library to generate a charging pile fault diagnosis report.

[0019] In this embodiment, the charging pile operation status data specifically includes output voltage data, output current data, charging power data, power ripple signal data, current ripple signal data, voltage ripple signal data, power module temperature data, gun wire temperature data, insulation resistance data, contactor operation status data, communication status data, charging stage identification data, and collection timestamp data.

[0020] In this embodiment, generating the segmented state sequence of the charging stage includes: The charging pile operation status data is sorted by time according to the collected timestamp data. Missing data is processed by linear interpolation. Bandpass filtering is performed on the power ripple, current ripple, and voltage ripple signal data. Status encoding is performed on the contactor operation status and communication status data to generate a time-aligned status sequence, where: Perform linear interpolation on the missing data, specifically: Read the values ​​of the previous and next sampling points in the same data field at the missing position, calculate the time interval between the two sampling points, calculate the time difference between the missing position and the previous sampling point, multiply the difference between the values ​​of the previous and next sampling points by the proportion of the time difference, and add it to the value of the previous sampling point to obtain the interpolation result corresponding to the missing position. When the length of consecutive missing points is greater than 5 sampling points, keep the missing mark and stop the interpolation calculation. The bandpass filtering process is performed as follows: Read power ripple signal data, current ripple signal data, and voltage ripple signal data. Set the lower cutoff frequency of the bandpass filter to 300Hz and the upper cutoff frequency to 30000Hz. Perform frequency filtering on each ripple signal, retain the ripple frequency components in the range of 300Hz to 30000Hz, filter out low-frequency drift components below 300Hz and high-frequency noise components above 30000Hz, and generate a filtered ripple signal sequence. The execution of status coding processing is as follows: Read the closed state, open state, and jitter state from the contactor action status data and encode them as 1, 0, and 2 respectively. Read the normal communication state, delayed communication state, and interrupted communication state from the communication status data and encode them as 1, 2, and 0 respectively. Write the encoding results into the corresponding sampling positions according to the order of the collected timestamp data to generate a status encoding sequence. Read the charging stage identifier data in the time-aligned state sequence, divide the time-aligned state sequence into stages according to different charging stages, and arrange the charging pile operation status data and sampling time difference in the corresponding stage according to the order of the collected timestamp data to generate a segmented state sequence of charging stages.

[0021] In this embodiment, the generation of topologically coupled Chirplet atoms includes: Read the power ripple signal data, current ripple signal data, voltage ripple signal data, power module temperature data, gun wire temperature data, insulation resistance data, contactor operation status data, and communication status data from the segmented state sequence of the charging stage, and construct a topology diagram of the charging pile components, where: The topology diagram of the charging pile components is constructed as follows: Read the power module node corresponding to the power module temperature data, the gun wire node corresponding to the gun wire temperature data, the insulation node corresponding to the insulation resistance data, the contactor node corresponding to the contactor action status data, and the communication node corresponding to the communication status data. When there is a power transmission relationship between two nodes, establish a power transmission connection edge; when there is a control trigger relationship between two nodes, establish a control connection edge; when there is a communication interaction relationship between two nodes, establish a communication connection edge. Write the node number according to the component type to which the node belongs, and write the connection edge direction according to the connection relationship to generate a charging pile component topology diagram. Power ripple signal data, current ripple signal data, and voltage ripple signal data are divided into corresponding node ripple signals according to component nodes. Based on the node ripple signals, potential distribution features corresponding to center frequency parameters, frequency modulation parameters, and window width parameters are generated. Based on the potential distribution features, probabilistic reconstruction is performed on the center frequency parameters, frequency modulation parameters, and window width parameters to generate adaptive Chirplet atoms, where: The power ripple signal data, current ripple signal data, and voltage ripple signal data are divided into corresponding node ripple signals according to component nodes, specifically: Read the node number in the charging pile component topology diagram, extract the power ripple signal data, current ripple signal data and voltage ripple signal data within the time range corresponding to the node number, write the ripple signal belonging to the power module node into the power module node data area, write the ripple signal belonging to the gun line node into the gun line node data area, write the ripple signal belonging to the contactor node into the contactor node data area, and combine the ripple amplitude values ​​in the corresponding nodes according to the acquisition time order to generate node ripple signals. The potential distribution characteristics corresponding to the center frequency parameter, frequency modulation parameter, and window width parameter are generated based on the nodal ripple signal, specifically: Read the ripple amplitude of the node ripple signal at each sampling time, count the frequency position with the largest ripple amplitude, and determine the corresponding frequency value as the center frequency parameter. Read the difference of the center frequency parameter between adjacent sampling times, and divide it by the corresponding time difference to obtain the frequency modulation parameter. Read the range of positions in the node ripple signal where the ripple amplitude is greater than the maximum ripple amplitude of 0.60, calculate the corresponding time length, and obtain the window width parameter. Statistically count the average value and fluctuation range of the center frequency parameter, frequency modulation parameter, and window width parameter in all sampling windows according to the sampling time order, and generate the potential distribution characteristics corresponding to the center frequency parameter, frequency modulation parameter, and window width parameter. Based on the latent distribution characteristics, probabilistic reconstruction is performed on the center frequency parameter, frequency modulation parameter, and window width parameter, specifically as follows: Read the average value and fluctuation range corresponding to the center frequency parameter, and generate the reconstructed center frequency parameter by adding a random fluctuation amount to the average value. The random fluctuation amount ranges from -0.15 times to +0.15 times the fluctuation range. Read the average value and fluctuation range corresponding to the frequency modulation parameter, and generate the reconstructed frequency modulation parameter by adding a random fluctuation amount to the average value. The random fluctuation amount ranges from -0.10 times to +0.10 times the fluctuation range. Read the average value and fluctuation range corresponding to the window width parameter, and generate the reconstructed window width parameter by adding a random fluctuation amount to the average value. The random fluctuation amount ranges from -0.12 times to +0.12 times the fluctuation range. Write the reconstructed center frequency parameter, frequency modulation parameter, and window width parameter to the corresponding time-frequency response position to generate an adaptive Chirplet atom. A phylogenetic set of atoms is constructed based on adaptive Chirplet atoms. Using the nodal ripple signal reconstruction error as the fitness index, iterative selection, crossover, and mutation are performed on the center frequency parameter, frequency modulation parameter, and window width parameter to generate evolving Chirplet atoms, where: The phylogenetic set of atoms is constructed based on adaptive Chirplet atoms, specifically as follows: Read all adaptive Chirplet atoms, calculate the difference in center frequency parameter, frequency modulation parameter, and window width parameter between any two adaptive Chirplet atoms. When the difference in center frequency parameter is less than 50Hz, the difference in frequency modulation parameter is less than 0.08, and the difference in window width parameter is less than 5ms, the corresponding adaptive Chirplet atoms are assigned to the same lineage group. Write the corresponding adaptive Chirplet atoms according to the lineage group number to generate a lineage evolution atom set. Iterative selection, crossover, and mutation are performed on the center frequency parameter, frequency modulation parameter, and window width parameter, specifically as follows: Read the adaptive Chirplet atoms in the phylogenetic atom set, read the ripple amplitude of the node ripple signal at each sampling time, read the time-frequency response value of the adaptive Chirplet atom at the corresponding sampling time, calculate the difference between the ripple amplitude and the time-frequency response value, square all the differences and sum them, then divide by the total number of sampling points to obtain the node ripple signal reconstruction error. Sort the node ripple signal reconstruction errors from smallest to largest, retain the top 30% of the adaptive Chirplet atoms, and select two groups of atoms from the retained adaptive Chirplet atoms each time. The center frequency parameter of the first group of atoms is compared with that of the second group of atoms. The frequency modulation parameter is swapped, and the window width parameter of the first group of atoms is swapped with the center frequency parameter of the second group of atoms to generate a cross parameter combination. A random number between 0 and 1 is read. When the random number is less than 0.05, parameter mutation is performed. When the current iteration round is even, the center frequency parameter is increased by 20Hz, the frequency modulation parameter is increased by 0.02, and the window width parameter is increased by 2ms. When the current iteration round is odd, the center frequency parameter is decreased by 20Hz, the frequency modulation parameter is decreased by 0.02, and the window width parameter is decreased by 2ms. The cross parameter combination and mutation parameter combination are written into the lineage evolution atom set to generate the evolutionary Chirplet atom. Chirplet transform is performed on the node ripple signal based on the evolved Chirplet atom pair, generating time-frequency domain response results and instantaneous phase-frequency domain response results, respectively. A hypersphere rotation time-frequency space is constructed based on these results, and a joint rotation mapping is performed on the center frequency parameter, tuning frequency parameter, and ripple energy distribution within this space to generate the node time-frequency energy vector, where: The Chirplet transform is performed on the node ripple signal based on the evolution of Chirplet atoms, specifically as follows: Read the ripple amplitude corresponding to the node ripple signal, read the center frequency parameter, frequency modulation parameter, and window width parameter in the evolved Chirplet atom, perform frequency alignment between the center frequency parameter and the frequency component in the node ripple signal, perform change alignment between the frequency modulation parameter and the frequency change trend in the node ripple signal, perform time alignment between the window width parameter and the duration in the node ripple signal, calculate the product value between the node ripple amplitude and the atom amplitude, and accumulate it along the sampling time sequence. Determine the accumulation result as the time-frequency response amplitude, generate the time-frequency domain response result containing the sampling time position, frequency position, and time-frequency response amplitude, calculate the phase difference between the node ripple signal and the evolved Chirplet atom, as well as the phase difference change between adjacent sampling times, and generate the instantaneous phase-frequency domain response result containing the phase value and the phase change value. The time-frequency space of the hypersphere rotation is constructed based on the time-frequency domain response results and the instantaneous phase-frequency domain response results, specifically as follows: The sampling time position, frequency position, time-frequency response amplitude, and phase change value are read. The ripple energy value is obtained by squaring the time-frequency response amplitude and summing them. The sampling time position, frequency position, ripple energy value, and phase change value are combined to form a four-dimensional time-frequency position point, which is then arranged in the order of acquisition time to generate a hypersphere rotating time-frequency space. In the hypersphere's rotating time-frequency space, a joint rotational mapping is performed on the center frequency parameter, the frequency modulation parameter, and the ripple energy distribution, specifically as follows: Read the frequency position, ripple energy value, phase change value, center frequency parameter, and frequency modulation parameter. Calculate the frequency offset between the frequency position and the center frequency parameter. Multiply the frequency offset by 0.50, the frequency modulation parameter by 0.30, and the phase change value by 0.20, and then sum them to obtain the rotation adjustment value. Add the frequency position to the rotation adjustment value to obtain the rotation frequency position. Multiply the ripple energy value by 1 and add the absolute value of the rotation adjustment value to determine the rotation energy value. Combine the sampling time position, rotation frequency position, rotation energy value, and phase change value according to the sampling time sequence to generate the node time-frequency energy vector. The similarity of the node time-frequency energy vectors between corresponding component nodes is calculated based on the connecting edges in the charging pile component topology graph, and the node time-frequency energy vector similarity is written as the ripple propagation association weight into the corresponding connecting edge, where: The similarity of the time-frequency energy vectors between corresponding component nodes is calculated based on the connecting edges in the charging pile component topology graph, specifically as follows: Read the node time-frequency energy vectors of the corresponding component nodes at both ends of the connection edge, calculate the vector difference at the corresponding position, square all the vector differences and sum them up, then divide by the vector length to get the average difference. Calculate the node time-frequency energy vector similarity by dividing by 1 and summing the average difference, and write it into the corresponding connection edge to generate ripple propagation association weights. Based on the ripple propagation correlation weights in the charging pile component topology graph, graph convolution updates are performed on the center frequency parameter, tuning frequency parameter, and window width parameter of the Chirplet atom. Then, topologically coupled Chirplet atoms are generated based on the updated center frequency parameter, tuning frequency parameter, and window width parameter, where: Based on the ripple propagation correlation weights in the charging pile component topology graph, graph convolution updates are performed on the center frequency parameter, tuning frequency parameter, and window width parameter of the Chirplet atom. Specifically: Read the ripple propagation association weights and the center frequency parameters, frequency modulation parameters, and window width parameters corresponding to adjacent component nodes. Multiply each parameter by its corresponding ripple propagation association weight and sum them up. Then divide by the sum of all ripple propagation association weights to obtain the center frequency update parameters, frequency modulation update parameters, and window width update parameters. Blend the center frequency update parameters with the current center frequency parameters at a weight ratio of 0.70 to 0.30, the frequency modulation update parameters with the current frequency modulation parameters at a weight ratio of 0.65 to 0.35, and the window width update parameters with the current window width parameters at a weight ratio of 0.60 to 0.40 to generate the graph convolution update parameters. Based on the updated center frequency parameter, frequency modulation parameter, and window width parameter, topologically coupled Chirplet atoms are generated, specifically as follows: Read the center frequency update parameters, frequency modulation update parameters, and window width update parameters. Calculate the time distance between each sampling moment and the center of the time window. When the time distance is 0, set the time weight to 1. When the time distance is half the time length corresponding to the window width parameter, set the time weight to 0.60. When the time distance is equal to the time length corresponding to the window width parameter, set the time weight to 0.14. Add the results of multiplying the center frequency update parameters and frequency modulation update parameters by the time distance to obtain the instantaneous frequency value. Multiply the instantaneous frequency value by the time weight to obtain the time-frequency response value. Arrange all time-frequency response values ​​in the order of acquisition time to generate topologically coupled Chirplet atoms.

