A charging pile health diagnosis adaptive optimization method

Abstract

The application discloses a charging pile health diagnosis self-adaptive optimization method, comprising: acquiring charging pile multi-source data and generating a standardized contact resistance sequence; by identifying specific stages of the charging process and using adaptive modeling, the dynamic degradation rate is accurately evaluated; when a common-mode electrical event is detected, the fault source is accurately located by analyzing the morphological differences of the response signals of each charging pile; by fusing technical diagnosis results such as dynamic degradation rate and fault source probability distribution with operation economic data such as real-time queuing vehicle number and time-of-use electricity price, dynamic decision is made by quantifying time-varying operation loss to generate an optimized maintenance strategy considering technical reliability and economic benefit. The application improves the diagnosis accuracy of early faults of the charging pile and realizes intelligent and optimized decision of the maintenance time.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent operation and maintenance of charging piles, and particularly relates to a health diagnosis and adaptive optimization method for charging piles. BACKGROUND

[0002] Stable and efficient operation of charging piles is the basis for the healthy development of new energy vehicle ecology. Among the many components of charging piles, the contact resistance of charging interfaces, connectors and other parts is a core parameter affecting charging safety and efficiency. If the slight deterioration of contact resistance is not discovered in time, it may lead to a decrease in charging power and an increase in energy loss, or even cause the connector to overheat and even melt. Therefore, it is of practical significance and application value to develop an advanced technology method that can accurately diagnose the health status of charging piles, especially the contact resistance and other key components, and can adaptively optimize maintenance combined with actual operating conditions, so as to improve the reliability, safety and economy of charging infrastructure.

[0003] Currently, there are many technical means for health status monitoring and fault diagnosis of power electronic devices such as charging piles. In engineering practice, the most traditional method is preventive maintenance based on a preset time period, that is, maintenance personnel perform inspection, cleaning and component replacement on charging piles at fixed time intervals (such as every six months or a year), which is disconnected from the actual health status of the equipment. With the development of technology, online monitoring methods have begun to be popular, one of the common ways is the alarm mechanism based on fixed threshold. This method monitors key parameters such as the temperature of the charging interface or the total resistance of the loop in real time during the charging process, and when the monitoring value exceeds a pre-set fixed safety threshold, the system triggers an alarm or automatically stops. In further research, some solutions begin to apply data-driven ideas, by collecting historical data of charging pile operation, and using offline statistical analysis tools or conventional machine learning algorithms (such as support vector machines, isolation forests, etc.) to perform batch processing on the data to identify abnormal data points that deviate significantly from the normal state, thereby assisting maintenance personnel in troubleshooting and diagnosis decision-making.

[0004] However, the existing technology still has the core problems of insufficient diagnosis accuracy and low decision-making efficiency in dealing with the complex and variable actual working conditions of charging piles. Specifically, the existing technology mainly faces two deep technical challenges: first, it is difficult to accurately model the weak degradation characteristics under non-stationary and multi-condition; second, there is a lack of adaptive decision-making mechanism combining technical diagnosis with dynamic operation economy. SUMMARY

[0005] The present application belongs to the field of intelligent operation and maintenance of charging piles, and particularly relates to a health diagnosis and adaptive optimization method for charging piles, in order to solve at least one technical problem existing in the prior art.

[0006] The technical scheme is a charging pile health diagnosis adaptive optimization method, comprising the following steps:

[0007] Obtaining original current, original voltage and environmental temperature sequence of the charging pile, performing data preprocessing, and generating a standardized contact resistance sequence;

[0008] Based on the standardized contact resistance sequence and the original current sequence, the dynamic degradation rate is evaluated by identifying the specific stage of the charging process and using adaptive modeling;

[0009] Based on the standardized contact resistance sequence of the predetermined charging pile and the pre-stored charging pile topology connection relationship, when a common mode electrical event is detected, the fault source probability distribution is located by analyzing the morphological difference of the response signal of each charging pile;

[0010] Obtaining real-time queuing vehicle number and time-of-use electricity price, combining the dynamic degradation rate and the fault source probability distribution, and making a dynamic decision by quantifying the operation economic loss to generate a maintenance strategy.

[0011] Beneficial effects, the present application not only improves the diagnosis accuracy of early faults of the charging pile, and combines real-time operation economic data to realize intelligent maintenance timing and optimal decision. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 A step flowchart of a charging pile health diagnosis adaptive optimization method provided by the embodiment of the present application.

[0013] Figure 2 A step flowchart of locating the fault source probability distribution provided by the embodiment of the present application.

[0014] Figure 3 A step flowchart of calculating the morphological difference degree provided by the embodiment of the present application.

[0015] Figure 4 A step flowchart of extracting the response signal sub-sequence provided by the embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to enable personnel in the art to better understand the present application scheme, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0017] It is to be understood that the terms "including", "comprising", "having" and their conjugates mean "including but not limited to", "comprising but not limited to", "having but not limited to", and conjugates thereof respectively.

[0018] It is found in the research that the charging process of the charging pile is a typical non-stationary process in terms of diagnosis accuracy, which will experience constant current, constant voltage, trickle current and other stages with completely different electrical characteristics. The degradation characteristics of the contact resistance also have different forms in these different stages, for example, it may be the increase of high-frequency noise in the constant current stage, and it may be the slow rise of nonlinearity in the constant voltage stage. The fixed threshold alarm or the analysis method based on global characteristics adopted in the prior art essentially flattens the whole complex process. This one-size-fits-all mode is easy to drown out the weak degradation characteristics of the specific stage, leading to missed reports of early failures or misjudgments of fluctuations caused by normal working conditions as failures, causing false alarms, and it is difficult to achieve accurate early warning and remaining life prediction.

[0019] In addition, in terms of decision-making benefits, the prior art generally has the limitation of technology-based, that is, the diagnosis and decision-making process is completely decoupled from the actual operation and economic conditions of the charging station. The ultimate purpose of maintenance decision is to maximize the overall operation benefit. Simply deciding whether to stop maintenance based on the technical health threshold is economically blind. For example, during the peak power consumption period, a charging pile with a full queue of waiting vehicles, even if its contact resistance is slightly increased, but the direct income loss and opportunity cost of user loss caused by immediate shutdown for maintenance may be much higher than the risk of allowing it to continue to serve for a few hours with a slight illness. The prior art cannot quantify this dynamic economic impact, so it cannot answer the key question of when is the best time for maintenance, leading to maintenance behavior being either too conservative and losing revenue or too aggressive and increasing unnecessary costs.

[0020] Therefore, an adaptive optimization method for health diagnosis of a charging pile is proposed. It should be noted that in the implementation of the present application, some basic data can be pre-configured or obtained through a standard interface as the basis for subsequent operation of each embodiment. For example, the charging pile topology connection relationship, which is used as input in the scheme several times, can be configured by technical personnel through a background management system during the construction or reconstruction stage of the charging station, and stored in the server in the form of an adjacency matrix or a graph data structure describing the electrical upstream and downstream relationship between the charging piles. The number of real-time queuing vehicles can be collected in real time by a vehicle detection sensor (such as a ground coil or a video analysis camera) installed at the entrance of the charging station, and the data is uploaded to the server in real time through the Internet gateway.

[0021] AsFigure 1 As shown, a charging pile health diagnosis adaptive optimization method includes the following steps:

[0022] Obtain the original current sequence, original voltage sequence and environmental temperature sequence of the charging pile, perform data preprocessing, and generate a standardized contact resistance sequence.

