Charging pile load test data analysis method and system based on intelligent perception

Abstract

The present invention relates to the field of new energy vehicles, and discloses a charging pile load test data analysis method and system based on intelligent perception, including dynamic mapping of charging pile groups in public parking lots through multi-source heterogeneous perception components, and staggered annotation of handshake waveforms, thermal image hotspots and harmonic pulses based on a scene semantic segmentation network; performing multi-scale sliding window cross-domain correlation tracking on data priority sequences, combining historical waveform fingerprints, power factor drift trajectories and environmental electromagnetic interference matrices; extracting load evolution trajectories using a priority adaptive model based on a multi-dimensional time series feature tensor, by identifying cluster density mutations and entropy drift inflection points; calibrating charging entities and coupling ranges corresponding to candidate safety hazard events as candidate areas; combining potential instability nodes and global risk entropy fluctuation trends in monitoring data based on local load priority interference maps. The present invention has the advantage of improving load safety accuracy in complex environments.

Description

Technical Field

[0001] The present invention relates to the field of new energy vehicles, and in particular to a charging pile load test data analysis method and system based on intelligent perception. Background Art

[0002] With the rapid growth of electric vehicle ownership, charging stations are being deployed on a large scale in diverse scenarios, including shopping malls, office building basements, and intercity highway service areas. Load test data has evolved from static voltage and current sampling to a multi-source, heterogeneous sensing model that integrates handshake protocol logs, conductor thermal imaging, harmonic noise, power factor, ambient temperature and humidity, and user identity tags. Existing load test systems typically use fixed or device-number-based sampling strategies, with no dynamic adjustment of sensing priorities. Thermal imaging and current waveforms are uniformly sampled every 5 seconds, making it impossible to increase the thermal imaging frame rate even when a high-temperature warning is encountered. Handshake protocol logs are only triggered at the start and end of charging, making it difficult to capture transient arcing caused by repeated plugging and unplugging. Some studies have proposed threshold-triggered priority switching, but these methods still rely on single-metric thresholds and fail to uniformly model cross-domain data correlations, scenario-based load characteristics, and risk levels. This can lead to high-value data streams being squeezed out of bandwidth by low-risk scenarios, or critical anomalies being missed due to sparse sampling. Therefore, it is necessary to design a charging pile load test data analysis method and system based on intelligent perception to improve the load safety accuracy in complex environments. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides a charging pile load test data analysis method and system based on intelligent perception, which has the advantages of improving load safety and accuracy in complex environments and solves the problems in the above-mentioned background technology.

[0004] To achieve the above-mentioned purpose of improving load safety and accuracy in complex environments, the present invention provides the following technical solution: a charging pile load test data analysis method based on intelligent perception, comprising the following steps:

[0005] Dynamically map charging piles in public parking lots through multi-source heterogeneous perception components, and perform staggered annotation of handshake waveforms, thermal image hotspots, and harmonic pulses based on a scene semantic segmentation network, generating a data priority sequence based on risk sensitivity.

[0006] Perform multi-scale sliding window cross-domain correlation tracking on the data priority sequence, combining historical waveform fingerprints, power factor drift trajectory and environmental electromagnetic interference matrix to construct a multi-dimensional time series feature tensor of charging load behavior;

[0007] Based on the multi-dimensional time series feature tensor, a priority adaptive model is used to extract the load evolution trajectory. By identifying the cluster density mutation and entropy drift inflection points, candidate safety hazard events are detected.

[0008] The charging entities and coupling ranges corresponding to candidate safety hazard events are marked as candidate areas, and the connection end nodes, thermal sensing nodes, and power control nodes in the candidate areas are integrated to construct a local load priority interference map.

[0009] Based on the local load priority interference map, combined with the potential unstable nodes and the global risk entropy fluctuation trend in the monitoring data, the perception weight and sampling frame rate are dynamically reconstructed, and charging safety response instructions are generated.

[0010] Preferably, the process of generating a data priority sequence according to risk sensitivity is:

[0011] The perception data stream is formed by collecting multi-source heterogeneous perception components;

[0012] Based on the scene semantic segmentation network, various perception data streams are processed in time and space synchronously, and feature labels are annotated on the perception data streams using staggered annotation.