[0022] In this embodiment, the generation of charging ripple time-frequency evolution characteristics includes: Read the topologically coupled Chirplet atoms, power ripple signal data, current ripple signal data, voltage ripple signal data, power module temperature data, gun wire temperature data, and acquisition timestamp data. Divide the power ripple signal data, current ripple signal data, and voltage ripple signal data into ripple signal segments within the corresponding stages according to the segmented state sequence of the charging stages. Randomly mask a portion of the power module temperature data and gun wire temperature data to generate a temperature rise mask sequence, where: Generate the temperature rise mask sequence as follows: Read the power module temperature data and gun wire temperature data. Select one sampling point every 10 sampling points according to the acquisition time sequence as the masking position. Replace the temperature value corresponding to the masking position with a fixed mask value of -1. Keep the original temperature value of the other sampling positions unchanged. Arrange the processed power module temperature data and gun wire temperature data according to the acquisition time sequence to generate a temperature rise mask sequence. Based on topologically coupled Chirplet atoms, Chirplet time-frequency decomposition is performed on each ripple signal segment to obtain the power ripple time-frequency response, current ripple time-frequency response, and voltage ripple time-frequency response. Then, based on the unmasked power module temperature data and gun wire temperature data, temperature rise reconstruction is performed on the temperature rise mask sequence to generate a temperature rise reconstruction sequence, where: Chirplet time-frequency decomposition is performed on each ripple signal segment based on topologically coupled Chirplet atoms, specifically as follows: Read the ripple amplitude in the ripple signal segment, read the center frequency parameter, frequency modulation parameter and window width parameter in the topologically coupled Chirplet atom, multiply the ripple amplitude by the atom response value at the corresponding sampling time of the topologically coupled Chirplet atom, and accumulate them sequentially along the sampling time to obtain the time-frequency response value at the corresponding frequency position. The time-frequency response value corresponding to the power ripple signal segment is determined as the power ripple time-frequency response, the time-frequency response value corresponding to the current ripple signal segment is determined as the current ripple time-frequency response, and the time-frequency response value corresponding to the voltage ripple signal segment is determined as the voltage ripple time-frequency response. Temperature rise reconstruction is performed on the temperature rise mask sequence based on the unmasked power module temperature data and gun wire temperature data, specifically as follows: Read the power module temperature and gun wire temperature values ​​corresponding to the unmasked positions in the temperature rise mask sequence. Read the temperature change trends of the three sampling points before and after the corresponding position. Calculate the average temperature change of the first three sampling points and the average temperature change of the last three sampling points. Add the two and divide by 2 to obtain the predicted temperature change value. Add the temperature value corresponding to the sampling point before the masked position to the predicted temperature change value to obtain the reconstructed temperature value corresponding to the masked position. Arrange all the reconstructed temperature values ​​in the order of acquisition time to generate a temperature rise reconstruction sequence. The temperature rise distribution is calculated based on the temperature rise reconstruction sequence, which determines the temperature changes of the power module and the gun wire between adjacent sampling windows. A temperature rise distribution is then constructed based on these temperature changes. Finally, a Chirplet time-frequency energy distribution is constructed based on the power ripple time-frequency response, current ripple time-frequency response, and voltage ripple time-frequency response. The calculation of power module temperature change and gun wire temperature change between adjacent sampling windows is based on the temperature rise reconstruction sequence, specifically: Read the average temperature value of the power module corresponding to the current sampling window and the average temperature value of the power module corresponding to the previous sampling window, calculate the difference between the two, and obtain the change in power module temperature. Read the average temperature value of the gun wire corresponding to the current sampling window and the average temperature value of the gun wire corresponding to the previous sampling window, calculate the difference between the two, and obtain the change in gun wire temperature. The temperature rise distribution is constructed based on the temperature changes of the power module and the gun wire, specifically as follows: Read the temperature change of the power module and the temperature change of the gun wire corresponding to each sampling window, multiply the temperature change of the power module by 0.60, multiply the temperature change of the gun wire by 0.40, and then perform weighted summation to obtain the comprehensive temperature rise value corresponding to each sampling window; arrange the sampling windows in ascending order of comprehensive temperature rise value to generate the temperature rise distribution. The Chirplet time-frequency energy distribution is constructed based on the power ripple time-frequency response, current ripple time-frequency response, and voltage ripple time-frequency response, specifically as follows: Read the time-frequency response values ​​from the power ripple time-frequency response, current ripple time-frequency response, and voltage ripple time-frequency response. Square each time-frequency response value and sum them to obtain the time-frequency energy value of the corresponding sampling window. Multiply the time-frequency energy value corresponding to the power ripple by 0.40, the time-frequency energy value corresponding to the current ripple by 0.35, and the time-frequency energy value corresponding to the voltage ripple by 0.25. Then perform weighted summation to obtain the Chirplet time-frequency energy value corresponding to each sampling window. Arrange all Chirplet time-frequency energy values ​​in the order of sampling time to generate the Chirplet time-frequency energy distribution. An optimal transmission mapping relationship is constructed based on the Chirplet time-frequency energy distribution and the temperature rise distribution. The optimal transmission cost from the Chirplet time-frequency energy distribution to the temperature rise distribution is calculated. A temperature rise coupling weight is generated based on the optimal transmission cost. Modulation of the Chirplet time-frequency energy is then performed based on the temperature rise coupling weight to generate temperature rise coupled Chirplet time-frequency energy, where: The optimal transport mapping relationship is constructed based on the Chirplet time-frequency energy distribution and temperature rise distribution, specifically as follows: Read the Chirplet time-frequency energy value and comprehensive temperature rise value corresponding to each sampling window. Divide the Chirplet time-frequency energy value by the sum of the Chirplet time-frequency energy values ​​of all sampling windows, and divide the comprehensive temperature rise value by the sum of the comprehensive temperature rise values ​​of all sampling windows to obtain the normalized distribution of time-frequency energy and the normalized distribution of temperature rise. Calculate the absolute value of the acquisition time difference and the absolute value of the distribution difference between any time-frequency energy sampling window and any temperature rise sampling window. Multiply the absolute value of the acquisition time difference by 0.60 and the absolute value of the distribution difference by 0.40 and then sum them up to obtain the corresponding mapping cost value. Perform total cost value calculation on the mapping combination between all time-frequency energy sampling windows and all temperature rise sampling windows. Select the mapping combination with the smallest total cost value as the transmission correspondence between the time-frequency energy sampling window and the temperature rise sampling window to generate the optimal transmission mapping relationship. The optimal transport cost from the Chirplet time-frequency energy distribution to the temperature rise distribution is calculated as follows: Read the transmission correspondence corresponding to each time-frequency energy sampling window in the optimal transmission mapping relationship, read the corresponding mapping cost value and the corresponding time-frequency energy normalization value respectively, multiply the mapping cost value and the corresponding time-frequency energy normalization value to obtain the corresponding transmission weighted cost, and sum all the transmission weighted costs to obtain the optimal transmission cost corresponding to the mapping combination with the minimum total cost. The temperature rise coupling weight is generated based on the optimal transmission cost, specifically as follows: Read the optimal transmission cost, divide the optimal transmission cost by the total number of sampling windows to obtain the average transmission cost; when the average transmission cost is less than or equal to 0.10, set the temperature rise coupling weight to 1.20; when the average transmission cost is greater than 0.10 and less than or equal to 0.30, set the temperature rise coupling weight to 1.00; when the average transmission cost is greater than 0.30, set the temperature rise coupling weight to 0.80. The time-frequency energy of Chirplet is modulated based on the temperature rise coupling weight, specifically as follows: Read the Chirplet time-frequency energy value corresponding to each sampling window, multiply the Chirplet time-frequency energy value with the temperature rise coupling weight to obtain the temperature rise coupling Chirplet time-frequency energy value of the corresponding sampling window, arrange all temperature rise coupling Chirplet time-frequency energy values ​​according to the order of the collected timestamp data, and generate the temperature rise coupling Chirplet time-frequency energy. Based on temperature rise coupled Chirplet time-frequency energy extraction, the dominant frequency position of the ripple, the peak value of the ripple energy, the frequency band spread range, and the stage frequency offset are extracted and combined according to the order of the collected timestamp data to generate the charging ripple time-frequency evolution characteristics, where: The time-frequency evolution characteristics of the generated charging ripple are as follows: Read the temperature rise coupled Chirplet time-frequency energy value corresponding to each sampling window, select the frequency position with the largest temperature rise coupled Chirplet time-frequency energy value as the ripple dominant frequency position, take the corresponding maximum temperature rise coupled Chirplet time-frequency energy value as the ripple energy peak value, read the continuous frequency range where the temperature rise coupled Chirplet time-frequency energy value is greater than 0.50 times the ripple energy peak value, subtract the lowest frequency position from the corresponding highest frequency position to obtain the frequency band spread range, read the ripple dominant frequency position corresponding to the current sampling window and the ripple dominant frequency position corresponding to the first sampling window of the current charging stage, calculate the difference between the two to obtain the stage frequency offset, combine the ripple dominant frequency position, ripple energy peak value, frequency band spread range and stage frequency offset according to the order of the collected timestamp data to generate the charging ripple time-frequency evolution characteristics.

