A method and system for fault detection of smart charging piles

Abstract

This invention belongs to the field of charging pile fault detection technology, specifically relating to an intelligent charging pile fault detection method and system. The method includes: acquiring topological parameters and time-series monitoring data of the charging pile, including thermal capacity, thermal resistance, and line impedance; using a variational inference network to transform discrete variables into a continuous distribution to construct a fully continuous feature tensor; constructing a probability density flow model and rearranging the feature tensor using a double random matrix; generating parameters through a transformation network and transforming them using spline functions, combining physical residuals to construct a regularization term to optimize the loss function, and mapping the tensor to a Gaussian mixture latent space; inputting the test data to determine the most probable operating mode, calculating the optimal transmission distance between covariance matrices as the anomaly score, and determining a fault if the value exceeds a threshold. This invention integrates physical priors and time-series data, improving the sensitivity of weak anomaly identification and the overall detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of charging pile fault detection technology. More specifically, this invention relates to a smart charging pile fault detection method and system. Background Technology

[0002] Charging piles operate under high voltage and high current conditions and are exposed to outdoor environments for extended periods. They are susceptible to electrical or thermodynamic faults due to factors such as internal component aging, external environmental interference, and non-standard operating procedures. These faults can not only disrupt charging services but also potentially cause fires or damage to electric vehicle batteries, resulting in safety hazards and property losses. Modern charging piles generate multi-dimensional time-series monitoring data during operation, including voltage, current, temperature, connection status, and battery charging status. The system itself also possesses defined thermal capacitance, thermal resistance, and line impedance topology attributes. Currently, with the development of artificial intelligence technology, the industry widely employs machine learning and deep neural network models driven by large amounts of operational data to mine time-series data features, thereby enabling the monitoring of charging equipment status and fault identification.

[0003] However, existing fault detection schemes sever the inherent connection between data and device topology, mostly relying on purely data-driven models or simple threshold judgments without integrating and constraining patterns. Charging piles are essentially systems involving multiple interconnected physical processes, including electrical, thermal, and mechanical aspects. Due to the lack of integration with inherent topological attributes and physical laws, existing networks cannot express the strong physical correlation mechanisms when processing heterogeneous data and multivariate joint representations. Furthermore, in practical applications, the operating status of charging piles typically exhibits multimodal distribution characteristics, and the monitoring data contains a mixture of continuous and discrete variables. This further exacerbates the difficulty of low-level feature mapping, and conventional methods often cannot detect the continuous change trajectory of data distribution in high-dimensional feature spaces.

[0004] Referring to Chinese patent document CN113989550B, a method for predicting the operating status of electric vehicle charging piles based on a hybrid CNN and LSTM network is disclosed. This patent document first obtains relevant historical typical fault data, information parameters, and operating environment status data of the charging pile as input sources, and then builds and trains a hybrid model combining a convolutional neural network (CNN) and a long short-term memory network (LSTM). This method can mine the spatiotemporal dependencies of time-series information through deep learning, thereby improving the accuracy of fault classification based on pure data mapping to a certain extent, and automatically analyzing some typical equipment anomalies.

[0005] However, existing technologies fail to transform the inherent topological priors of charging devices, such as thermal capacity, thermal resistance, and line impedance, into physical boundaries and regularization constraints for model forward inference. This results in the model's inability to collaboratively uncover and effectively represent the deep-seated connections between internally strongly physically coupled variables when faced with multivariate inputs. Consequently, when dealing with complex multimodal processes, this approach struggles to establish continuous and accurate state distribution evolution trajectories in the potentially high-dimensional space, leading to insufficient accuracy in assessing early minor faults or potential anomalies. This makes it highly susceptible to false alarms or missed alarms due to mode switching or transient noise interference. Summary of the Invention

[0006] To address the problems of fragmented data and insufficient accuracy in assessing early, minor faults in existing technologies, this invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides a method for fault detection of intelligent charging piles, comprising: S1, acquiring topological parameters of the charging pile including thermal capacity, thermal resistance, and line impedance, and time-series monitoring data including voltage, current, temperature, and connection status; using a variational random noise matrix of a variational inference network for reparameterized sampling to transform discrete variables in the time-series monitoring data into continuous probability distributions, and constructing a fully continuous feature tensor; S2, constructing a probability density flow model, rearranging the fully continuous feature tensor using an alternating normalized double random matrix and dividing it into a first feature part and a second feature part; inputting the first feature part into a transform network to generate node and boundary derivative parameters, extracting the reciprocal of the product of thermal resistance and thermal capacity as a scaling factor to constrain the lower limit of the boundary derivative parameters, and based on the node... A spline function is constructed using the boundary derivative parameters of the points and constraints. The second feature part is transformed using the spline function. The thermodynamic and pressure drop residuals calculated by the topological parameters are combined to construct a physical regularization term to optimize the loss function. The fully continuous feature tensor is mapped to the Gaussian mixture latent space of the probability density flow model. S3: The test data is converted into a fully continuous feature tensor and input into the probability density flow model to obtain the latent space trajectory. The most probable working mode is determined based on the maximum a posteriori probability. The sliding covariance matrix of the latent space trajectory is calculated, and the normal covariance benchmark matrix matching the most probable working mode is extracted. The optimal transmission distance between the two covariance matrices is calculated as the anomaly score. When the anomaly score is greater than the threshold in multiple consecutive time windows, a fault is determined to exist.