[0023] Specifically, the original data during operation can be collected or batch read in real time from the controller or background monitoring system of the charging pile. These multi-source data provide a basis for subsequent health state evaluation and fault diagnosis. In this embodiment, in addition to the data reflecting the physical state such as current, voltage and temperature, the operation data including the charging pile topology connection relationship, the real-time number of queued vehicles and the time-of-use electricity price are also obtained, which aims to combine technical diagnosis with economic operation and provide data support for higher-level intelligent decision-making.

[0024] Based on the standardized contact resistance sequence and the original current sequence, the dynamic degradation rate is evaluated by identifying the specific stage of the charging process and using adaptive modeling.

[0025] In this embodiment, when charging a typical electric vehicle battery, the charging pile will go through different stages such as constant current, constant voltage and trickle current, and at different stages, the degradation behavior and signal noise characteristics of the contact resistance are different. Therefore, according to the morphological characteristics of the original current sequence, the specific stage of the charging process is identified, and then according to the signal characteristics of different stages, the feature extraction and modeling strategy suitable for them is adopted. Breakthrough the limitations of traditional fixed window analysis method, can realize the accurate capture of the small change of contact resistance.

[0026] Based on the standardized contact resistance sequence of a predetermined number of charging piles and the pre-stored charging pile topology connection relationship, when a common mode electrical event is detected, the fault source probability distribution is located by analyzing the morphological differences of the response signals of each charging pile.

[0027] It can be understood that when a common mode event occurs in the charging station network, for example, voltage fluctuation on the grid side or the start of a high-power charging pile in the station, theoretically, the morphologies of the electrical response curves of all healthy charging piles should be highly similar. The response curve of the charging pile with excessive contact resistance and other faults will show obvious distortion. This embodiment compares the differences in the response signal morphologies of all charging piles under the same common mode event to identify the abnormal fault source. The complex fault diagnosis problem is transformed into a pattern recognition problem with better robustness.

[0028] Obtain the real-time number of queued vehicles and the time-of-use electricity price, combine the dynamic degradation rate and the fault source probability distribution, make a dynamic decision by quantifying the operation economic loss, and generate a maintenance strategy.

[0029] In this embodiment, the goal of diagnosis is not only to find out the fault, but also to determine the optimal maintenance timing. The pure technical level diagnosis results obtained by the foregoing steps (such as how fast the degradation is, which pile is the fault source) are deeply fused with the economic indicators of real-time operation (such as whether there are many vehicles in the queue at present, whether the electricity price is high or not). By dynamically calculating the economic impact that different decisions (immediate repair, delayed repair, etc.) may bring, an optimized maintenance strategy that can maximize the operation benefit and minimize the comprehensive loss is finally generated.

[0030] According to one aspect of the present application, data preprocessing is performed to generate a standardized contact resistance sequence, including:

[0031] Based on the original current sequence and the original voltage sequence, the original contact resistance sequence is calculated through Ohm's law.

[0032] In this embodiment, the original current sequence, the original voltage sequence and the ambient temperature sequence are read from the local controller of each charging pile or the cloud monitoring platform. Since there may be deviations in the sampling frequency and time reference of different data sources, time stamp alignment and resampling processing (for example, linear interpolation or nearest neighbor interpolation) are needed to obtain a synchronous sampling data set with a unified time axis. For each time point in the synchronous sampling data set, the Ohm's law R = U / I is applied to calculate the instantaneous resistance value. It should be noted that the calculated value is the total resistance of the entire charging loop including the charging gun, cable, connector, etc. In order to more accurately reflect the health status of the contact point, the line inherent resistance (which can be obtained by factory calibration or offline measurement) can be subtracted to obtain the original contact resistance sequence.

[0033] An algorithm based on local outlier factor is used to identify and repair measurement outliers in the original contact resistance sequence to obtain a cleaned contact resistance sequence.

[0034] In actual operation, the original contact resistance sequence may contain measurement outliers caused by sensor instantaneous failure, electromagnetic interference or communication packet loss, etc. In this embodiment, an anomaly detection algorithm based on local outlier factor (LOF) is used to identify these abnormal points. This algorithm compares the density of each data point with its neighborhood points to determine whether it is an outlier, and is suitable for identifying non-Gaussian distribution anomalies. After identifying the abnormal points, a spline interpolation method is used to repair them to ensure the continuity and smoothness of the signal, and finally a cleaned contact resistance sequence is obtained.

[0035] A temperature-resistance compensation model is applied to eliminate the influence of the ambient temperature sequence on the cleaned contact resistance sequence to generate a standardized contact resistance sequence.

[0036] Specifically, the contact resistance of the charging connector changes with temperature, which is not caused by device degradation, and if not eliminated, it will interfere with the subsequent degradation trend analysis. To this end, a temperature-resistance compensation model is applied to process the contact resistance sequence after cleaning. For example, a linear compensation model can be used: R std (t) = R clean (t) - β · (T env (t) - T ref ); where R clean (t) is the contact resistance sequence after cleaning, R std (t) is the standardized contact resistance sequence obtained after compensation; β is the resistance temperature coefficient of the contact point material, which is a constant, and its value can be obtained by experiment calibration or consulting material manual; T env (t) is the ambient temperature collected synchronously with the resistance value; T ref is a preset reference temperature, for example, 25 degrees Celsius. Effectively eliminate the impact of temperature fluctuations on contact resistance, so that the final standardized contact resistance sequence can more truly reflect the long-term degradation trend caused by physical wear, oxidation and other factors.

[0037] According to an aspect of the present application, the dynamic degradation rate is evaluated, comprising:

[0038] Based on the morphological changes of the original current sequence, the constant current, constant voltage and trickle charging stages in the charging process are identified, and a charging stage label sequence is generated.

[0039] In the present embodiment, by analyzing the original current sequence, for example, calculating its first derivative and variance in a sliding window, the current mutation points and stable segments can be identified. Accordingly, the entire charging process is divided into different stages, such as: constant current charging stage where the current value remains basically unchanged; constant voltage charging stage where the current value presents a nonlinear downward trend; and trickle charging stage at the end of the charging process where the current value is maintained at a very low level. The phase type corresponding to each time point is recorded to form a charging stage label sequence.

[0040] Optionally, the charging stage label sequence is generated, specifically comprising:

[0041] Read the original current sequence, calculate the current rate of change, current variance and current spectral features through the sliding window to obtain a multi-dimensional charging feature matrix. The matrix preferably contains the following features: current first derivative dI / dt: this feature is used to capture the rising, falling or stable trend of the current, and is a direct indicator to distinguish the constant current and constant voltage phases. Current short-time variance: this feature is used to quantify the degree of fluctuation of the current signal in a short time. In the constant current or trickle current phase, a healthy charging process should have very low variance, and a significant increase in variance may indicate unstable connection or mode switching. Current main frequency component: by performing short-time Fourier transform (STFT) on the current signal, the main frequency component is extracted, which helps to identify specific noise patterns caused by the switching frequency of the power electronic converter itself or external interference. Combine these features into a vector to form a row in the multi-dimensional charging feature matrix.

