Charging pile fault diagnosis method based on state time sequence division

Abstract

The invention relates to a charging pile fault diagnosis method based on dynamic state time sequence division, and the method comprises the steps: obtaining the current, voltage, temperature and charging mode signals of a charging pile in real time through a multi-source synchronous collection module, and constructing multi-mode data; based on a dynamic window optimization algorithm of reinforcement learning, time sequence window parameters are adaptively generated according to charging mode switching and a power change rate, and the problem of data segmentation failure of a fixed window in a fast charging / slow charging scene is solved; nonlinear transformation is carried out on the time sequence data segment, and weak fault features submerged by noise are extracted; event triggering synchronization is performed on asynchronous multi-modal data, so that time dislocation interference is eliminated; the fault probability is output by fusing the multi-dimensional features, a diagnosis report is generated by combining time continuity combination and confidence rating, the fault occurrence time period and type are accurately positioned, the weak fault detection capability and the multi-modal data fusion precision are improved, the omission ratio and the false alarm rate are reduced, and the method is suitable for intelligent operation and maintenance of the charging pile under the complex working condition.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault detection, and in particular to a charging pile fault diagnosis method based on state time series division. Background Art

[0002] With the rapid development of new energy vehicles, charging piles, as core infrastructure, have become increasingly important for the user's charging experience and grid security. Charging pile fault diagnosis technology, through real-time monitoring of electrical parameters, temperature, communication status, and other data, identifies abnormal behavior and locates the source of the fault, becoming a key means of ensuring stable equipment operation. Traditional diagnostic methods typically rely on segmented analysis of operating data within a fixed time window or identify faults based on a single signal threshold. However, in real-world scenarios, charging piles frequently switch between various operating modes, such as fast and slow charging. Parameters such as current and voltage exhibit dynamic, time-varying characteristics, placing higher demands on the capture of fault characteristics and the timeliness of diagnosis.

[0003] In the existing technology, the time series division method based on fixed time windows is difficult to adapt to the drastic differences in data changes under different charging modes. For example, during fast charging, the current transient fluctuations are frequent, and the fixed window may cause the key fault signals to be divided into different time periods, while the long-term steady-state operation in the slow charging stage is prone to introduce redundant data interference. In addition, the asynchrony of multimodal sensor data causes time misalignment during feature fusion, further reducing the diagnostic accuracy. Traditional methods also have limitations in the detection of weak faults. Early fault signals are often submerged by noise, and conventional filtering methods cannot effectively separate effective information, resulting in a high rate of missed detection of hidden faults. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide a charging pile fault diagnosis method based on state time series division that can dynamically adapt to changes in charging modes, accurately divide fault time periods, and integrate multi-source data.

[0005] The purpose of the present invention is achieved by the following scheme:

[0006] In a first aspect, the present invention provides a charging pile fault diagnosis method based on state time series division, comprising the following steps:

[0007] S1: Perform multimodal acquisition of the charging pile's operating data through a multi-source synchronous acquisition module to generate multimodal data containing current data, voltage data, temperature data, and charging mode signals;

[0008] S2: Dynamically optimize window parameters for multimodal data based on a reinforcement learning framework to generate a window parameter set containing real-time time series window length and overlap ratio;

[0009] S3: Based on the real-time time series window length and overlap rate of the window parameter set, the multimodal data is divided into time series segments to generate time series data segments. The time series data segments are then feature extracted based on quantum-derived enhancement technology to generate a feature matrix containing weak fault features.

[0010] S4: Perform asynchronous multimodal feature fusion on the feature matrix through a bidirectional time alignment network and output the fault probability;

[0011] S5: Mark the fault time interval of the time series data segment according to the fault probability, and generate a diagnostic report including the fault type, timestamp and confidence rating. The diagnostic report is used to indicate the area where the charging pile fault occurs.

[0012] In one embodiment, S1 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0013] S11: The current of the charging pile is collected and processed based on current transformer technology. The scaled current analog signal is output through the secondary winding. The signal conditioning circuit then filters and calibrates the amplitude to obtain current data.

[0014] S12: Based on the voltage sampling circuit technology, the voltage of the charging pile is divided and isolated, and the divided signal is quantized by a high-precision analog-to-digital converter to obtain voltage data;

[0015] S13: Based on multi-node thermocouple array technology, the charging gun connector contact temperature, contactor coil temperature, and power module heat sink temperature are synchronously collected and processed to generate temperature data;

[0016] S14: Based on the CAN bus message parsing technology, the SAE J1939 protocol data sent by the battery management system is filtered for identifiers and data domain extraction to generate a charging mode signal;

[0017] S15: Arrange the current data, voltage data, temperature data, and charging mode signal to generate multi-modal data.

[0018] In one embodiment, S2 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0019] S21: Calculate power based on the current data and voltage data of the multimodal data to generate real-time power data including power change rate;

[0020] S22: performing data matching on the multimodal data based on a preset historical database to obtain a difference in the frequency of failures of the charging piles during the same period. The difference in the frequency of failures during the same period is used to indicate a temporal distribution regularity of historical failure events during the same period.

[0021] S23: Feature extraction is performed on the charging mode signal of the multimodal data to generate the charging mode state of the charging pile. The state space of the reinforcement learning agent is constructed based on the charging mode state, real-time power data, and the difference in fault frequency during the same period. The expression of the state space is:

[0022]

[0023] Where S represents the state space, M(t) represents the charging mode state, Indicates the power change rate of real-time power data, ΔF hist Indicates the difference in fault frequency during the same period;

[0024] S24: Define the reward function of the reinforcement learning agent based on the state space. The expression of the reward function is:

[0025] R=-(λ1*R1+λ2*R2+λ3*R3)

[0026] Where R represents the reward function, λ1, λ2, and λ3 represent weight coefficients, λ1+λ2+λ3=1, R1, R2, and R3 represent the missed detection rate, false alarm rate, and computational delay in the state space, respectively.

[0027] S25: Optimize the policy network parameters of the reinforcement learning agent based on the PPO algorithm and reward function, and optimize the multimodal data based on the optimized reinforcement learning agent to generate a window parameter set containing the real-time time series window length and overlap rate.

[0028] In one embodiment, S3 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0029] S31: performing time series division on the multimodal data based on the real-time time series window length and overlap rate of the window parameter set to generate time series data segments;

[0030] S32: Perform quantum tunneling simulation on the current data in the time series data segment to generate a quantum enhanced signal. The calculation formula of the quantum enhanced signal is:

[0031]

[0032] Among them, I quantum (t) is the quantum enhanced signal, i.e., the quantum enhanced signal generated after the quantum tunneling effect simulation at time t, I(t) is the original current data, i.e., the current value collected at time t without the quantum tunneling effect simulation, γ is the tunneling gain coefficient, and e is a natural constant;

[0033] S33: Perform complex Morlet wavelet transform on the quantum enhanced signal to extract the energy ridge features in the time-frequency domain, and process the energy ridge features in the time-frequency domain to generate a feature matrix containing weak fault features.

[0034] In one embodiment, S31 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0035] S311: Dynamically calibrate the real-time time series window length and overlap rate based on the window parameter set to generate window configuration parameters;

[0036] S312: Segmenting the multimodal data based on a sliding window algorithm and window configuration parameters to generate an initial set of time series data segments;

[0037] S313: Perform boundary data compensation processing on the initial time series data segment set, fill the remaining data at the end with zero values or splice it with the previous data segment to generate a complete time series data segment.

[0038] In one embodiment, S4 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0039] S41: Time synchronization is performed on the current data and the temperature data based on the event trigger mechanism to generate an alignment timestamp sequence;

[0040] S42: Perform feature fusion on the feature matrix through a bidirectional long short-term memory network and aligned timestamp sequence to generate a multimodal fusion matrix;

[0041] S43: Based on the multi-head attention mechanism, weights are assigned to the multimodal fusion matrix to calculate the failure probability.

[0042] In one embodiment, S5 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0043] S51: Threshold screening is performed on the time series data segments based on the failure probability to generate a set of candidate failure time periods;

[0044] S52: Merging the candidate fault time period set based on time continuity to generate a continuous fault time interval;

[0045] S53: Convert the continuous fault time intervals into fault type labels based on a preset fault type mapping table, and generate a diagnosis report including the fault type, timestamp, and confidence rating.

