A remote interactive and intelligent operation and maintenance system and method for electric two-wheeled vehicle charging piles

Abstract

This invention belongs to the field of electric vehicle charging management technology, specifically a remote interactive and intelligent operation and maintenance system and method for electric two-wheeled vehicle charging piles. The system includes a charging pile terminal, an edge computing device, a cloud server, and a user interaction terminal. The edge computing device runs an LSTM neural network model to perform real-time anomaly detection on the frequency domain feature vector. The edge computing device also includes a feature extraction unit and a data encryption module. The feature extraction unit generates a frequency domain feature vector from the current spectrum data collected by the charging pile terminal using a fast Fourier transform. This invention utilizes the feature extraction unit in the edge computing device to generate a frequency domain feature vector from the current spectrum data collected by the charging pile terminal using a fast Fourier transform, and then runs an LSTM neural network model to perform real-time anomaly detection on the frequency domain feature vector. This enables real-time anomaly monitoring of the current during charging of two-wheeled electric vehicles, thereby improving the safety of two-wheeled electric vehicles during charging.

Description

Technical Field

[0001] This invention belongs to the field of electric vehicle charging management technology, specifically a remote interactive and intelligent operation and maintenance system and method for electric two-wheeled vehicle charging piles. Background Technology

[0002] Currently, most intelligent charging pile systems for electric two-wheelers adopt an independently operating localized control architecture. Their current monitoring modules only have basic waveform acquisition capabilities and are limited by the following technical shortcomings: First, the charging pile itself lacks a spectrum feature extraction unit and a wireless communication module, resulting in the inability to upload time-frequency domain characteristic parameters such as unsteady AC ripple and harmonic distortion rate generated during charging to the cloud analysis platform in real time. This data link break means that the battery equivalent series resistance (ESR) prediction model based on deep learning can only rely on historical data for offline training, failing to achieve online correlation analysis between charging conditions and polarization impedance. Ultimately, this leads to the technical bottleneck of maintaining a battery state of health (SOH) assessment error rate of 12%-15% for a long time.

[0003] Secondly, the discrete threshold-triggered alarm mechanism used in existing charging piles has significant limitations. Its preset fixed current threshold cannot effectively capture waveform envelope distortion caused by contact impedance degradation during dynamic charging, nor can it identify intermittent pulse spikes caused by individual battery cell imbalances. Experimental data shows that when contact resistance exceeds 30mΩ due to contactor oxidation, the traditional alarm system has a false negative rate of up to 42.7% for abnormal waveforms. More seriously, during the constant current-constant voltage transition phase, existing systems lack real-time analytical capabilities for the differential current characteristic vector, making it impossible to distinguish between normal third-order Bessel filtering and genuine contact failures. This directly results in an average overcharge warning delay of 8.5 seconds, significantly increasing the risk of thermal runaway.

[0004] Therefore, existing electric two-wheeler charging pile systems operate in an independent mode, which cannot upload current spectrum characteristics in real time. Furthermore, traditional threshold alarm mechanisms cannot identify dynamically changing current waveform characteristics, resulting in significant limitations. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a remote interactive and intelligent operation and maintenance system and method for electric two-wheeled vehicle charging piles. This invention primarily addresses the problems of existing electric two-wheeled vehicle charging pile systems operating in an independent mode, which cannot upload current spectrum characteristics in real time, and the traditional threshold alarm mechanism's inability to identify dynamically changing current waveform characteristics, thus exhibiting significant limitations.

[0006] The technical solution adopted by the present invention to solve its technical problem is as follows: The present invention provides a remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles, including: charging pile terminal, edge computing device, cloud server and user interaction terminal;

[0007] The charging pile terminal collects charging current spectrum data during the charging process through a Hall sensor;

[0008] The edge computing device is communicatively connected to the charging pile terminal and is used to run an LSTM neural network model to perform real-time anomaly detection on the frequency domain feature vector and generate anomaly detection results; specifically, it can be connected to the charging pile terminal via the LoRaWAN protocol.

[0009] The cloud server is used to store a fault knowledge base and generate operation and maintenance instructions; the fault knowledge base is a decision tree model containing more than 20,000 fault cases, which can then generate operation and maintenance instructions based on the anomaly detection results.