[0042] Read the multi-dimensional charging feature matrix, identify the mutation points in the feature space by calculating the Mahalanobis distance between adjacent time window feature vectors, obtain the candidate change point set and change point significance score. Read the multi-dimensional charging feature matrix and the candidate change point set, and identify the constant current, constant voltage, trickle current and other charging phases by dynamic time warping matching with the pre-stored charging mode template library, to obtain the preliminary phase label. It should be noted that the charging mode template library can be constructed in the following way: theoretical templates can be generated based on the definitions of constant current, constant voltage and other phases in the general charging protocol (such as CHAdeMO, CCS, GB / T); more preferably, a large number of real historical charging process data can be collected, the multi-dimensional charging feature matrix of each process can be extracted, and unsupervised clustering algorithms (such as K-Means or DBSCAN) can be used to automatically learn and generate feature templates that can represent typical charging modes from the data. The library can be updated regularly online to adapt to the changes in charging behavior caused by different vehicle models and battery aging. Read the preliminary phase label, change point significance score and original current sequence, and accurately locate the phase switching time by a local search algorithm, considering the current continuity constraint, to obtain the charging phase label sequence and phase switching time set.

[0043] According to the charging phase label sequence, segment the normalized contact resistance sequence, and use the pre-set phase-specific feature extractor for different charging phase segments to extract the phase-specific degradation feature vector.

[0044] In further embodiments, this step can specifically include at least one of the following:

[0045] For the segment of the constant current charging phase, the micro-fluctuation components of a specific frequency band in the normalized contact resistance sequence are extracted by wavelet decomposition to obtain the constant current phase degradation feature.

[0046] Specifically, since in the constant current stage, the ideal contact resistance should remain stable, its slight fluctuation often contains information of early degradation. In this embodiment, the normalized contact resistance sequence in this stage is decomposed by wavelet decomposition, for example, using db4 wavelet basis, and the micro-fluctuation component in the 3-5Hz frequency band is analyzed. By calculating the rising slope of the envelope of this fluctuation component, the degradation trend of the contact point in this stage can be quantified, and the constant current stage degradation feature is obtained.

[0047] For the segment of the constant voltage charging stage, the nonlinear rising trend of the normalized contact resistance sequence is captured by piecewise polynomial fitting, and the constant voltage stage degradation feature is obtained.

[0048] Specifically, in the constant voltage stage, as the battery resistance changes, the current decreases, and the contact resistance usually shows a nonlinear rise due to heating and electrochemical effects. A third-order polynomial is used to fit the normalized contact resistance sequence in this segment, and the curvature change rate of the fitted curve is extracted as a degradation indicator, which reflects the acceleration degree of resistance rise, and the constant voltage stage degradation feature is obtained.

[0049] Further, in some preferred embodiments, the following can also be included:

[0050] For the segment of the trickle charging stage, the trickle stage degradation feature is obtained by calculating its stability index. In the trickle stage, the resistance value of a healthy contact point should be highly stable. The long-range correlation of the normalized contact resistance sequence in this segment is evaluated by calculating its local Hurst Exponent. A Hurst Exponent close to 0.5 indicates that the randomness of the signal is enhanced and the stability is decreased, which can be used as a sign of contact point degradation, and the trickle stage degradation feature is obtained.

[0051] The degradation features of the above multiple stages are fused, for example, by a time-varying weighting mechanism to smoothly transition at stage switching, and finally a continuous stage-specific degradation feature vector is obtained.

[0052] Based on the stage-specific degradation feature vector, a non-stationary degradation process model is constructed, and the dynamic degradation rate is calculated from the non-stationary degradation process model.

[0053] In further embodiments, constructing a non-stationary degradation process model and calculating a dynamic degradation rate includes:

[0054] Using empirical mode decomposition, the stage-specific degradation feature vector is processed to separate the degradation trend component reflecting the long-term degradation law.

[0055] Specifically, the stage-specific degradation feature vector is decomposed into several intrinsic mode functions (IMFs) and a residual component by empirical mode decomposition (EMD). The residual component is usually monotonic and represents the long-term trend of the signal, which is defined as the degradation trend component.

[0056] Based on the degradation trend component, an optimal basis function is adaptively selected to construct a degradation model.

[0057] In an optional embodiment, the degradation trend component is segmented according to the charging stage label sequence. For each segment, an optimal basis function is automatically selected from a predetermined set of candidate basis functions by Akaike Information Criterion (AIC), and a segmented basis function parameter set is obtained. For example, the set of candidate basis functions can include linear functions, exponential functions, and power-law functions. The fitting effect of each candidate function on each segment is evaluated by AIC, and the function with the smallest AIC value is selected as the optimal basis function for the segment.

[0058] A nonlinear optimization method with continuity constraints is used to adjust the segmented basis function parameter set so that the functions of adjacent segments are continuous at the switching points, and a degradation model is constructed. Specifically, it is required that the function values of the two adjacent functions are equal at the switching points, and the first-order derivatives are equal, so as to construct a globally smooth degradation model.

[0059] The degradation model is differentiated (e.g., first-order derivative) to obtain a dynamic degradation rate. Further, by extrapolating the constructed degradation model to the future until it reaches a preset failure threshold (e.g., the contact resistance reaches 150% of the initial value), the remaining useful life prediction value of the charging pile can be predicted.

[0060] The embodiment solves the problem of insufficient diagnostic accuracy caused by the inability to process non-stationary signals by first identifying different stages such as constant current, constant voltage, and trickle current during the charging process, and then using stage-specific feature extractors that match the signal characteristics of each stage (such as micro-fluctuation analysis for the constant current stage and non-linear trend modeling for the constant voltage stage) to build a refined adaptive degradation model. Specifically, the physical degradation process of the contact resistance will exhibit different characteristics under different electrical stresses. The embodiment divides the entire degradation process and accurately captures early signs of degradation corresponding to specific working conditions from the signal details that are averaged out by traditional methods. In the charging pile diagnosis scenario, it means that the system can detect the slight degradation of the contact resistance earlier than traditional methods, such as capturing high-frequency fluctuations caused by small arcs in the constant current stage, which indicate poor contact, giving more accurate health status assessment and remaining useful life (RUL) prediction, providing a solid data foundation for achieving true predictive maintenance.

[0061] As shown in Figure 2 According to one aspect of the present application, the probability distribution of locating the fault source includes:

[0062] For the time window of the common-mode electrical event, the response signal sub-sequence of each charging pile is extracted from the normalized contact resistance sequence.

[0063] In the embodiment, the common-mode electrical event is detected and the response signal sub-sequence is extracted. Specifically, the common-mode electrical event can be selected from at least one of the following group: a charging start event in which the charging power of any charging pile in the charging station is greater than a threshold value; an event in which any charging pile in the charging station switches from the constant current charging phase to the constant voltage charging phase; a voltage sag or voltage flicker event caused by the external power grid and affecting at least two charging piles. Specifically, these events can be detected by the power quality monitoring unit deployed on the common AC bus or DC bus of the charging station: for example, when the bus voltage drop is monitored to exceed the preset threshold value (such as 10% of the rated voltage) and the duration is between 20ms and 2s, it can be determined that a voltage sag event has occurred; when the starting current change rate (di / dt) of any charging pile exceeds the preset threshold value (for example, 500A / s), it can be determined that a high-power charging start event has occurred. When such an event is detected, a time window is determined.

[0064] As shown in Figure 4 Preferably, the step of extracting the response signal sub-sequence includes:

[0065] The multi-scale differential processing is performed on the normalized contact resistance sequence in the time window to enhance the transient change characteristics caused by the common-mode electrical event, and a multi-scale differential sequence is obtained. For example, the first-order, second-order and third-order differential operations can be performed simultaneously to capture the change information of different time scales.

[0066] The instantaneous features (such as instantaneous amplitude and instantaneous phase) of the multi-scale differential sequence are extracted by using the Hilbert transform, and the instantaneous features are defined as the response signal sub-sequences for morphological comparison. The advantage of this is that the instantaneous features can highlight the dynamic response characteristics caused by the event more than the original signal.