[0023] In this embodiment, constructing the ripple dynamic degradation feature sequence includes: Read the ripple dominance frequency position, ripple energy peak, frequency band spread range, stage frequency offset and acquisition timestamp data from the charging ripple time-frequency evolution characteristics. Calculate the ratio between the ripple dominance frequency position difference and the sampling time difference between adjacent sampling windows according to the acquisition timestamp data order to generate ripple frequency drift change characteristics. The ratio between the peak ripple energy difference and the sampling time difference between adjacent sampling windows is calculated according to the order of the collected timestamp data to generate ripple energy fluctuation characteristics. The ratio between the frequency band spread range difference and the sampling time difference between adjacent sampling windows is also calculated to generate frequency band spread change characteristics. The ripple frequency drift characteristics, ripple energy fluctuation characteristics, frequency band spread characteristics, and stage frequency offset corresponding to the same sampling window are combined in the order of the acquisition timestamp data to generate a ripple dynamic degradation characteristic sequence.

[0024] In this embodiment, generating the continuous degradation state vector of the charging pile includes: An improved liquid time constant network is constructed, comprising a ripple dynamic degradation input layer, a liquid state encoding layer, a continuous-time liquid update layer, and a multi-state fusion layer, wherein: Construct an improved liquid time constant network, specifically as follows: Based on the input mapping layer of the traditional liquid time constant network, time jitter enhancement, amplitude occlusion enhancement, and dual-path time sequence comparison mapping are added to obtain the ripple dynamic degradation input layer. Before the recursive update structure of the liquid state in the traditional liquid time constant network, state coupling mapping of ripple dynamic degradation features, insulation resistance change features, and contactor action state features is added to obtain the liquid state encoding layer. Based on the continuous time state update structure of the traditional liquid time constant network, energy conservation-Hamilton liquid block, Koopman spectral decomposition mechanism, graph-ODE cooperative diffusion coupling, and symplectic integral time update block are added to obtain the continuous time liquid update layer. Based on the output mapping layer of the traditional liquid time constant network, state splicing and weight fusion of Hamilton-constrained liquid state sequence, spectral decomposition liquid state sequence, diffusion-coupled liquid state sequence, continuously degraded liquid state sequence, and random diffusion liquid state sequence are added to obtain the multi-state fusion layer, generating an improved liquid time constant network. The ripple dynamic degradation input layer performs temporal jitter enhancement and amplitude occlusion enhancement on the ripple dynamic degradation feature sequences within the same sampling window, generating a dual-path temporal enhancement sequence. Based on this dual-path temporal enhancement sequence, it performs temporal comparison self-supervised enhancement to generate a liquid input state sequence, where: The ripple dynamic degradation input layer includes: Time jitter buffer: Stores a four-dimensional ripple feature sequence after time jitter with a length of 512 sampling points; Amplitude occlusion buffer: Stores a four-dimensional ripple feature sequence after amplitude occlusion with a length of 512 sampling points; Difference register: Stores the feature difference vector between two sequences point by point; Weight register: Write the scalar weight according to the maximum absolute value in the difference register. Write 1.20 when the difference is not greater than 0.10, write 0.80 when the difference is greater than 0.30, and write 1.00 for the rest. Sequence splicing buffer: splices the first and last parts of two weighted sequences and outputs a liquid input state sequence; The execution of time jitter enhancement and amplitude occlusion enhancement is as follows: Read the ripple frequency drift change value, ripple energy fluctuation value, bandwidth spread change value, and stage frequency offset value from the same sampling window; shift all feature values ​​one sampling position backward according to the acquisition time sequence, and copy the feature value of the last sampling position to the end of the current sampling window to generate a time jitter enhancement sequence; determine the occlusion position by selecting one sampling point every 8 sampling points, replace the ripple frequency drift change value, ripple energy fluctuation value, bandwidth spread change value, and stage frequency offset value corresponding to the occlusion position with 0, generate an amplitude occlusion enhancement sequence, and combine the time jitter enhancement sequence and the amplitude occlusion enhancement sequence to generate a dual-path time-series enhancement sequence; The timing comparison self-supervised enhancement based on dual-path timing enhancement sequences is performed as follows: Read the time jitter enhancement feature value and amplitude occlusion enhancement feature value corresponding to the same sampling window in the dual-path temporal enhancement sequence, calculate the absolute value of the feature difference at the corresponding position, square all the absolute values ​​of feature differences and sum them up, then divide by the total number of features to obtain the path difference value. The dual-path temporal enhancement sequence with a path difference value less than 0.10 is determined as the positive sample sequence, and the dual-path temporal enhancement sequence with a path difference value greater than 0.30 is determined as the negative sample sequence. Multiply the feature value corresponding to the positive sample sequence by 1.20, and multiply the feature value corresponding to the negative sample sequence by 0.80 to generate the liquid input state sequence. The liquid state encoding layer reads the liquid input state sequence, extracts the ripple dynamic degradation features of adjacent sampling windows, calculates the time decay coefficient corresponding to the sampling time difference, and performs state coupling mapping on the ripple dynamic degradation features, insulation resistance change features, and contactor action state features to generate the initial liquid state sequence and ripple abrupt change features, where: The liquid state coding layer includes: Short-time buffer block: stores the state coupling values ​​of the most recent 16 sampling windows; Long-term cache block: stores the state coupling values ​​of the most recent 64 sampling windows; Residual register: Stores the intermediate encoded value after adding the short-time buffer block and the long-time buffer block.