[0010] The user interaction terminal is used to receive and visually display the operation and maintenance instructions;

[0011] The edge computing device further includes a feature extraction unit and a data encryption module;

[0012] The feature extraction unit generates a frequency domain feature vector from the current spectrum data collected by the charging pile terminal through a fast Fourier transform.

[0013] The data encryption module uses the SM4 algorithm to encrypt the frequency domain feature vector.

[0014] The data processing logic steps of the system are as follows:

[0015] Step S1: The Hall sensor of the charging pile terminal captures the charging current waveform data at a sampling rate of 10kHz;

[0016] Step S2: The feature extraction unit in the edge computing device calls the FFT algorithm to perform a fast Fourier transform on the original waveform and generate a 128-dimensional frequency domain feature vector.

[0017] Step S3: The environmental sensing module synchronously collects temperature, humidity and smoke concentration data to generate an equipment operating environment index;

[0018] Step S4: The edge computing device performs data obfuscation on the feature vector and uses a dynamic key seed to call the SM4 encryption algorithm to generate encrypted data packets;

[0019] Step S5: Upload {{Device ID, timestamp, encrypted data}} to the edge gateway via the LoRaWAN communication protocol;

[0020] Step S6: The edge computing device receives the encrypted data packet and performs SM4 decryption using a preset key seed;

[0021] Step S7: Perform Z-score normalization on the frequency domain feature vectors to generate a normalized input matrix;

[0022] Step S8: The edge computing device inputs the standardized matrix into the pre-trained LSTM neural network model to generate anomaly detection results.

[0023] Preferably, the real-time anomaly detection of the frequency domain feature vector includes:

[0024] Training an LSTM neural network model;

[0025] The input layer receives the standardized frequency domain feature vector; specifically, it performs Z-score standardization on the frequency domain feature vector to generate a standardized input matrix.

[0026] The standardized matrix is ​​input into the pre-trained LSTM neural network model, and the output layer generates the prediction error rate based on the LSTM neural network model.

[0027] Warning trigger condition: An alarm is triggered when the error rate is >15% for three consecutive sampling periods. Specifically, a JSON alarm package containing {{fault code, confidence level, recommended action}} can be generated.

[0028] Preferably, the training method for the LSTM neural network model includes:

[0029] A training dataset containing normal / abnormal current waveforms was constructed, and time series samples were generated using the sliding window method;

[0030] The spectral feature weights are dynamically allocated through an attention mechanism; high-frequency noise features in the 500-800Hz range are suppressed with a weight of less than 0.3; and effective features in the 100-300Hz range are enhanced with a weight of more than 0.7.

[0031] An early stopping mechanism is set up to terminate training when the loss function on the validation set decreases by less than 1% for five consecutive epochs.

[0032] Preferably, the cloud server is also used to train a demand prediction model based on historical charging data and to optimize the power allocation strategy of the regional charging pile group using the NSGA-II algorithm.

[0033] The demand forecasting model:

[0034] Input: {{historical charging amount, weather data, user behavior characteristics}};

[0035] Output: Heat map of demand distribution for the next 2 hours.

[0036] Preferably, the parameter configuration of the NSGA-II algorithm includes:

[0037] Set the population size to three times the number of charging stations;

[0038] An adaptive crossover probability adjustment mechanism is adopted, ranging from 0.7 to 0.9;

[0039] An elite retention strategy is introduced, which retains the top 10% of the best solutions for the next iteration.

[0040] Preferably, the workflow of the data encryption module includes:

[0041] Before SM4 algorithm encryption, the frequency domain feature vector is subjected to data obfuscation processing;

[0042] A dynamic key allocation mechanism is adopted, and the encryption key seed is updated every 24 hours; the dynamic key allocation mechanism can be an elliptic curve cryptography algorithm.

[0043] Add the metadata field {{checksum, protocol version, encryption identifier}} to the header of the data packet.

[0044] Preferably, the user interaction terminal further includes a charging reservation module, an emergency command channel, and a billing strategy generation unit;

[0045] The charging reservation module generates optimal charging time slot recommendations based on historical charging data analysis.

[0046] The emergency command channel triggers an audible and visual alarm when a risk of thermal runaway is detected.