[0067] The abnormality degree of the response behavior is quantified by calculating the morphological difference degree between the response signal sub-sequence of any charging pile and the response signal sub-sequences of the remaining charging piles.

[0068] As shown in Figure 3 , in the optional embodiment, the morphological difference degree is calculated, including:

[0069] The dynamic time warping algorithm is used to align the time axis of the response signal sub-sequences of the predetermined charging piles, to generate the aligned response signal sub-sequences; the distance between the aligned response signal sub-sequence of any charging pile and the aligned response signal sub-sequences of the remaining charging piles is calculated, and the distance is used to determine the morphological difference degree.

[0070] Specifically, the dynamic time warping (DTW) algorithm can find the optimal matching path between two time sequences, even if they have non-linear distortion in the time axis. After alignment, the average Euclidean distance between the response signal sub-sequence of any charging pile i and the response signal sub-sequences of all other charging piles can be calculated, and the average distance is defined as the morphological difference degree Di of the charging pile i.

[0071] The charging pile with the largest morphological difference degree is identified as the potential fault source, and the fault source probability distribution is generated based on the relative size of the morphological difference degree.

[0072] In this embodiment, the larger the morphological difference degree Di of a charging pile, the more abnormal its response behavior is, and the higher the possibility of being a fault source is. Preferably, in order to convert a group of difference degrees D1, D2,..., DN (N is the total number of charging piles) of relative sizes into a normalized probability distribution, the Softmax function can be used for processing: P source，i = (e Di / τ ) / ∑ j＝1 N e Dj / τ ; where P source，i is the probability of the i-th charging pile being judged as a fault source, which constitutes the fault source probability distribution P sourceτ is a temperature parameter, used to adjust the sharpness of the probability distribution, the smaller τ is, the more concentrated the probability is on the charging pile with the largest difference.

[0073] The embodiment improves the reliability and practicability of fault location by using a common-mode electrical event (such as a high-power charging start) as a natural and full-station synchronous excitation source, and converting the fault location problem from a time delay measurement problem with high requirements for hardware and synchronization into a pattern recognition problem that is more robust to noise and jitter. Specifically, when a common-mode event occurs, all charging piles will produce an electrical response, but the response curve of the pile with a fault such as high contact resistance will be distorted due to abnormal impedance. By using algorithms such as dynamic time warping (DTW), the response curve patterns of all piles can be quantitatively compared to accurately identify the abnormal fault source. In the charging station scenario, where the electromagnetic environment is complex and the cost of precise time synchronization is high, without additional precise clock synchronization equipment, high-precision fault source location can be achieved using only existing sensor data, reducing the implementation complexity and deployment cost of the system. At the same time, its principle of shape comparison makes it insensitive to random noise in the data, and the diagnosis result is more stable and reliable.

[0074] Further, after generating the fault source probability distribution, a fault authenticity discrimination step is further included, specifically:

[0075] Based on the potential fault source determined by the fault source probability distribution, the charging pile topology connection relationship, and the response signal sub-sequence of each charging pile, the amplitude attenuation law of the response signal with the propagation path is analyzed, and the disturbance attenuation feature is extracted; specifically, the attenuation rate can be quantified by calculating the ratio of the signal amplitude to the propagation electrical distance (which can be obtained from the topology relationship G);

[0076] The disturbance attenuation feature and the dynamic degradation rate of the potential fault source itself are fused to determine the fault type of the potential fault source, and a fault type label that distinguishes between real faults and affected states is generated.

[0077] Specifically, the system identifies the charging pile with the highest probability in the fault source probability distribution as the potential fault source. Based on the charging pile topology connection relationship, the disturbance signal emitted by the potential fault source is analyzed, and how its signal amplitude attenuates during propagation to other charging piles. If the signal attenuates slowly and the dynamic degradation rate of the potential fault source itself is also high, it is labeled as a real fault. Conversely, if the signal attenuates quickly or its own degradation rate is low, it may only be disturbed by other real fault sources and is labeled as an affected state. Finally, a fault type label is generated.

[0078] According to another aspect of the present application, locating the fault source probability distribution can also use a disturbance propagation time delay analysis method, specifically:

[0079] extracting a disturbance signal sequence caused by a common-mode electrical event from the standardized contact resistance sequences.

[0080] Specifically, a common-mode electrical event needs to be identified. After the event occurs, the standardized contact resistance sequences of each charging pile are processed, for example, high-pass filtering or band-pass filtering is used to filter out the slowly changing degradation trend, and only the high-frequency transient disturbance caused by the event is retained, to obtain the disturbance signal sequence.

[0081] Optionally, the step of extracting the disturbance signal sequence comprises:

[0082] Reading the standardized contact resistance sequences of the plurality of charging piles, extracting change information of different time scales through 1st, 2nd and 3rd order difference operations to obtain a multi-scale difference matrix.

[0083] Reading the multi-scale difference matrix, calculating local statistical features through a sliding window, setting a dynamic threshold using an improved version of the 3σ principle (considering non-Gaussian distribution), identifying disturbance events that exceed the normal fluctuation range, and obtaining disturbance event markers.

[0084] Specifically, in order to accurately detect disturbance events from the multi-scale difference matrix, an adaptive threshold method based on Box Plot statistics can be used as an improvement over the traditional 3σ principle to better adapt to non-Gaussian distributed electrical noise data. The specific algorithm is as follows: In a sliding window, calculate the first quartile (Q1) and the third quartile (Q3) of the difference sequence, and calculate the interquartile range (IQR = Q3-Q1). The abnormal judgment threshold of this window is set as: upper limit Th upper = Q3+k*IQR and lower limit Th lower = Q1-k*IQR; where the coefficient k is an adjustable parameter, usually taking a value of 1.5 (corresponding to regular outliers) or 3.0 (corresponding to extreme outliers). Any data point that exceeds this upper and lower limit range is marked as a valid disturbance event. This embodiment does not rely on the assumption that the data is Gaussian distributed, and has stronger robustness.

[0085] Reading the disturbance event markers and the standardized contact resistance sequences, extracting the instantaneous amplitude and instantaneous phase of the disturbance signal through Hilbert transform, constructing the disturbance fingerprint feature, obtaining the disturbance signal sequence and the disturbance occurrence time.

[0086] Based on the disturbance signal sequence, a disturbance propagation time delay matrix is calculated between different charging piles.

[0087] Optionally, the step of calculating the disturbance propagation time delay matrix comprises:

[0088] The generalized cross-correlation method is used to process the disturbance signal sequence to obtain a preliminary time delay estimation matrix.

[0089] Specifically, for the disturbance signal sequences S dist，i (t) and S dist，j (t), the generalized cross-correlation with phase transform (GCC-PHAT) method is used to calculate the time delay therebetween. The GCC-PHAT method can obtain a sharp correlation peak in a low signal-to-noise ratio environment by whitening the cross-power spectrum of the signals, thereby obtaining a high-precision preliminary time delay estimation value. This calculation is performed between all charging piles two by two, and a preliminary time delay estimation matrix is obtained.

