Charging pile remote online monitoring method and system based on internet of things

Abstract

The application relates to the technical field of charging pile monitoring, and discloses a charging pile remote online monitoring method and system based on the Internet of Things, the method comprising the following steps: collecting charging pile electrical signal data in real time and preprocessing to obtain a first signal data set; performing time-frequency analysis on a time domain signal to convert the time domain signal into a time-frequency spectrum diagram, and integrating to obtain a first spectrum diagram set; combining a preset health state benchmark spectrum diagram library to calculate a preliminary residual error, adjusting the preliminary residual error through a preset correction model to obtain a second residual error distribution; comparing a preset threshold interval to obtain a preliminary early warning sign, and performing secondary verification in combination with abnormal mode characteristics in historical operation data; and generating a remote monitoring report according to an aging state judgment result and pushing the remote monitoring report to a management platform. The method can realize accurate aging early warning under dynamic working conditions and environmental interference, and improves the reliability and real-time performance of charging pile remote monitoring.

Description

Technical Field

[0001] This invention relates to the field of charging pile monitoring technology, and in particular to a method and system for remote online monitoring of charging piles based on the Internet of Things. Background Technology

[0002] With the rapid development of the new energy vehicle industry, charging piles, as a core supporting infrastructure, have become a core research direction in the field of charging pile operation and maintenance management, ensuring their long-term stable operation and promoting fault prediction and health management.

[0003] Currently, existing remote monitoring technologies for charging piles mainly rely on fixed parameter threshold judgment or periodic manual inspection. Fixed threshold mode uses preset safety ranges for electrical parameters such as current and voltage, triggering an alarm when parameters exceed the threshold; manual inspection involves maintenance personnel periodically checking the equipment's operating status on-site. However, charging piles face dynamic factors during operation, such as frequent switching of charging power, fluctuations in ambient temperature, and unstable grid voltage. Fixed thresholds are difficult to adapt to complex operating conditions, easily leading to false alarms or missed alarms, and cannot meet the needs of accurate monitoring for fault prediction and health management. Manual inspection is limited by the wide distribution and large number of charging piles, making high-frequency, full-coverage monitoring difficult. Often, aging problems are only discovered after equipment failure, resulting in insufficient foresight in fault prediction and health management.

[0004] In summary, existing technologies have the problem of being unable to adapt to the complex operating conditions of charging piles, thus failing to achieve accurate fault prediction for charging piles. Summary of the Invention

[0005] This invention provides a remote online monitoring method and system for charging piles based on the Internet of Things, in order to solve the problem that existing technologies are unable to adapt to the complex operating conditions of charging piles and cannot achieve accurate fault prediction of charging piles.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a remote online monitoring method for charging piles based on the Internet of Things, comprising: Electrical signal data during the operation of the charging pile is collected in real time, and the electrical signal data is preprocessed to obtain a first signal dataset; Time-frequency analysis is performed on the time-domain signals of the first signal dataset, the time-domain signals are converted into time-spectrum graphs, the time-spectrum graphs are summarized, and the first spectrum set is obtained. Obtain a preset health benchmark matrix from a preset health status benchmark spectrum library, calculate the preliminary residual between the first spectrum set and the preset health benchmark matrix, and obtain the first residual distribution; The first residual distribution is adjusted by a preset correction model to obtain the second residual distribution; The second residual distribution is compared with a preset threshold range. If the corrected residual value in the second residual distribution exceeds the preset threshold range, a preliminary warning sign is obtained. By combining the abnormal pattern characteristics in historical operation data, the preliminary warning signs are verified a second time to determine the aging status judgment result; Based on the aging status assessment results, a remote monitoring report is generated and pushed to the relevant management platform to complete real-time early warning and feedback on the aging status of the charging pile.

[0007] Secondly, the present invention provides a remote online monitoring system for charging piles based on the Internet of Things, comprising: The signal acquisition and preprocessing module is used to acquire electrical signal data during the operation of the charging pile in real time, and preprocess the electrical signal data to obtain a first signal dataset. The time-frequency analysis module is used to perform time-frequency analysis on the time-domain signal of the first signal dataset, convert the time-domain signal into a time-frequency spectrum, summarize the time-frequency spectrum, and integrate them to obtain a first spectrum set. The residual calculation module is used to obtain a preset health benchmark matrix from a preset health status benchmark spectrum library, calculate the preliminary residual between the first spectrum set and the preset health benchmark matrix, and obtain the first residual distribution. The residual correction module is used to adjust the first residual distribution through a preset correction model to obtain the second residual distribution; The preliminary warning module is used to compare the second residual distribution with a preset threshold range. If the corrected residual value in the second residual distribution exceeds the preset threshold range, a preliminary warning sign is obtained. The secondary verification module is used to perform secondary verification on the preliminary warning sign by combining the abnormal pattern characteristics in historical operation data to determine the aging status judgment result. The early warning and feedback module is used to generate a remote monitoring report based on the aging status judgment results and push it to the relevant management platform to complete the real-time early warning and feedback of the aging status of the charging pile.

[0008] Thirdly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the IoT-based remote online monitoring method for charging piles described above.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention collects and preprocesses electrical signal data during the operation of charging piles in real time, removes external noise interference, and forms a standardized first signal dataset, which provides reliable data support for subsequent monitoring and analysis, and solves the problem that noise components in the original signal affect the state judgment.

[0010] (2) This invention converts the time-domain signal into a time-frequency spectrum through time-frequency analysis, calculates the residual by combining it with a preset health status benchmark spectrum library, and introduces a preset correction model to adjust the residual value, eliminating non-fault deviations caused by dynamic working conditions and environmental interference, thus solving the problem that fixed standards are difficult to accurately reflect the true state of equipment under complex working conditions.

[0011] (3) After obtaining the initial warning sign through threshold comparison, the present invention performs secondary verification by combining the abnormal pattern characteristics in historical operation data and confirms the aging status through frequency component feature matching, forming a complete closed loop from data collection, processing, analysis to warning feedback. This solves the problem of false alarms and missed alarms that are easy to occur with a single judgment method, and provides an effective basis for the operation and maintenance management of charging piles. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a remote online monitoring method for charging piles based on the Internet of Things provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a remote online monitoring system for charging piles based on the Internet of Things, provided in an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides a method for remote online monitoring of charging piles based on the Internet of Things, including the following steps: S101, Real-time acquisition of electrical signal data during the operation of the charging pile, preprocessing of the electrical signal data to obtain a first signal dataset; S102, perform time-frequency analysis on the time-domain signal of the first signal dataset, convert the time-domain signal into a time-frequency spectrum, summarize the time-frequency spectrum, and integrate to obtain the first spectrum set; S103, obtain a preset health benchmark matrix from a preset health status benchmark spectrum library, calculate the preliminary residual between the first spectrum set and the preset health benchmark matrix, and obtain the first residual distribution; S104, the first residual distribution is adjusted by a preset correction model to obtain the second residual distribution; S105, compare the second residual distribution with the preset threshold range. If the corrected residual value in the second residual distribution exceeds the preset threshold range, a preliminary warning sign is obtained. S106, Combining the abnormal pattern characteristics in historical operating data, the preliminary warning sign is verified a second time to determine the aging status judgment result; S107. Based on the aging status judgment result, generate a remote monitoring report and push it to the relevant management platform to complete the real-time early warning and feedback of the aging status of the charging pile.

[0015] In step S101, electrical signal data during the operation of the charging pile is collected in real time, and the electrical signal data is preprocessed to obtain a first signal dataset, including: Discretize the electrical signal data of the charging pile during operation to obtain the original sequence containing timing information; The original sequence is filtered and denoised to obtain a clean signal amplitude; The standardized parameters are obtained by calculating the mean shift and variance distribution data of the pure signal amplitude. The amplitude of the pure signal is scaled using the standardized parameters to obtain the first signal dataset.

[0016] It should be noted that electrical signal data is collected through high-precision sensors installed inside the charging pile. The collected data includes voltage and current signals during operation. The sensor sampling frequency is set to 1000 times per second. This sampling frequency was determined based on experimental statistics. Signal tests on charging piles of different power levels show that 1000 samples per second can completely capture the instantaneous fluctuations of voltage and current. When the sampling frequency is lower than 800 times per second, the signal detail loss rate reaches 15.6%. When it is higher than 1000 times per second, the data redundancy increases by more than 30% without improving monitoring accuracy. Therefore, the sampling frequency is set to 1000 times per second. Timestamps are recorded synchronously during the collection process, with timestamp accuracy down to the millisecond level, ensuring that each data point contains clear timing information.