[0047] The billing strategy generation unit dynamically adjusts the tiered electricity price coefficient based on the grid load status.

[0048] The billing strategy generation unit calculates the real-time electricity price using the following formula:

[0049] P_t=P_base×(1+α·(L_t / L_max))

[0050] Where P_base is the base electricity price, which is the initial electricity price when the grid load is at the baseline state, α is the adjustment factor (0.2-0.5), L_t is the current regional grid load value, and L_max is the grid carrying capacity threshold.

[0051] P_base can be trained on historical data through the demand forecasting model, thereby optimizing the initial value setting of P_base; and by displaying the comparison between P_base and real-time electricity price to users through the user interaction terminal, charging behavior can be guided, and users can reduce costs by choosing to charge during low-load periods, while effectively alleviating the pressure on the power grid.

[0052] Preferably, the user interaction terminal optimizes the interface interaction in the following ways:

[0053] The charging status visualization uses HSV color space mapping, so that the overload warning color gamut covers the wavelength of 540-580nm;

[0054] The emergency interface should be set to high contrast mode, with a font size of at least 18pt.

[0055] The voice control interface supports semantic parsing and execution of voice commands. For example, it supports semantic parsing of commands such as "pause charging" and "extend duration".

[0056] Preferably, the charging pile terminal further includes a mechanical locking mechanism, an anti-theft detection unit, and an anti-theft linkage module;

[0057] The mechanical locking mechanism is controlled by an electromagnetic drive device to extend and retract the locking tongue.

[0058] The anti-theft detection unit collects the three-dimensional displacement signal of the device through a vibration sensor and generates a smooth displacement trajectory through Kalman filtering;

[0059] When unauthorized movement is detected, an audible and visual alarm is triggered, and the device ID and displacement trajectory are uploaded to the cloud server.

[0060] The anti-theft linkage module compares displacement trajectories and pushes an early warning to maintenance terminals within a 5-kilometer radius when the similarity between the displacement trajectory and the theft case database is greater than 85%.

[0061] The displacement trajectory comparison uses the Hausdorff distance algorithm to compare the similarity between the currently obtained displacement trajectory and the historical theft case database.

[0062] A remote interaction and intelligent operation and maintenance method for electric two-wheeled vehicle charging piles includes the following steps:

[0063] S1: The Hall sensor at the charging pile terminal captures charging current waveform data at a sampling rate of 10kHz.

[0064] S2: The feature extraction unit in the edge computing device calls the FFT algorithm to perform a fast Fourier transform on the original waveform and generate a 128-dimensional frequency domain feature vector;

[0065] S3: The environmental sensing module synchronously collects temperature, humidity and smoke concentration data to generate an equipment operating environment index;

[0066] S4: The edge computing device performs data obfuscation on the feature vector and uses a dynamic key seed to call the SM4 encryption algorithm to generate encrypted data packets;

[0067] S5: Uploads {{Device ID, timestamp, encrypted data}} to the edge gateway via the LoRaWAN communication protocol;

[0068] S6: The edge computing device receives encrypted data packets and performs SM4 decryption using a preset key seed;

[0069] S7: Perform Z-score normalization on the frequency domain feature vectors to generate a normalized input matrix;

[0070] S8: Edge computing devices input a standardized input matrix into a pre-trained LSTM neural network model; the LSTM neural network model dynamically allocates frequency domain feature weights through an attention mechanism: applying a weight of less than 0.3 to suppress high-frequency noise features in the 500-800Hz range; and applying a weight of more than 0.7 to strengthen effective features in the 100-300Hz range.

[0071] S9: The edge computing device calculates the prediction error rate between the output value and the actual value of the LSTM neural network model. When the error rate exceeds 15% for three consecutive cycles, a level three alarm signal is triggered.

[0072] S10: The cloud server first trains a demand prediction model based on historical charging data to predict the regional charging demand for the next 2 hours.

[0073] S11: The cloud server then uses the NSGA-II algorithm to optimize the power allocation strategy of the regional charging pile group;

[0074] S12: The cloud server generates a set of scheduling instructions containing {{device ID, power limit, time window}} and sends it to the edge gateway.