[0090] Preferably, the generalized cross-correlation with phase transform (GCC-PHAT) method is used. Specifically, this method whitens or normalizes the amplitude spectrum of the cross-power spectrum of two signals in the frequency domain, thereby retaining only the phase information for correlation calculation. In the charging station diagnosis scenario of the present application, since there is a large amount of colored noise with relatively fixed frequency generated by power electronic switching devices, the correlation peak calculated by the traditional cross-correlation method is easily broadened and covered by these strong noise components, resulting in a decrease in the accuracy and reliability of time delay estimation. The GCC-PHAT method greatly suppresses the influence of signal amplitude on the calculation result through phase transformation, so that the cross-correlation function in the time domain tends to be a sharp impulse function (Dirac delta function) at the true time delay point. This characteristic makes the time delay estimation highly robust to amplitude changes and colored noise, and even in the case of low signal-to-noise ratio, a preliminary time delay estimation value without ambiguity and with high resolution can be obtained, providing a solid and reliable data foundation for subsequent accurate fault source reverse tracking through the time delay equation set.

[0091] Based on real-time load data, the preliminary time delay estimation matrix is corrected by constructing an electrical distance-physical time delay mapping model to obtain a disturbance propagation time delay matrix.

[0092] It can be understood that the propagation speed of the disturbance signal in the cable is not constant, and it will be affected by the size of the real-time load (i.e. charging current) on the line. For this purpose, the embodiment constructs an electrical distance-physical time delay mapping model in advance through circuit simulation or experimental measurement. It should be noted that the mapping model can be modeled by offline circuit simulation software (such as SPICE), input different load and line parameters, and obtain the mapping relationship between time delay and load; or it can also be established by actively injecting test signals for on-site calibration during the on-site debugging stage after the deployment of the charging station. The model takes the electrical distance between the charging piles (which can be obtained from the charging pile topology connection relationship G) and the real-time load data as input, and outputs a time delay correction coefficient. The preliminary time delay estimation matrix is corrected using the coefficient to compensate for the change in propagation speed caused by the change in load, and a more accurate disturbance propagation time delay matrix is finally obtained.

[0093] According to the disturbance propagation time delay matrix and the charging pile topology connection relationship, a time delay equation set is constructed and solved to trace back and determine the fault source probability distribution.

[0094] In the embodiment, the time required for the disturbance generated by the fault source to propagate to each charging pile is proportional to the length of the propagation path between them. Assuming that the position of the fault source is (x s , y s ) and the occurrence time is t s . For any charging pile i with position (x i , y i ), the time when it receives the disturbance is t i . Then, for any two charging piles i and j, the time difference t i -t j (the value can be obtained from the disturbance propagation time delay matrix) should be equal to the difference in time required for the disturbance to propagate from the source to them respectively. This can construct an over-determined linear equation set about the unknown quantities (x s , y s , t s ). By using a robust solving algorithm such as iterative reweighted least squares (IRLS), the equation set can be solved to backtrace the position of the fault source. Finally, based on the stability of the positioning result evaluated by the residual of the solving result or by the Bootstrap resampling method, the fault source probability distribution can be generated. It does not rely on the analysis of signal morphology, but traces back by accurately measuring the time delay of the fault disturbance propagating in the charging pile network.

[0095] The embodiment provides an alternative method capable of directly physically positioning the fault source by accurately measuring the propagation time difference of the disturbance signal caused by the common mode event between different charging piles and constructing a time delay equation set for solving, and realizes high precision and positioning certainty. Specifically, by adopting a generalized cross-correlation with phase transform (GCC-PHAT) method, a sharp correlation peak value can be extracted from a strong noise background, and a high-reliability preliminary time delay estimation is obtained. Further, the propagation speed of the signal in the power cable is affected by the real-time load (charging current), and the preliminary time delay is corrected by constructing an electrical distance-physical time delay mapping model. The timing error caused by the load change is eliminated, and the accuracy of the final positioning result is improved. In the charging station scene, not only the specific fault pile can be positioned, but theoretically the fault point on the cable between two charging piles can also be positioned, which provides more accurate fault position guidance for the operation and maintenance personnel and shortens the fault troubleshooting time.

[0096] According to an aspect of the present application, a maintenance strategy is generated, comprising:

[0097] The dynamic degradation rate, the fault source probability distribution, the real-time queuing vehicle number and the time-of-use electricity price are comprehensively considered to quantify the expected operation economic loss caused by the potential fault, and a time-varying loss function is constructed.

[0098] In this embodiment, the time-varying loss function L(t) is not a fixed formula, but is dynamically calculated according to the real-time input. Preferably, the construction of L(t) takes into account the following aspects: direct income loss: based on the dynamic degradation rate and the remaining service life, the probability of potential failure is predicted, and combined with the probability distribution of the failure source, the expected probability of failure of each pile in the future period of time is calculated. The probability is multiplied by the number of service vehicles lost due to downtime (which can be estimated by the real-time queue vehicle number Q(t)) and the unit service income determined by the time-of-use electricity price to obtain the expected direct income loss. Opportunity cost of user loss: based on historical operation data, a relationship model (for example, a survival analysis model) between user waiting time and loss probability is established. When a charging pile has a risk of failure, the potential downtime increases the waiting time of other vehicles, resulting in user loss. This part of the loss is the loss of future income that these lost users may bring. Cascading failure risk loss: further, considering that the failure of a single charging pile may cause its load to be transferred to adjacent charging piles (according to the charging pile topology connection relationship G), thereby increasing the overload risk and failure probability of adjacent piles, triggering cascading failure. The probability of potential failure indicated by the failure source probability distribution and the possible loss of a larger range of downtime caused by cascading failure can be evaluated by Monte Carlo simulation. Maintenance cost: including planned maintenance and emergency maintenance cost, the latter is usually much higher than the former. The weighted sum of the above losses can construct a quantitative, time-varying loss function L(t) that changes dynamically over time.

[0099] In addition, when constructing the time-varying loss function L(t), the additional risk brought by decision delay needs to be quantified. Decision delay refers to the time difference between the generation of maintenance suggestions by the system and the actual execution of maintenance operations by the operation and maintenance personnel, for example, the system suggests maintenance at 10 am, but the maintenance window may not be available until 10 pm, there is a 12-hour delay. In this embodiment, this risk is reflected by adding a delay risk term to the loss function. The calculation method of this term is: using the dynamic degradation rate and failure model to calculate the cumulative probability of failure within the time window from the initial time t0 to t0+T delay (T delay , the delay duration). Multiply the cumulative probability by the maximum economic loss caused by the failure (such as emergency maintenance cost and loss of all queuing users), to get the expected value of the delay risk. Add the expected value to the time-varying loss function L(t), so that the system can take into account the worst consequences that may be caused by the execution delay when optimizing the decision threshold, and make safer and more forward-looking decisions.

[0100] Based on the time-varying loss function, the adaptive decision threshold is optimized to generate an optimized maintenance strategy by comparing the adaptive decision threshold with the current health status of the charging pile.

[0101] In an optional embodiment, the step of optimizing the adaptive decision threshold comprises:

[0102] Based on the historical diagnosis records, a time-varying Receiver Operating Characteristic (ROC) curve is constructed, and a set of time-varying ROC parameters is obtained.

[0103] Specifically, by employing a sliding time window on the historical data, the True Positive Rate (TPR) and False Positive Rate (FPR) under different diagnosis thresholds can be continuously calculated, and a series of ROC curves evolving over time can be constructed. This constitutes a time-varying decision space, reflecting the performance of the diagnosis system at different periods.

[0104] The time-varying loss function is mapped into the decision space defined by the set of time-varying ROC parameters, and by performing multi-objective Pareto front optimization between minimizing economic loss and maximizing diagnosis stability, the optimal working point is determined, and the adaptive decision threshold is generated.