[0008] This invention transforms discrete states into continuous distributions through variational networks, achieving a unified representation of heterogeneous data. It optimizes the loss function using thermodynamic and pressure drop residuals as physical regularization terms, and constrains the lower limit of the derivative with the inverse of the product of thermal resistance and thermal capacity, forcing the feature transformation to follow the inherent thermoelectric response boundary of the charging pile, thus enhancing physical interpretability. It utilizes a double-random matrix to rearrange the feature tensor, enabling joint representation of physically coupled variables. By mapping the data to a Gaussian mixture latent space and determining the most probable operating mode, it calculates the optimal transmission distance between the sliding covariance and the normal baseline as the anomaly score. This distance metric keenly captures distribution deviations in high-dimensional trajectories, effectively eliminating mode switching interference and improving the sensitivity of weak anomaly identification and overall detection accuracy.

[0009] Preferably, the step of using a variational random noise matrix obtained by reparameterization sampling via a variational inference network to transform discrete variables in time-series monitoring data into continuous probability distributions and constructing a fully continuous feature tensor includes: inputting the time-series monitoring data into a variational inference network, and outputting mean and variance parameters for smoothing the inverse quantization noise distribution at jump points of discrete variables; generating a variational random noise matrix with the same dimension as the discrete variables through reparameterization sampling based on the mean and variance parameters; adding the data matrix of the discrete variables in the time-series monitoring data to the variational random noise matrix, and constructing the fully continuous feature tensor after zero-mean normalization.

[0010] This invention uses a variational network to adaptively learn the noise distribution characteristics and perform sampling compensation, smoothly transforming the probability density distribution of the original discrete variables into a continuous probability density distribution, thus ensuring the stable convergence of the subsequent probability density flow model during gradient optimization.

[0011] Preferably, the step of rearranging the fully continuous feature tensor using alternating normalized double random matrices includes: initializing a learnable parameter matrix corresponding to the feature dimensions; performing an exponential operation on the elements of the learnable parameter matrix; alternatingly normalizing each row and column of the learnable parameter matrix to obtain a continuously relaxed permutation matrix as the double random matrix; performing dimension rearrangement by multiplying the double random matrix with the fully continuous feature tensor during the model training phase; and applying the Hungarian algorithm to discretize the double random matrix into a permutation matrix for dimension rearrangement during the model inference phase.

[0012] This invention preserves the differentiability of the dimension rearrangement mechanism during the training period to learn the optimal sorting method of variables autonomously. During the inference period, it transforms the result into a deterministic permutation matrix using the Hungarian algorithm, thereby improving the efficiency of manifold recombination while avoiding the loss of feature space information.

[0013] Preferably, the step of extracting the reciprocal of the product of thermal resistance and thermal capacity as a scaling factor to constrain the lower limit of the boundary derivative parameter includes: using the reciprocal of the product of the thermal resistance and the thermal capacity as the thermal response rate; extracting the thermal response rate as a scaling factor, and multiplying the scaling factor by a basic minimum value after normalization mapping to obtain the lower limit of the dynamic boundary derivative parameter; and applying a nonlinear activation function to ensure that the boundary derivative parameter is always positive, so as to ensure that the generated spline function is monotonically increasing.

[0014] This invention transforms the inherent thermal response time delay of physical topology parameters into a numerical constraint boundary of the mapping space, ensuring that the neural network strictly adheres to the monotonically increasing property and the thermodynamic inertia of the objective device when fitting spline interpolation.

[0015] Preferably, mapping the fully continuous feature tensor to the Gaussian mixture latent space of the probability density flow model includes: pre-establishing corresponding Gaussian components of the Gaussian mixture latent space for multiple working modes of the charging pile, and configuring a mean vector and a diagonal covariance matrix for each Gaussian component; calculating the log-likelihood value based on the output of the fully continuous feature tensor during forward propagation of the probability density flow model, extracting the logarithmic value of the determinant of the Jacobian matrix during the model transformation process, and taking the opposite of the log-likelihood value after adding the log-likelihood value to the logarithmic value of the determinant of the Jacobian matrix; obtaining the connection state in the time-series monitoring data as a supervision label to calculate the classification cross-entropy, and constructing the main loss function by combining the opposite; combining the main loss function with the physical regularization term to construct the total loss function, calculating the gradient with this as the optimization objective and performing backpropagation optimization, and mapping the fully continuous feature tensor to the Gaussian distribution cluster corresponding to the Gaussian component.