[0075] The beneficial effects of this invention are as follows:

[0076] 1. This invention utilizes a feature extraction unit in an edge computing device to generate a frequency domain feature vector from the current spectrum data collected by the charging pile terminal using Fast Fourier Transform. Then, an LSTM neural network model is run to perform real-time anomaly detection on the frequency domain feature vector, thereby achieving real-time anomaly monitoring of the current during charging of two-wheeled electric vehicles and improving their charging safety. Furthermore, by storing a decision tree model with over 20,000 historical fault cases on a cloud server, it can match the frequency domain feature vector and quickly output a repair plan, thus reducing the average maintenance response time from 4.2 hours to 18 minutes.

[0077] 2. The LSTM neural network model in this invention dynamically allocates spectral feature weights through an attention mechanism, applying a weight of less than 0.3 to suppress high-frequency noise features in the 500-800Hz range and applying a weight of more than 0.7 to strengthen effective features in the 100-300Hz range. This results in a fault prediction accuracy of up to 92.3% for the LSTM neural network model, which greatly improves the real-time anomaly monitoring capability of two-wheeled electric vehicles during charging and further ensures the safety of two-wheeled electric vehicles during charging.

[0078] 3. This invention trains a demand prediction model based on historical charging data using a cloud server, and optimizes the power allocation strategy of regional charging pile groups using the NSGA-II algorithm. Specifically, by introducing an elite retention strategy, the top 10% of the optimal solutions are retained for the next generation of iteration, thereby enabling the charging energy in the region to be fully utilized, and thus improving the regional energy utilization rate by more than 40%.

[0079] 4. This invention uses the Hausdorff distance algorithm in the anti-theft linkage module to compare the similarity between the currently obtained displacement trajectory and the historical theft case database. When the similarity between the displacement trajectory and the theft case database is greater than 85%, an early warning is pushed to maintenance terminals within 5 kilometers, further shortening the maintenance response time. Attached Figure Description

[0080] The invention will now be further described with reference to the accompanying drawings.

[0081] Figure 1 This is an overall framework diagram of the system of the present invention;

[0082] Figure 2 This is the data flow topology diagram of the system of the present invention;

[0083] Figure 3 This is a data processing flowchart of the system of the present invention;

[0084] Figure 4 This is a flowchart of the LSTM model training process in this invention;

[0085] Figure 5 This is a flowchart illustrating the data encryption process from the perspective of various terminals in this invention;

[0086] Figure 6 This is a flowchart of the data encryption process from the perspective of the data end in this invention;

[0087] Figure 7 This is a flowchart of the power optimization algorithm in this invention;

[0088] Figure 8 This is a diagram of the user interface in this invention;

[0089] Figure 9This is a timing diagram of the anti-theft detection in this invention. Detailed Implementation

[0090] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0091] like Figures 1 to 3 As shown, a remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles includes: a charging pile terminal, an edge computing device, a cloud server, and a user interaction terminal.

[0092] The charging pile terminal collects charging current spectrum data during the charging process through a Hall sensor;

[0093] The edge computing device is communicatively connected to the charging pile terminal and is used to run an LSTM neural network model to perform real-time anomaly detection on the frequency domain feature vector and generate anomaly detection results; specifically, it can be connected to the charging pile terminal via the LoRaWAN protocol.

[0094] The cloud server is used to store a fault knowledge base and generate operation and maintenance instructions; the fault knowledge base is a decision tree model containing more than 20,000 fault cases, which can then generate operation and maintenance instructions based on the anomaly detection results.

[0095] The user interaction terminal is used to receive and visually display the operation and maintenance instructions;

[0096] The edge computing device further includes a feature extraction unit and a data encryption module;

[0097] The feature extraction unit generates a frequency domain feature vector from the current spectrum data collected by the charging pile terminal through a fast Fourier transform.

[0098] The data encryption module uses the SM4 algorithm to encrypt the frequency domain feature vector.

[0099] The data processing logic steps of the system are as follows:

[0100] Step S1: The Hall sensor of the charging pile terminal captures the charging current waveform data at a sampling rate of 10kHz;

[0101] Step S2: The feature extraction unit in the edge computing device calls the FFT algorithm to perform a fast Fourier transform on the original waveform and generate a 128-dimensional frequency domain feature vector.