[0105] Specifically, the system needs to optimize two mutually exclusive objectives simultaneously: minimize the expected economic loss defined by the time-varying loss function L(t); maximize the stability of the diagnosis, for example, avoid frequent fluctuations in the threshold value leading to repeated changes in maintenance recommendations within a short period of time. In this embodiment, a multi-objective optimization algorithm such as NSGA-II (Non-dominated Sorting Genetic Algorithm II) is used to find a set of Pareto optimal solutions in this decision space. Each solution represents a different trade-off between economic loss and diagnosis stability. The system can select an optimal working point from this Pareto front according to the pre-set operating preferences (e.g., conservative, aggressive). This working point uniquely corresponds to a diagnosis threshold, which is the adaptive decision threshold.

[0106] The current health status of the charging pile (e.g., a health index calculated from the standardized contact resistance sequence and dynamic degradation rate) is compared with the adaptive decision threshold. If the health index is below the threshold, a maintenance recommendation is triggered, and an optimized maintenance strategy is generated, for example, recommending maintenance within 4 hours after the peak period ends.

[0107] The embodiment couples the device health management and business operation management by constructing a time-varying loss function that combines technical diagnosis results (such as dynamic degradation rate and fault source probability) and real-time operation economic data (such as the number of queued vehicles and time-of-use electricity price), and performing multi-objective Pareto frontier optimization based on the time-varying loss function to generate adaptive decision thresholds. The maintenance decision is no longer a response to a fixed technical indicator, but serves a dynamic strategy that aims to maximize overall operation efficiency. In the charging station operation scenario, it means that the system can intelligently determine that for a charging pile with a slight fault, in the business peak period with many queued vehicles and high electricity price, the value created by its continued service is much greater than its potential risk, so the system will adaptively increase the tolerance (increase the decision threshold) to avoid unnecessary downtime loss; conversely, in the night idle period, the system will reduce the tolerance to take advantage of the low opportunity cost window period and actively recommend preventive maintenance. The problem of economically blind maintenance decision is solved, the leap from can repair to will repair is realized, and the asset operation efficiency and intelligent level of the charging station are improved.

[0108] According to another aspect of the application, the process of obtaining the time-varying loss function can further include:

[0109] Reading real-time queued vehicle number, time-of-use electricity price and historical charging records, predicting the charging demand distribution in the next 24 hours through a SARIMA model, considering the influence of working days / holidays, weather and other factors, obtaining a charging demand prediction sequence;

[0110] Reading fault type labels, charging demand prediction sequence and user behavior history data, estimating user churn probability under different waiting times through a survival analysis model, considering the difference between user types (fast charging / slow charging), obtaining a user churn probability curve;

[0111] Reading fault source probability distribution, charging pile topology connection relationship and dynamic degradation rate, evaluating the probability and influence range of cascading faults caused by single-point faults through Monte Carlo simulation, calculating the loss in the worst case, obtaining a cascading loss risk value;

[0112] Reading the user churn probability curve, the cascading loss risk value, the expected income curve and the maintenance cost data, constructing a comprehensive loss function including direct loss, indirect loss and opportunity cost through weighted summation, obtaining a time-varying loss function

[0113] Further, the construction of the time-varying loss function L(t) involves the following modules:

[0114] Charging demand forecasting module: In order to scientifically calculate the expected revenue loss, the future charging demand needs to be forecasted. In this embodiment, a SARIMA (Seasonal AutoRegressive Integrated Moving Average) model is preferred. The input of the model is the historical hourly total charging amount time series of the charging station, and exogenous variables such as whether it is a weekday / holiday, weather conditions, oil prices, etc. can be introduced. By analyzing the historical time series, for example, observing its autocorrelation function (ACF) and partial autocorrelation function (PACF) plots, the parameters (p, d, q) of the model and the seasonal parameters (P, D, Q, s) are determined, where the seasonal period s is usually 24 (hours) or 168 (weeks), p is the autoregressive order, d is the difference order, q is the moving average order, P is the seasonal autoregressive order, D is the seasonal difference order, and Q is the seasonal moving average order. The trained SARIMA model can predict the charging demand for a future period of time (e.g. 24 hours), obtaining a charging demand prediction sequence, which is the basis for subsequent calculation of the expected revenue loss.

[0115] User churn probability dynamic modeling module: In order to quantify the opportunity cost caused by the prolonged waiting time of users due to potential faults, the probability of user churn needs to be modeled. In this embodiment, a survival analysis model such as the Cox proportional hazards model is preferred. User abandonment of the queue is regarded as a death event, and the duration from the user starting to queue to the final charging or abandonment is regarded as the survival time. The covariates of the model can include: the current queue length, user type (such as member / non-member), user historical charging frequency, current time period, etc. By fitting the historical operation data, the hazard ratio (Hazard Ratio) of each covariate can be obtained. In real-time decision-making, the current operating conditions (such as queue length) are substituted into the trained Cox model, and the user churn probability curve under different waiting times can be dynamically calculated. The curve shows how the probability of user abandonment of the queue rises as the expected waiting time caused by potential faults increases.

[0116] Cascade failure risk assessment module: In order to assess the risk that a failure of a single charging pile can trigger a larger scale loss, a cascade failure needs to be assessed. In this embodiment, Monte Carlo simulation method is preferred. The simulation procedure is as follows: based on the probability distribution of failure sources and their failure probabilities, at the beginning of a simulation iteration, first randomly determine whether the potential failure source fails. If it is determined that it fails, according to the charging pile topology connection relationship G, the original charging load of it is transferred to its adjacent charging piles that are still working normally according to a preset strategy (such as average allocation or capacity allocation). For the adjacent piles that receive additional load, their own failure probability will rise as a result. The magnitude of the rise can be calculated according to the overload degree and the health model of itself. The system continues to determine whether these piles that receive additional load also fail due to the rise in failure rate. If so, the next round of load redistribution and risk transmission is continued. After one simulation iteration, the total economic loss caused by the cascade failure in this simulation is counted. The above steps are repeated thousands of times (for example, 10000 times). Finally, the average value of the total loss of all simulation iterations is taken to obtain a quantitative cascade loss risk value. The risk value will be an important part of the time-varying loss function L(t) for decision-making.

[0117] In further embodiments, when performing the fault authenticity discrimination, a fault type determination can also be performed based on Bayesian inference, specifically: a Bayesian network can be constructed to fuse multi-source evidence and perform probabilistic inference on the true type of the potential fault source. The structure of the network can include: Evidence Nodes: including variables describing the state of the potential fault source from various dimensions, for example: disturbance decay characteristics (whose state can be divided into fast, medium, and slow), self-dynamic degradation rate (whose state can be divided into high, medium, and low), and location confidence (if the time delay method is used for positioning, its state can be divided into high and low). Target Node: the fault type Type, whose state can be divided into real fault, affected state, and normal. Conditional probability table (CPTs): the dependency relationship between nodes in the network is defined by the conditional probability table. These probabilities can be set based on the experience knowledge of domain experts, or obtained by statistical learning on a large amount of historical fault data. For example, the probability value of P (fault type = real fault | disturbance decay = slow, degradation rate = high) will be very high. When the system completes the preliminary diagnosis of a potential fault source, it will input the obtained evidence (such as its decay characteristic is slow and degradation rate is high) into the Bayesian network. The network then uses a standard inference algorithm (such as the joint tree algorithm) to calculate the posterior probability of each state of the fault type node under the current evidence. The state with the highest posterior probability is selected as the final fault type label of the charging pile. Using Bayesian network for inference can better handle the uncertainty of information compared to simple rule judgment, and provide a more reliable diagnosis conclusion with probabilistic significance.