[0016] This invention drives deep learning models to actively separate and organize mixed time-series data, enabling multi-stage operational data to align and reduce dimensionality in the feature space, effectively avoiding the fluctuation of evaluation indicators caused by the change of operating conditions.

[0017] Preferably, the step of calculating the optimal transmission distance between the two covariance matrices as the outlier score includes: substituting the sliding covariance matrix and the normal covariance reference matrix into the second-order Bures-Wasserstein metric analytical formula to calculate the optimal transmission distance between the sliding covariance matrix and the normal covariance reference matrix in the manifold space; smoothing the time series of the optimal transmission distance using an exponentially weighted moving average filter; and normalizing the filtered result to a preset numerical range using a nonlinear activation function to generate the outlier score.

[0018] This invention utilizes second-order metrics to accurately quantify the degree of variation in the overall scattering structure of high-dimensional data in Riemann space, and integrates time-dimensional smoothing and activation mapping filtering to output intuitive quantitative feature indicators that are resistant to jitter and have consistent scale.

[0019] Preferably, the acquisition of topological parameters of the charging pile, including thermal capacity, thermal resistance, and line impedance, and time-series monitoring data including voltage, current, temperature, and connection status, includes: acquiring the voltage, current, temperature, and connection status through a hardware communication interface and sensors; timestamping the acquired data and setting an alignment tolerance time window; applying a resampling mechanism to align data sources with different acquisition frequencies; and stitching the aligned data of each dimension into the time-series monitoring data.

[0020] This invention overcomes the limitation of the inherent sampling rate difference between the sensor loop and the hardware communication bus, and uses a preset tolerance window and interpolation method to ensure the tight alignment of the data slice time axis and maintain the synchronization of the associated feature vector splicing.

[0021] Preferably, the loss function is optimized by constructing a physical regularization term based on the thermodynamic and voltage drop residuals calculated from the topology parameters, including: calculating the theoretical voltage drop using the current and line impedance in the time-series monitoring data, and calculating the voltage drop residual using the voltage in the time-series monitoring data; calculating the Joule heating power using the current, and calculating the heat dissipation power using the temperature, ambient temperature, and thermal resistance; calculating the theoretical temperature derivative of the difference between the Joule heating power and the heat dissipation power divided by the total heat capacity, and comparing the theoretical temperature derivative with the actual time derivative of the temperature to generate a thermodynamic residual; multiplying the thermodynamic residual and the voltage drop residual by a preset dimensionless weighting coefficient, and summing the absolute values, which are then added as a penalty term to the physical regularization term.

[0022] This invention abstracts the established topological Ohm's law and heat transfer equations into prior constraints and incorporates them into the target optimization space, thereby suppressing the blindness of the model in blindly fitting numerical laws and deviating from the actual hardware operating mechanism.

[0023] Preferably, the step of determining that a fault exists when the abnormal score is greater than the threshold in multiple consecutive time windows includes: setting a warning scalar threshold as the threshold; determining that a fault exists when the abnormal score exceeds the warning scalar threshold limit in multiple consecutive sliding time windows; recording the timestamp of the fault occurrence and reporting the blocking control signal.

[0024] This invention filters out isolated false alarms caused by communication packet loss and electromagnetic pulses based on the continuous over-limit rules of the sliding sequence, enhances the anti-interference capability of the equipment-level anomaly judgment decision, and facilitates the closed loop of blocking intervention for abnormal states.

[0025] Secondly, the present invention provides an intelligent charging pile fault detection system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent charging pile fault detection method is implemented.

[0026] By adopting the above technical solution, a computer program for the above-mentioned intelligent charging pile fault detection method is generated and stored in a memory so that it can be loaded and executed by a processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0027] The beneficial effects of this invention are as follows: This invention constructs a fault detection system that deeply integrates physical priors with temporal data manifolds. After eliminating the barriers to the representation of heterogeneous data by using inverse quantization, the feature transformation step utilizes the lower limit constraint mechanism of the thermal response derivative extracted based on internal topological parameters and the system residual loss penalty term to directly block the algorithm's data fitting path to non-physical realities. This ensures that the dimensionality reduction and rearrangement process fully follows the objectively existing thermoelectric evolution boundary of the hardware, significantly improving the reliability of the strongly physically coupled data representation.