[0025] Attenuation Register: Generates a time attenuation coefficient based on the ratio between the current sampling window time difference and the average time difference of all sampling windows, and writes the time attenuation coefficient into the residual register to obtain the initial liquid state value; Mutation flag: Write 1 when the absolute value of the difference between the current coupling value and the coupling value of the previous window is greater than 1.50 times the average value of all state coupling values; otherwise write 0. State output buffer: Sequentially outputs the initial liquid state sequence and ripple abrupt change characteristics; The time decay coefficient corresponding to the sampling time difference is calculated as follows: Read the timestamp data corresponding to the current sampling window and the timestamp data corresponding to the previous sampling window, calculate the time difference between the two, divide the time difference by the average time difference of all sampling windows to obtain the time difference ratio, and calculate the time decay coefficient by dividing 1 by 1 and the sum of the time difference ratio. The execution of state coupling mapping is as follows: Read the ripple frequency drift change value, ripple energy fluctuation value, bandwidth spread change value, stage frequency offset, insulation resistance change characteristic, and contactor operation status characteristic. Multiply the ripple frequency drift change value by 0.30, the ripple energy fluctuation value by 0.25, the bandwidth spread change value by 0.20, the stage frequency offset by 0.10, the insulation resistance change characteristic by 0.10, and the contactor operation status characteristic by 0.05, and then perform weighted accumulation to obtain the state coupling value. Arrange the state coupling values ​​corresponding to all sampling windows in the order of acquisition time to generate an initial liquid state sequence. Read the state coupling values ​​corresponding to adjacent sampling windows and calculate the absolute value of the difference between the current state coupling value and the previous state coupling value. When the absolute value of the difference is greater than 1.50 times the average value of all state coupling values, mark the corresponding sampling window as the ripple abrupt change position and generate the ripple abrupt change characteristic. The continuous-time liquid state update layer, based on output voltage data, output current data, charging power data, and ripple energy fluctuation characteristics, introduces an energy conservation-Hamilton liquid block to construct a Hamiltonian energy state and executes liquid time constant constraints, generating a Hamiltonian-constrained liquid state sequence, where: The continuous-time liquid renewal layer includes: Energy Register: Writes ripple energy fluctuation value at a ratio of 0.60, writes power deviation value at a ratio of 0.40, and accumulates them to generate instantaneous energy state value; Time constant register: Adjusts the liquid time constant based on the instantaneous energy state value. The liquid time constant is limited to 0.05 seconds to 5 seconds, and the variation between adjacent sampling windows does not exceed 30% of the liquid time constant of the previous sampling window. Spectral register: The liquid state is decomposed into stable components, oscillatory components, and abrupt change components, and then multiplied by 0.95, 0.80, and 0.60 respectively to generate the spectral decomposition of the liquid state; Diffusion buffer: Stores the product of the state difference between the nodes at both ends of the connection edge in the topology graph and the ripple propagation association weight; Random disturbance register: Write time disturbance values ​​at a ratio of 0.40, communication disturbance values ​​at a ratio of 0.35, and sudden change disturbance values ​​at a ratio of 0.25, and accumulate them to generate random disturbance values; State Differential Register: Stores the state changes after spectral decomposition of liquid state, diffusion buffer output, and accumulation of random perturbation values; Symmetric Integral Sequencer: Performs half-step Hamiltonian correction, random perturbation writing, and quadratic half-step Hamiltonian correction on the state change quantity, and outputs a continuous degenerate liquid state sequence and a random diffusion liquid state sequence; Energy conservation-Hamilton liquid block refers to a continuous-time liquid update structure composed of electrical energy state quantity, ripple energy state quantity and liquid state quantity. The electrical energy state quantity is characterized by output voltage data, output current data and charging power data, the ripple energy state quantity is characterized by ripple energy fluctuation characteristics, and the liquid state quantity is characterized by the initial liquid state sequence. Liquid time constant constraint refers to the limitation relationship between the range and variation of liquid time constant. The minimum value of liquid time constant is set to 0.05 seconds, the maximum value is set to 5 seconds, and the variation of liquid time constant between adjacent sampling windows shall not exceed 0.30 times the liquid time constant of the previous sampling window. Introducing energy conservation – Hamiltonian liquid block to construct Hamiltonian energy state and applying liquid time constant constraint, specifically: Read the output voltage and output current data. Multiply the output voltage and output current values ​​to obtain the calculated power value. Calculate the absolute value of the difference between the calculated power value and the charging power data to obtain the energy deviation value. Multiply the ripple energy fluctuation characteristic by 0.60, multiply the energy deviation value by 0.40, and then sum them to obtain the Hamiltonian energy state value. Multiply the Hamiltonian energy state value by the current initial liquid state value to obtain the liquid time constant adjustment value. Add the liquid time constant adjustment value to the liquid time constant of the previous sampling window to obtain the current liquid time constant. When the current liquid time constant is less than 0.05 seconds, it is corrected to 0.05 seconds. When the current liquid time constant is greater than 5 seconds, it is corrected to 5 seconds. When the difference between the current liquid time constant and the liquid time constant of the previous sampling window is greater than 0.30 times the liquid time constant of the previous sampling window, the current liquid time constant is corrected to 1.30 times the liquid time constant of the previous sampling window. The corrected liquid time constant is multiplied by the current initial liquid state value to obtain the Hamiltonian-constrained liquid state value. All Hamiltonian-constrained liquid state values ​​are arranged in the order of the collected timestamp data to generate the Hamiltonian-constrained liquid state sequence. The initial liquid state sequence is mapped to the observable function space. Based on the Koopman spectral decomposition mechanism, spectral operator decomposition and exponential mapping time advancement are performed to generate the spectral decomposed liquid state sequence, where: The observable function space refers to a high-dimensional state representation space composed of the initial liquid state sequence, ripple energy fluctuation characteristics, time decay coefficient, and Hamiltonian energy state in the order of the fields. The initial liquid state sequence is mapped to the observable function space, specifically as follows: Read the state coupling value, ripple energy fluctuation characteristics, time decay coefficient and Hamiltonian energy state from the initial liquid state sequence, write the state coupling value into the first position, write the ripple energy fluctuation characteristics into the second position, write the time decay coefficient into the third position, and write the Hamiltonian energy state into the fourth position, arrange them according to the order of the collected timestamp data, and obtain the liquid state representation in the observable function space. Koopman spectral decomposition mechanism refers to the spectral domain processing relationship that decomposes the liquid state representation in the observable function space into stable, oscillatory, and abrupt change components. The spectral operator decomposition and exponential mapping time advancement are performed based on the Koopman spectral decomposition mechanism, specifically as follows: The liquid state representation in the observable function space is read, the state difference between adjacent sampling windows is calculated, the portion of the state difference with an absolute value less than 0.10 is written into the stable change component, the portion of the state difference with an absolute value greater than or equal to 0.10 and less than 0.35 is written into the oscillating change component, and the portion of the state difference with an absolute value greater than or equal to 0.35 is written into the abrupt change component. The stable change component is multiplied by 0.95, the oscillating change component is multiplied by 0.80, and the abrupt change component is multiplied by 0.60 and then summed to obtain the time-progressed state value. The values ​​are then arranged in the order of the collected timestamp data to generate the spectral decomposition liquid state sequence. Based on the topology diagram of charging pile components, node diffusion connections are constructed. A state differential equation encompassing liquid state evolution, node diffusion propagation, and random disturbance diffusion is built through graph-ODE collaborative diffusion coupling. Random diffusion intensity is generated based on sampling time difference, communication state data, and ripple mutation characteristics. Continuous-time diffusion integrals are then applied to the state differential equation to generate a diffusion-coupled liquid state sequence, where: The node diffusion connection relationship is constructed based on the topology diagram