[0102] Step S3: The environmental sensing module synchronously collects temperature, humidity and smoke concentration data to generate an equipment operating environment index;

[0103] Step S4: The edge computing device performs data obfuscation on the feature vector and uses a dynamic key seed to call the SM4 encryption algorithm to generate encrypted data packets;

[0104] Step S5: Upload {{Device ID, timestamp, encrypted data}} to the edge gateway via the LoRaWAN communication protocol;

[0105] Step S6: The edge computing device receives the encrypted data packet and performs SM4 decryption using a preset key seed;

[0106] Step S7: Perform Z-score normalization on the frequency domain feature vectors to generate a normalized input matrix;

[0107] Step S8: The edge computing device inputs the standardized matrix into the pre-trained LSTM neural network model to generate anomaly detection results.

[0108] like Figures 1 to 3 As shown, the real-time anomaly detection of the frequency domain feature vector includes:

[0109] Training an LSTM neural network model;

[0110] The input layer receives the standardized frequency domain feature vector; specifically, it performs Z-score standardization on the frequency domain feature vector to generate a standardized input matrix.

[0111] The standardized matrix is ​​input into the pre-trained LSTM neural network model, and the output layer generates the prediction error rate based on the LSTM neural network model.

[0112] Warning trigger condition: An alarm is triggered when the error rate is >15% for three consecutive sampling periods. Specifically, a JSON alarm package containing {{fault code, confidence level, recommended action}} can be generated.

[0113] like Figure 4 As shown, the training method for the LSTM neural network model includes:

[0114] A training dataset containing normal / abnormal current waveforms was constructed, and time series samples were generated using the sliding window method;

[0115] The spectral feature weights are dynamically allocated through an attention mechanism; high-frequency noise features in the 500-800Hz range are suppressed with a weight of less than 0.3; and effective features in the 100-300Hz range are enhanced with a weight of more than 0.7.

[0116] An early stopping mechanism is set up to terminate training when the loss function on the validation set decreases by less than 1% for five consecutive epochs.

[0117] like Figure 7As shown, the cloud server is also used to train a demand prediction model based on historical charging data and to optimize the power allocation strategy of the regional charging pile group using the NSGA-II algorithm.

[0118] The demand forecasting model:

[0119] Input: {{historical charging amount, weather data, user behavior characteristics}};

[0120] Output: Heat map of demand distribution for the next 2 hours.

[0121] like Figure 7 As shown, the parameter configuration of the NSGA-II algorithm includes:

[0122] Set the population size to three times the number of charging stations;

[0123] An adaptive crossover probability adjustment mechanism is adopted, ranging from 0.7 to 0.9;

[0124] An elite retention strategy is introduced, which retains the top 10% of the best solutions for the next iteration.

[0125] like Figures 5 to 6 As shown, the workflow of the data encryption module includes:

[0126] Before SM4 algorithm encryption, the frequency domain feature vector is subjected to data obfuscation processing;

[0127] A dynamic key allocation mechanism is adopted, and the encryption key seed is updated every 24 hours; the dynamic key allocation mechanism can be an elliptic curve cryptography algorithm.

[0128] Add the metadata field {{checksum, protocol version, encryption identifier}} to the header of the data packet.

[0129] like Figure 1 As shown, the user interaction terminal also includes a charging reservation module, an emergency command channel, and a billing strategy generation unit;

[0130] The charging reservation module generates optimal charging time slot recommendations based on historical charging data analysis.

[0131] The emergency command channel triggers an audible and visual alarm when a risk of thermal runaway is detected.

[0132] The billing strategy generation unit dynamically adjusts the tiered electricity price coefficient based on the grid load status.

[0133] The billing strategy generation unit calculates the real-time electricity price using the following formula:

[0134] P_t=P_base×(1+α·(L_t / L_max))

[0135] Where P_base is the base electricity price, which is the initial electricity price when the grid load is at the baseline state, α is the adjustment factor (0.2-0.5), L_t is the current regional grid load value, and L_max is the grid carrying capacity threshold.