[0028] Furthermore, the multimodal monitoring status is smoothly stretched and mapped into a combination of independent cluster morphologies within a Gaussian mixture latent space layer, establishing a clear decoupling structure for operating conditions. Based on the optimal transmission distance of the distribution sequence, the spatial shape variation of the high-dimensional covariance scatter plot is extracted. This enables the detection indicators to not only smoothly avoid the normalized deviation disturbances caused by the switching of operating phases, but also to perform high-sensitivity identification of hidden and subtle trajectory lesions, thereby improving the success rate of comprehensive detection and early warning of various potential risk characteristics. Attached Figure Description

[0029] Figure 1 This is a flowchart of a fault detection method for smart charging piles; Figure 2 This is a schematic diagram of rational quadratic spline transformation; Figure 3 This is a schematic diagram of an ablation experiment to test performance. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0031] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] This invention discloses a fault detection method for smart charging piles, referring to... Figure 1 This includes steps S1-S3: S1. Collect topology parameters and time series data and perform inverse quantization processing.

[0033] Real-time voltage, current, and temperature data are collected using voltage sensors, current sensors, and thermocouples built into the charging pile. Communication messages are read via the controller local area network bus to obtain connection status and battery charging status data, which are then loaded into a data frame structure. Combined with the charging pile hardware design drawings, topology parameters including thermal capacity, thermal resistance, and line impedance are extracted and constructed into a dictionary stored in memory.

[0034] For discrete variables such as connection status and battery charging status, a variational inference network incorporating a multilayer perceptron is constructed. Continuous and discrete variables are used as inputs, and the outputs are mean and variance parameters of the inverse quantization noise distribution used to smooth the transition points of discrete variables. Based on this, reparameterization is used to sample from Gaussian or other variational posterior distributions to generate a variational random noise matrix with the same dimension as the discrete variables. The data matrix of the discrete variables is superimposed on this noise matrix and smoothed to reduce discrete transition points, transforming the probability mass of the original discrete variables into a probability density distribution.

[0035] Zero-mean normalization is performed on all dequantized continuous variables and the original voltage, current and temperature continuous variables to unify the dimensions and obtain the input feature tensor for later use. In this way, the original mixed continuous and discrete multi-source heterogeneous data are incorporated into the same continuous probability space, eliminating the obstacle of dimension and discrete jumps to unified modeling.

[0036] In one embodiment, topology parameters are extracted from the charging pile's hardware BOM and factory test report. For example, the specific heat capacity parameter is limited to between 0.5 J / (g·K), preferably 0.85 J / (g·K), and the equivalent heat transfer mass parameter is extracted simultaneously; the thermal resistance is between 0.1 K / W and 1 K / W, preferably 0.25 K / W; and the line impedance is between 5 mΩ and 50 mΩ, preferably 15 mΩ. The vehicle handshake message is parsed in real time using a CAN bus communication interface conforming to the GB / T27930 protocol to extract discrete connection states (0 indicating disconnection, 1 indicating connection) and continuous battery charge state of charge (SOC), ranging from 0% to 100%, with an example of 45%.

[0037] Since the sampling frequencies of each data source are different, an alignment tolerance window of ±5ms is set based on microsecond-level timestamps. For channels with inconsistent frequencies, linear interpolation or zero-order hold is applied for resampling and alignment. Then, voltage, current, temperature, connection status and battery charging status are seamlessly stitched together according to the same timestamp to construct a multi-dimensional time-series monitoring data matrix with dimensions of N×15, where N is the time step and the 15 dimensions cover each voltage, current, temperature channel and connection and battery charging status variables.

[0038] Therefore, when a sensor has no valid sample within the alignment window, the system fills the channel with the previous valid value using a zero-order hold and marks the missing state until the channel resumes valid sampling, and then automatically switches back to normal alignment to avoid missing points contaminating subsequent modeling.

[0039] S2. Construct and train a probability density flow model that incorporates physical constraints.

[0040] A probability density flow model that can integrate topological mechanisms is established. Specifically, the probability density flow model is constructed based on a neural network module base class and consists of eight sequentially stacked coupling layers alternating with learnable permutation matrices. The number of coupling layers can be increased or decreased within a reasonable range. The input is a multidimensional temporal monitoring feature vector, and the output is a Gaussian distribution vector of the same dimension mapped to the latent space.

[0041] Within each coupling layer, a learnable parameter matrix of size 15×15, corresponding to the square of the feature dimension, is initialized. Its elements follow a normal distribution with a mean of 0 and a variance of 0.01. Introducing a temperature scaling factor Initially set to 1 and exponentially annealed to 0.1 over the training cycle. Perform the following on the parameter matrix: After scaling, the matrix is ​​converted to non-negativity using the natural exponential function. Then, Sinkhorn iteration is performed, which normally divides each row and column of the matrix by its respective row and column sums. The number of iterations is 10 to 20, preferably 15, to make it converge into a double random matrix where the row and column sums are close to 1, which serves as the permutation matrix for continuous relaxation.