of the charging pile components, specifically as follows: Read the power module node, charging gun line node, contactor node, insulation detection node, and communication node in the topology diagram of the charging pile components. Read the connection edges between each node. Write the connection edge between the power module node and the charging gun line node into the power diffusion connection relationship. Write the connection edge between the contactor node and the power module node into the motion disturbance diffusion connection relationship. Write the connection edge between the insulation detection node and the power module node into the insulation degradation diffusion connection relationship. Write the connection edge between the communication node and other nodes into the communication disturbance diffusion connection relationship. Generate node diffusion connection relationships. Graph-ODE cooperative diffusion coupling refers to a coupling structure in which the changes in the liquid state itself, the state diffusion between nodes, and the diffusion of random disturbances are all written into the continuous-time state differential relationship within the node diffusion connection relationship defined by the topology diagram of the charging pile components. A state differential equation incorporating liquid state evolution, nodal diffusion propagation, and random perturbation diffusion is constructed through graph-ODE cooperative diffusion coupling, specifically: The spectral decomposition liquid state sequence is read as the liquid state self-change term. The state difference between the two nodes of the connecting edge in the node diffusion connection relationship is read and multiplied by the ripple propagation association weight corresponding to the connecting edge to obtain the node diffusion propagation term. The random diffusion intensity is read and multiplied by the encoding value corresponding to the communication state data to obtain the random disturbance diffusion term. The liquid state self-change term, node diffusion propagation term and random disturbance diffusion term are added together to obtain the state change amount corresponding to each sampling window. The state change amounts are arranged in the order of the acquisition timestamp data to form the state differential equation. The random diffusion intensity is generated based on the sampling time difference, communication status data, and ripple abrupt change characteristics, specifically as follows: Read the sampling time difference, divide the sampling time difference by the average sampling time difference of all sampling windows to obtain the time disturbance ratio, read the communication status data, normal communication status is coded as 0, delayed communication status is coded as 0.50, interrupted communication status is coded as 1, read the ripple mutation feature, non-mutation position is coded as 0, mutation position is coded as 1, multiply the time disturbance ratio by 0.40, the communication status code value by 0.35, and the ripple mutation code value by 0.25, and then sum them up to obtain the random diffusion intensity; The state differential equation is solved by performing a continuous-time diffusion integral, specifically: Read the spectral decomposition liquid state value, the state change corresponding to the state differential equation, and the sampling time difference of the current sampling window. Multiply the state change by the sampling time difference to obtain the state increment. Add the spectral decomposition liquid state value of the current sampling window to the state increment to obtain the diffusion-coupled liquid state value of the next sampling window. Repeat the same calculation according to the order of the collected timestamp data to generate the diffusion-coupled liquid state sequence. A symplectic integral time update block is introduced to perform split integral updates, generating a continuously degenerate liquid state sequence and a randomly diffused liquid state sequence, wherein: A symptotic integral time update block is a continuous-time state update structure consisting of half-step state advancement, random diffusion writing, and half-step state correction. The execution of the split score update is as follows: Read the diffusion-coupled liquid state sequence, Hamilton-constrained liquid state sequence, and random diffusion intensity. Multiply the difference between the diffusion-coupled liquid state value and the Hamilton-constrained liquid state value by 0.50 to obtain the half-step Hamilton correction. Add the half-step Hamilton correction to the diffusion-coupled liquid state value to obtain the first half-step state value. Multiply the random diffusion intensity by the sampling time difference and add it to the first half-step state value to obtain the random diffusion liquid state value. Calculate the difference between the random diffusion liquid state value and the Hamilton-constrained liquid state value again and multiply it by 0.50 to obtain the second half-step Hamilton correction. Add the second half-step Hamilton correction to the random diffusion liquid state value to obtain the continuously degraded liquid state value. Arrange the random diffusion liquid state value and the continuously degraded liquid state value according to the order of the collected timestamp data to generate the random diffusion liquid state sequence and the continuously degraded liquid state sequence. The multi-state fusion layer performs state concatenation and weight fusion on the Hamilton-constrained liquid state sequence, the spectral decomposition liquid state sequence, the diffusion-coupled liquid state sequence, the continuously degraded liquid state sequence, and the random diffusion liquid state sequence to generate a continuously degraded state vector for the charging pile, where: The multi-state fusion layer includes: The splicing register sequentially writes Hamilton-constrained liquid states, spectral decomposition liquid states, diffusion-coupled liquid states, continuously degenerate liquid states, and randomly diffused liquid states; Weight table register: Stores five sets of fusion weights with fixed values ​​of 0.30, 0.20, 0.20, 0.20, and 0.10; Fusion Register: Multiply the 5 states by their corresponding fusion weights and then sum them to generate a fusion state value; Timing buffer: Writes the fused state value in the order of sampling time and outputs the continuous degradation state vector of the charging pile; The execution of state concatenation and weight fusion is as follows: Read the Hamilton-constrained liquid state value, spectral decomposition liquid state value, diffusion-coupled liquid state value, continuously degraded liquid state value, and random diffusion liquid state value corresponding to the same sampling window. Write the Hamilton-constrained liquid state value into the first state position, the spectral decomposition liquid state value into the second state position, the diffusion-coupled liquid state value into the third state position, the continuously degraded liquid state value into the fourth state position, and the random diffusion liquid state value into the fifth state position to generate a state splicing sequence. Multiply the Hamilton-constrained liquid state value by 0.30, the spectral decomposition liquid state value by 0.20, the diffusion-coupled liquid state value by 0.20, the continuously degraded liquid state value by 0.20, and the random diffusion liquid state value by 0.10, and then sum them to obtain the fused state value. Arrange the fused state values ​​corresponding to all sampling windows in the order of the acquisition timestamp data to generate the continuous degradation state vector of the charging pile. The improved liquid time constant network is trained using the prediction error of the continuous degenerate state vector, the reconstruction error of the ripple dynamic degradation feature sequence, and the continuous-time liquid state evolution error as joint optimization objectives. The parameters of the ripple dynamic degradation input layer, the liquid state encoding layer, the continuous-time liquid update layer, and the multi-state fusion layer are continuously optimized. Training stops when the change in joint loss is less than the convergence threshold for five consecutive training epochs. The improved liquid time constant network is trained as follows: Read the ripple dynamic degradation feature sequence and the training target continuous degradation state vector of the corresponding sampling window, input the ripple dynamic degradation feature sequence into the ripple dynamic degradation input layer, the liquid state encoding layer, the continuous time liquid update layer and the multi-state fusion layer, and output the charging pile continuous degradation state vector. The prediction error of the continuous degradation state vector is obtained by calculating the mean square difference between the continuous degradation state vector of the charging pile and the continuous degradation state vector of the training target. The absolute value of the feature difference between the reconstructed ripple dynamic degradation feature sequence and the original ripple dynamic degradation feature sequence is calculated to obtain the reconstruction error of the ripple dynamic degradation feature sequence. The absolute value of the state difference between the continuous degradation liquid state values ​​between adjacent sampling windows is calculated to obtain the continuous time liquid state evolution error. The prediction error of the continuous degradation state vector is multiplied by 0.50, the reconstruction error of the ripple dynamic degradation feature sequence is multiplied by 0.30, and the continuous time liquid state evolution error is multiplied by 0.20, and then summed to obtain the joint loss value. Calculate the joint loss value for each layer after increasing or decreasing the parameters by 0.001, and update the parameters in the direction of decreasing the joint loss value. Repeat the parameter update for all training data. Stop training when the change in the joint loss value for 5 consecutive training periods is less than 0.0001.