[0046] In a second aspect, the present invention provides a charging pile fault diagnosis system based on state time series division, which is configured with the following modules:

[0047] The data acquisition and processing module is used to perform multimodal acquisition of the operating data of the charging pile through the multi-source synchronous acquisition module to generate multimodal data containing current data, voltage data, temperature data and charging mode signals;

[0048] The data window optimization module is used to dynamically optimize window parameters for multimodal data based on a reinforcement learning framework, generating a window parameter set containing real-time time series window length and overlap ratio;

[0049] The fault feature generation module is used to perform time series segmentation and feature extraction on multimodal data based on the real-time time series window length and overlap rate of the window parameter set, and generate a feature matrix containing weak fault features;

[0050] Multimodal fusion module, which is used to perform asynchronous multimodal feature fusion on the feature matrix through a bidirectional time alignment network and output the fault probability;

[0051] The diagnostic tag generation module is used to mark the fault time interval of the time series data segment according to the fault probability and generate a diagnostic report containing the fault type, timestamp and confidence rating. The diagnostic report is used to indicate the area where the charging pile fault occurs.

[0052] In a third aspect, the present application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements any of the above-mentioned charging pile fault diagnosis methods based on state timing division.

[0053] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements any of the above-mentioned charging pile fault diagnosis methods based on state time series division.

[0054] In summary, the charging pile fault diagnosis method based on state time series division provided by the present invention can achieve precise operation and maintenance under complex working conditions through multi-dimensional technological innovation. In order to solve the problem of time series segmentation failure in traditional fixed time window division methods in multi-mode switching scenarios such as fast charging and slow charging, this method introduces a dynamic window optimization mechanism driven by reinforcement learning, which adaptively adjusts the window length and overlap rate according to the real-time power change rate and historical fault patterns. It can effectively avoid the omission of key signals under transient working conditions and redundant data processing under steady-state scenarios, thereby significantly improving the adaptability of time series division to dynamic charging scenarios. In order to solve the time domain misalignment problem caused by sampling rate differences in multi-modal sensor data, an event-triggered bidirectional time alignment network is used to intelligently synchronize asynchronous signals such as current and temperature, thereby solving the pseudo-abnormal interference in cross-modal feature fusion and ensuring the accurate association of fault features under a unified time reference. To address the problem that early weak fault characteristics such as relay contact oxidation and insulation aging are easily drowned out by noise, quantum-derived enhancement technology is used to simulate the nonlinear tunneling effect of the current signal, breaking through the limitations of traditional filtering methods on the separation of high-frequency noise and low-frequency signals, and improving the time-frequency domain identifiability of weak fault components. By integrating dynamic time series division, multimodal alignment and quantum-enhanced feature extraction technology, this method can achieve full-link optimization from data acquisition to decision output. It can not only accurately locate the time period of occurrence of typical faults such as overvoltage, overcurrent, and poor contact in charging piles, but also quantify the reliability of diagnostic results through confidence ratings, providing operation and maintenance personnel with intuitive fault area indications and repair priority recommendations, thereby solving the systemic defects of traditional threshold judgment methods such as high false alarm rates and difficulty in tracing hidden faults under complex working conditions, and providing efficient and reliable technical support for the intelligent operation and maintenance of charging piles.

[0055] For better understanding and implementation, the present invention is described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flow chart of a charging pile fault diagnosis method based on state time series division provided in an embodiment of the present application;

[0057] Figure 2 A schematic diagram of a process for generating a window parameter set containing a real-time time series window length and overlap ratio provided in an embodiment of the present application;

[0058] Figure 3 A schematic diagram of a process for generating a feature matrix containing weak fault features provided in an embodiment of the present application;

[0059] Figure 4 A schematic structural diagram of a charging pile fault diagnosis system based on state time series division is provided in another embodiment of the present application. DETAILED DESCRIPTION

[0060] To facilitate understanding of the present invention, the present invention will be described more fully below with reference to the accompanying drawings. The drawings illustrate preferred embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the disclosure.

[0061] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0062] In one embodiment, Figure 1 As shown, a charging pile fault diagnosis method based on state time series division is provided. This embodiment uses the method applied to a terminal as an example for illustration. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0063] S1: The multi-source synchronous acquisition module performs multi-modal acquisition on the operating data of the charging pile to generate multi-modal data containing current data, voltage data, temperature data and charging mode signals.

[0064] Specifically, the system achieves multimodal acquisition of charging pile operation data through various high-precision sensors deployed on the charging pile itself, its associated circuits, and thermal management components. Current data acquisition utilizes high-sampling-rate Hall-effect sensors, whose sampling frequency can be dynamically adjusted, automatically adapting to the 10kHz to 20kHz range according to different charging modes such as fast charging or slow charging, ensuring accurate capture of transient current fluctuations that may occur during fast charging and steady-state current characteristics during the slow charging phase. Voltage data is acquired through isolated voltage sensors, which have wide-range and high-precision characteristics, covering the charging voltage requirements of different vehicle models and capable of high-precision measurement over the full range of the charging pile output voltage. At the same time, the sampling timing is strictly synchronized with the current data to ensure timing consistency during subsequent feature analysis. Temperature data acquisition uses a distributed thermistor network, covering key heat-generating areas such as power semiconductor devices and transformers within the charging pile. The thermistors have fast response characteristics and can adapt to changes in temperature rise rate under different charging loads.

[0065] Charging mode signal acquisition is obtained by parsing charging pile communication bus protocols such as CAN bus and Modbus TCP / IP. The system can identify mode switching states such as fast charging, slow charging and standby in real time, and accurately align the time stamps of the mode signal with the collected current, voltage and temperature data to form a unified multimodal data sequence. This multimodal data sequence takes time series as the core dimension, and each modal data carries accurate time identification information, providing basic data support for subsequent time series-based diagnostic analysis. The system adopts a redundant design in the data acquisition stage. Each sensor has an independent power supply and signal conditioning circuit to ensure that the failure of a single sensor will not affect the continuity and integrity of the overall data acquisition, thereby ensuring the reliability of the diagnostic system.

[0066] S2: Dynamically optimize the window parameters of multimodal data based on the reinforcement learning framework to generate a window parameter set containing the real-time time series window length and overlap rate.

[0067] Specifically, the system uses its built-in reinforcement learning framework to conduct in-depth analysis of multimodal data to dynamically optimize window parameters. The reinforcement learning agent uses the time series characteristics of the multimodal data as input and, through continuous interaction with the environment, learns how to adjust the window parameters based on the varying operating states of the charging pile. The environmental state is defined by key characteristic indicators such as the current charging mode, the frequency of current and voltage fluctuations, and the rate of temperature rise.

[0068] The agent evaluates the effectiveness of the selected window parameters based on a constructed reward function that comprehensively considers multiple factors, including the timeliness of fault detection, feature integrity, and data redundancy. For example, in fast-charging mode, where current and voltage fluctuate dramatically, the agent will shorten the time series window length and increase the window overlap rate through learning to capture transient fault features in a timely manner. In slow-charging mode, the system will appropriately increase the window length to reduce redundant data while maintaining a certain overlap rate to ensure feature continuity. Through multiple rounds of iterative training, the agent can output a set of optimal window parameters, including the real-time time series window length and overlap rate for the current operating state. These parameters will guide the subsequent data time series segmentation process, enabling it to accurately adapt to the real-time operating conditions of the charging pile.

[0069] S3: Based on the real-time time series window length and overlap rate of the window parameter set, the multimodal data is divided into time series to generate time series data segments. The time series data segments are then feature extracted based on quantum derivative enhancement technology to generate a feature matrix containing weak fault features.

[0070] Specifically, the system uses a dynamically optimized set of window parameters to precisely time-sequence the raw multimodal data, dividing the continuous data stream into multiple time-series data segments with specific time lengths and overlap rates. Each data segment contains complete time series information for current, voltage, temperature, and charging mode signals, providing the foundation for subsequent feature extraction.