[0136] P_base can be trained on historical data through the demand forecasting model, thereby optimizing the initial value setting of P_base; and by displaying the comparison between P_base and real-time electricity price to users through the user interaction terminal, charging behavior can be guided, and users can reduce costs by choosing to charge during low-load periods, while effectively alleviating the pressure on the power grid.

[0137] like Figure 1 As shown, the user interaction terminal optimizes the interface interaction in the following ways:

[0138] The charging status visualization uses HSV color space mapping, so that the overload warning color gamut covers the wavelength of 540-580nm;

[0139] The emergency interface should be set to high contrast mode, with a font size of at least 18pt.

[0140] The voice control interface supports semantic parsing and execution of voice commands. For example, it supports semantic parsing of commands such as "pause charging" and "extend duration".

[0141] like Figure 9 As shown, the charging pile terminal also includes a mechanical locking mechanism, an anti-theft detection unit, and an anti-theft linkage module;

[0142] The mechanical locking mechanism is controlled by an electromagnetic drive device to extend and retract the locking tongue.

[0143] The anti-theft detection unit collects the three-dimensional displacement signal of the device through a vibration sensor and generates a smooth displacement trajectory through Kalman filtering;

[0144] When unauthorized movement is detected, an audible and visual alarm is triggered, and the device ID and displacement trajectory are uploaded to the cloud server.

[0145] The anti-theft linkage module compares displacement trajectories and pushes an early warning to maintenance terminals within a 5-kilometer radius when the similarity between the displacement trajectory and the theft case database is greater than 85%.

[0146] The displacement trajectory comparison uses the Hausdorff distance algorithm to compare the similarity between the currently obtained displacement trajectory and the historical theft case database.

[0147] like Figures 1 to 9 As shown, a remote interaction and intelligent operation and maintenance method for electric two-wheeled vehicle charging piles includes the following steps:

[0148] S1: The Hall sensor at the charging pile terminal captures charging current waveform data at a sampling rate of 10kHz.

[0149] S2: The feature extraction unit in the edge computing device calls the FFT algorithm to perform a fast Fourier transform on the original waveform and generate a 128-dimensional frequency domain feature vector;

[0150] S3: The environmental sensing module synchronously collects temperature, humidity and smoke concentration data to generate an equipment operating environment index;

[0151] S4: The edge computing device performs data obfuscation on the feature vector and uses a dynamic key seed to call the SM4 encryption algorithm to generate encrypted data packets;

[0152] S5: Uploads {{Device ID, timestamp, encrypted data}} to the edge gateway via the LoRaWAN communication protocol;

[0153] S6: The edge computing device receives encrypted data packets and performs SM4 decryption using a preset key seed;

[0154] S7: Perform Z-score normalization on the frequency domain feature vectors to generate a normalized input matrix;

[0155] S8: Edge computing devices input a standardized input matrix into a pre-trained LSTM neural network model; the LSTM neural network model dynamically allocates frequency domain feature weights through an attention mechanism: applying a weight of less than 0.3 to suppress high-frequency noise features in the 500-800Hz range; and applying a weight of more than 0.7 to strengthen effective features in the 100-300Hz range.

[0156] S9: The edge computing device calculates the prediction error rate between the output value and the actual value of the LSTM neural network model. When the error rate exceeds 15% for three consecutive cycles, a level three alarm signal is triggered.

[0157] S10: The cloud server first trains a demand prediction model based on historical charging data to predict the regional charging demand for the next 2 hours.

[0158] S11: The cloud server then uses the NSGA-II algorithm to optimize the power allocation strategy of the regional charging pile group;

[0159] S12: The cloud server generates a set of scheduling instructions containing {{device ID, power limit, time window}} and sends it to the edge gateway.

[0160] Example 1: Current Spectrum Anomaly Detection

[0161] Specific implementation steps:

[0162] Data acquisition: An HX10-NP Hall sensor is configured at the charging pile terminal to collect charging current waveform data at a sampling rate of 10kHz, and simultaneously acquire environmental parameters such as temperature (±0.5℃ accuracy) and smoke concentration (0-20% LEL range);

[0163] Feature extraction: The edge device uses an STM32F407 chip to perform 1024-point FFT operations to generate a 128-dimensional feature vector containing the fundamental component (50Hz) and the 3rd / 5th / 7th harmonics (150 / 250 / 350Hz).