[0017] It's worth noting that discretization is the process of converting continuous analog electrical signals into digital signals, achieved through the sensor's built-in analog-to-digital converter module. A fixed sampling interval of 1 millisecond is set, and voltage or current values ​​are collected every millisecond, forming a raw sequence containing timing information. For example, within 10 seconds of a charging pile's operation, 10,000 data points will be collected, each corresponding to a specific time point and a corresponding voltage or current value. This discretization process facilitates subsequent digital signal analysis while fully preserving the signal's temporal characteristics, providing a foundation for identifying abnormal patterns. The raw sequence must be bound to the charging pile's unique device identifier during storage to ensure data traceability to the specific device.

[0018] It should be noted that when filtering and denoising the original sequence, a time-frequency domain eigenvalue decomposition method is used to identify external noise. The original sequence is decomposed into signal components of different frequencies using a short-time Fourier transform. High-frequency components unrelated to the operation of the charging pile are selected as external noise. This high-frequency noise mostly originates from electromagnetic interference from surrounding equipment, and its frequency range is typically higher than 5000 Hz, determined based on extensive statistical data. The signal frequencies generated during normal operation of the charging pile are mainly concentrated in the 0-500 Hz range; this frequency difference allows for precise separation of noise components. When removing noise, a frequency threshold filtering mechanism is used to directly remove signal components with frequencies higher than 5000 Hz. This frequency threshold is determined through analysis of signals from multiple charging piles operating in different scenarios, filtering external electromagnetic interference while avoiding the removal of useful signal components.

[0019] It should be noted that when calculating the statistical characteristics of the pure signal amplitude, the statistical period is first set to 1 hour, and the pure signal amplitude is statistically analyzed once per hour. The mean μ and standard deviation σ of the current statistical period are calculated for subsequent Z-score standardization. For example, the mean voltage signal in a certain period is 5.25 volts, and the standard deviation is 0.529 volts. Simultaneously, the mean drift and variance (or standard deviation) are calculated as auxiliary indicators for equipment condition monitoring. The mean drift is calculated by comparing the signal mean of the current statistical period with the signal mean of the previous statistical period. For example, if the voltage signal mean of the previous period was 5.18 volts, then the mean drift is 0.07 volts. The variance is obtained by calculating the degree of deviation of all pure signal amplitudes from the mean in the current statistical period. For example, the variance of the voltage signal amplitude in a certain period is 0.3.

[0020] It is worth noting that when scaling the amplitude of the pure signal using the standardized parameters, a piecewise standardization strategy is employed. The signal is first converted to a standard normal distribution, and then linearly mapped to the target interval. The scaling process specifically includes the following steps.

[0021] The first step is to calculate the Z-score standardized value for the current statistical period: z = (x - μ) / σ; where z represents the Z-score value, which is dimensionless; x represents the original amplitude of the pure signal; μ represents the mean of the pure signal amplitude within the current statistical period; and σ represents the standard deviation of the pure signal amplitude within the current statistical period.

[0022] The second step is to determine the reasonable range threshold Z_max of the Z-score based on historical statistical data. This is achieved by analyzing the absolute value sequence of the Z-scores from the clean signals of the same model charging pile over the past 30 days (i.e., the sequence formed by taking the absolute value of each Z-score). The 99.7th percentile of this absolute value sequence is taken as Z_max. This means that approximately 99.7% of the Z-score values ​​in the historical data have an absolute value less than Z_max, falling within the interval [-Z_max, Z_max], thus ensuring that this interval covers the vast majority of normal operating conditions.

[0023] The third step is to linearly map the Z-score value to the target interval [0,10]: y=5×(z / Z_max)+5; where y represents the scaled value, and the range is [0,10]; when z=0, y=5, and when z=±Z_max, y corresponds to 0 and 10 respectively.

[0024] For example, when a charging station in a residential community is running, the sensor collects voltage signals 1,000 times per second, forming a raw sequence of 3,600,000 data points per hour. After removing high-frequency noise above 5000 Hz through time-frequency domain feature decomposition, the amplitude of the pure voltage signal is obtained. Statistical analysis shows that the mean of the pure voltage signal for that hour is 5.25 volts, and the standard deviation is 0.529 volts.

[0025] Based on historical data from the past 30 days, Z_max is determined to be 3.0. Therefore, the scaling formula is: y = 5 × (z / 3.0) + 5; Take a data point x = 5.25 volts, calculate z = (5.25 - 5.25) / 0.529 = 0, then y = 5 × 0 + 5 = 5; Let x = 5.25 + 0.529 = 5.779 volts, calculate z = 1.0, then y = 5 × (1.0 / 3.0) + 5 ≈ 6.67; Let x = 5.25 - 0.529 = 4.721 volts, calculate z = -1.0, then y = 5 × (-1.0 / 3.0) + 5 ≈ 3.33. All scaled values ​​fall within the interval [0, 10], forming a unified first signal dataset.

[0026] It should be noted that the entire preprocessing process must be executed in real time within the local data processing module of the charging pile, with processing latency controlled within 1 second to avoid data backlog affecting real-time monitoring. In addition to being stored in the local database, the processed first signal dataset also needs to be uploaded to the cloud server for backup in real time via the IoT module. The backup uses an incremental upload method, uploading only newly added data to reduce network bandwidth consumption. The fields of the first signal dataset include device identifier, data acquisition timestamp, scaled signal value, mean drift, variance distribution data, mean μ, standard deviation σ, and range threshold Z_max, ensuring traceability of data processing details during subsequent analysis.

[0027] In step S102, time-frequency analysis is performed on the time-domain signal of the first signal dataset, converting the time-domain signal into a time-spectrum graph. The time-spectrum graphs are then summarized and integrated to obtain a first spectrum set, including: Based on the preset truncation length and preset overlap rate, the time-domain signal of the first signal dataset is subjected to frame-slicing operation to obtain multiple local time-domain segments with temporal continuity. The local time-domain segment is windowed and smoothed to obtain a windowed time-domain sequence; The windowed time-domain sequence is subjected to a frame-by-frame Fast Fourier Transform calculation to obtain complex frequency domain response data; The square of the modulus is calculated based on the complex frequency domain response data, and a two-dimensional time-frequency matrix reflecting the frequency evolution over time is constructed. The two-dimensional time-frequency matrix is ​​visualized to form a time-frequency spectrum. All the time-frequency spectrums are then aggregated to obtain the first spectrum set.

[0028] It should be noted that before performing frame slicing, a preset truncation length and a preset overlap rate must be set. The preset truncation length is determined based on the duration and variation characteristics of the time-domain signal in the first signal dataset, typically selected within the range of 0.5 to 2 seconds. The preset overlap rate is selected within the range of 30% to 70%. The combination of these two ensures that local time-domain segments retain both independent features and temporal continuity. For example, if the first signal dataset contains a voltage signal lasting 10 seconds, setting the preset truncation length to 1 second and the preset overlap rate to 50%, then starting from the beginning of the signal, a 1-second segment is extracted every 0.5 seconds. The 10-second signal can be divided into 19 local time-domain segments. Each subsequent segment overlaps with the previous segment by 0.5 seconds of signal data, avoiding feature breaks in the temporal dimension due to slicing.

[0029] It is worth noting that a Hamming window function is used when performing windowing smoothing on the local time-domain segments. The Hamming window function is multiplied point-by-point with the signal data of each local time-domain segment, smoothing and attenuating the signal data at both ends of the segment to reduce the spectral leakage caused by abrupt changes at the signal edges. For example, in a local time-domain segment containing 1000 data points, after Hamming window weighting, the amplitude of the data at the beginning and end of the segment gradually approaches 0, while the amplitude of the data in the middle region remains relatively stable. The resulting windowed time-domain sequence better meets the requirements of subsequent frequency domain analysis, making the frequency calculation results more closely match the true characteristics of the signal.

[0030] It should be noted that when performing Fast Fourier Transform (FFT) calculations frame by frame on the windowed time-domain sequence, the length of each windowed time-domain sequence is first padded to an integer power of 2, typically set to 256, 512, or 1024 points, to ensure the efficiency and stability of the transform calculation. For example, a windowed time-domain sequence containing 1000 data points is padded to 1024 data points, and each windowed time-domain sequence is converted from the time domain to the frequency domain using FFT, yielding complex frequency domain response data containing amplitude and phase information. Each data point corresponds to the response of a specific frequency component. It is important to note that the 1024 complex frequency domain response data obtained after FFT calculation do not involve the selection of effective components. Since the windowed time-domain sequence is a real-valued signal, its FFT output satisfies conjugate symmetry, and the number of independent effective frequency components in the range from 0 to the Nyquist frequency should be 513.