[0026] In this embodiment, generating a charging pile fault diagnosis report includes: The system reads the charging pile's continuous degradation state vector, collects timestamp data and charging stage identifier data, and reads power module aging prototypes, gun contact degradation prototypes, contactor jitter prototypes, insulation attenuation prototypes, and communication anomaly prototypes from a preset fault prototype library. Each fault prototype includes a fault prototype state vector, a fault type label, a fault location label, and a fault cause label. The preset fault prototype library refers to a set of fault state data consisting of power module aging prototypes, gun wire contact degradation prototypes, contactor jitter prototypes, insulation attenuation prototypes, and communication anomaly prototypes. Each fault prototype includes a fault prototype state vector, fault type label, fault location label, and fault cause label. Calculate the state matching distance between the continuously degraded state vector of the charging pile and the state vector of each fault prototype. Identify the fault prototype with the smallest state matching distance as the target fault prototype. Calculate the fault confidence based on the state matching distance. Generate a corresponding fault identification result when the fault confidence is greater than or equal to 0.75; otherwise, generate a fault identification result to be reviewed when the fault confidence is less than 0.75. The state matching distance between the continuously degraded state vector of the charging pile and the state vector of each fault prototype is calculated as follows: Read the continuous degradation state vector of the charging pile and the state vector of a fault prototype. Calculate the difference between the corresponding state values ​​for each item at the same vector position. Square all the differences and sum them up. Take the square root of the sum to obtain the state matching distance corresponding to the current fault prototype. Calculate the state matching distance between the continuous degradation state vector of the charging pile and the power module aging prototype, the gun wire contact degradation prototype, the contactor jitter prototype, the insulation attenuation prototype, and the communication abnormality prototype in the same way. Determine the fault prototype with the smallest state matching distance as the target fault prototype. The fault confidence is calculated based on the state matching distance, specifically as follows: Read the state matching distance corresponding to the target fault prototype, divide the state matching distance by the sum of the state matching distance and the fixed value 1 to obtain the distance normalization value, and subtract the distance normalization value from 1 to obtain the fault confidence. When the fault confidence is greater than or equal to 0.75, combine the fault type label, fault location label and fault cause label corresponding to the target fault prototype to form the fault identification result. When the fault confidence level is greater than or equal to 0.75, a corresponding fault identification result is generated, specifically: Read the fault type label, fault location label, and fault cause label corresponding to the target fault prototype. Combine the fault type label, fault location label, fault cause label, and fault confidence to form a fault identification record. Write the value 1 in the fault confirmation mark field. Write the fault identification record into the fault identification result sequence according to the data collection timestamp order to generate the fault identification result. When the fault confidence level is less than 0.75, a fault identification result to be reviewed is generated, specifically as follows: Read the fault type label, fault location label and fault cause label corresponding to the target fault prototype, combine the fault type label, fault location label, fault cause label and fault confidence to form a fault record to be reviewed, write the value 1 in the to be reviewed mark field, and write the fault record to be reviewed into the fault result sequence according to the data collection timestamp data order to generate the fault identification result to be reviewed. A charging pile fault diagnosis report is generated based on the fault identification results, the fault identification results to be reviewed, the target fault prototype, the collected timestamp data, and the charging stage identifier data. The charging pile fault diagnosis report includes the fault occurrence stage, fault status, fault type, fault location, fault cause label, and fault confidence level, wherein: A charging pile fault diagnosis report is generated based on the fault identification results, the fault identification results to be reviewed, the target fault prototype, the collected timestamp data, and the charging stage identifier data. Specifically: Read the fault type label, fault location label, fault cause label, fault confidence, fault confirmation mark, and pending review mark from the fault identification result or the fault identification result to be reviewed. Read the collection timestamp data and charging stage identifier data of the corresponding sampling window. When the fault confirmation mark is 1, write the fault status field as confirmed fault. When the pending review mark is 1, write the fault status field as pending review fault. Write the charging stage identifier data into the fault occurrence stage field. Write the fault type label into the fault type field. Write the fault location label into the fault location field. Write the fault cause label into the fault cause label field. Write the fault confidence into the fault confidence field. Write the collection timestamp data into the fault occurrence time field. Generate a charging pile fault diagnosis report.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a simulated operation and maintenance scenario of DC charging piles. This scenario included 24 DC charging piles, each with a rated power of 120 kW. The sampling interval was 1 second, and each analysis window consisted of 60 sampling points with a sliding step of 10 seconds. During the training phase, a total of 12,000 samples were collected, including 6,000 normal samples, 1,800 power module aging samples, 1,400 gun wire contact degradation samples, 1,200 contactor jitter samples, 900 insulation attenuation samples, and 700 communication anomaly samples. Each sample included output voltage, output current, charging power, power ripple signal, current ripple signal, voltage ripple signal, power module temperature, gun wire temperature, insulation resistance, contactor operation status, communication status, charging stage identifier, and collection timestamp data. In normal samples, the power ripple frequency is concentrated between 1180 Hz and 1260 Hz, the power module temperature is concentrated between 42 degrees Celsius and 61 degrees Celsius, and the gun wire temperature is concentrated between 35 degrees Celsius and 52 degrees Celsius. In aging samples, the power ripple frequency drifts from 1210 Hz to 1450 Hz within a continuous window, and the peak power ripple energy increases from 0.39 to 0.75, but the output voltage, current, and temperature still do not exceed the traditional alarm threshold.

[0028] In a certain continuous charging segment, charging pile C-08 was in the constant current charging stage. Traditional threshold methods showed an output voltage of 607.8 volts, an output current of 141.6 amps, a charging power of 86.1 kilowatts, a power module temperature of 58.4 degrees Celsius, a gun wire temperature of 50.9 degrees Celsius, an insulation resistance of 2.9 megohms, a closed contactor status, and normal communication status, thus classifying it as operating normally. However, the method of this invention, after reading the ripple signal of the same segment, revealed that the power ripple frequency continuously drifted from 1218 Hz to 1412 Hz, the current ripple frequency drifted from 962 Hz to 1096 Hz, the voltage ripple frequency drifted from 1491 Hz to 1608 Hz, the peak power ripple energy increased from 0.41 to 0.78, and the bandwidth expansion range increased from 220 Hz to 506 Hz. In this segment, there are 2 missing points out of 60 sampling points. The system performs linear interpolation based on the previous and next sampling values. The output current of the 17th sampling point is interpolated from 141.2 A and 141.8 A to 141.5 A. The power module temperature of the 43rd sampling point is interpolated from 57.6 degrees Celsius and 57.9 degrees Celsius to 57.75 degrees Celsius. The ripple frequency components in the range of 300 Hz to 30000 Hz are retained.

[0029] Subsequently, the system establishes a topology diagram of the charging pile components, consisting of power module nodes, gun line nodes, contactor nodes, insulation nodes, and communication nodes. Taking the power module node as an example, the initial Chirplet atom's center frequency parameter is 1224 Hz, the tuning frequency parameter is 0.06, and the window width parameter is 20 milliseconds. After probabilistic reconstruction and spectral evolution, the center frequency parameter is adjusted to 1241 Hz, the tuning frequency parameter is adjusted to 0.09, and the window width parameter is adjusted to 22 milliseconds. The node ripple signal reconstruction error decreases from 0.134 to 0.074. After performing time-frequency decomposition based on the evolved Chirplet atom, the system calculates the node time-frequency energy vector similarity between the power module node and the gun line node to be 0.82, and the similarity between the power module node and the contactor node to be 0.76. In normal samples, the above similarities are usually below 0.50. The system writes 0.82 and 0.76 into the corresponding connection edges as ripple propagation association weights and updates the topology coupling Chirplet atoms, so that the center frequency parameter of the power module node is adjusted from 1241 Hz to 1276 Hz, and the center frequency parameter of the gun line node is adjusted from 1008 Hz to 1033 Hz.