[0164] Encrypted transmission: The feature vector is dynamically obfuscated (interpolated with random noise ±5%), encrypted using SM4-CTR mode, and the key seed is updated every 24 hours via the ECDH protocol;

[0165] Anomaly detection: The LSTM model has a sliding window length of 30 seconds. When the prediction error is greater than 15% for 3 consecutive cycles, a JSON alarm packet containing a fault location code (such as E102 representing an overcurrent fault) is generated.

[0166] The following technical effects were obtained from Example 1: compared with the traditional threshold detection method, the false alarm rate was reduced by 62% (measured data), and the high-frequency noise suppression effect was improved by 40%.

[0167] Example 2: Dynamic Key

[0168] Specific implementation plan:

[0169] Key management: A two-level key architecture is adopted. The master key is stored in the HSM encryption machine through the PKI system, and the session key is dynamically generated through the ECIES algorithm.

[0170] Data packet structure:

[0171]

[0172] Anti-theft linkage: When the triaxial accelerometer detects a displacement velocity > 0.5 m / s 2 At that time, the trajectory comparison algorithm is activated, and the similarity with the theft case database is calculated using Hausdorff distance.

[0173] The following test data was obtained from Example 2: In the simulated attack test, 99.3% of replay attacks and man-in-the-middle attacks were successfully blocked.

[0174] Example 3: Power Grid Cooperative Optimization

[0175] Detailed implementation process:

[0176] Demand forecasting: Input historical charging volume (time series decomposition), weather data (temperature / humidity / wind speed), and user behavior (charging duration preference) to train the TCN time series model and output a 2-hour granular demand heatmap;

[0177] NSGA-II optimization algorithm:

[0178] Objective function: min(grid load fluctuation rate) & max(charging pile utilization rate)

[0179] Constraints: Single pile power ≤ 3kW, total load ≤ 80% of the area transformer capacity.

[0180] Dynamic billing: Tiered electricity pricing coefficients are generated based on the optimization results, and the value of α in the formula is dynamically adjusted (0.5 during peak hours, 0.3 during off-peak hours, and 0.1 during valley hours).

[0181] The following application effects were obtained from Example 3: In a pilot area (50 charging piles), the peak load of the power grid was reduced by 27%, and the average charging cost for users decreased by 18%.

[0182] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A remote interactive and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles, characterized in that, include: Charging pile terminals, edge computing devices, cloud servers, and user interaction terminals; The charging pile terminal collects charging current spectrum data during the charging process through a Hall sensor; The edge computing device is communicatively connected to the charging pile terminal and is used to run an LSTM neural network model to perform real-time anomaly detection on the frequency domain feature vector. The cloud server is used to store a fault knowledge base and generate operation and maintenance instructions. The user interaction terminal is used to receive and visually display the operation and maintenance instructions; The edge computing device further includes a feature extraction unit and a data encryption module; The feature extraction unit generates a frequency domain feature vector from the current spectrum data collected by the charging pile terminal through a fast Fourier transform. The data encryption module uses the SM4 algorithm to encrypt the frequency domain feature vector; The remote interaction and intelligent operation and maintenance method for electric two-wheeler charging piles includes the following steps: S1: The Hall sensor at the charging pile terminal captures charging current waveform data at a sampling rate of 10kHz. S2: The feature extraction unit in the edge computing device calls the FFT algorithm to perform a fast Fourier transform on the original waveform and generate a 128-dimensional frequency domain feature vector; S3: The environmental sensing module synchronously collects temperature, humidity and smoke concentration data to generate an equipment operating environment index; S4: The edge computing device performs data obfuscation on the feature vector and uses a dynamic key seed to call the SM4 encryption algorithm to generate encrypted data packets; S5: Uploads {{Device ID, timestamp, encrypted data}} to the edge gateway via the LoRaWAN communication protocol; S6: The edge computing device receives encrypted data packets and performs SM4 decryption using a preset key seed; S7: Perform Z-score normalization on the frequency domain feature vectors to generate a normalized input matrix; S8: Edge computing devices input a standardized input matrix into a pre-trained LSTM neural network model; the LSTM neural network model dynamically allocates frequency domain feature weights through an attention mechanism: applying a weight of less than 0.3 to suppress high-frequency noise features in the 500-800Hz range; and applying a weight of more than 0.7 to strengthen effective features in the 100-300Hz range. S9: The edge computing device calculates the prediction error rate between the output value and the actual value of the LSTM neural network model. When the error rate exceeds 15% for three consecutive cycles, a level three alarm signal is triggered. S10: The cloud server first trains a demand prediction model based on historical charging data to predict the regional charging demand for the next 2 hours. S11: The cloud server then uses the NSGA-II algorithm to optimize the power allocation strategy of the regional charging pile group; S12: The cloud server generates a set of scheduling instructions containing {{device ID, power limit, time window}} and sends it to the edge gateway.