[0031] It should be noted that when calculating the square of the modulus based on the complex frequency domain response data, the square of the real part and the square of the imaginary part are added together for each complex frequency domain response data to obtain the energy intensity value corresponding to each frequency component. Since the windowed time-domain sequence is a real-valued signal, its fast Fourier transform output satisfies conjugate symmetry, and the number of independent effective frequency components in the range of 0 to Nyquist is 513. Therefore, with time as the horizontal axis and independent effective frequency as the vertical axis, the energy intensity values ​​corresponding to each windowed time-domain sequence are arranged in chronological and frequency order to construct a two-dimensional time-frequency matrix. For example, after processing, each of the 19 local time-domain segments corresponds to an energy intensity vector of 513 rows (frequency dimension, i.e., the number of independent effective frequency components) × 1 column (time segment). The 19 vectors are concatenated in chronological order to form a two-dimensional time-frequency matrix of 513 rows × 19 columns. The value of each element in the matrix directly reflects the signal energy intensity of the corresponding time point and the corresponding frequency component.

[0032] It is worth noting that when visualizing the two-dimensional time-frequency matrix, a color mapping method is used to map different energy intensity values ​​in the matrix to different shades or hues of color. The larger the energy intensity value, the darker the corresponding area's color or the closer it is to a preset bright hue; the smaller the energy intensity value, the lighter the corresponding area's color or the closer it is to a preset low bright hue, forming a time-frequency spectrum that intuitively reflects the frequency's characteristics over time. For example, energy intensity values ​​from 0 to 100 are mapped to a gradient from blue to red, with a value of 0 displayed as dark blue, a value of 100 as dark red, and intermediate values ​​corresponding to transitional colors. Each two-dimensional time-frequency matrix generates a corresponding time-frequency spectrum.

[0033] For example, the first signal dataset of charging piles along a highway contains a current signal lasting 20 seconds. A preset truncation length of 2 seconds and a preset overlap rate of 50% are set, resulting in 19 local time-domain segments. Each segment is smoothed using a Hamming window and padded to 1024 data points. A fast Fourier transform is performed frame-by-frame to obtain complex frequency domain response data. The square of the modulus is calculated to construct a 513-row × 19-column two-dimensional time-frequency matrix. This two-dimensional matrix data is converted into a time-spectrum graph using color mapping. The horizontal axis represents the 19 time frames (covering the entire 20-second time period), and the vertical axis represents the 513 independent effective frequency components. The color intensity represents the energy intensity at the corresponding time-frequency point. This time-spectrum graph fully presents the frequency evolution of the current signal over 20 seconds, providing visual data support for subsequent comparison with health benchmarks.

[0034] It should be noted that when compiling all the aforementioned time-spectrum graphs, they must be sorted in chronological order, and the time interval information corresponding to each graph must be labeled to ensure the temporal continuity and traceability of the first spectrum set. The sorted time-spectrum graphs are stored in a local storage unit, and are associated with the corresponding device identifier and data acquisition timestamp, facilitating retrieval and comparison with a preset health status benchmark spectrum library in subsequent steps.

[0035] In step S103, a preset health benchmark matrix is ​​obtained from a preset health status benchmark spectrum library, and the preliminary residual between the first spectrum set and the preset health benchmark matrix is ​​calculated to obtain a first residual distribution, including: Convert the time-spectrum graph in the first spectrum set into the current spectrum matrix, and retrieve the preset health benchmark matrix in the preset health state benchmark spectrum library; Perform pointwise difference operations on the current spectrum matrix and the preset health baseline matrix to generate a residual matrix, and count the frequency of occurrence of elements in each numerical interval of the residual matrix; A frequency distribution histogram is constructed based on the frequency of occurrence of the element, and a probability density curve is generated by fitting the frequency distribution histogram. Analyze the probability density characteristics of the probability density curve, and determine the first residual distribution of the degree of deviation of equipment health based on the probability density characteristics.

[0036] It should be noted that when converting the time-spectrum image into the current spectrum matrix, the pixels of the time-spectrum image are used as the basis, and the grayscale value of each pixel is directly used as the element value of the corresponding position in the matrix. The horizontal pixels of the time-spectrum image correspond to the time dimension, and the vertical pixels correspond to the frequency dimension. The number of rows in the matrix is ​​the same as the number of vertical pixels in the time-spectrum image, and the number of columns in the matrix is ​​the same as the number of horizontal pixels in the time-spectrum image, ensuring that the time and frequency features are completely preserved in the matrix. For example, a time-spectrum image with a resolution of 200 pixels (vertical, frequency) × 300 pixels (horizontal, time) will be converted into a current spectrum matrix with 200 rows × 300 columns. The value of each element in the matrix ranges from 0 to 255, corresponding to the grayscale depth of the original pixel.

[0037] It is worth noting that a preset health status benchmark spectrum library needs to be constructed in advance before retrieving the preset health benchmark matrix. Charging piles of the same model and specifications, in either brand new or normal operating condition, are selected. Their electrical signals under different operating conditions are collected, preprocessed in S101, and converted into time-frequency spectrograms through S102 time-frequency analysis. These are then converted into health spectrum matrices using the same method. Multiple sets of health spectrum matrices are averaged to form the preset health benchmark matrix, which is then stored in the preset health status benchmark spectrum library. During retrieval, the corresponding preset health benchmark matrix is ​​matched according to the model and specifications of the currently monitored charging pile to ensure the relevance and rationality of the comparison and avoid deviations in judgment caused by equipment differences.

[0038] It should be noted that when performing point-by-point difference operations on the current spectral matrix and the preset health benchmark matrix, the difference between corresponding elements of the two matrices is taken. That is, the value of an element at a certain position in the current spectral matrix is ​​subtracted from the value of the element at the same position in the preset health benchmark matrix, generating an original difference matrix containing positive and negative deviation information. To focus on the degree of deviation rather than the direction of deviation, absolute value operations are performed on all elements in the original difference matrix to obtain a residual matrix. All elements in the residual matrix are non-negative, directly reflecting the magnitude of the deviation at the corresponding position. For example, if the value of an element at a certain position in the current spectral matrix is ​​80, and the value of the corresponding element at the preset health benchmark matrix is ​​70, the original difference is 10, and the value of the element at that position in the residual matrix after absolute value processing is 10; if the current element value is 60, the original difference is -10, and the value after absolute value processing is still 10, uniformly representing the degree of deviation with non-negative values.

[0039] For example, if the residual matrix element values ​​range from 0 to 50, the intervals are divided with a fixed step size of 5, using a left-closed, right-open approach to avoid boundary repetition, resulting in 10 numerical intervals: [0, 5), [5, 10), [10, 15), ... up to [45, 50). All elements in the residual matrix are traversed, and the numerical interval to which each element belongs is counted. For example, an element value of 5 is assigned to the interval [5-10], and an element value of 45 is assigned to the interval [45-50]. The number of elements in each interval is recorded to generate frequency statistics. This interval division method clearly presents the distribution of different degrees of deviation, providing a foundation for subsequent distribution analysis.

[0040] It's worth noting that when constructing the frequency distribution histogram, the horizontal axis is the numerical interval, and the vertical axis is the frequency of occurrence of elements within the corresponding interval. Each numerical interval corresponds to a rectangular bar, and the height of the rectangular bar is proportional to the frequency. When fitting the probability density curve based on this frequency distribution histogram, a smooth fitting method is used to ensure the curve continuously reflects the changing trend of the frequency distribution. The peak position of the curve corresponds to the numerical interval with the highest element frequency, and the width of the curve reflects the distribution range of the degree of deviation. For example, in the frequency distribution histogram, the frequency of the 0-5 interval is 1200, the 5-10 interval is 800, and the 10-15 interval is 300. The fitted probability density curve forms a peak in the 0-5 interval and then gradually decreases as the numerical interval increases, intuitively presenting the distribution characteristic that most elements are concentrated in the low deviation interval.

[0041] For example, the first spectrum set of a charging pile in a certain community contains 5 time-spectrum graphs, each with a resolution of 150×200 pixels. After conversion, they form 5 current spectrum matrixes of 150 rows × 200 columns. A preset health benchmark matrix (150×200) of the same model of charging pile is retrieved. Pointwise differences are performed between each current spectrum matrix and the preset health benchmark matrix, and the absolute values ​​are taken to obtain 5 residual matrices. The element values ​​of each residual matrix are counted, dividing it into 10 intervals, such as 0-5 and 5-10. A certain residual matrix has 1200 elements in the 0-5 interval, 800 elements in the 5-10 interval, 300 elements in the 10-15 interval, and fewer elements in the remaining intervals. A frequency distribution histogram was constructed using these frequency data. The fitted probability density curve showed a significant peak in the 0-5 interval. The curve was generally right-skewed. By analyzing the peak position (0-5 interval) and the curve decay rate (slow decay in the low interval and fast decay in the high interval), the first residual distribution of the current health deviation of the charging pile was determined. That is, most positions have small deviations, and only a few positions have large deviations.