[0118] In further embodiments, in a charging station with complex topology, when a disturbance signal generated by a fault source propagates from the source to a measurement point (another charging pile), there can be multiple electrical paths. This can result in multiple correlation peaks in the calculation results when using the generalized cross-correlation method, each of which corresponds to a possible propagation time delay, resulting in ambiguity in time delay estimation. For this purpose, a weighted least squares method can be used to fuse multiple possible time delay estimates. Specifically, the cross-correlation calculation results between any two charging piles are peak detected, and the time delay estimation values corresponding to all significant correlation peaks are extracted. Each extracted time delay estimation value is assigned a weight. The weight can be determined based on the physical properties of the peak, such as the amplitude, signal-to-noise ratio, or sharpness of the peak. Generally, a stronger and sharper peak corresponds to a more reliable time delay estimate and should be given a higher weight. Using the weighted least squares method, all weighted time delay estimation values are fitted to obtain a unique, optimal fused time delay value. This fused value takes into account information from all possible paths, and its result is more robust and accurate than any single estimate. This fused value is used as the final element in the disturbance propagation time delay matrix.

[0119] In further embodiments, the adaptive decision threshold directly generated by the Pareto frontier optimization can jump frequently and unnecessarily due to transient noise or small fluctuations in real-time queue vehicle number Q(t) and other operational data. This can result in the system repeatedly giving contradictory instructions to repair and continue observation in a short period of time, reducing the usability of the decision. To solve this problem, a Kalman filter can be introduced to smooth the original decision threshold after it is generated. Specifically, a state space model is constructed to describe the variation of the decision threshold. The state of the system is defined as the true, smoothly changing decision threshold, and the original adaptive decision threshold sequence output by the Pareto optimization is considered as noisy observations of this true state. The Kalman filter processes the threshold sequence through its classic prediction-update cycle. At each time step, the filter first predicts the true threshold at the current time based on the state at the previous time, and then corrects this prediction using the current observed threshold to obtain the optimal posterior estimate. After Kalman filtering, the smoothed threshold sequence is output. This sequence filters out the high-frequency noise in the original sequence and changes more smoothly, but still effectively follows the true, continuous trend of the operational conditions. The smoothed threshold is used to compare with the health index of the charging pile to generate a more stable and reliable optimized maintenance strategy.

[0120] In one specific embodiment, assume that at a specific time point t0(e.g., 10:00 am on Tuesday, which is a peak operating period), the method described in this application is being applied to a charging station with three DC fast charging piles (labeled as pile 1, pile 2, and pile 3, respectively).

[0121] At time point t0, the system obtains the following technical diagnosis results and real-time operating data: Technical diagnosis inputs include: fault source localization results: after diagnosis, the system highly suspects that pile 2 is the potential fault source. Its fault source probability distribution P source is: {pile 1: 0.05, pile 2: 0.90, pile 3: 0.05}. Degradation state assessment results: the dynamic degradation rate V deg (t0) of pile 2 is 0.02 mΩ / charging cycle. Based on its historical degradation trend, the system calculates that its current comprehensive health index HI is 0.88 (where 1.0 represents a brand-new state, and less than 0.80 is defined as complete failure). According to the dynamic degradation rate and the current HI value, the model predicts that the probability P fail of pile 2 failing to operate within the next 24 hours is 15%. Operating economic inputs include: real-time operating data: the real-time number of queued vehicles Q(t0) is 4 vehicles; electricity price and cost data: the time-of-use electricity price P hour (t0) is 1.1 yuan / kWh; the charging service fee is 0.4 yuan / kWh. The average charging amount per vehicle is 40 kWh. The comprehensive cost C plan of planned maintenance (e.g., scheduled at night) is 300 yuan; while the comprehensive cost C emg of emergency maintenance during peak hours is 800 yuan.

[0122] The system needs to evaluate the expected economic loss corresponding to the two decisions of taking immediate action (emergency maintenance) and postponing action (delaying maintenance) at the current time t0. The expected loss L no-action of the postponing action decision: this loss mainly includes the operating loss caused by the failure of pile 2 before the next decision cycle arrives. The expected direct income loss L revenue : total service income per vehicle = average charging amount × (electricity price + service fee) = 40 kWh × (1.1 + 0.4) yuan / kWh = 60 yuan. L revenue = P source (pile 2) × P fail × Q(t0) × total service income per vehicle = 0.90 × 0.15 × 4 × 60 yuan = 32.4 yuan. The expected user churn opportunity cost L churn : assume that the model predicts that if pile 2 fails during the peak period, it will cause at least one queued user to give up charging due to excessive waiting time. Assume that the lifetime value of this lost user is estimated to be 50 yuan. L churn = Psource (Stake 2) x P fail = 6.75 yuan. Total expected loss L no-action = L revenue + L churn = 32.4 + 6.75 = 39.15 yuan. Expected loss L action-now if immediate action is taken. L action-now = cost of emergency repair C emg = 800 yuan.

[0123] The loss function information calculated above is input into the multi-objective optimization module. At the current time t0, since it is in the peak operating period (Q(t0) = 4, P hour (t0) = 1.1), the potential economic risk (39.15 yuan) of not taking action is much smaller than the cost of emergency repair (800 yuan), but this risk is real and accumulates continuously. When the system performs multi-objective Pareto frontier optimization, it will tend to improve the sensitivity of diagnosis to avoid this potential loss amplified by high operating load. In other words, the system will choose a more conservative or stricter decision point. After optimization, the system selects an optimal working point from the Pareto optimal solution set. This working point corresponds to a more stringent health index threshold. For example, the system determines that the adaptive decision threshold Th adapt (t0) under the current high load operating state is 0.90. It should be noted that if it is in the early morning period at night (for example, Q(t) = 0), when the expected loss of not taking action is close to 0, the system will choose a more relaxed decision point, which corresponds to an adaptive decision threshold that may drop to, for example, 0.82, allowing the device to continue operating in a lower health state.

[0124] Compare the current health index HI of Stake 2 (0.88) with the adaptive decision threshold Th adapt (t0) just determined (0.90). The result is 0.88 < 0.90; this indicates that the current health state of Stake 2 has fallen below the minimum health level that the system can tolerate under the current high operating load. The system generates the final optimized maintenance strategy accordingly. Instead of simply immediate repair, the strategy takes into account the cost and operating rhythm, and outputs: the health index of charging stake 2 (0.88) is lower than the adaptive decision threshold (0.90) under the current high operating load. To avoid potential emergency shutdown losses, it is recommended to arrange a planned maintenance with lower cost (expected cost 300 yuan) after the end of the peak period in the next 2 hours.

[0125] Further, the application is applied to the load step response analysis based on high-power charging start, and a most common operation behavior of the charging pile, i.e., high-power start, is used as a natural disturbance source. In a specific embodiment, in a charging station, a battery with a low state of charge (SoC) is connected to a 180kW direct-current fast charging pile and starts charging. At the start moment, the charging pile requests a huge step current from the power grid, and a clear, millisecond-level voltage drop is generated on the station-level power distribution network, which constitutes an ideal common-mode electrical event. The station-level monitoring system monitors the current on the public bus, and when the current change rate exceeds the preset threshold, it is determined that a high-power charging start event occurs, and the time window of the event occurrence (for example, from t0 to t0+500ms) is accurately marked. The system then instructs all charging piles (including those being charged and idle) to synchronously collect and upload the voltage sequence in the time window. These voltage sequences are then processed to generate respective response signal subsequences. The dynamic time warping (DTW) algorithm is used to align all response signal subsequences, and the morphological difference degree of each sequence relative to all other sequences is calculated. It can be understood that a faulty pile with a high contact resistance will have an abnormal impedance characteristic, and in the same voltage drop event, its own terminal voltage will drop deeper or recover more slowly, or additional oscillations will be generated during the recovery process. Therefore, the morphology of its response signal will be significantly different from that of a healthy charging pile, and the calculated morphological difference degree will be significantly higher than that of other healthy piles. The system locks the charging pile with the maximum difference degree as the potential fault source, and can further start the true and false discrimination process.