[0042] The convergence criterion involves calculating the residuals between the row and column sum vectors and the all-1 vector after each round of row and column normalization. The convergence criterion is established when the maximum residual is less than 1 × 10⁻⁶. -4 Convergence is determined when the tolerance is minimal; alternatively, it can be determined directly when the preset number of iterations is reached. During the training phase, a pass-through estimator is used to maintain the computational graph closed-loop. Specifically, during forward propagation, the Hungarian algorithm is applied to the continuous double-random matrix to obtain a discrete hard-permuted matrix to complete the dimensional rearrangement. Since the hard-permuted matrix is ​​an orthogonal matrix with a Jacobian logarithmic determinant of 0, it can be included in the log-likelihood calculation without error. During graph construction, the permutation matrix used is expressed as the sum of the discrete hard-permuted matrix and the continuous double-random matrix, and then the truncated gradient value of the double-random matrix is ​​subtracted. Thus, backpropagation bypasses the discretization non-differentiable step, directly propagating the gradient back to the continuous double-random matrix to update the parameters. In this way, the probability density flow model can learn the optimal ordering of variables with strong thermoelectric coupling in the topology by combining regularization loss, allowing the strong physical coupling features to be cross-assigned to different feature parts for collaborative representation.

[0043] The rearranged features are hard-divided into dimensions using cross or continuous index masks. The first feature part and dimension The second feature part of the first feature part both bypasses the transformation and is directly used as the output, and is also fed into the transformation network. This transformation network is a multilayer perceptron, consisting of one input layer, three hidden layers, and one output layer. Each hidden layer has 64 neurons activated using ReLU. Its layer-by-layer forward propagation relationship is as follows: ; ; ; Without an activation function, the output layer calculates the following relationship:

[0044] in, The first feature part of the input, , , These are the output feature tensors of the first hidden layer, the second hidden layer, and the third hidden layer, respectively; It is a linear rectified activation function. , , The hidden layer weight matrix is... , , For the corresponding bias vector , For the output layer weights and biases, These are the output parameters. The transform network outputs each dimension of the second feature part. Transformation parameters for each internal interval, The value can be between 5 and 10, and in this embodiment, 8 is preferred. Therefore, the total output dimension is... The parameters.

[0045] The network output parameters undergo physical and geometric constraint processing using numerical functions. Node width and height parameters are processed using the Softmax function, and a residual guarantee mechanism is introduced during the computation: the mapped interval value equals the minimum constraint width plus the Softmax output value multiplied by the total interval span minus... The minimum width limit is set to 10 times the minimum width limit within a preset range of -5 to 5 and a total span of 10, ensuring that the minimum width of each segment is limited to 10. -3 To prevent numerical singularities, the boundary derivative parameters and internal node derivative parameters are ensured to be always positive using the Softplus function. Simultaneously, the thermal response rate calculated from the physical topology parameters (i.e., the reciprocal of the product of thermal resistance and heat capacity) is extracted as a scaling factor. This scaling factor, after being normalized and dimensionlessly mapped, is multiplied by the fundamental minimum value to serve as the lower bound of the dynamic minimum derivative constraint. The minimum value can be 10. -4The resulting spline function is monotonically increasing, and its rate of change of derivative is controlled by the inherent response speed of the physical system. The input-output mapping relationship of the rational quadratic spline transform in different dimensions is as follows: Figure 2 As shown, the trend of the mapping values ​​corresponding to each dimension feature after transformation is presented when the input value is in the range of [-5, 5].

[0046] In the application phase, for each scalar input of the second feature part, the interval segment it falls into is located by binary search. Combined with the extracted node coordinates and boundary derivative parameters, the mapping is completed by substituting into the rational quadratic polynomial analytical formula. The first and second parts are then spliced ​​together to form a 15-dimensional complete feature and sent into the deep stack, such as the next coupling layer with a total of 10 layers.

[0047] A continuous Gaussian mixture latent space is constructed at the top level of the model, consisting of a mixture of four multivariate Gaussian distributions with different means and covariance matrices. Based on the current and voltage characteristics of the charging pile control protocol, the charging cycle is divided into four operating modes: standby no-load, constant current charging, constant voltage charging, and trickle charging, each corresponding to a Gaussian component. A fixed prior mean vector is assigned to each component, configured on an orthogonal hypersphere centered at the origin with a radius R=10. For example, the mean of the constant current mode is 10 in the first dimension and 0 in the rest, while the mean of the constant voltage mode is 10 in the second dimension and 0 in the rest. A diagonal covariance matrix initialized as a 15-dimensional identity matrix is ​​assigned to each component, with a mixing weight of 0.25.

[0048] During the training of the probability density flow network model, an initial learning rate of 1×10⁻⁶ is used. -4 In conjunction with the Adam optimizer using cosine annealing, the data batch size is set to 256. Forward propagation calculates the sample log-likelihood values ​​in real time using a weighted sum of Gaussian mixture latent spaces. The negative value of this sum is used as the negative log-likelihood loss, obtained by adding it to the logarithm of the Jacobian matrix determinant. Furthermore, a supervised classification cross-entropy penalty term is introduced, incorporating connection status and battery charging status labels parsed from handshake messages. Simultaneously, the thermodynamic heat transfer equation and Ohm's law are constructed as lumped-parameter ordinary differential equations. Physical regularization is introduced, using the total heat capacity calculated from the product of thermal resistance, specific heat capacity, and equivalent mass to calculate the temperature derivative with respect to time (i.e., the difference between Joule heating power and heat dissipation power divided by the total heat capacity). The residual of the voltage-current voltage drop relationship is calculated using line impedance, and its absolute value is used as the regularization term.