2. The remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles according to claim 1, characterized in that: The real-time anomaly detection of the frequency domain feature vector includes: Training an LSTM neural network model; The input layer receives the standardized frequency domain feature vector; Based on the LSTM neural network model, the output layer generates the prediction error rate; Warning trigger condition: An alarm is triggered when the error rate is >15% for 3 consecutive sampling periods.

3. The remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles according to claim 2, characterized in that: The training method for the LSTM neural network model includes: A training dataset containing normal / abnormal current waveforms was constructed, and time series samples were generated using the sliding window method; The spectral feature weights are dynamically allocated through an attention mechanism; high-frequency noise features in the 500-800Hz range are suppressed with a weight of less than 0.3; and effective features in the 100-300Hz range are enhanced with a weight of more than 0.

7. An early stop mechanism is set up to terminate training when the loss function on the validation set decreases by less than 1% for five consecutive epochs.

4. The remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles according to claim 1, characterized in that: The cloud server is also used to train a demand prediction model based on historical charging data and to optimize the power allocation strategy of regional charging pile groups using the NSGA-II algorithm.

5. The remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles according to claim 4, characterized in that: The parameter configuration of the NSGA-II algorithm includes: Set the population size to three times the number of charging stations; An adaptive crossover probability adjustment mechanism is adopted, ranging from 0.7 to 0.9; An elite retention strategy is introduced, retaining the top 10% of the best solutions for the next iteration.

6. The remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles according to claim 1, characterized in that: The workflow of the data encryption module includes: Before SM4 algorithm encryption, the frequency domain feature vector is subjected to data obfuscation processing; A dynamic key distribution mechanism is adopted, and the encryption key seed is updated every 24 hours; Add the metadata field {{checksum, protocol version, encryption identifier}} to the header of the data packet.

7. The remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles according to claim 1, characterized in that: The user interaction terminal also includes a charging reservation module, an emergency command channel, and a billing strategy generation unit; The charging reservation module generates optimal charging time slot recommendations based on historical charging data analysis. The emergency command channel triggers an audible and visual alarm when a risk of thermal runaway is detected. The billing strategy generation unit dynamically adjusts the tiered electricity price coefficient based on the grid load status.

8. The remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles according to claim 1, characterized in that: The user interaction terminal optimizes the interface interaction in the following ways: The charging status visualization uses HSV color space mapping, so that the overload warning color gamut covers the wavelength of 540-580nm; The emergency interface uses high contrast mode. The voice control interface supports semantic parsing and execution of voice commands.

9. The remote interaction and intelligent operation and maintenance system for electric two-wheeled vehicle charging piles according to claim 1, characterized in that: The charging pile terminal also includes a mechanical locking mechanism, an anti-theft detection unit, and an anti-theft linkage module; The mechanical locking mechanism is controlled by an electromagnetic drive device to extend and retract the locking tongue. The anti-theft detection unit collects the three-dimensional displacement signal of the device through a vibration sensor and generates a smooth displacement trajectory through Kalman filtering; When unauthorized movement is detected, an audible and visual alarm is triggered, and the device ID and displacement trajectory are uploaded to the cloud server. The anti-theft linkage module compares displacement trajectories and pushes an early warning to maintenance terminals within 5 kilometers when the similarity between the displacement trajectory and the theft case database is greater than 85%. The displacement trajectory comparison uses the Hausdorff distance algorithm to compare the similarity between the currently obtained displacement trajectory and the historical theft case database.

Images