[0126] Further, the application is applied to the micro-disturbance response analysis based on charging phase switching, which is not only suitable for strong disturbance events, but also suitable for micro-disturbance events in the charging process, thereby widening the applicability of the method. In a specific embodiment, a charging pile that is being charged in constant current mode (CC) switches to constant voltage mode (CV) under its charging control strategy. At the moment of CC / CV switching, the power control strategy of the charging pile changes, which also generates a common-mode electrical disturbance on the power grid, which is weaker than the start but can still be captured by high-precision sensors. The accuracy of signal collection (for example, a higher sampling rate is required) and the sensitivity of the subsequent feature extraction algorithm are higher in this scenario. The application of this scenario proves the sensitivity of the method of the application, which can use various electrical events generated during the entire charging process for diagnosis.

[0127] Further, the application is applied to common-mode response analysis based on external grid swing, which proves that the protection scope of the application is not limited to the disturbance caused by the charging pile itself, and is also applicable to external grid disturbance, which consolidates the support basis of the upper concept of common-mode electrical event. Specifically, the entire charging station encounters a voltage sag or swing caused by the external grid (for example, the start of a large industrial equipment nearby). This is an ideal common-mode event with a wider coverage that is not caused by the charging pile itself. The station-level power quality monitoring system detects the abnormal grid voltage and marks the event window. In this scenario, the focus of the analysis is the voltage recovery phase, and the voltage of each charging pile follows the recovery curve. The faulty pile with high contact resistance will have a significantly different voltage recovery curve shape from the healthy pile due to its abnormal impedance characteristics, such as slower recovery or additional damped oscillation at the end of the recovery. By quantifying the shape difference of these recovery curves using the application, the faulty pile with abnormal response can also be accurately located.

[0128] The preferred embodiments of the application are described in detail above, but the application is not limited to the specific details in the above-described embodiments. Within the technical concept of the application, various equivalent transformations of the technical solutions of the application can be made, and these equivalent transformations all belong to the protection scope of the application.

Claims

1. An adaptive optimization method for health diagnosis of charging piles, characterized in that, include: The raw current, raw voltage, and ambient temperature sequences of the charging pile are obtained, and the data is preprocessed to generate a standardized contact resistance sequence. Based on the standardized contact resistance sequence and the original current sequence, the dynamic degradation rate is evaluated by identifying the specific stages of the charging process and using adaptive modeling. Based on the standardized contact resistance sequence of a predetermined number of charging piles and the pre-stored topological connection relationship of the charging piles, when a common-mode electrical event is detected, the probability distribution of the fault source is located by analyzing the morphological differences of the response signals of each charging pile. By acquiring the real-time number of vehicles in the queue and the time-of-use electricity price, and combining the dynamic degradation rate and the probability distribution of fault sources, dynamic decision-making is made by quantifying operational economic losses, and maintenance strategies are generated. The assessment yielded the dynamic degradation rate, including: Based on the morphological changes of the original current sequence, the constant current, constant voltage and trickle charging stages in the charging process are identified, and a charging stage tag sequence is generated. Based on the charging stage label sequence, the standardized contact resistance sequence is segmented, and a preset stage-specific feature extractor is used for different charging stage segments to extract stage-specific degradation feature vectors. A non-stationary degradation process model is constructed based on stage-specific degradation feature vectors, and the dynamic degradation rate is calculated accordingly. Construct a model of the non-stationary degradation process and calculate the dynamic degradation rate, including: Empirical mode decomposition is used to process the stage-specific degradation feature vectors and separate the degradation trend component that reflects the long-term degradation pattern. Based on the degradation trend component, a degradation model is constructed by adaptively selecting the optimal basis function, and the dynamic degradation rate is obtained by differentiating the degradation model. Degenerative models are constructed by adaptively selecting the optimal basis functions, including: The degradation trend component is segmented according to the charging stage tag sequence. For each segment, the optimal basis function is automatically selected from a predetermined family of candidate basis functions using the Akaike information content criterion to obtain the segmented basis function parameter set. A nonlinear optimization method with continuity constraints is used to adjust the parameter set of piecewise basis functions to ensure that the functions of adjacent segments are continuous at the switching points, thereby constructing a degradation model.

2. The method according to claim 1, characterized in that, Locating the probability distribution of the fault source includes: For the time window of common-mode electrical events, the response signal subsequence is extracted from the standardized contact resistance sequence of each charging pile; The degree of abnormality in response behavior is quantified by calculating the morphological difference between the response signal subsequence of any charging pile and the response signal subsequence of the other charging piles. The charging pile with the greatest morphological difference is identified as a potential source of failure, and a probability distribution of the failure source is generated based on the relative magnitude of this morphological difference.

3. The method according to claim 2, characterized in that, The calculation of morphological differences includes: A dynamic time warping algorithm is used to align the time axis of the response signal subsequences of a predetermined number of charging piles, generating aligned response signal subsequences. The distance between the aligned response signal subsequence of any charging pile and the aligned response signal subsequence of the other charging piles is calculated, and the morphological difference degree is determined by this distance.

4. The method according to claim 2, characterized in that, Extract the response signal subsequence, including: The standardized contact resistance sequence within the time window is subjected to multi-scale differential processing to obtain a multi-scale differential sequence; The Hilbert transform is used to extract the instantaneous features of the multi-scale difference sequence, and the instantaneous features are defined as response signal subsequences for morphological comparison.

5. The method according to claim 2, characterized in that, After generating the fault source probability distribution, the process also includes a fault authenticity determination step, specifically: Based on the potential fault sources determined by the fault source probability distribution, the topological connection relationship of charging piles, and the response signal subsequence of each charging pile, the attenuation law of the amplitude of the response signal with the propagation path is analyzed, and the disturbance attenuation characteristics are extracted. By integrating disturbance attenuation characteristics with the dynamic degradation rate of the potential fault source itself, the fault type of the potential fault source is determined, and a fault type label that distinguishes between the actual fault and the affected state is generated.

6. The method according to claim 1, characterized in that, A stage-specific feature extractor is used to extract stage-specific degenerate feature vectors, including at least one of the following: For the constant current charging stage, wavelet decomposition is used to extract the micro-fluid components of a specific frequency band in the standardized contact resistance sequence to obtain the degradation characteristics of the constant current stage, where the specific frequency band refers to the 3-5Hz frequency band. For the constant voltage charging stage, piecewise polynomial fitting is used to capture the nonlinear upward trend of the standardized contact resistance sequence and obtain the degradation characteristics of the constant voltage stage.

7. The method according to claim 1, characterized in that, Generate maintenance strategies, including: By combining dynamic degradation rate, fault source probability distribution, real-time queued vehicle number and time-of-use electricity price, the expected operational economic loss caused by potential faults is quantified, and a time-varying loss function is constructed. Based on the time-varying loss function, an adaptive decision threshold is generated and compared with the current health status of the charging pile to generate an optimized maintenance strategy.

Images