[0071] During the feature extraction phase, the system utilizes quantum-derived enhancement technology. Based on quantum computing principles, this technology leverages the superposition and entanglement properties of quantum bits to amplify and isolate subtle fault signatures in time series data. Specifically, the system first quantizes the time series data, mapping it into a quantum state space. Using operations such as the quantum Fourier transform, it reveals implicit periodicity and correlation within the data. Quantum entanglement operations are then used to enhance potential weak fault signatures within the data, making fault signals hidden within noise more prominent. For example, quantum-derived enhancement technology can effectively improve the detectability of weak current fluctuations caused by poor contact or slight leakage current variations due to insulation aging. Finally, quantum measurement is used to convert the enhanced characteristic quantum states into classical eigenvectors, generating a feature matrix containing the weak fault signature information. This feature matrix not only preserves the time series characteristics of the original data but also highlights fault-related feature information. Each eigenvector element corresponds to a specific feature dimension, such as current harmonic components, voltage fluctuation amplitude, and temperature change rate, providing a more discriminative feature foundation for subsequent fault diagnosis models.

[0072] During the feature extraction process, the system can use an adaptive feature selection algorithm to dynamically adjust the feature weights based on the feature distribution in the historical fault sample library to ensure that the extracted feature matrix can reflect the fault information to the greatest extent possible, while reducing the interference of redundant features on the diagnostic results, thereby improving the accuracy and timeliness of fault diagnosis.

[0073] S4: Perform asynchronous multimodal feature fusion on the feature matrix through a bidirectional time alignment network and output the fault probability.

[0074] Specifically, the network structure consists of two time-aligned branches, forward and backward. The forward branch processes the feature sequence from left to right, capturing the cumulative feature evolution trend of the time series through multiple layers of long short-term memory (LSTM) units. This effectively identifies the gradual development of fault features over time, for example, the characteristic change path from a mild temperature anomaly to a significant overheating fault. The backward branch processes the feature sequence from right to left, also using a multi-layer LSTM structure, to extract the lag effects and long-term dependencies of the features. This helps capture the residual effects of features after a fault occurs and the delayed correlation between different modal features. Through multiple layers of time alignment units and based on the dynamic time warping (DTW) algorithm, the network accurately aligns different modal features such as current, voltage, and temperature in the time dimension. This eliminates asynchronous deviations caused by factors such as sampling frequency differences and signal transmission delays during multimodal sensor data acquisition, ensuring that the modal features have a strict temporal correspondence when fused.

[0075] During the fusion process, the network can use the attention mechanism to learn the weight coefficients of each modal feature in the fault diagnosis task and automatically perform weighted fusion of different modal features. The attention mechanism dynamically adjusts the weight distribution by calculating the correlation between each modal feature and the potential fault type. For example, during fast charging, current and voltage features may have a higher contribution to fault diagnosis, while during slow charging, temperature features may be more critical. The final output is a fused fault probability vector, in which each element corresponds to the probability value of a specific fault type. The probability value is normalized by the softmax function to ensure that the sum of all probability values is 1, providing a quantitative basis for the fault diagnosis results. At the same time, the output probability vector also contains uncertainty measurement information, which is used to evaluate the reliability of the diagnosis results. When the uncertainty exceeds the preset threshold, the system will trigger a further diagnosis verification process to ensure the accuracy of the diagnosis results.

[0076] S5: Mark the fault time interval of the time series data segment according to the fault probability, and generate a diagnostic report including the fault type, timestamp and confidence rating. The diagnostic report is used to indicate the area where the charging pile fault occurs.

[0077] Specifically, the system determines the fault time interval by setting a fault probability threshold, judging the start and end timestamps of the fault occurrence. The threshold setting can adopt an adaptive method, based on the statistical analysis of historical fault data, combined with the current operating conditions of the charging pile for dynamic adjustment to balance the sensitivity and false alarm rate of fault detection. After determining the fault time interval, the system identifies the most likely fault type based on the peak position and corresponding probability value of the fault probability vector, and finely classifies the fault type in combination with the historical fault sample library. For example, if the probability of the current feature in the probability vector is high, it may indicate an electrical connection fault or a power module fault; and if the probability of the temperature feature is high, it may point to a heat dissipation system fault or an insulation material overheating fault.

[0078] At the same time, the system assigns a confidence rating to the diagnostic results based on the confidence assessment mechanism of the fault diagnosis algorithm. The rating indicators comprehensively consider factors such as feature matching, probability stability and model generalization ability. The confidence rating adopts a quantitative scoring system and is divided into three levels: high, medium and low. High confidence means that the diagnostic results are highly reliable and can be directly used to guide operation and maintenance operations; medium confidence means that further confirmation is required in combination with other auxiliary diagnostic information; low confidence triggers the system to perform a secondary diagnosis or prompts operation and maintenance personnel to conduct on-site verification. The final generated diagnostic report contains detailed fault information: the fault type is clearly identified in professional terms, covering short circuits, electrical faults such as overcurrent, overheating, thermal faults due to poor heat dissipation, communication faults and other categories; the timestamp accurately records the time period when the fault occurred, accurate to the second level, providing a time basis for fault tracing; the confidence rating is presented in the form of a quantitative score, which intuitively reflects the credibility of the diagnostic results.

[0079] The diagnostic report is output in a structured data format, including XML and JSON formats, to facilitate data interaction with other operation and maintenance management systems. The report also contains a detailed process record of the fault diagnosis, including feature extraction results, fusion process parameters, and probability calculation basis. This provides operation and maintenance personnel with a comprehensive fault diagnosis basis, helps quickly locate the fault area, ensures the stable operation of the charging pile, and improves the safety of the power grid and the user's charging experience.

[0080] In summary, the charging pile fault diagnosis method based on state time series division provided by the present invention can achieve precise operation and maintenance under complex working conditions through multi-dimensional technological innovation. In order to solve the problem of time series segmentation failure in the traditional fixed time window division method in multi-mode switching scenarios such as fast charging and slow charging, this method introduces a dynamic window optimization mechanism driven by reinforcement learning, which adaptively adjusts the window length and overlap rate according to the real-time power change rate and historical fault rules. It can effectively avoid the omission of key signals under transient working conditions and redundant data processing under steady-state scenarios, thereby significantly improving the adaptability of time series division to dynamic charging scenarios. In order to solve the time domain misalignment problem caused by sampling rate differences in multi-modal sensor data, an event-triggered bidirectional time alignment network is used to intelligently synchronize asynchronous signals such as current and temperature, so as to solve the pseudo-abnormal interference in cross-modal feature fusion and ensure the accurate association of fault features under a unified time reference. To address the problem that early weak fault characteristics such as relay contact oxidation and insulation aging are easily drowned out by noise, quantum-derived enhancement technology is used to simulate the nonlinear tunneling effect of the current signal, breaking through the limitations of traditional filtering methods on the separation of high-frequency noise and low-frequency signals, and improving the time-frequency domain identifiability of weak fault components. By integrating dynamic time series division, multimodal alignment and quantum-enhanced feature extraction technology, this method can achieve full-link optimization from data acquisition to decision output. It can not only accurately locate the time period of occurrence of typical faults such as overvoltage, overcurrent, and poor contact in charging piles, but also quantify the reliability of diagnostic results through confidence ratings, providing operation and maintenance personnel with intuitive fault area indications and repair priority recommendations, thereby solving the systemic defects of traditional threshold judgment methods such as high false alarm rates and difficulty in tracing hidden faults under complex working conditions, and providing efficient and reliable technical support for the intelligent operation and maintenance of charging piles.

[0081] In one embodiment, S1 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0082] S11: The current of the charging pile is collected and processed based on the current transformer technology, and the scaled current analog signal is output through the secondary winding. The current data is then obtained by filtering and amplitude calibration through the signal conditioning circuit.

[0083] Specifically, a current transformer is an electrical device that converts large currents into small currents based on the principle of electromagnetic induction. Its primary winding is connected in series in the main circuit of the charging pile. When current flows through the primary winding, a current proportional to the primary current is induced in the secondary winding. The scaled current analog signal output by the secondary winding can truly reflect the current changes in the main circuit of the charging pile, and the signal is isolated from the primary current, which has high safety. Subsequently, the current analog signal is sent to the signal conditioning circuit. In this circuit, high-frequency interference noise is removed through a filtering link. For example, a low-pass filter is used to filter out noise signals above a specific cutoff frequency. At the same time, the amplitude of the signal is accurately adjusted using an amplitude calibration circuit to ensure that it meets the input requirements of the subsequent analog-to-digital converter, thereby obtaining accurate, stable and interference-resistant current data, providing reliable current information for operation monitoring and fault diagnosis of the charging pile.