[0042] It should be noted that when analyzing the probability density characteristics of the probability density curve, the focus is on the peak interval, peak height, and curve extension range. The smaller the value corresponding to the peak interval, the less the overall deviation of the equipment from a healthy state; the higher the peak height, the more concentrated the deviation distribution; the narrower the curve extension range, the smaller the fluctuation in the degree of deviation, and the more stable the equipment state. By integrating these characteristics, the numerical distribution range, peak interval, and proportion of the probability density curve are combined to form a complete first residual distribution, providing a quantitative basis for subsequent residual correction and early warning judgment.

[0043] In step S104, the first residual distribution is adjusted using a preset correction model to obtain a second residual distribution, including: The power load data and temperature sensor sequence of the device operation cycle are read synchronously. Gradient calculation is performed on the power load data to extract transient response features. The thermal drift coefficient is determined based on the fluctuation slope of the temperature sensor sequence. The transient response characteristics and the thermal drift coefficient are mapped to a preset correction model to generate a comprehensive influence vector; Based on the comprehensive influence vector, calculate the interference weight matrix for different frequency ranges and generate the correction model compensation factor; The correction model compensation factor is used to perform a point-by-point correction operation on each data point in the first residual distribution to eliminate non-faulty deviations and obtain the adjusted residual value. The adjusted residual values ​​are statistically binned again, and a new probability density curve is generated based on the binning results. The statistical characteristic parameters of the residual data are then updated to obtain the updated statistical characteristic parameters. Based on the updated statistical characteristic parameters, the second residual distribution after removing the influence of dynamic operating conditions and environmental disturbances is determined.

[0044] It should be noted that when synchronously reading power load data and temperature sensor sequences, the power load data is obtained from the charging pile's power monitoring unit, and the temperature sensor sequence is obtained from the charging pile's built-in temperature sensor. The sampling frequency of both is consistent with the statistical period of the first residual distribution, typically set to once per second based on experimental data statistics. This ensures accurate data correspondence in the time dimension and avoids misjudgments of interference factors due to asynchronous sampling. For example, if the statistical period of the first residual distribution is one hour, both the power load data and temperature sensor sequence are collected at a frequency of once per second, acquiring 3600 data points each within one hour. Each data point is accompanied by a timestamp of the same precision, ensuring that subsequent analysis can correspond one-to-one with interference factors and residual data at the same time point.

[0045] It is worth noting that when performing gradient calculations to extract transient response features from the power load data, the difference between two adjacent power load data points is taken and divided by the time interval between the two data points (1 second) to obtain the instantaneous rate of change of the power load. This rate is the transient response feature value. When the power load changes rapidly, the instantaneous rate of change will fluctuate significantly. By continuously calculating this rate, the dynamic impact of power switching on the spectrum morphology can be captured. For example, if the power load changes from 50 kW to 80 kW, the difference between adjacent data points is 30 kW, and the time interval is 1 second, the corresponding transient response feature value is 30 kW / s. This value directly reflects the severity of the power switching.

[0046] It should be noted that when determining the thermal drift coefficient based on the fluctuation slope of the temperature sensor sequence, the difference between two adjacent temperature data points is calculated, divided by the time interval to obtain the temperature change rate, and the average value of the temperature change rate over one hour is taken as the thermal drift coefficient. The more drastic the temperature fluctuation, the larger the thermal drift coefficient, quantifying the distortion effect of temperature changes on the signal amplitude and providing a basis for subsequent correction of model quantization interference.

[0047] It is worth noting that the preset correction model is a hybrid neural network that integrates fully connected layers and frequency attention mechanisms. It is specifically used to quantify the correlation between transient response characteristics, thermal drift coefficients and spectral morphology changes, so as to achieve accurate compensation for the influence of interference factors on the residual.

[0048] The input layer of this hybrid neural network receives two types of normalized data: one is the time sequence of the transient response features, segmented according to the device's operating cycle, with each segment having a fixed length of 100 data points; the other is the numerical sequence of the thermal drift coefficient, synchronized with the transient response feature time sequence, with each segment also having a length of 100 data points. Simultaneously, a frequency range partitioning code is input, including codes corresponding to low, medium, and high frequency ranges. The low-frequency range is encoded as [1,0,0], the medium-frequency range as [0,1,0], and the high-frequency range as [0,0,1]. These three types of data are concatenated to form a multi-dimensional input vector. The input layer is followed by two fully connected layers. The first fully connected layer has 128 neurons, and the second layer has 64 neurons, both using the ReLU activation function to extract the nonlinear correlation features between transient response and thermal drift. After the fully connected layers, a frequency attention layer is connected, which learns the importance of interference in different frequency ranges and assigns specific weights to the features of each range. Finally, the output layer is connected, with 6 neurons corresponding to the frequency drift and amplitude distortion values ​​of the low, medium, and high frequency ranges, respectively. The output dimension corresponds one-to-one with the number of frequency ranges and the quantization target.

[0049] The model was trained using a supervised learning approach. Training data was derived from historical operating data of the same model of charging piles under different power loads and ambient temperatures, collecting 100,000 valid samples. The input to each sample was a transient response feature sequence, thermal drift coefficient sequence, and frequency range encoding under specific operating conditions. The output was the actual frequency drift and amplitude distortion values ​​under those conditions, obtained through experimental measurements (averaged after multiple repeated measurements to ensure accuracy). During training, mean squared error was used as the loss function, and the Adam optimizer was employed to adjust network parameters. The number of training iterations was set to 500 rounds. After each round, the validation set loss value was calculated. Training was stopped when the fluctuation of the validation set loss value was less than 0.001 for 10 consecutive rounds, and the network parameters at this point were saved as the final preset modified model.

[0050] The specific process of mapping the transient response features and thermal drift coefficients to the preset correction model is as follows: First, the transient response features and thermal drift coefficients are normalized to adjust their numerical range to the 0-1 interval to avoid the influence of dimensional differences on model calculation. Then, the normalized transient response feature sequence, thermal drift coefficient sequence, and corresponding frequency interval codes are concatenated to form a multi-dimensional vector that meets the model input requirements. This vector is then input into the trained hybrid neural network, where nonlinear correlation features are extracted through a fully connected layer, and after interval weights are assigned by a frequency attention layer, the output layer directly outputs each frequency interval. The quantization results are the frequency drift and amplitude distortion values ​​in the low-frequency range, the mid-frequency range, and the high-frequency range. For each frequency range, the frequency drift and amplitude distortion values ​​are arranged in the order of "frequency drift first, amplitude distortion value second". Then, the arrangement results of the low-frequency range, the mid-frequency range, and the high-frequency range are continuously concatenated in the order of "low-frequency range first, mid-frequency range second, and high-frequency range last" to form a comprehensive influence vector. Each pair of consecutive elements in the vector corresponds to the comprehensive interference-related parameters of a frequency range.

[0051] It should be noted that the preset correction model is pre-constructed based on a large amount of historical data. The preset correction model includes the correspondence between different transient response characteristic values, thermal drift coefficients, and spectral morphology changes. The transient response characteristics and thermal drift coefficients are input into the preset correction model. Based on the built-in correspondence, the model quantifies the frequency drift caused by power switching and the amplitude distortion caused by temperature fluctuations. The two quantization results are combined according to frequency intervals to generate a comprehensive influence vector. Each component in the vector corresponds to the comprehensive interference level of a frequency interval.

[0052] It is worth noting that when calculating the interference weight matrix for different frequency ranges, the composition of the comprehensive influence vector is first defined. Each pair of consecutive elements in the comprehensive influence vector corresponds to the interference-related parameters for a frequency range, in the following order: frequency drift and amplitude distortion for the low-frequency range, frequency drift and amplitude distortion for the mid-frequency range, and frequency drift and amplitude distortion for the high-frequency range. For each frequency range, the two corresponding parameters are summed to obtain the comprehensive interference level value for that frequency range. The comprehensive interference level values ​​for the three frequency ranges are then normalized so that the range of each value is limited to 0-1. The normalized values ​​are the interference weights for the corresponding frequency ranges. The interference weight matrix is ​​formed by arranging the three interference weights in the order of low frequency, mid frequency, and high frequency. When generating the compensation factor for the correction model, the mean of the first residual distribution is first calculated. The interference weight for each frequency range in the interference weight matrix is ​​then multiplied by the mean of the first residual distribution. The result is the compensation factor for the corresponding frequency range.