[0030] In the temperature rise coupling analysis, the system reads the temperature change of the power module (3.3 degrees Celsius) and the temperature change of the gun wire (2.2 degrees Celsius), obtaining a combined temperature rise value of 2.86. The system performs optimal transmission matching between the Chirplet time-frequency energy distribution and the temperature rise distribution, obtaining an average transmission cost of 0.083, therefore setting the temperature rise coupling weight to 1.20. After modulation, the peak time-frequency energy of the power ripple changes from 0.78 to 0.936, the peak time-frequency energy of the current ripple changes from 0.69 to 0.828, and the peak time-frequency energy of the voltage ripple changes from 0.57 to 0.684. The system further obtains the ripple dominant frequency position (1412 Hz), the ripple energy peak value (0.936), the bandwidth spread range (506 Hz), and the stage frequency offset (194 Hz), and calculates the ripple frequency drift change value (7.4 Hz / s), the ripple energy fluctuation value (0.018), and the bandwidth spread change value (9.1 Hz / s), forming a dynamic degradation characteristic sequence of the ripple.

[0031] After the improved liquid time constant network reads the ripple dynamic degradation feature sequence, the maximum value of the dual-path difference between time jitter enhancement and amplitude occlusion enhancement is 0.086, which is less than 0.10. Therefore, the enhancement weight is set to 1.20. The liquid state encoding layer calculates that the state coupling value increases from 0.42 to 0.69. The difference in state coupling value between the 4th window and the 3rd window is 0.18, which is more than 1.50 times the average value of all state coupling values ​​(0.11). Therefore, it is marked as the ripple abrupt change location. The continuous-time liquid update layer reads the output voltage of 607.8 volts, the output current of 141.6 amps, and the charging power of 86.1 kilowatts, obtaining a power verification value of 86.07 kilowatts and an energy deviation value of 0.03 kilowatts. Combined with the ripple energy fluctuation characteristics, the Hamiltonian energy state value is obtained as 0.0228. The liquid time constant is adjusted from 0.92 seconds in the previous window to 1.15 seconds, with an adjustment range of 25.00%, which does not exceed the 30% limit. After spectral decomposition, graph-ODE co-diffusion, and symplectic integral time update, the system obtains Hamiltonian-constrained liquid state values ​​of 0.64, spectral decomposition liquid state values ​​of 0.59, diffusion-coupled liquid state values ​​of 0.67, continuously degraded liquid state values ​​of 0.71, and random diffusion liquid state values ​​of 0.48. After weighted fusion with 0.30, 0.20, 0.20, 0.20, and 0.10, the fused state value corresponding to the continuously degraded state vector is obtained as 0.626.

[0032] During the fault matching phase, the system calculates the state matching distance between the continuously degraded state vector and the power module aging prototype as 0.17, the distance to the gun line contact degradation prototype as 0.39, the distance to the contactor jitter prototype as 0.53, the distance to the insulation attenuation prototype as 0.68, and the distance to the communication anomaly prototype as 0.74. Since the power module aging prototype has the smallest distance, the system identifies it as the target fault prototype and calculates a fault confidence score of 0.85, which is higher than the confirmation threshold of 0.75. A fault diagnosis report is generated, recording the fault occurrence stage as constant current charging, the fault status as confirmed fault, the fault type as power module aging, the fault location as the power module, the fault cause label as continuous ripple frequency drift accompanied by ripple energy diffusion, and the fault confidence score as 0.85. Traditional threshold methods still output normal results in the same segment.

[0033] In another constant-voltage charging segment, numbered C-15, the conventional method reads an output voltage of 618.2 volts, a gun wire temperature of 52.1 degrees Celsius, and an insulation resistance of 3.1 megohms, classifying it as normal. This invention detects that the current ripple frequency increases from 968 Hz to 1114 Hz, the peak current ripple energy increases from 0.37 to 0.64, the bandwidth expansion range increases from 180 Hz to 332 Hz, and the ripple propagation correlation weight between the gun wire node and the power module node reaches 0.79. The system calculates that the matching distance between this state and the prototype of gun wire contact degradation is 0.41, and the fault confidence is 0.71, which is lower than 0.75. Therefore, a fault diagnosis report to be reviewed is generated. The report records the fault state as a fault to be reviewed, the fault type as gun wire contact degradation, the fault location as the gun wire connection end, and the fault confidence as 0.71. Simulated retesting showed that the contact resistance at the gun wire connection end increased from 0.2 milliohms to 0.7 milliohms, indicating that it has not yet formed an intermittent fault, but early contact degradation has occurred.