[0084] S12: Based on the voltage sampling circuit technology, the voltage of the charging pile is divided and isolated, and the divided signal is quantized by a high-precision analog-to-digital converter to obtain voltage data.

[0085] Specifically, the voltage sampling circuit usually includes key parts such as a voltage-dividing resistor network and an isolation module. The voltage-dividing resistor network divides the high voltage of the charging pile according to a certain ratio, so that the output voltage signal is within the range that can be safely processed by the subsequent processing circuit. For example, in a DC charging pile, facing an output voltage of up to 1000Vdc, by reasonably designing the resistance ratio of the voltage-dividing resistor, the voltage can be reduced to a level of several volts to more than ten volts. Subsequently, the isolation module isolates the divided signal, on the one hand protecting the precision circuits such as the back-end analog-to-digital converter from the potential damage risk of high voltage, and on the other hand ensuring electrical safety during the measurement process. The isolated voltage signal is sent to a high-precision analog-to-digital converter. The analog-to-digital converter accurately converts the analog voltage signal into a digital signal based on a specific quantization bit number and sampling frequency, thereby obtaining the voltage data of the charging pile. This data can accurately reflect the voltage output status of the charging pile under different operating modes, providing a key basis for subsequent fault diagnosis and operation analysis.

[0086] S13: Based on multi-node thermocouple array technology, the charging gun connector contact temperature, contactor coil temperature and power module heat sink temperature are synchronously collected and processed to generate temperature data.

[0087] Specifically, a thermocouple is a temperature sensor based on the Seebeck effect. When there is a temperature difference between the two ends of its two thermocouples made of different materials, a thermoelectric potential will be generated in a closed loop. This thermoelectric potential has a definite functional relationship with the temperature difference. The multi-node thermocouple array rationally arranges multiple thermocouple sensors at key heating parts such as the charging gun connector contacts, contactor coils, and power module heat sinks. Each thermocouple node can sense the temperature changes at the corresponding position in real time and convert them into corresponding thermoelectric potential signals. These signals are preliminarily processed by a dedicated signal acquisition circuit, including cold-end compensation and signal amplification. Cold-end compensation technology can effectively eliminate the influence of ambient temperature changes on the thermoelectric potential measurement results and improve the accuracy of temperature measurement; the signal amplification link amplifies the weak thermoelectric potential signal to a suitable amplitude for subsequent analog-to-digital conversion and data processing.

[0088] The processed temperature signal is converted into digital temperature data, which enables the system to fully understand the temperature conditions of key parts inside the charging pile, promptly discover potential overheating faults, and provide strong guarantees for the safe and stable operation of the charging pile.

[0089] S14: Based on the CAN bus message parsing technology, the SAE J1939 protocol data sent by the battery management system is subjected to identifier filtering and data domain extraction processing to generate a charging mode signal.

[0090] Specifically, the CAN bus, a serial communication protocol widely used in the automotive and industrial sectors, offers high reliability, strong real-time performance, and excellent immunity to electromagnetic interference, making it ideal for data transmission between components in electric vehicle charging systems. The battery management system, in accordance with the SAE J1939 protocol, sends messages containing key information such as the vehicle's battery status and charging requirements to the charging station via the CAN bus.

[0091] The CAN bus controller in the system parses received message data in real time. First, it selects specific messages related to the charging mode based on the identifier filtering mechanism specified in the protocol. Subsequently, through in-depth extraction and parsing of the data fields in the message, it obtains field information representing the current charging mode, such as charging power level, charging stages such as pre-charging, constant current charging, and constant voltage charging, as well as key parameters such as the battery charge state. This extracted charging mode-related information is integrated to generate a charging mode signal, providing an accurate charging status basis for the operation control and fault diagnosis of the charging pile. This enables the system to promptly adjust monitoring and diagnostic strategies based on different charging modes, ensuring the efficiency and safety of the charging process.

[0092] S15: Arrange the current data, voltage data, temperature data, and charging mode signal to generate multi-modal data.

[0093] Specifically, the system performs strict time synchronization alignment based on the timestamp information of each data to ensure that the correspondence between the current, voltage, temperature and charging mode signals in the time dimension is accurate, thus laying the foundation for subsequent comprehensive analysis based on time series. Then, system 1 normalizes and integrates different types of data according to the preset data structure and format. For example, analog data such as current, voltage, and temperature are uniformly converted into a standardized digital format, and organically integrated with digital status information such as charging mode signals. The multimodal data finally formed not only covers key information in multiple aspects such as the electrical, thermal and communication status of the charging pile during operation, but also has good format consistency and time coherence. It can comprehensively and accurately reflect the operating status of the charging pile, and provides a rich and reliable data foundation for subsequent advanced functions such as fault diagnosis, performance evaluation and intelligent decision-making.

[0094] In one embodiment, S2 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0095] S21: Calculate power based on the current data and voltage data of the multimodal data to generate real-time power data including power change rate.

[0096] Specifically, the system uses the power calculation formula based on the current and voltage information in the multimodal data to calculate the output power of the charging pile in real time. The output power calculation formula is:

[0097] P(t)=I(t)*U(t)

[0098] Among them, P(t) is the output power of the charging pile, I(t) is the current data, and U(t) is the voltage data. During the calculation process, the system uses a high-precision analog-to-digital converter to sample the current and voltage analog signals to ensure the accuracy and stability of the sampling. To reduce short-term fluctuation interference, the system uses a sliding average filter algorithm to smooth the power data. On this basis, by taking the derivative of the power data, the power change rate is obtained. This data is then integrated with power data to generate real-time power data. This real-time power data not only accurately reflects the output power level of the charging pile at different points in time, but also, through the key indicator of power change rate, promptly captures the dynamic changes in the charging pile's output power. This provides a key basis for subsequent fault diagnosis and charging strategy adjustments, and helps the system identify potential fault precursors in advance, such as sudden power changes that may indicate load anomalies or electrical faults.

[0099] S22: Perform data matching on the multimodal data based on a preset historical database to obtain the difference in fault frequencies of the charging piles during the same period. The difference in fault frequencies during the same period is used to indicate the temporal distribution regularity of historical fault events during the same period.

[0100] Specifically, the system compares the collected multimodal data with the preset historical database to select historical data of the same period that are similar to the current operating conditions. The historical database covers a large amount of operating data and fault records accumulated by charging piles in different operating stages and under different environmental conditions. These data have been strictly screened and sorted to ensure their representativeness and reliability. Through advanced data matching algorithms, the system calculates the difference in fault frequency ΔF during the same period. hist This indicator reflects the degree of deviation between the fault frequency during the current operating cycle and the historical average for the same period. A positive value indicates that the current fault frequency is higher than the historical average, indicating potential risks; a negative value indicates that the current fault frequency is lower and the system is relatively stable. This process leverages big data analysis technology to deeply explore patterns and trends in historical data, providing strong support for fault diagnosis, helping the system identify fault risks beyond the normal range, and providing important reference for subsequent diagnostic decisions.

[0101] S23: Extract features of the charging mode signal of the multimodal data to generate the charging mode state of the charging pile, and construct the state space of the reinforcement learning agent based on the charging mode state, real-time power data, and the difference in fault frequency during the same period.

[0102] Specifically, the system uses deep learning algorithms, such as Convolutional Neural Network (CNN) or Long Short-Term Memory Network (LSTM), to extract features from the charging mode signal. The charging mode signal contains a wealth of information, including key parameters such as charging current, charging voltage, charging status, and battery temperature. These parameters present unique characteristic patterns in different charging stages, such as the constant current charging stage where the current is stable and the voltage gradually rises, and the constant voltage charging stage where the current gradually decreases and the voltage remains constant. Through deep learning and analysis of the algorithm, the characteristic pattern of the output current of the charging pile is identified to form the charging mode state M(t). Combined with the real-time power change rate The difference between the fault frequency and the same period ΔF hist , construct the state space of the reinforcement learning agent, and the expression of the state space is:

[0103]

[0104] Where S represents the state space, M(t) represents the charging mode state, Indicates the power change rate of real-time power data, ΔF hist The state space integrates multi-dimensional information such as the current charging mode, power change trends, and historical failure patterns of the charging pile. This provides a comprehensive description of the environmental state for the reinforcement learning agent, enabling it to accurately perceive the environmental state and lay the foundation for subsequent precise decision-making and optimization of the charging process.

[0105] S24: Define the reward function of the reinforcement learning agent based on the state space.