[0053] For example, if the comprehensive influence vector is [0.2, 0.1, 0.3, 0.15, 0.1, 0.08], then the parameters corresponding to the low-frequency range are 0.2 and 0.1, and summing them gives a comprehensive interference level value of 0.3 for the low-frequency range; the parameters corresponding to the mid-frequency range are 0.3 and 0.15, and summing them gives a comprehensive interference level value of 0.45 for the mid-frequency range; and the parameters corresponding to the high-frequency range are 0.1 and 0.08, and summing them gives a comprehensive interference level value of 0.18 for the high-frequency range. The sum of the three comprehensive interference level values ​​is 0.3 + 0.45 + 0.18 = 0.93, and the interference weight matrix formed after normalization is [[0.32], [0.48], [0.19]]. If the mean of the first residual distribution is 5, then the compensation factor in the low-frequency range is 0.32×5=1.6, the compensation factor in the mid-frequency range is 0.48×5=2.4, and the compensation factor in the high-frequency range is 0.19×5=0.95. This process ensures that the compensation factor in each frequency range accurately matches its interference level and the overall residual level, providing a precise basis for subsequent residual correction.

[0054] It should be noted that when performing point-by-point correction operations on each data point in the first residual distribution using the compensation factor of the correction model, the frequency range to which each data point belongs is first determined, the compensation factor for the corresponding range is called, and the original value of the data point is subtracted from the compensation factor to obtain the adjusted residual value. For example, if a data point in the first residual distribution has a value of 8, the compensation factor for its low-frequency range is 4, and the adjusted residual value after correction is 4; another data point has a value of 6, the compensation factor for its mid-frequency range is 2.5, and the adjusted residual value after correction is 3.5. This operation eliminates non-fault deviations caused by power switching and temperature fluctuations.

[0055] For example, the first residual distribution statistical period of a charging pile in a highway service area is 1 hour. In the synchronously read power load data, the power switches from 60 kW to 90 kW at a certain moment, with a transient response characteristic value of 30 kW / s. Based on the historical operating data of this charging pile, the historical maximum value of the transient response characteristic is determined to be 50 kW / s, and the minimum value is 0 kW / s. The normalized transient response characteristic value is 0.6. A temperature sensor sequence shows that the temperature rises from 28 degrees Celsius to 36 degrees Celsius within 1 hour, and the calculated thermal drift coefficient is 0.02 degrees Celsius / s. Based on historical data, the historical maximum value of the thermal drift coefficient is determined to be 0.05 degrees Celsius / s, and the minimum value is 0 degrees Celsius. The normalized thermal drift coefficient value is 0.4.

[0056] The normalized transient response eigenvalues ​​and thermal drift coefficients are mapped to a preset correction model. The normalized frequency drift and amplitude distortion values ​​are output for each frequency range in the order of low frequency, mid frequency, and high frequency. The frequency drift is 0.5 and the amplitude distortion is 0.3 for the low frequency range, 0.24 and 0.2 for the mid frequency range, and 0.12 and 0.1 for the high frequency range. For each frequency range, the corresponding normalized frequency drift and normalized amplitude distortion are summed to obtain the comprehensive interference level value for each range: 0.5 + 0.3 = 0.8 for the low frequency range, 0.24 + 0.2 = 0.44 for the mid frequency range, and 0.12 + 0.1 = 0.22 for the high frequency range. The comprehensive interference level value of each interval is divided by the sum and normalized to obtain the interference weight of each interval: 0.55 for low frequency interval, 0.30 for mid frequency interval, and 0.15 for high frequency interval. The interference weight matrix is ​​formed by arranging these three interference weights in the order of low frequency, mid frequency, and high frequency.

[0057] The mean of the first residual distribution is 6. The interference weights of each interval in the interference weight matrix are multiplied by this mean to obtain the compensation factor for the corresponding frequency interval: 0.55 × 6 = 3.3 for the low-frequency interval, 0.30 × 6 = 1.8 for the mid-frequency interval, and 0.15 × 6 = 0.9 for the high-frequency interval. The data points in each frequency interval of the first residual distribution are then corrected by subtracting the corresponding frequency interval compensation factor from the original residual value: For example, a data point in the low-frequency interval with an original value of 9 is corrected to 9 - 3.3 = 5.7; a data point in the mid-frequency interval with an original value of 7 is corrected to 7 - 1.8 = 5.2; and a data point in the high-frequency interval with an original value of 5 is corrected to 5 - 0.9 = 4.1. All adjusted residual values ​​were re-binded statistically according to the intervals of 0-2, 2-4, 4-6, and 6-8. Based on the frequency data after binning, a new probability density curve was generated. The updated statistical characteristic parameters showed that the adjusted residual values ​​were mainly concentrated in the 4-6 interval. Based on this, the second residual distribution after removing the influence of dynamic operating conditions and environmental interference was determined.

[0058] It should be noted that when updating the statistical characteristic parameters of the residual data, the mean, variance, median, and other parameters of the adjusted residual values ​​are calculated and compared with the statistical characteristic parameters of the first residual distribution to verify the correction effect. If the adjusted mean is closer to 0 and the variance is smaller, it indicates that non-fault deviations have been effectively eliminated, and the corrected second residual distribution better reflects the true degree of equipment health deviation. The statistical characteristic parameters and the second residual distribution are stored together to provide data support for subsequent threshold comparison and early warning judgment.

[0059] In step S105, the second residual distribution is compared with a preset threshold interval. If the corrected residual value in the second residual distribution exceeds the preset threshold interval, a preliminary warning sign is obtained, including: A continuous probability density function is constructed for the second residual distribution using kernel density estimation; Based on the continuous probability density function, a preset threshold interval is defined. The corrected residual value in the second residual distribution is compared with the preset threshold interval. The Manhattan distance of the corrected residual value exceeding the preset threshold interval is calculated to obtain the set of outlier deviations. An aging feature vector is constructed based on the set of outlier deviations, and an aging risk probability value is calculated based on the aging feature vector. If the aging risk probability value exceeds the preset safety margin baseline, a risk assessment report is generated, and the risk level information in the risk assessment report is analyzed to obtain a preliminary warning sign.

[0060] It should be noted that when constructing a continuous probability density function for the second residual distribution using kernel density estimation, all corrected residual values ​​in the second residual distribution are first collected, the kernel function type is determined to be a Gaussian kernel function, and the kernel function bandwidth is set. The bandwidth value is determined according to the number of corrected residual values. Based on experimental statistics, the bandwidth is usually set to 0.5 when there are 1000-5000 corrected residual values ​​and 0.3 when there are 5000-10000 values. This ensures that the function can smoothly reflect the characteristics of the residual distribution and avoids the distribution becoming blurred due to excessive bandwidth or the fluctuation becoming drastic due to excessively small bandwidth.

[0061] It is worth noting that when defining the preset threshold interval based on the continuous probability density function, a 95% confidence interval is used as the dividing criterion. This criterion can cover the corrected residual values ​​of most normal fluctuations, only including a small number of extreme values ​​in the anomaly consideration range. The residual range corresponding to the 95% area under the continuous probability density function is calculated by integration, and this range is the preset threshold interval.

[0062] It should be noted that when comparing the corrected residual value with the preset threshold range, each corrected residual value is checked individually to see if it falls within the range of -4 to 4. If the value is greater than 4 or less than -4, it is determined to be an outlier value exceeding the preset threshold range. When calculating the Manhattan distance, for outliers greater than 4, the distance is obtained by subtracting 4 from the value; for outliers less than -4, the distance is obtained by subtracting the value from -4. The distance directly quantifies the degree to which an outlier value deviates from the threshold boundary. For example, if a corrected residual value is 6.5, exceeding the upper threshold of 4, the Manhattan distance is 2.5; if another corrected residual value is -5.8, exceeding the lower threshold of -4, the Manhattan distance is 1.8. The distances of all outliers are summed to form an outlier deviation set.

[0063] It should be noted that when constructing the aging feature vector based on the outlier deviation set, three core statistics are extracted from the set: mean deviation, maximum deviation, and number of outliers. These three statistics are combined in sequence to form a three-dimensional aging feature vector, which comprehensively reflects the overall deviation level, the most severe deviation, and the scale of the number of outliers.