[0034] To verify the diagnostic effectiveness, 2400 test samples were input into the traditional threshold method, the traditional time-frequency analysis combined with a classification model, and the method of this invention, respectively. The test samples included 1000 normal samples, 420 power module aging samples, 360 gun wire contact degradation samples, 280 contactor jitter samples, 200 insulation attenuation samples, and 140 communication anomaly samples. The traditional threshold method achieved an overall accuracy of 82.4%, a false negative rate of 13.6%, a false positive rate of 8.9%, and an average early detection time of 1.8 sampling windows. The traditional time-frequency analysis combined with a classification model achieved an overall accuracy of 88.7%, a false negative rate of 8.4%, a false positive rate of 6.1%, and an average early detection time of 3.2 sampling windows. The method of this invention achieved an overall accuracy of 94.6%, a false negative rate of 3.2%, a false positive rate of 3.9%, and an average early detection time of 5.6 sampling windows. The simulation results above demonstrate that the present invention can identify hidden faults such as power module aging, gun wire contact degradation, and contactor jitter when the traditional threshold is not triggered, thereby improving the accuracy of fault diagnosis and early warning capability.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for diagnosing a fault of a charging pile based on machine learning, characterized in that, include: Collect charging pile operation status data, perform preprocessing on the charging pile operation status data, and generate a segmented state sequence for the charging stage; A topology graph of charging pile components is constructed based on the segmented state sequence of the charging stage. Chirplet transformation is performed on the power ripple signal, current ripple signal and voltage ripple signal to establish the ripple propagation correlation between the nodes of the charging pile components. The center frequency parameter, tuning frequency parameter and window width parameter of the Chirplet atom are updated by graph convolution to generate topologically coupled Chirplet atoms. Chirplet time-frequency decomposition is performed on the power ripple signal, current ripple signal and voltage ripple signal based on topologically coupled Chirplet atoms, and temperature rise coupled weight modulation is performed on the Chirplet time-frequency energy according to the temperature change of the power module and the temperature change of the gun wire to generate charging ripple time-frequency evolution characteristics. Based on the time-frequency evolution characteristics of charging ripple, the ripple frequency drift characteristics, ripple energy fluctuation characteristics, and frequency band diffusion characteristics are calculated, and a dynamic degradation characteristic sequence of ripple is constructed. An improved liquid time constant network is constructed based on the ripple dynamic degradation feature sequence. An energy conservation-Hamilton liquid block is introduced to construct the Hamiltonian energy state and execute the liquid time constant constraint. Random diffusion intensity is generated based on sampling time difference, communication state and ripple mutation features, and continuous time liquid state update is performed to generate the continuous degradation state vector of the charging pile. The charging pile's continuous degradation state vector is matched with the execution state of a preset fault prototype library to generate a charging pile fault diagnosis report. 2.The machine learning based charging pile fault diagnosis method of claim 1, wherein, The charging pile operation status data specifically includes output voltage data, output current data, charging power data, power ripple signal data, current ripple signal data, voltage ripple signal data, power module temperature data, gun wire temperature data, insulation resistance data, contactor operation status data, communication status data, charging stage identification data, and collection timestamp data. 3.The machine learning based charging pile fault diagnosis method of claim 1, wherein, The generation of the segmented state sequence for the charging stage includes: The charging pile operation status data is sorted by time according to the collected timestamp data, linear interpolation is performed on missing data, bandpass filtering is performed on power ripple signal data, current ripple signal data and voltage ripple signal data, and status encoding is performed on contactor action status data and communication status data to generate a time-aligned status sequence. Read the charging stage identifier data in the time-aligned state sequence, divide the time-aligned state sequence into stages according to different charging stages, and arrange the charging pile operation status data and sampling time difference in the corresponding stage according to the order of the collected timestamp data to generate a segmented state sequence of charging stages. 4.The machine learning based charging pile fault diagnosis method of claim 1, wherein, The generation of topologically coupled Chirplet atoms includes: Read the power ripple signal data, current ripple signal data, voltage ripple signal data, power module temperature data, gun wire temperature data, insulation resistance data, contactor operation status data, and communication status data from the segmented state sequence of the charging stage, and construct the topology diagram of the charging pile components; The power ripple signal data, current ripple signal data, and voltage ripple signal data are divided into corresponding node ripple signals according to the component nodes. Based on the node ripple signals, the potential distribution features corresponding to the center frequency parameter, frequency modulation parameter, and window width parameter are generated. Based on the potential distribution features, the center frequency parameter, frequency modulation parameter, and window width parameter are probabilistically reconstructed to generate adaptive Chirplet atoms. Based on adaptive Chirplet atoms, a set of phylogenetic evolutionary atoms is constructed. The nodal ripple signal reconstruction error is used as the fitness index. Iterative selection, crossover and mutation are performed on the center frequency parameter, frequency modulation parameter and window width parameter to generate evolutionary Chirplet atoms. Chirplet transform is performed on the node ripple signal based on the evolved Chirplet atom pair to generate time-frequency domain response results and instantaneous phase-frequency domain response results, respectively. A hypersphere rotation time-frequency space is constructed based on the time-frequency domain response results and instantaneous phase-frequency domain response results. In the hypersphere rotation time-frequency space, a joint rotation mapping is performed on the center frequency parameter, the tuning frequency parameter and the ripple energy distribution to generate the node time-frequency energy vector. Based on the connection edges in the charging pile component topology graph, the node time-frequency energy vector similarity between corresponding component nodes is calculated, and the node time-frequency energy vector similarity is written as the ripple propagation association weight into the corresponding connection edge. Based on the ripple propagation association weights in the charging pile component topology graph, graph convolution updates are performed on the center frequency parameter, tuning frequency parameter, and window width parameter of the Chirplet atom, and topologically coupled Chirplet atoms are generated based on the updated center frequency parameter, tuning frequency parameter, and window width parameter. 5.The machine learning based charging pile fault diagnosis method of claim 1, wherein, The time-frequency evolution characteristics of the generated charging ripple include: Read the topologically coupled Chirplet atom, power ripple signal data, current ripple signal data, voltage ripple signal data, power module temperature data, gun wire temperature data, and acquisition timestamp data. Divide the power ripple signal data, current ripple signal data, and voltage ripple signal data into ripple signal segments within the corresponding stages according to the segmented state sequence of the charging stage. Randomly mask part of the power module temperature data and gun wire temperature data to generate a temperature rise mask sequence. Chirplet time-frequency decomposition is performed on each ripple signal segment based on topologically coupled Chirplet atoms to obtain the power ripple time-frequency response, current ripple time-frequency response and voltage ripple time-frequency response. Temperature rise reconstruction is then performed on the temperature rise mask sequence based on the unmasked power module temperature data and gun wire temperature data to generate the temperature rise reconstruction sequence. The temperature rise reconstruction sequence is used to calculate the temperature change of the power module and the temperature change of the gun wire between adjacent sampling windows. The temperature rise distribution is constructed based on the temperature change of the power module and the temperature change of the gun wire. The Chirplet time-frequency energy distribution is constructed based on the power ripple time-frequency response, the current ripple time-frequency response and the voltage ripple time-frequency response. The optimal transmission mapping relationship is constructed based on the Chirplet time-frequency energy distribution and the temperature rise distribution. The optimal transmission cost from the Chirplet time-frequency energy distribution to the temperature rise distribution is calculated. The temperature rise coupling weight is generated based on the optimal transmission cost. The Chirplet time-frequency energy is modulated based on the temperature rise coupling weight to generate the temperature rise coupled Chirplet time-frequency energy. Based on temperature rise coupling, Chirplet time-frequency energy extraction is used to extract the ripple main frequency position, ripple energy peak value, frequency band spread range, and stage frequency offset, and these are combined in the order of the collected timestamp data to generate charging ripple time-frequency evolution characteristics. 6.The machine learning based charging pile fault diagnosis method of claim 1, wherein, The construction of the ripple dynamic degradation feature sequence includes: Read the ripple dominance frequency position, ripple energy peak, frequency band spread range, stage frequency offset and acquisition timestamp data from the charging ripple time-frequency evolution characteristics. Calculate the ratio between the ripple dominance frequency position difference and the sampling time difference between adjacent sampling windows according to the acquisition timestamp data order to generate ripple frequency drift change characteristics. The ratio between the peak ripple energy difference and the sampling time difference between adjacent sampling windows is calculated according to the order of the collected timestamp data to generate ripple energy fluctuation characteristics. The ratio between the frequency band spread range difference and the sampling time difference between adjacent sampling windows is also calculated to generate frequency band spread change characteristics. The ripple frequency drift characteristics, ripple energy fluctuation characteristics, frequency band spread characteristics, and stage frequency offset corresponding to the same sampling window are combined in the order of the acquisition timestamp data to generate a ripple dynamic degradation characteristic sequence.

7. The machine learning-based charging pile fault diagnosis method of claim 1, wherein, The generation of the charging pile's continuous degradation state vector includes: An improved liquid time constant network is constructed, comprising a ripple dynamic degradation input layer, a liquid state encoding layer, a continuous-time liquid update layer, and a multi-state fusion layer; The ripple dynamic degradation input layer performs time jitter enhancement and amplitude occlusion enhancement on the ripple dynamic degradation feature sequence within the same sampling window to generate a dual-path temporal enhancement sequence. Based on the dual-path temporal enhancement sequence, it performs temporal comparison self-supervised enhancement to generate a liquid input state sequence. The liquid state coding layer reads the liquid input state sequence, extracts the ripple dynamic degradation features of adjacent sampling windows, calculates the time decay coefficient corresponding to the sampling time difference, and performs state coupling mapping on the ripple dynamic degradation features, insulation resistance change features and contactor action state features to generate the initial liquid state sequence and ripple abrupt change features. The continuous-time liquid update layer is based on output voltage data, output current data, charging power data and ripple energy fluctuation characteristics. It introduces energy conservation-Hamilton liquid block to construct Hamiltonian energy state and executes liquid time constant constraint to generate Hamiltonian constrained liquid state sequence. The initial liquid state sequence is mapped to the observable function space, and spectral operator decomposition and exponential mapping time advancement are performed based on the Koopman spectral decomposition mechanism to generate the spectral decomposed liquid state sequence. Based on the topology of charging pile components, node diffusion connection relationships are constructed. Through graph-ODE collaborative diffusion coupling, a state differential equation containing liquid state evolution, node diffusion propagation and random disturbance diffusion is constructed. Random diffusion intensity is generated based on sampling time difference, communication state data and ripple mutation characteristics. The state differential equation is solved by continuous-time diffusion integral to generate a diffusion-coupled liquid state sequence. A symptotic integral time update block is introduced to perform split integral updates, generating a continuously degenerate liquid state sequence and a randomly diffused liquid state sequence. The multi-state fusion layer performs state splicing and weight fusion on Hamilton-constrained liquid state sequences, spectral decomposition liquid state sequences, diffusion-coupled liquid state sequences, continuously degraded liquid state sequences, and random diffusion liquid state sequences to generate a continuously degraded state vector for the charging pile. The improved liquid time constant network was trained, and the prediction error of the continuous degenerate state vector, the reconstruction error of the ripple dynamic degenerate feature sequence, and the continuous time liquid state evolution error were used as joint optimization objectives. The parameters of the ripple dynamic degenerate input layer, the liquid state encoding layer, the continuous time liquid update layer, and the multi-state fusion layer were continuously optimized. Training was stopped when the change value of the joint loss was less than the convergence threshold for five consecutive training cycles.

8. The charging pile fault diagnosis method based on machine learning according to claim 1, characterized in that, The generated charging pile fault diagnosis report includes: Read the charging pile continuous degradation state vector, collect timestamp data and charging stage identification data, and read the power module aging prototype, gun line contact degradation prototype, contactor jitter prototype, insulation attenuation prototype and communication abnormality prototype in the preset fault prototype library. Each fault prototype includes a fault prototype state vector, fault type label, fault location label and fault cause label. Calculate the state matching distance between the continuous degradation state vector of the charging pile and the state vector of each fault prototype. Determine the fault prototype with the smallest state matching distance as the target fault prototype. Calculate the fault confidence based on the state matching distance. When the fault confidence is greater than or equal to 0.75, generate the corresponding fault identification result. When the fault confidence is less than 0.75, generate the fault identification result to be reviewed. A charging pile fault diagnosis report is generated based on the fault identification results, the fault identification results to be reviewed, the target fault prototype, the collected timestamp data, and the charging stage identification data. The charging pile fault diagnosis report includes the fault occurrence stage, fault status, fault type, fault location, fault cause label, and fault confidence level.

Images