[0106] Specifically, the reward function is expressed as:

[0107] R=-(λ1*R1+λ2*R2+λ3*R3)

[0108] Here, R represents the reward function, λ1, λ2, and λ3 represent weight coefficients, where λ1+λ2+λ3=1. R1, R2, and R3 represent the missed detection rate, false alarm rate, and computational delay, respectively, in the state space. The missed detection rate reflects the proportion of actual faults that the system fails to detect in a timely manner, the false alarm rate indicates the probability that the system mistakenly misjudges a normal state as a faulty state, and the computational delay reflects the time lag in the system's data processing and diagnosis process. By properly setting the weight coefficients, the system can achieve a dynamic balance between the missed detection rate, false alarm rate, and computational delay. For example, in fast-charging scenarios with strict real-time requirements, the weight of computational delay can be appropriately increased to encourage the agent to optimize the computational process and improve diagnostic speed. In special charging scenarios with extremely high accuracy requirements, the weights of the missed detection rate and false alarm rate can be increased to encourage the agent to prioritize the accuracy of the diagnostic results. The reward function provides a clear behavioral guide by quantitatively evaluating the effectiveness of the agent's behavior, motivating it to continuously adjust and optimize its strategy to achieve optimal fault diagnosis performance.

[0109] S25: Optimize the policy network parameters of the reinforcement learning agent based on the PPO algorithm and reward function, and optimize the multimodal data based on the optimized reinforcement learning agent to generate a window parameter set containing the real-time time series window length and overlap rate.

[0110] Specifically, during the parameter update process, a reward function is used to evaluate and guide the agent's behavior. The Proximal Policy Optimization (PPO) algorithm, an advanced reinforcement learning optimization algorithm, incorporates trust region constraints during policy updates to ensure the stability and effectiveness of policy updates, avoiding performance fluctuations and divergence during the update process. The optimized agent processes multimodal data based on the current state space information, dynamically adjusting the length and overlap of the time window to generate the optimal set of window parameters. The system comprehensively considers the characteristics and requirements of different charging stages. For example, during fast charging, the time window length is shortened based on the power change rate and charging mode to improve the detection of transient faults; during slow charging, the window length is appropriately extended to reduce data redundancy. Furthermore, the window parameters are further optimized based on differences in fault frequency during the same period, ensuring efficient and accurate fault diagnosis and performance optimization under different operating conditions. This process optimizes the entire process, from data acquisition and feature extraction to intelligent decision-making, significantly improving the intelligence level and practical application effectiveness of the charging pile fault diagnosis system.

[0111] The proposed charging pile fault diagnosis method based on state-time series segmentation effectively addresses the lack of adaptability to dynamic scenarios caused by fixed time window segmentation in traditional charging pile fault diagnosis by constructing a dynamic window parameter optimization mechanism based on reinforcement learning. Furthermore, the system calculates the power change rate based on real-time current and voltage data and constructs a multidimensional state space based on historical fault frequency differences. This enables the reinforcement learning agent to deeply perceive multi-dimensional operating conditions such as charging mode switching, load fluctuations, and historical fault patterns, overcoming the limitations of traditional methods that rely on manual experience to preset window parameters. By defining a composite reward function that integrates the missed detection rate, false alarm rate, and computational latency, the agent is guided to balance diagnostic accuracy and real-time performance requirements during the dynamic optimization process, avoiding systematic bias caused by single-metric optimization. By continuously iterating the policy network parameters based on a proximal policy optimization algorithm, the generated real-time time series window length and overlap ratio are adaptively adapted to capture high-frequency transient signals in fast-charging scenarios and track long-term degradation trends in slow-charging scenarios, significantly improving the rationality of data segmentation under different operating conditions. This method embeds historical fault patterns into the state space to enhance the model's ability to predict periodic hidden faults. At the same time, it combines the real-time feedback mechanism of the power change rate to ensure the close coupling of dynamic adjustment of window parameters with the actual operating status of the equipment. It fundamentally solves the problem of feature truncation or redundant calculation caused by parameter rigidity of fixed windows in complex charging scenarios, lays a high-quality data foundation for subsequent quantum-enhanced feature extraction and multimodal fusion, and realizes full-process optimization from data-driven to intelligent decision-making.

[0112] In one embodiment, Figure 3As shown, S3 of the charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0113] S31: Perform time series division on the multimodal data based on the real-time time series window length and overlap rate of the window parameter set to generate time series data segments.

[0114] Specifically, the purpose of time series segmentation is to divide a continuous data stream into multiple data segments with temporal continuity and correlation for subsequent analysis and processing. The window length determines the time span of each data segment, while the overlap rate controls the degree of overlap between adjacent data segments. This division method can ensure that each data segment contains enough information, and at the same time, the overlapping parts of adjacent data segments help capture the dynamic changes of the data, avoiding the omission of important information due to excessive window division, or the lack of correlation between data due to excessive window division. Preferably, the time series data segments are generated by the following steps:

[0115] S311: Dynamically calibrate the real-time timing window length and overlap rate based on the window parameter set to generate window configuration parameters.

[0116] Specifically, the system calibrates the window length, taking into account the rate of change of the electrical characteristics of the charging pile under different operating modes and the period of manifestation of fault characteristics. For example, in fast charging mode, in order to capture high-frequency transient fluctuations in parameters such as current and voltage, the window length is set to a shorter millisecond level; in slow charging mode, due to the relatively slow data changes, the window length is appropriately extended to seconds or even longer to reduce data redundancy and improve analysis efficiency. The calibration of the overlap rate ensures that there is a certain degree of overlap between adjacent data segments, avoiding the omission of important information due to excessive window division, or the lack of correlation between data due to excessive window division. After dynamic calibration, the system generates window configuration parameters, providing a precise parameter basis for subsequent time series division operations.

[0117] S312: Segment the multimodal data based on the sliding window algorithm and window configuration parameters to generate an initial set of time series data segments.

[0118] Specifically, the moving window algorithm is an efficient data processing technique that can segment a continuous data stream into multiple data segments of fixed length and overlap while maintaining the temporal continuity of the data. During this process, the system first sets the length and step size of the sliding window based on the window configuration parameters. The sliding window slides point by point along the time axis of the multimodal data, intercepting a data segment equal to the window length with each step. These data segments not only contain information from multiple modalities such as current, voltage, temperature, and charging mode signals, but also, due to the overlapping nature of the sliding window, adjacent data segments share some data, thus ensuring feature continuity and integrity. The generation of the initial set of time series data segments is a key step in the fault diagnosis process, providing the foundation for subsequent data processing and feature extraction. In this way, the system can convert the original multimodal data into a series of data segments that are easy to process and analyze, laying the foundation for subsequent boundary data compensation processing and feature extraction.

[0119] S313: Perform boundary data compensation processing on the initial time series data segment set, fill the remaining data at the end with zero values or splice it with the previous data segment to generate a complete time series data segment.

[0120] Specifically, in the process of generating the initial set of time series data segments, since the data length may not fully match the window length and overlap rate requirements, there may be residual data at the end of the data that cannot form a complete data segment. To solve this problem, the system adopts a boundary data compensation processing strategy. When the amount of residual data at the end is not enough to form a complete data segment, the system will process it according to the actual situation; if the amount of residual data is small, the system will fill the remaining data segment with zero values so that its length meets the window length requirement; if the amount of residual data is large, the system will splice it with the previous data segment to form a complete data segment. The time series data segments processed in this way not only maintain temporal coherence, but are also more complete in features, ensuring the integrity and validity of the data, providing a high-quality data foundation for subsequent feature extraction and fault diagnosis, and avoiding the problem of inaccurate diagnostic results or reduced reliability due to missing data.

[0121] S32: Perform quantum tunneling effect simulation on the current data in the time series data segment to generate a quantum enhanced signal.