[0064] It is worth noting that when calculating the probability value of aging risk based on the aging feature vector, the weights of each item in the weighted mapping rule are first set. The weight determination process is as follows: historical fault data and corresponding operation data of the same type of charging pile are collected, and three statistics related to aging faults—mean deviation, maximum deviation, and number of outliers—are extracted from the historical data. The correlation between the three statistics and the actual occurrence of aging faults is statistically analyzed, and the correlation coefficient between each statistic and the occurrence of faults is calculated. Then, multiple sets of weight combinations with different proportions are designed, and each set of weight combinations is applied to historical data for verification. The weight proportions that best fit the logic of the actual impact of aging risk are selected, and finally, the weight of mean deviation is 40%, the weight of maximum deviation is 30%, and the weight of number of outliers is 30%.

[0065] The normalization process employs linear normalization. The required maximum and minimum values ​​for each statistic are obtained as follows: Historical operational data of the same model of charging piles are collected throughout the entire process from normal operation to aging failure. From the historical data, the mean deviation, maximum deviation, and number of outliers are extracted. The historical maximum and minimum values ​​of the three statistics are then calculated and used as the baseline parameters for normalization. The normalization calculation method is as follows: the original value of each statistic is subtracted from the corresponding historical minimum value, and then divided by the difference between the historical maximum and historical minimum values ​​of that statistic, resulting in a normalized value within the 0-1 range.

[0066] After normalization, a weighted sum is calculated according to the set weights. The weighted sum is the aging risk probability value, with a value between 0 and 1. For example, in the historical data of the same type of charging pile collected, the historical maximum value of the mean deviation is 3.0 and the historical minimum value is 0; the historical maximum value of the maximum deviation is 4.0 and the historical minimum value is 0; and the historical maximum value of the number of outliers is 7.5 and the historical minimum value is 0. In a certain aging feature vector, the original value of the mean deviation is 2.34, and the normalized calculation is (2.34-0) ÷ (3.0-0) = 0.78; the original value of the maximum deviation is 3.2, and the normalized calculation is (3.2-0) ÷ (4.0-0) = 0.8; the original value of the number of outliers is 5, and the normalized calculation is (5-0) ÷ (7.5-0) ≈ 0.67. The weighted sum calculated using the set weights is 0.78×40%+0.8×30%+0.67×30%=0.312+0.24+0.201=0.753, corresponding to an aging risk probability value of 0.753.

[0067] It should be noted that the preset safety margin baseline is set based on a large amount of historical data from charging pile operations, typically 0.6. This value balances the risks of missed and false alarms, neither ignoring potential aging issues nor exaggerating anomalies caused by normal fluctuations. If the aging risk probability value exceeds 0.6, a risk assessment report is generated. The report categorizes risk levels according to probability values: 0.6-0.7 is medium risk, 0.7-0.8 is relatively high risk, and above 0.8 is high risk. Each level corresponds to a clear risk description and preliminary maintenance recommendations. The risk level information in the risk assessment report is analyzed: medium risk corresponds to a yellow warning sign, relatively high risk to an orange warning sign, and high risk to a red warning sign. The warning signs include the risk level, outlier statistics, and preliminary maintenance directions, which constitute the final preliminary warning sign.

[0068] For example, the second residual distribution of charging piles in a city's commercial area contains 3600 corrected residual values. With a kernel function bandwidth of 0.5, after constructing a continuous probability density function, the 95% confidence interval is defined as a preset threshold range of -4 to 4. After comparison, five corrected residual values ​​were found to exceed this range: 6.5, 7.2, -5.8, -6.3, and 5.9. The corresponding Manhattan distances were calculated to be 2.5, 3.2, 1.8, 2.3, and 1.9, forming an outlier deviation set. An aging feature vector of [2.34, 3.2, 5] was constructed and normalized based on historical statistical parameters (mean deviation, maximum value 3.0, minimum value 0; maximum deviation, minimum value 4.0, minimum value 0; number of outliers, maximum value 7.5, minimum value 0). The normalized value of the mean deviation was (2.34-0) / (3.0-0)=0.78; the normalized value of the maximum deviation was (3.2-0) / (4.0-0)=0.8; and the normalized value of the number of outliers was (5-0) / (7.5-0)≈0.67. Subsequently, a weighted sum was calculated according to the set weights to obtain an aging risk probability value of 0.753, which exceeds the preset safety margin baseline of 0.6. A risk assessment report was generated, and the risk level in the report was analyzed as high risk, corresponding to an orange warning sign. The sign noted "5 outliers, maximum deviation 3.2, it is recommended to check the operating status of the power module", thus completing the generation of the preliminary warning sign.

[0069] In step S106, the preliminary warning indicator is further verified by combining the abnormal pattern characteristics in historical operating data to determine the aging state judgment result, including: Extract current and voltage time-series waveform data and abnormal pattern characteristics from historical operating data; The current and voltage time-series waveform data are subjected to Fast Fourier Transform to obtain the current frequency domain spectrum data. The main frequency component features of the current frequency domain spectrum data are extracted using principal component analysis. The current abnormal frequency component set is constructed based on the main frequency component features. Combining the current abnormal frequency component set with the abnormal pattern features, a cosine similarity value is calculated, and the historical aging pattern with the highest cosine similarity value is selected as the matching benchmark pattern. If the cosine similarity value is greater than the preset verification confidence threshold, extract the fault type label and aging level information corresponding to the matching benchmark pattern, and generate a secondary verification confirmation signal. Based on the secondary verification confirmation signal, the status attributes of the preliminary warning flag are corrected, and the aging status judgment result is output.

[0070] It should be noted that the historical operating data extracted comes from long-term operating records of charging piles of the same model, covering various scenarios with an operating period of 1-5 years, an average daily charging frequency of 2-15 times, and an ambient temperature of -10℃ to 45℃, ensuring comprehensive data coverage. The data includes complete current and voltage time-series waveform data and corresponding abnormal mode features. These abnormal mode features are reverse-labeled using post-fault disassembly and inspection results. Each feature is bound to a specific aging type; for example, insulation aging corresponds to harmonic peaks in a specific frequency band, and switching device degradation corresponds to amplitude abrupt changes at a fixed frequency, avoiding mismatches between features and aging types. Before constructing the historical aging feature spectrum library, the historical raw waveform data undergoes the same preprocessing steps as the current monitoring data, including filtering and denoising, standardization and scaling, time-frequency analysis (using the same windowing and FFT parameters), and feature extraction, to obtain a one-dimensional feature vector with the same physical meaning and scale as the current abnormal frequency component vector. The extraction process is segmented according to the equipment's operating cycle, with each segment's duration consistent with the statistical cycle of the current monitoring data, typically one hour. This ensures the comparability of the two data segments in the time dimension and avoids feature distortion caused by cycle differences. In the pre-set historical aging feature spectrum library, each historical aging feature sample is stored as a one-dimensional vector. The vector construction method is consistent with the current abnormal frequency component vector, and each sample is bound to a corresponding fault type label and aging degree level information, providing a foundation for similarity comparison and result matching.

[0071] It's worth noting that before performing a Fast Fourier Transform (FFT) on the current and voltage time-series waveform data, the data is padded to a length that is an integer power of 2, commonly set to 1024 or 2048 points, to ensure the stability and efficiency of the transformation calculation. The FFT converts the time-domain current and voltage waveform data into frequency-domain data, resulting in a current frequency spectrum containing amplitudes of different frequency components. Each frequency point in the data corresponds to an amplitude, visually presenting the frequency distribution characteristics of the signal.

[0072] When extracting the main frequency component features using principal component analysis, the covariance matrix of the current frequency domain spectrum data is first calculated. The eigenvalues ​​and eigenvectors of the covariance matrix are then solved. The eigenvalues ​​are sorted from largest to smallest, and the frequency components corresponding to the top N eigenvectors with a cumulative contribution rate of over 90% are selected. These components are the main frequency component features, which retain the core information reflecting the aging state while eliminating irrelevant and redundant components. When constructing the current abnormal frequency component set, the main frequency components are arranged in ascending order of frequency. The frequency value and corresponding amplitude of each component are recorded. The amplitudes of each main frequency component are then sequentially arranged according to the sorting results to form a one-dimensional vector of the current abnormal frequency components. This ensures that the correspondence between amplitude and frequency in the vector is fixed and unique, providing input data of a unified dimension for subsequent cosine similarity calculations.

[0073] The cosine similarity calculation takes the current abnormal frequency component vector and the historical aging feature sample vector as input. Since the construction rules of the two types of vectors are consistent, they are both sorted by frequency from low to high and their amplitudes are arranged sequentially to ensure that the two have the same dimension and the corresponding meaning of the elements. The similarity value is obtained by calculating the cosine of the angle between the two types of vectors. The larger the value, the higher the degree of matching between the current abnormal frequency component and the historical aging features, and the better it reflects the correlation between the current aging mode of the charging pile and the historical failure mode.