[0049] Ultimately, the total loss equals the negative log-likelihood loss plus the penalty coefficient multiplied by the regularization loss. The Adam algorithm is used to backpropagate and update the weights of the entire network until convergence. Through this training, the probability density flow network model removes topological coupling redundancy and stretches the originally mixed and overlapping continuous time-series data of different modes in a directional manner, converging them to the center of Gaussian clusters around the corresponding mean vector, thus completing mode separation and the construction of a structured feature space.

[0050] In calculating the log-likelihood loss, the model forward propagation outputs the Jacobian matrix of the current sample in real time, calculates the logarithm of its determinant, and adds it to the log-likelihood value and inverts it. Simultaneously, the labeled real connection states and battery charging states from historical time-series monitoring data are extracted via the communication bus as supervision labels. The classification cross-entropy between the predicted modality output by the network and the supervision labels is calculated, serving as a penalty compensation term in the main loss function, thereby enhancing the accuracy of multimodal boundary separation.

[0051] S3. Determine charging pile faults based on distance metrics of potential trajectories.

[0052] The tensor of the time series data to be detected, which is collected in real time and has undergone the same dequantization and normalization processes as described above, is input into the trained probability density flow network model. Its forward mapping function is called to obtain the latent variable vectors of continuous time steps, which constitute the latent space trajectory. Then, the posterior probability of each latent vector on the four Gaussian components is calculated, and the component index corresponding to the maximum posterior probability is taken as the current most probable working mode.

[0053] A sliding window with a window length of 100 and a step size of 10 is set in the time dimension. The interdimensional covariance of the potential data within the window is calculated to obtain the current sliding covariance matrix. The corresponding mode is extracted from the HDF5 format benchmark library using a key-value index based on the most probable working mode index, and the normal covariance benchmark matrix is ​​calculated offline based on a large amount of normal historical data.

[0054] The second-order Wasserstein distance between the current sliding covariance matrix and the normal covariance benchmark matrix is ​​used as the anomaly score at the current moment, and compared with a preset warning scalar threshold. Therefore, when the anomaly score exceeds the threshold, the system triggers an alarm subroutine, determines that the charging pile is faulty, records the fault timestamp, and outputs a blocking control signal. When the reported link response times out without confirmation, the system caches the fault record locally and periodically retryes reporting until successful confirmation, clearing the cache to avoid missed reports.

[0055] In this way, the data to be detected is restored to a continuous running trajectory in the potential space, and the deviation of its distribution pattern relative to the normal baseline is quantified into a single anomaly score, so that even early and weak deviations can be captured.

[0056] In the inference phase of real-time anomaly detection for charging piles, a high-dimensional time series trajectory is obtained from the output of the probability density flow model, and its length is set to... Each time step corresponds to a 2-second duration of data, and a sliding time window with a step size S=10.

[0057] For a set of 100 potential vector samples extracted within each time window, an empirical covariance matrix of size 15×15 is calculated to represent the correlation of local feature fluctuations.

[0058] Meanwhile, based on the posterior probability density of each preset Gaussian component, the most probable working mode to which the window sequence belongs is calculated, for example, it is determined that the current mode is in constant voltage charging mode with a 95% probability.

[0059] From the offline pre-built validation knowledge base, the normal covariance benchmark matrix that matches the modality and has ideal fault-free performance is extracted. The empirical mean of the samples within the truncated window is forcibly aligned to the preset benchmark Gaussian component mean through a mean shift operation. Based on the premise that the multivariate Gaussian distribution means have been aligned, the extracted empirical covariance matrix and normal covariance benchmark matrix are substituted into the second-order Bures-Wasserstein metric analytical formula to calculate the optimal transmission distance of the two in the characteristic Riemannian manifold space.

[0060] Because hardware sampling jitter can cause high-frequency transient changes in the distance sequence, an exponentially weighted moving average filter with a window length of 5 and a decay factor of 0.3 is used to smooth the calculated distance value sequence over time. The filtered result is nonlinearly normalized to a preset numerical range, such as 0 to 1, using the Sigmoid activation function to generate an anomaly score. A fault is only detected and a blocking control signal is reported when this score exceeds a preset alarm threshold of 0.75 for three consecutive sliding time windows.