[0122] Specifically, the system simulates the quantum tunneling effect on the current data in the time series data segment to enhance the weak fault characteristics therein. The quantum tunneling effect simulation is based on the principles of quantum mechanics and processes the current data through a specific mathematical model to generate a quantum enhanced signal. The calculation formula of the quantum enhanced signal is:

[0123]

[0124] Among them, I quantum (t) is the quantum-enhanced signal, generated at time t after quantum tunneling simulation. I(t) is the original current data, the current value collected at time t without quantum tunneling simulation. γ is the tunneling gain coefficient, and e is a natural constant. During the simulation, the system treats the current data as a particle flow and amplifies weak fluctuations in the current using the probabilistic properties of quantum tunneling. When certain fluctuations in the current data represent potential faults, these features are enhanced by the quantum tunneling effect, making them easier to distinguish from background noise. For example, if partial discharge occurs during charging, the resulting tiny current pulses are typically drowned out by noise. However, after quantum tunneling simulation, the amplitudes of these weak pulses are significantly increased, significantly different from the current characteristics under normal operating conditions, providing key clues for subsequent fault detection. This process not only improves the detectability of fault characteristics but also effectively ensures the accuracy of the system in early fault diagnosis.

[0125] S33: Perform complex Morlet wavelet transform on the quantum enhanced signal to extract the energy ridge features in the time-frequency domain, and process the energy ridge features in the time-frequency domain to generate a feature matrix containing weak fault features.

[0126] Specifically, the complex Morlet wavelet transform is a powerful time-frequency analysis tool that can convert signals from the time domain to the time-frequency domain, providing information about the signal's energy distribution across time and frequency. During the transformation process, the system uses the complex Morlet wavelet as the mother wavelet function. Through scaling and translation operations, it convolves the complex wavelet with the quantum-enhanced signal to obtain complex wavelet coefficients. These coefficients contain detailed information about the signal at different frequencies and times. Time-frequency domain energy ridge features are derived by analyzing the modulus of the complex wavelet coefficients. They represent the concentrated distribution path of the signal energy in the time-frequency plane and reflect the main energy variation trends and characteristics of the signal. These energy ridge features can effectively capture key information such as sudden changes and fluctuations in the signal. In particular, when a fault occurs, the signal's energy distribution undergoes significant changes. Energy ridge features can keenly reflect these changes, providing an important basis for fault diagnosis. After extracting the time-frequency domain energy ridge features, the system further processes and integrates them to remove redundant information, enhance their independence and discriminability, and ultimately generate a feature matrix containing weak fault signatures.

[0127] This feature matrix organizes feature information in matrix form, with each row representing a specific feature dimension and each column corresponding to a time-frequency point. The matrix element values quantify the strength of the feature at that time-frequency point. In this way, the system can transform complex quantum-enhanced signals into a structured feature matrix, providing subsequent fault diagnosis models with high-dimensional, high-information input data that accurately describes fault characteristics. This ensures that the system can accurately identify and locate potential faults in charging piles, thereby improving the accuracy and reliability of fault diagnosis.

[0128] The charging pile fault diagnosis method based on state time series division provided above can significantly improve the accuracy and reliability of charging pile weak fault detection through the synergy of dynamic time series division and quantum enhanced feature extraction technology. Based on the window parameters dynamically optimized by reinforcement learning, the time series window length and overlap rate are adaptively adjusted for different modes such as fast charging and slow charging, effectively overcoming the limitations of traditional fixed windows in transient signal capture and steady-state trend tracking, ensuring that key fault features are not truncated or submerged in redundant data. The current signal is transformed nonlinearly through quantum tunneling effect simulation, and the high-amplitude noise component is suppressed by using the exponential decay function. At the same time, the signal-to-noise ratio of the low-amplitude weak fault signal is amplified, thereby solving the problem of separation of the noise and fault spectrum overlapping areas of conventional filtering methods, making the time domain characteristics of early latent faults such as relay contact oxidation and insulation aging explicit. Combining the time-frequency analysis capabilities of the complex Morlet wavelet transform, this method accurately extracts energy ridge features from quantum-enhanced signals, breaking through the bottleneck of traditional frequency-domain analysis in ambiguously locating transient anomalies. It effectively identifies microscopic anomalies such as current waveform distortion and sudden increases in high-frequency harmonics, and reconstructs multidimensional time-frequency features into a highly discriminative feature matrix, providing a high-information-density input source for subsequent multimodal data fusion. This method achieves full-link optimization from data denoising and feature enhancement to time-frequency analysis through cascaded processing of dynamic windows and quantum enhancement, providing an innovative technical path for early fault warning and precise diagnosis of charging piles under complex operating conditions.

[0129] In one embodiment, S4 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0130] S41: Performing time synchronization alignment on the current data and the temperature data based on an event trigger mechanism to generate an alignment timestamp sequence.

[0131] Specifically, during the operation of the charging pile, current data and temperature data are collected by different sensors respectively. These sensors may have different sampling frequencies and time bases, resulting in time deviations between the data. The event triggering mechanism establishes a time association between the two by identifying key events in the data, such as sudden changes in current and sudden temperature rises. When these key events are detected, the system uses them as reference points for time alignment and adjusts the timestamps of the data so that data of different modalities are accurately synchronized in time. For example, during the charging process, when transient fluctuations in current (such as slight changes in current caused by poor contact) are detected, the temperature data may rise slightly. The system regards these two events as related events and uses this as a benchmark to correct the timestamps. In this way, the system can effectively eliminate the time deviation between data from different sensors, generate a unified aligned timestamp sequence, ensure the consistency of current data and temperature data in the time dimension, and provide an accurate time basis for subsequent feature fusion.

[0132] S42: Feature fusion is performed on the feature matrix through a bidirectional long short-term memory network and aligned timestamp sequences to generate a multimodal fusion matrix.

[0133] Specifically, the Bidirectional Long Short-Term Memory (BiLSTM) network is a neural network structure capable of processing sequential data. It combines forward and backward long short-term memory networks to capture long-term dependencies and bidirectional contextual information in sequential data. During the feature fusion process, aligning the timestamp sequences ensures temporal consistency between current and temperature data, enabling the system to input feature vectors from the same time point into the BiLSTM network. The forward layer of the BiLSTM network processes data from the start to the end of the sequence, capturing the cumulative evolution of features; while the backward layer processes data from the end to the start of the sequence, extracting feature lags and long-term dependencies. In this way, the system fully utilizes the time series information of current and temperature data, exploring the inherent connections between them, and generating a multimodal fusion matrix containing rich time series features and multimodal information. This matrix not only preserves the feature information of the original data but also integrates the complementary information between different modal data, providing a more comprehensive and accurate feature representation for subsequent fault diagnosis.

[0134] S43: Based on the multi-head attention mechanism, weights are assigned to the multimodal fusion matrix to calculate the failure probability.

[0135] Specifically, the multi-head attention mechanism is a technology that effectively captures the correlations between different features in data. It uses multiple attention heads to calculate weights between features, then concatenates and fuses these weights to produce a final weight distribution. In the multimodal fusion matrix, each element represents information at a specific time point and on a specific feature dimension. By calculating the correlations between these elements, the multi-head attention mechanism assigns a weight to each feature dimension, reflecting its importance in fault diagnosis. For example, if a feature dimension in current data is highly correlated with fault occurrence, it will be assigned a larger weight.

[0136] Based on these weights, the system performs a weighted summation on the multimodal fusion matrix to produce a comprehensive feature vector. This comprehensive feature vector is then processed using a fault probability calculation model, such as logistic regression or neural networks, to ultimately output the fault probability. The fault probability represents the likelihood of a charging pile failure, with values ranging from 0 to 1, with values closer to 1 indicating a higher probability of failure. In this way, the system effectively utilizes the information in the multimodal fusion matrix, combined with the weight allocation of the multi-head attention mechanism, to accurately calculate the fault probability, providing a reliable basis for fault diagnosis and early warning of charging piles.

[0137] In one embodiment, S5 of a charging pile fault diagnosis method based on state time series division provided by the present invention specifically includes the following steps:

[0138] S51: Perform threshold screening on the time series data segments based on the failure probability to generate a set of candidate failure time periods.

[0139] Specifically, the system first evaluates each time series data segment based on a preset failure probability threshold. This threshold is typically adjusted dynamically based on historical failure data and the operating characteristics of the charging station to ensure the accuracy and reliability of the screening results. When the failure probability of a time series data segment exceeds the preset threshold, the system marks it as a potential failure period and records its start and end timestamps.

[0140] To improve screening accuracy, the system also considers the continuity and stability of failure probabilities. If the failure probability of multiple consecutive time series data segments exceeds a threshold, the system merges these segments into a single, continuous candidate failure segment. Furthermore, the system performs preliminary verification of the screening results, comparing them with historical failure patterns to eliminate possible false positives. Ultimately, the system generates a set of candidate failure segments, each with a clear start and end timestamp, providing a foundation for subsequent fault merging and precise diagnosis.