[0074] It should be noted that the historical aging feature spectrum library pre-stores a large number of verified historical aging patterns. Each historical aging pattern includes a corresponding set of abnormal frequency components, fault type labels, and aging severity level information. The fault type labels cover common aging types of charging piles, such as "power module insulation aging," "switching device degradation," "capacitor attenuation," and "poor line contact." The aging severity level information is divided into three categories: "mild," "moderate," and "severe," each corresponding to different maintenance priorities. When calculating the cosine similarity between the current abnormal frequency component set and the historical aging pattern set, the amplitude data in both sets are first normalized to the 0-1 range to eliminate the influence of dimensional differences. Then, the dot product of the two normalized amplitude vectors is calculated and divided by the product of the magnitudes of the two vectors. The result is the cosine similarity value, which ranges from 0 to 1. The larger the value, the higher the feature matching degree between the two sets. When selecting the matching benchmark mode, all modes in the historical aging feature spectrum library are traversed, and the cosine similarity value between each mode and the current set is recorded. The mode with the highest value is selected as the matching benchmark mode, which is the historical aging case closest to the current equipment state.

[0075] It is worth noting that the preset verification confidence threshold was determined through statistical analysis of a large number of historical cases. Historical data containing cases of normal operation and confirmed aging types were collected. The cosine similarity between the feature vector of each normal operation case and the feature vectors of all failure modes in the historical aging feature spectrum library was calculated. The upper limit of all such similarity values ​​was taken, for example, the 95th percentile, as the preset verification confidence threshold, such as 0.85. This setting method ensures that when the similarity between the features of a new case and the failure mode exceeds this threshold, the possibility that it belongs to normal fluctuation is extremely low, thus effectively confirming the aging state. The setting of the preset verification confidence threshold can balance the risks of missed detection and false detection. When the similarity value is greater than 0.85, it means that the current feature matches the historical aging mode sufficiently, and the current aging state can be confirmed based on the historical mode; when the value is less than or equal to 0.85, it means that the matching degree is insufficient, and it cannot be directly associated with a clear aging type. It needs to be marked as "insufficient matching degree" and handled with caution.

[0076] It should be noted that when generating the secondary verification confirmation signal, if the cosine similarity value is greater than the preset verification confidence threshold, the fault type label and aging level information corresponding to the matching baseline pattern are directly extracted, and the two types of information are combined to form the secondary verification confirmation signal. The signal must contain sufficient details to ensure that there is a clear basis for subsequent correction of the warning flag. If the similarity value is less than or equal to the preset verification confidence threshold, no secondary verification confirmation signal is generated, and it is directly marked as "insufficient matching degree".

[0077] When correcting the status attribute of the initial warning flag, it is executed according to fixed rules. The trigger condition for "insufficient matching degree" is that the cosine similarity value between the current abnormal frequency component vector and all sample vectors in the historical aging feature spectrum library is less than or equal to the preset verification confidence threshold. At this time, it is marked as "insufficient matching degree".

[0078] The specific correction rules are as follows: If the initial warning sign is yellow (medium risk), and the secondary verification confirms the signal as "moderate aging," it will be upgraded to orange (higher risk); if it is "severe aging," it will be upgraded to red (high risk); and if it is "mild aging," it will be downgraded to blue (low risk). If the initial warning sign is orange (higher risk), and the secondary verification confirms the signal as "severe aging," it will be upgraded to red (high risk); and if it is "mild aging," it will be downgraded to yellow (medium risk). If the initial warning sign is red (high risk), and the secondary verification confirms the signal as "severe aging," it will remain red; if it is "moderate aging," it will be downgraded to orange (higher risk); and if it is "mild aging," it will be downgraded to yellow (medium risk). If it is marked as "insufficient matching," regardless of the initial warning level, it will be downgraded to blue (low risk), with the added explanation "Continuous monitoring is required, and re-verification will be performed after accumulating more data."

[0079] For example, a suburban charging pile is initially identified as yellow (medium risk) by a preliminary warning. One hour of its current and voltage time-series waveform data is extracted, padded to 1024 points, and then subjected to a Fast Fourier Transform. Principal component analysis is used to extract the top eight main frequency components, and the amplitudes are arranged from low to high to obtain the current abnormal frequency component vector [0.35, 0.32, 0.28, 0.24, 0.20, 0.17, 0.14, 0.11]. The historical aging feature spectrum library is retrieved, and the cosine similarity between this vector and all historical sample vectors in the library is calculated. The highest similarity value is 0.78, and the preset verification confidence threshold is 0.85. This similarity value is less than the threshold and is marked as "insufficient matching". According to the correction rules, the initial yellow warning sign is downgraded to blue (low risk), with the additional note "Continuous monitoring is required; re-verification will be performed after accumulating more data," thus completing the correction of the initial warning sign's status attribute.

[0080] In step S107, a remote monitoring report is generated based on the aging status judgment result and pushed to the relevant management platform to complete the real-time early warning and feedback of the aging status of the charging pile.

[0081] It should be noted that after obtaining the aging status judgment result, the structured monitoring data packet is packaged using JSON format to organize the information, ensuring that the fields are clear and easy to parse. The data packet must include the following core fields: a unique device serial number, which consists of 8 digits + 4 letters, corresponding one-to-one with the charging pile hardware identifier; the judgment time, accurate to the second, in the format "year-month-day hour:minute:second"; a fault type label, which directly uses the label content confirmed by the secondary verification; an aging degree level, which is divided into "mild", "moderate", and "severe"; the initial warning indicator of the original state; the secondary verification similarity value; and recommended maintenance actions. The recommended maintenance actions are bound to the aging level: mild corresponds to "regular inspection, recording operating data every 30 days"; moderate corresponds to "on-site inspection within 48 hours, focusing on checking the corresponding faulty components"; and severe corresponds to "immediate shutdown, prohibition of use, replacement of faulty components within 24 hours", ensuring that the maintenance direction is clear. The data packet also needs to include a data source identifier, indicating key information such as the signal acquisition period and time-frequency analysis parameters on which the judgment result is based, so as to provide a basis for the remote platform to trace and analyze the process.

[0082] In summary, this invention provides a remote online monitoring method for charging piles based on the Internet of Things (IoT). It collects electrical signals from charging piles in real time, performs noise reduction and standardization preprocessing, converts the time-domain signals into a time-frequency spectrum, calculates preliminary residuals using a pre-set health baseline spectrum library, introduces a correction model to eliminate the influence of dynamic operating conditions and environmental interference, generates preliminary early warning indicators through threshold comparison, and then confirms the aging state through secondary verification using abnormal pattern characteristics in historical operating data. Finally, a monitoring report is generated and pushed to the management platform, forming a closed-loop mechanism of "signal acquisition and preprocessing - time-frequency analysis - residual calculation and correction - preliminary early warning - secondary verification - early warning feedback." This method can solve the problem that existing technologies are unable to adapt to the complex operating conditions of charging piles, thus failing to achieve accurate fault prediction. refer to Figure 2 The second embodiment of the present invention provides a remote online monitoring system for charging piles based on the Internet of Things, comprising: The signal acquisition and preprocessing module is used to acquire electrical signal data during the operation of the charging pile in real time, and preprocess the electrical signal data to obtain a first signal dataset. The time-frequency analysis module is used to perform time-frequency analysis on the time-domain signal of the first signal dataset, convert the time-domain signal into a time-frequency spectrum, summarize the time-frequency spectrum, and integrate them to obtain a first spectrum set. The residual calculation module is used to obtain a preset health benchmark matrix from a preset health status benchmark spectrum library, calculate the preliminary residual between the first spectrum set and the preset health benchmark matrix, and obtain the first residual distribution. The residual correction module is used to adjust the first residual distribution through a preset correction model to obtain the second residual distribution; The preliminary warning module is used to compare the second residual distribution with a preset threshold range. If the corrected residual value in the second residual distribution exceeds the preset threshold range, a preliminary warning sign is obtained. The secondary verification module is used to perform secondary verification on the preliminary warning sign by combining the abnormal pattern characteristics in historical operation data to determine the aging status judgment result. The early warning and feedback module is used to generate a remote monitoring report based on the aging status judgment results and push it to the relevant management platform to complete the real-time early warning and feedback of the aging status of the charging pile.

[0083] It should be noted that the IoT-based remote online monitoring system for charging piles provided in this embodiment of the invention is used to execute all the process steps of the IoT-based remote online monitoring method for charging piles described in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0084] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a model training program. When the processor executes the computer program, it implements the steps described in the various machine learning-based multi-turn electric actuator fault prediction method embodiments above, for example... Figure 1 The step S101 is shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the data acquisition module.