[0061] The model was validated using a real-world dataset from a large charging station over a year. The dataset comprised multi-dimensional time-series monitoring data from 300 DC charging piles, with approximately 90% showing normal operation and about 10% showing abnormalities such as poor contact, converter board overheating, and line aging. This data was divided into training, validation, and test sets in a 7:2:1 ratio. The basic network used the Adam optimizer with an initial learning rate of 1×10⁻⁶. -4 The batch size was 256, and the training cycle was 50. The evaluation indicators were anomaly detection accuracy, recall, F1 score, and false positive rate.

[0062] The baseline model using only conventional affine transformation and Euclidean distance evaluation achieved an accuracy of 82.5%, recall of 78.1%, F1 score of 80.2%, and false positive rate of 6.3%. Replacing the coupling layer data transformation with a physically constrained rational quadratic spline transformation improved accuracy to 88.6% and the F1 score to 86.4%. This transformation enhanced the mapping fitting ability, allowing the model to better learn local mutation features, and improved the F1 score by approximately 6.2%. Further feature rearrangement using a Sinkhorn relaxation-optimized permutation matrix increased accuracy to 92.3% and the F1 score to 91.5%. This eliminated the topological information loss caused by fixed-dimensional sorting, enabling strong coupling variables to establish sufficient cross-mapping and reducing false negatives. Finally, deploying a complete scheme including Gaussian mixture latent space mapping and Wasserstein distance scoring with a sliding covariance matrix, the accuracy rose to 98.1%, recall to 97.4%, F1 score to 97.7%, and the false positive rate decreased to 0.8%. The performance changes after progressively stacking each module are shown below. Figure 3 As shown, the differences in four core evaluation indicators were compared between the baseline model, the rational quadratic spline transformation with physical constraints added sequentially, the Sinkhorn feature rearrangement module, and the complete scheme deployment.

[0063] Therefore, the physically constrained rational quadratic spline transformation enhances the fitting ability of the mapping, enabling the model to better learn local mutation features and improving the F1 score by 6.2%. The permutation matrix optimized by the Sinkhorn algorithm eliminates the topological information loss caused by fixed-dimensional sorting, promotes the establishment of sufficient cross-condition mappings for strongly coupled variables, and allows the model to focus on strongly coupled variables, reducing the false negative rate. The use of Gaussian mixture latent space and Wasserstein metric mechanism successfully decouples the multimodal working states in the latent space, reduces transient noise interference caused by mode switching, suppresses false alarms caused by single-point data jitter, and achieves a reduction in false positive rate and an improvement in the overall detection accuracy of the system.

[0064] This invention also discloses an intelligent charging pile fault detection system, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, an intelligent charging pile fault detection method according to the present invention is implemented.

[0065] The aforementioned intelligent charging pile fault detection system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0066] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A method for detecting faults in intelligent charging piles, characterized in that, include: S1. Obtain the topology parameters of the charging pile, including thermal capacity, thermal resistance and line impedance, as well as the time-series monitoring data including voltage, current, temperature and connection status. S1. Using the variational random noise matrix of the reparameterized sampling of the variational inference network, discrete variables in the time-series monitoring data are transformed into continuous probability distributions, and a fully continuous feature tensor is constructed. S2. A probability density flow model is constructed, and the fully continuous feature tensor is rearranged using an alternating normalized double random matrix and divided into a first feature part and a second feature part. The first feature part is input into the transformation network to generate node and boundary derivative parameters. The inverse of the product of thermal resistance and thermal capacity is extracted as a scaling factor to constrain the lower limit of the boundary derivative parameters. A spline function is constructed based on the node and the constrained boundary derivative parameters. The second feature part is transformed using the spline function. The physical regularization term is constructed by combining the thermodynamic and pressure drop residuals calculated by the topological parameters to optimize the loss function. The fully continuous feature tensor is mapped to the Gaussian mixture latent space of the probability density flow model. S3. The data to be measured is transformed into a fully continuous feature tensor and input into the probability density flow model to obtain the latent space trajectory. The most probable working mode is determined based on the maximum a posteriori probability. Calculate the sliding covariance matrix of the potential spatial trajectory, and extract the normal covariance reference matrix that matches the most probable working mode; calculate the optimal transmission distance between the two covariance matrices as the anomaly score, and determine that there is a fault when the anomaly score is greater than the threshold in multiple consecutive time windows.

2. The method for fault detection of a smart charging pile according to claim 1, characterized in that, The variational random noise matrix, obtained by reparameterizing the sampling using a variational inference network, transforms discrete variables in time-series monitoring data into continuous probability distributions, constructing a fully continuous feature tensor, including: The time-series monitoring data is input into the variational inference network, which outputs mean and variance parameters for smoothing the inverse quantization noise distribution at the jump points of discrete variables. Based on the mean parameter and the variance parameter, a variational random noise matrix with the same dimension as the discrete variable is generated by reparameterized sampling; The data matrix of discrete variables in the time-series monitoring data is added to the variational random noise matrix, and the fully continuous feature tensor is constructed after zero-mean normalization.