[0141] S52: Merge the candidate fault time period set based on time continuity to generate a continuous fault time interval.

[0142] Specifically, the system sorts each time period in the set of candidate fault time periods to ensure that they are arranged in chronological order. The system then checks the time interval between adjacent time periods. If the time interval between two adjacent time periods is less than a preset time threshold, the two time periods are considered to belong to the same fault event and are merged into a continuous fault time interval. The preset time threshold is adjusted based on the operating characteristics of the charging pile and the type of fault to ensure the rationality of the merging result. For example, for transient faults, the time threshold can be set to a smaller value; for persistent faults, the time threshold can be appropriately increased. During the merging process, the system also records the start and end timestamps of each continuous fault time interval and calculates its duration. In addition, the system verifies the merged fault time interval and ensures the accuracy and reliability of the merging result by comparing it with historical fault data. Ultimately, the system generates a set of continuous fault time intervals, providing an accurate time range for subsequent fault type identification and diagnostic report generation.

[0143] S53: Convert the continuous fault time intervals into fault type labels based on a preset fault type mapping table, and generate a diagnosis report including the fault type, timestamp, and confidence rating.

[0144] Specifically, the system extracts a feature matrix within each consecutive fault time interval, including multimodal feature information such as current, voltage, and temperature. The system then matches this feature information with a pre-set fault type mapping table and determines the most likely fault type by calculating feature similarity or using classification algorithms such as support vector machines and random forests. During the matching process, the system considers comprehensive information from multiple feature dimensions to improve the accuracy of fault type identification. After determining the fault type, the system generates a confidence rating based on the confidence level of the match. The rating scale is generally categorized as high, medium, and low, reflecting the reliability of the diagnosis result. A high confidence rating indicates a close match with historical fault patterns, a medium confidence rating indicates some uncertainty, and a low confidence rating indicates a possible false alarm or data anomaly. Finally, the system generates a detailed diagnostic report that includes the fault type, fault timestamp, fault duration, and confidence rating. Outputting this report in a standardized format, the diagnostic report provides operations and maintenance personnel with clear, intuitive, and comprehensive fault information, enabling them to quickly and accurately locate the fault and implement targeted remedial measures promptly, effectively ensuring the normal operation of the charging station and the safety and stability of the power grid.

[0145] Preferably, if Figure 4 As shown, the present invention provides a charging pile fault diagnosis system 600 based on state time series division, which is configured with the following modules:

[0146] The data acquisition and processing module 610 is used to perform multimodal acquisition of the operating data of the charging pile through the multi-source synchronous acquisition module to generate multimodal data containing current data, voltage data, temperature data and charging mode signals;

[0147] The data window optimization module 620 is used to dynamically optimize the window parameters of multimodal data based on the reinforcement learning framework to generate a window parameter set containing the real-time time series window length and overlap rate;

[0148] The fault feature generation module 630 is used to perform time series segmentation and feature extraction on the multimodal data based on the real-time time series window length and overlap rate of the window parameter set to generate a feature matrix containing weak fault features;

[0149] A multimodal fusion module 640 is configured to perform asynchronous multimodal feature fusion on the feature matrix through a bidirectional time alignment network and output a fault probability;

[0150] The diagnostic tag generation module 650 is used to mark the fault time interval of the time series data segment according to the fault probability and generate a diagnostic report including the fault type, timestamp and confidence rating. The diagnostic report is used to indicate the area where the charging pile fault occurs.

[0151] In summary, the charging pile fault diagnosis system based on state time series division provided by the present invention can achieve precise operation and maintenance under complex working conditions through multi-dimensional technological innovation. In order to solve the problem of time series segmentation failure in the traditional fixed time window division method in multi-mode switching scenarios such as fast charging and slow charging, this method introduces a dynamic window optimization mechanism driven by reinforcement learning, which adaptively adjusts the window length and overlap rate according to the real-time power change rate and historical fault patterns. It can effectively avoid the omission of key signals under transient working conditions and redundant data processing under steady-state scenarios, thereby significantly improving the adaptability of time series division to dynamic charging scenarios. In order to solve the time domain misalignment problem caused by sampling rate differences in multi-modal sensor data, an event-triggered bidirectional time alignment network is used to intelligently synchronize asynchronous signals such as current and temperature, thereby solving the pseudo-abnormal interference in cross-modal feature fusion and ensuring the precise association of fault features under a unified time reference. To address the problem that early weak fault characteristics such as relay contact oxidation and insulation aging are easily drowned out by noise, quantum-derived enhancement technology is used to simulate the nonlinear tunneling effect of the current signal, breaking through the limitations of traditional filtering methods on the separation of high-frequency noise and low-frequency signals, and improving the time-frequency domain identifiability of weak fault components. By integrating dynamic time series division, multimodal alignment and quantum-enhanced feature extraction technology, this system can achieve full-link optimization from data acquisition to decision output. It can not only accurately locate the occurrence time period of typical faults such as overvoltage, overcurrent, and poor contact in charging piles, but also quantify the reliability of diagnostic results through confidence ratings, providing operation and maintenance personnel with intuitive fault area indications and repair priority recommendations, thereby solving the systemic defects of traditional threshold judgment methods such as high false alarm rates and difficulty in tracing hidden faults under complex working conditions, providing efficient and reliable technical support for the intelligent operation and maintenance of charging piles.

[0152] Preferably, the data acquisition and processing module 610 is configured with the following units:

[0153] The current data acquisition unit is used to collect and process the current of the charging pile based on current transformer technology. The scaled current analog signal is output through the secondary winding, and then filtered and amplitude calibrated by the signal conditioning circuit to obtain current data.

[0154] The voltage data acquisition unit is used to divide and isolate the voltage of the charging pile based on the voltage sampling circuit technology, and quantize the divided signal through a high-precision analog-to-digital converter to obtain voltage data;

[0155] The temperature data acquisition unit is used to synchronously collect and process the charging gun connector contact temperature, contactor coil temperature, and power module heat sink temperature based on multi-node thermocouple array technology to generate temperature data;

[0156] A charging mode signal generation unit is used to perform identifier filtering and data field extraction processing on the SAE J1939 protocol data sent by the battery management system based on CAN bus message parsing technology to generate a charging mode signal;

[0157] The multimodal data generation unit is used to organize current data, voltage data, temperature data and charging mode signals to generate multimodal data.

[0158] Preferably, the data window optimization module 620 is configured with the following units:

[0159] A real-time power data generating unit, configured to perform power calculation based on the current data and voltage data of the multimodal data, and generate real-time power data including the power change rate;

[0160] A unit for obtaining a difference in fault frequency during the same period is used to match multimodal data based on a preset historical database to obtain a difference in the fault frequency of the charging pile during the same period. The difference in the fault frequency during the same period is used to indicate the temporal distribution regularity of historical fault events during the same period.

[0161] A state space construction unit is used to extract features from the charging mode signal of multimodal data, generate the charging mode state of the charging pile, and construct the state space of the reinforcement learning agent based on the charging mode state, real-time power data, and the difference in fault frequency during the same period;

[0162] Reward function definition unit, used to define the reward function of the reinforcement learning agent based on the state space;

[0163] The window parameter set generation unit is used to optimize the policy network parameters of the reinforcement learning agent based on the PPO algorithm and reward function, and optimize the multimodal data based on the optimized reinforcement learning agent to generate a window parameter set containing the real-time time series window length and overlap rate.

[0164] Preferably, the fault signature generation module 630 is configured with the following units:

[0165] A time series data segment generating unit, configured to perform time series segmentation on the multimodal data based on the real-time time series window length and overlap rate of the window parameter set to generate time series data segments;

[0166] A quantum enhanced signal generating unit, configured to simulate the quantum tunneling effect on the current data in the time series data segment to generate a quantum enhanced signal;

[0167] The characteristic matrix generating unit is used to perform complex Morlet wavelet transform on the quantum enhanced signal, extract the energy ridge features in the time-frequency domain, and process the energy ridge features in the time-frequency domain to generate a characteristic matrix containing weak fault features.