[0085] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0086] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0087] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0088] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0089] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0090] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0091] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A remote online monitoring method for charging piles based on the Internet of Things, characterized in that, include: Electrical signal data during the operation of the charging pile is collected in real time, and the electrical signal data is preprocessed to obtain a first signal dataset; Time-frequency analysis is performed on the time-domain signals of the first signal dataset, the time-domain signals are converted into time-spectrum graphs, the time-spectrum graphs are summarized, and the first spectrum set is obtained. Obtain a preset health benchmark matrix from a preset health status benchmark spectrum library, calculate the preliminary residual between the first spectrum set and the preset health benchmark matrix, and obtain the first residual distribution; The first residual distribution is adjusted using a preset correction model to obtain a second residual distribution. Specifically, this involves: synchronously reading power load data and temperature sensor sequences during the device's operating cycle; performing gradient calculation on the power load data to extract transient response features; determining the thermal drift coefficient based on the fluctuation slope of the temperature sensor sequence; inputting the transient response features and the thermal drift coefficient into the preset correction model to quantify the frequency drift caused by power switching and the amplitude distortion caused by temperature fluctuations; combining the two quantization results by frequency interval to generate a comprehensive influence vector; each pair of consecutive elements in the comprehensive influence vector corresponds to an interference-related parameter for a frequency interval; summing each pair of interference-related parameters to obtain the comprehensive interference level value for that frequency interval; and normalizing the comprehensive interference level values ​​for the three frequency intervals, with the normalized value serving as the interference weight for the corresponding frequency interval. The interference weight matrix is ​​formed by arranging three interference weights in the order of low frequency, medium frequency, and high frequency. The mean of the first residual distribution is calculated, and the interference weight of each frequency interval in the interference weight matrix is ​​multiplied by the mean of the first residual distribution. The result is the correction model compensation factor for the corresponding frequency interval. The correction model compensation factor is used to perform a point-by-point correction operation on each data point in the first residual distribution. First, the frequency interval to which each data point belongs is determined, and the compensation factor of the corresponding interval is called. The original value of the data point is subtracted from the compensation factor to obtain the adjusted residual value. The adjusted residual value is then statistically binned again. A new probability density curve is generated based on the binning result, and the statistical characteristic parameters of the residual data are updated to obtain the updated statistical characteristic parameters. Based on the updated statistical characteristic parameters, the second residual distribution after removing the influence of dynamic operating conditions and environmental interference is determined. The second residual distribution is compared with a preset threshold range. If the corrected residual value in the second residual distribution exceeds the preset threshold range, a preliminary warning sign is obtained. By combining the abnormal pattern characteristics in historical operation data, the preliminary warning signs are verified a second time to determine the aging status judgment result; Based on the aging status assessment results, a remote monitoring report is generated and pushed to the relevant management platform to complete the real-time early warning and feedback of the aging status of the charging pile; The preset correction model is pre-constructed based on a large amount of historical data. The preset correction model includes the correspondence between different transient response characteristic values, thermal drift coefficients and spectral morphology changes.

2. The method for remote online monitoring of charging piles based on the Internet of Things according to claim 1, characterized in that, The real-time acquisition of electrical signal data during the operation of the charging pile, and the preprocessing of the electrical signal data to obtain a first signal dataset, including: Discretize the electrical signal data of the charging pile during operation to obtain the original sequence containing timing information; The original sequence is filtered and denoised to obtain a clean signal amplitude; The standardized parameters are obtained by calculating the mean shift and variance distribution data of the pure signal amplitude. The amplitude of the pure signal is scaled using the standardized parameters to obtain the first signal dataset.

3. The method for remote online monitoring of charging piles based on the Internet of Things according to claim 1, characterized in that, The step involves performing time-frequency analysis on the time-domain signals of the first signal dataset, converting the time-domain signals into time-spectrum graphs, summarizing the time-spectrum graphs, and integrating them to obtain a first spectrum set, including: Based on the preset truncation length and preset overlap rate, the time-domain signal of the first signal dataset is subjected to frame-slicing operation to obtain multiple local time-domain segments with temporal continuity. The local time-domain segment is windowed and smoothed to obtain a windowed time-domain sequence; The windowed time-domain sequence is subjected to a frame-by-frame Fast Fourier Transform calculation to obtain complex frequency domain response data; The square of the modulus is calculated based on the complex frequency domain response data, and a two-dimensional time-frequency matrix reflecting the frequency evolution over time is constructed. The two-dimensional time-frequency matrix is ​​visualized to form a time-frequency spectrum. All the time-frequency spectrums are then aggregated to obtain the first spectrum set.

4. The method for remote online monitoring of charging piles based on the Internet of Things according to claim 1, characterized in that, The step of obtaining a preset health benchmark matrix from a preset health state benchmark spectrum library, calculating the preliminary residual between the first spectrum set and the preset health benchmark matrix, and obtaining a first residual distribution includes: Convert the time-spectrum graph in the first spectrum set into the current spectrum matrix, and retrieve the preset health benchmark matrix in the preset health state benchmark spectrum library; Perform pointwise difference operations on the current spectrum matrix and the preset health baseline matrix to generate a residual matrix, and count the frequency of occurrence of elements in each numerical interval of the residual matrix; A frequency distribution histogram is constructed based on the frequency of occurrence of the element, and a probability density curve is generated by fitting the frequency distribution histogram. Analyze the probability density characteristics of the probability density curve, and determine the first residual distribution of the degree of deviation of equipment health based on the probability density characteristics.

5. The method for remote online monitoring of charging piles based on the Internet of Things according to claim 1, characterized in that, The step of comparing the second residual distribution with a preset threshold range, and if the corrected residual value in the second residual distribution exceeds the preset threshold range, then a preliminary warning sign is obtained, including: A continuous probability density function is constructed for the second residual distribution using kernel density estimation; Based on the continuous probability density function, a preset threshold interval is defined. The corrected residual value in the second residual distribution is compared with the preset threshold interval. The Manhattan distance of the corrected residual value exceeding the preset threshold interval is calculated to obtain the set of outlier deviations. An aging feature vector is constructed based on the set of outlier deviations, and an aging risk probability value is calculated based on the aging feature vector. If the aging risk probability value exceeds the preset safety margin baseline, a risk assessment report is generated, and the risk level information in the risk assessment report is analyzed to obtain a preliminary warning sign.

6. The method for remote online monitoring of charging piles based on the Internet of Things according to claim 1, characterized in that, The process of combining abnormal pattern characteristics from historical operational data to perform secondary verification on the preliminary warning indicators and determine the aging status judgment result includes: Extract current and voltage time-series waveform data and abnormal pattern characteristics from historical operating data; The current and voltage time-series waveform data are subjected to Fast Fourier Transform to obtain the current frequency domain spectrum data. The main frequency component features of the current frequency domain spectrum data are extracted using principal component analysis. The current abnormal frequency component set is constructed based on the main frequency component features. Combining the current abnormal frequency component set with the abnormal pattern features, a cosine similarity value is calculated, and the historical aging pattern with the highest cosine similarity value is selected as the matching benchmark pattern. If the cosine similarity value is greater than the preset verification confidence threshold, extract the fault type label and aging level information corresponding to the matching benchmark pattern, and generate a secondary verification confirmation signal. Based on the secondary verification confirmation signal, the status attributes of the preliminary warning flag are corrected, and the aging status judgment result is output.

7. A remote online monitoring system for charging piles based on the Internet of Things, characterized in that, For implementing the method as described in any one of claims 1-6, comprising: The signal acquisition and preprocessing module is used to acquire electrical signal data during the operation of the charging pile in real time, and preprocess the electrical signal data to obtain a first signal dataset. The time-frequency analysis module is used to perform time-frequency analysis on the time-domain signal of the first signal dataset, convert the time-domain signal into a time-frequency spectrum, summarize the time-frequency spectrum, and integrate them to obtain a first spectrum set. The residual calculation module is used to obtain a preset health benchmark matrix from a preset health status benchmark spectrum library, calculate the preliminary residual between the first spectrum set and the preset health benchmark matrix, and obtain the first residual distribution. The residual correction module is used to adjust the first residual distribution through a preset correction model to obtain the second residual distribution; The preliminary warning module is used to compare the second residual distribution with a preset threshold range. If the corrected residual value in the second residual distribution exceeds the preset threshold range, a preliminary warning sign is obtained. The secondary verification module is used to perform secondary verification on the preliminary warning sign by combining the abnormal pattern characteristics in historical operation data to determine the aging status judgment result. The early warning and feedback module is used to generate a remote monitoring report based on the aging status judgment results and push it to the relevant management platform to complete the real-time early warning and feedback of the aging status of the charging pile.

Images