3. The method for fault detection of a smart charging pile according to claim 1, characterized in that, The rearrangement of the fully continuous feature tensor using alternating normalized double random matrices includes: Initialize the learnable parameter matrix corresponding to the feature dimension, and perform an exponential operation on the elements of the learnable parameter matrix; The learnable parameter matrix is ​​alternately normalized in each row and column to obtain a continuously relaxed permutation matrix, which is then used as the double random matrix. During the model training phase, the double random matrix is ​​multiplied by the fully continuous feature tensor to perform dimensional rearrangement. During the model inference phase, the double random matrix is ​​discretized into a permutation matrix using the Hungarian algorithm to perform dimensional rearrangement.

4. The method for fault detection of a smart charging pile according to claim 1, characterized in that, The extraction of the reciprocal of the product of thermal resistance and thermal capacity as a scaling factor to constrain the lower limit of the boundary derivative parameter includes: The reciprocal of the product of the thermal resistance and the thermal capacity is taken as the thermal response rate; The thermal response rate is extracted as a scaling factor. After normalization mapping, the scaling factor is multiplied by the basic minimum value to obtain the lower limit of the dynamic boundary derivative parameter. A nonlinear activation function is applied to ensure that the boundary derivative parameter is always positive, thereby guaranteeing that the generated spline function is monotonically increasing.

5. The method for fault detection of a smart charging pile according to claim 1, characterized in that, The process of mapping the fully continuous feature tensor to the Gaussian mixture latent space of the probability density flow model includes: pre-establishing corresponding Gaussian components of the Gaussian mixture latent space for multiple working modes of the charging pile, and configuring a mean vector and a diagonal covariance matrix for each Gaussian component; calculating the log-likelihood value based on the output of the fully continuous feature tensor during forward propagation of the probability density flow model, extracting the logarithmic value of the determinant of the Jacobian matrix during the model transformation process, and taking the opposite of the log-likelihood value and the logarithmic value of the Jacobian matrix determinant; obtaining the connection state in the time-series monitoring data as a supervision label to calculate the classification cross-entropy, and constructing the main loss function by combining the opposite; combining the main loss function with the physical regularization term to construct the total loss function, calculating the gradient with this as the optimization objective and performing backpropagation optimization, and mapping the fully continuous feature tensor to the Gaussian distribution cluster corresponding to the Gaussian component.

6. The method for fault detection of a smart charging pile according to claim 1, characterized in that, The calculation of the optimal transmission distance between the two covariance matrices as the anomaly score includes: Substituting the sliding covariance matrix and the normal covariance reference matrix into the second-order Bures-Wasserstein metric analytical formula, the optimal transmission distance of the sliding covariance matrix and the normal covariance reference matrix in the manifold space is calculated. An exponentially weighted moving average filter is used to smooth the time series of the optimal transmission distance. The filtered result is normalized to a preset numerical range using a nonlinear activation function to generate the outlier score.

7. The method for fault detection of a smart charging pile according to claim 1, characterized in that, The acquisition of charging pile topology parameters including thermal capacity, thermal resistance, and line impedance, as well as time-series monitoring data including voltage, current, temperature, and connection status, includes: The voltage, current, temperature, and connection status are obtained from the sensor via a hardware communication interface. Timestamp the acquired data and set an alignment tolerance time window; A resampling mechanism is applied to align data sources with different collection frequencies, and the aligned data from each dimension are then stitched together to form the time-series monitoring data.

8. The method for fault detection of a smart charging pile according to claim 1, characterized in that, The loss function is optimized by constructing a physical regularization term based on the thermodynamic and voltage drop residuals calculated from the topology parameters. This includes: calculating the theoretical voltage drop using the current and line impedance from the time-series monitoring data, and calculating the voltage drop residual using the voltage from the time-series monitoring data; calculating the Joule heating power using the current, and calculating the heat dissipation power using the temperature, ambient temperature, and thermal resistance; calculating the theoretical temperature derivative of the difference between the Joule heating power and the heat dissipation power divided by the total heat capacity, and comparing the theoretical temperature derivative with the actual time derivative of the temperature to generate a thermodynamic residual; multiplying the thermodynamic residual and the voltage drop residual by preset dimensionless weighting coefficients, and summing the absolute values, which are then added as a penalty term to the physical regularization term.

9. The method for fault detection of a smart charging pile according to claim 1, characterized in that, The determination of a fault when the abnormal score is greater than the threshold for multiple consecutive time windows includes: Set the warning scalar threshold as the threshold; When the abnormal score exceeds the warning scalar threshold limit within multiple consecutive sliding time windows, a fault is determined to exist; Record the timestamp of the fault occurrence and report the blocking control signal.

10. A smart charging pile fault detection system, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a smart charging pile fault detection method according to any one of claims 1-9.

Images