[0168] Preferably, the time series data segment generation unit includes a window configuration subunit, an initial time series set subunit, and a time series filling and splicing subunit. The window configuration subunit is used to dynamically calibrate the real-time time series window length and overlap rate based on the window parameter set to generate window configuration parameters; the initial time series set subunit is used to segment and cut the multimodal data using the sliding window algorithm and the window configuration parameters to generate an initial time series data segment set; the time series filling and splicing subunit is used to perform boundary data compensation processing on the initial time series data segment set, fill the remaining data at the end with zero values, or splice it with the previous data segment to generate a complete time series data segment.

[0169] Preferably, the multimodal fusion module 640 is configured with the following units:

[0170] A time synchronization and alignment unit is used to perform time synchronization alignment on current data and temperature data based on an event trigger mechanism to generate an alignment timestamp sequence;

[0171] A multimodal fusion matrix generation unit is used to perform feature fusion on the feature matrix through a bidirectional long short-term memory network and an aligned timestamp sequence to generate a multimodal fusion matrix;

[0172] The fault probability calculation unit is used to assign weights to the multimodal fusion matrix based on the multi-head attention mechanism and calculate the fault probability.

[0173] Preferably, the diagnostic marker generation module 650 is configured with the following units:

[0174] A candidate fault period set generation unit is used to perform threshold screening on the time series data segments based on the fault probability to generate a candidate fault period set;

[0175] A continuous fault time interval generating unit, configured to merge the candidate fault time period set based on time continuity to generate a continuous fault time interval;

[0176] The diagnostic report generation unit is used to convert continuous fault time intervals into fault type labels based on a preset fault type mapping table, and generate a diagnostic report including the fault type, timestamp and confidence rating.

[0177] In one embodiment, the present application further provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor implements the above-mentioned charging pile fault diagnosis method based on state timing division when executing the computer program.

[0178] In one embodiment, the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the above-mentioned charging pile fault diagnosis method based on state time series division is implemented.

[0179] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials, or characteristics described may be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and integrate different embodiments or examples described in this specification, as well as features of different embodiments or examples, unless they are mutually inconsistent.

[0180] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0181] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any person skilled in the art can easily conceive of various modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should be included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A charging pile fault diagnosis method based on state time series division, characterized in that: The following steps are involved: S1: Perform multimodal acquisition of the charging pile's operating data through a multi-source synchronous acquisition module to generate multimodal data containing current data, voltage data, temperature data, and charging mode signals; S2: dynamically optimizing window parameters of the multimodal data based on a reinforcement learning framework to generate a window parameter set containing a real-time time series window length and overlap ratio; S3: performing time series division on the multimodal data based on the real-time time series window length and overlap rate of the window parameter set to generate time series data segments, and performing feature extraction on the time series data segments based on quantum derivative enhancement technology to generate a feature matrix containing weak fault features; S4: performing asynchronous multimodal feature fusion on the feature matrix through a bidirectional time alignment network and outputting a fault probability; S5: Mark the fault time interval of the time series data segment according to the fault probability, and generate a diagnostic report including the fault type, timestamp and confidence rating, wherein the diagnostic report is used to indicate the area where the charging pile fails.

2. The method according to claim 1, characterized in that Said S1 comprises: S11: The current of the charging pile is collected and processed based on current transformer technology. The scaled current analog signal is output through the secondary winding. The signal conditioning circuit then filters and calibrates the amplitude to obtain current data. S12: Based on the voltage sampling circuit technology, the voltage of the charging pile is divided and isolated, and the divided signal is quantized by a high-precision analog-to-digital converter to obtain voltage data; S13: Based on multi-node thermocouple array technology, the charging gun connector contact temperature, contactor coil temperature, and power module heat sink temperature are synchronously collected and processed to generate temperature data; S14: Based on the CAN bus message parsing technology, the SAE J1939 protocol data sent by the battery management system is filtered for identifiers and data domain extraction to generate a charging mode signal; S15: Arrange the current data, the voltage data, the temperature data, and the charging mode signal to generate multi-modal data.

3. The method according to claim 1, characterized in that The S2 includes: S21: Calculating power based on the current data and voltage data of the multimodal data to generate real-time power data including power change rate; S22: performing data matching on the multimodal data based on a preset historical database to obtain a difference in fault frequency of charging piles during the same period, where the difference in fault frequency is used to indicate a temporal distribution regularity of historical fault events during the same period; S23: Extract features from the charging mode signal of the multimodal data to generate a charging mode state of the charging pile, and construct a state space of a reinforcement learning agent based on the charging mode state, the real-time power data, and the difference in fault frequencies during the same period. The state space is expressed as: Where S represents the state space, m(t) represents the charging mode state, Indicates the power change rate of real-time power data, ΔF hist Indicates the difference in fault frequency during the same period; S24: Based on the state space, a reward function of the reinforcement learning agent is defined. The expression of the reward function is: R=-(λ1*R1+λ2*R2+λ3*R3) Where R represents the reward function, λ1, λ2, and λ3 represent weight coefficients, λ1+λ2+λ3=1, R1, R2, and R3 represent the missed detection rate, false alarm rate, and computational delay of the state space, respectively; S25: Optimizing the policy network parameters of the reinforcement learning agent based on the PPO algorithm and the reward function, and optimizing the multimodal data based on the optimized reinforcement learning agent to generate a window parameter set containing a real-time time series window length and an overlap rate.

4. The method according to claim 1, wherein The S3 includes: S31: performing time series division on the multimodal data based on the real-time time series window length and overlap rate of the window parameter set to generate time series data segments; S32: Perform quantum tunneling simulation on the current data in the time series data segment to generate a quantum enhanced signal. The calculation formula of the quantum enhanced signal is: Among them, I quantum (t) is the quantum enhanced signal, i.e., the quantum enhanced signal generated after the quantum tunneling effect simulation at time t, I(t) is the original current data, i.e., the current value collected at time t without the quantum tunneling effect simulation, γ is the tunneling gain coefficient, and e is a natural constant; S33: Performing complex Morlet wavelet transform on the quantum enhanced signal to extract time-frequency domain energy ridge features, and processing the time-frequency domain energy ridge features to generate a feature matrix containing weak fault features.

5. The method according to claim 4, characterized in that The S31 includes: S311: Dynamically calibrate the real-time time series window length and overlap rate based on the window parameter set to generate window configuration parameters; S312: Segmenting the multimodal data based on a sliding window algorithm and the window configuration parameters to generate an initial time series data segment set; S313: performing boundary data compensation processing on the initial time series data segment set, filling the remaining data at the end with zero values or splicing it with the previous data segment to generate a complete time series data segment.

6. The method according to claim 1, characterized in that The S4 includes: S41: performing time synchronization alignment on the current data and the temperature data based on an event trigger mechanism to generate an alignment timestamp sequence; S42: performing feature fusion on the feature matrix using a bidirectional long short-term memory network and the aligned timestamp sequence to generate a multimodal fusion matrix; S43: Based on the multi-head attention mechanism, weight allocation is performed on the multimodal fusion matrix to calculate the failure probability.

7. The method according to any one of claims 1 to 6, characterized in that The S5 includes: S51: performing threshold screening on the time series data segments based on the failure probability to generate a set of candidate failure time periods; S52: merging the candidate fault time period set based on time continuity to generate a continuous fault time interval; S53: Convert the continuous fault time intervals into fault type labels based on a preset fault type mapping table, and generate a diagnosis report including the fault type, timestamp, and confidence rating.

8. A charging pile fault diagnosis system based on state time series division, characterized in that: The system comprises: The data acquisition and processing module is used to perform multimodal acquisition of the operating data of the charging pile through the multi-source synchronous acquisition module to generate multimodal data containing current data, voltage data, temperature data and charging mode signals; A data window optimization module is used to dynamically optimize the window parameters of the multimodal data based on a reinforcement learning framework to generate a window parameter set containing the real-time time series window length and overlap rate; A fault feature generation module, configured to perform time series segmentation and feature extraction on the multimodal data based on the real-time time series window length and overlap rate of the window parameter set, and generate a feature matrix containing weak fault features; A multimodal fusion module is used to perform asynchronous multimodal feature fusion on the feature matrix through a bidirectional time alignment network and output a fault probability; A diagnostic tag generation module is used to mark the fault time interval of the time series data segment according to the fault probability, and generate a diagnostic report including the fault type, timestamp and confidence rating, wherein the diagnostic report is used to indicate the area where the charging pile fault occurs.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images