[0013] Based on the risk contribution matrix constructed from the historical accident database, the correlation degree of typical accidents in each data stream is calculated;

[0014] The correlation degree and the real-time parking lot load density are coupled and scored to obtain the risk sensitivity coefficient reflecting the scene characteristics, and the final data priority sequence is output.

[0015] Preferably, the multi-scale sliding window cross-domain correlation tracking process is performed on the data priority sequence:

[0016] Prioritize the perceived data streams based on the risk sensitivity coefficient, defining data streams above the set threshold as high-weight data and data streams below the set threshold as low-weight data;

[0017] Assign short time windows to high-weight data and long time windows to low-weight data, and build a multi-scale sliding window set;

[0018] Synchronously read the perception data stream within each sliding window, and eliminate cross-domain data delay through timestamp alignment and phase correction methods;

[0019] A recursive cross-domain correlation matching algorithm is used to track the load waveform shape, thermal image temperature rise gradient, and harmonic amplitude changes between adjacent sliding windows to generate a cross-domain matching vector.

[0020] Output the integrated multi-scale cross-domain correlation results.

[0021] Preferably, the process of combining historical waveform fingerprints, power factor drift trajectory and environmental electromagnetic interference matrix is:

[0022] Retrieve a waveform fingerprint template group that matches the current charging pile device model from the historical database, match the template group with the real-time handshake waveform collected within the sliding window, and output a similarity score vector;

[0023] Exponentially weighted recursive smoothing is performed on the real-time collected power factor data to extract short-term oscillation features reflecting instantaneous fluctuations and long-term drift features describing trend changes, forming a dual-channel feature vector.

[0024] An environmental electromagnetic interference matrix based on sensor array data is constructed. The interference degree of each frequency band noise on the handshake waveform and harmonic pulse is calculated through frequency band analysis method, and the frequency domain interference influence weight vector is obtained.

[0025] The waveform similarity score vector, dual-channel power factor feature vector, and frequency domain interference weight vector are uniformly mapped to the multi-dimensional feature embedding space through the tensor embedding mechanism to form an embedding result tensor;

[0026] The multi-core kernel function fusion algorithm is used to perform nonlinear feature superposition on the embedded result tensor to generate a set of comprehensive weight coefficients for time series modeling and risk judgment.

[0027] Preferably, the process of constructing the multi-dimensional time series feature tensor of the charging load behavior is:

[0028] The perception data stream and comprehensive weight coefficient set in the cross-domain correlation results are stored in the tensor draft according to the three-dimensional index of time axis, data domain and risk weight;

[0029] Apply double threshold correction for missing anomalies to the tensor draft, use local interpolation to repair short-term missing samples, and use random forest discrimination to remove extreme outliers;

[0030] The redundant dimensions are reduced by tensor orthogonal projection mapping to form a multi-dimensional time series feature tensor.

[0031] Preferably, the process of extracting the load evolution trajectory using the priority adaptive model is:

[0032] Adopting an adaptive network combining gated recurrent units and attention mechanism to perform sequence encoding on multi-dimensional temporal feature tensors;

[0033] Embed a data priority sequence during the encoding process and use a priority gating unit to dynamically adjust the hidden state update rate of different perception data streams;

[0034] Perform aggregate analysis on continuous hidden states to generate several candidate evolution trajectory segments;

[0035] The similarity and risk mutation between candidate evolution trajectory segments are evaluated using a joint discriminant criterion of density peak identification and entropy drift detection.

[0036] Output load evolution trajectory with time and priority labels.

[0037] Preferably, the process of detecting a candidate safety hazard event is:

[0038] Perform dynamic time-warping comparison on the load evolution trajectory and the standard safe operation trajectory library to calculate the trajectory shape deviation and time synchronization error;

[0039] Construct a dual-threshold anomaly scoring function, where the first threshold is based on the cluster density mutation and the second threshold is based on the entropy drift inflection point amplitude;

[0040] If both the deviation and the anomaly score of a trajectory subsequence exceed the corresponding thresholds, the trajectory subsequence is marked as a potential safety hazard segment;

[0041] Adjacent potential hidden danger segments are merged in time and space to generate risk continuous segments and output candidate safety hidden danger events.

[0042] Preferably, the process of constructing the local load priority interference map is:

[0043] According to the list of candidate safety hazard events, the corresponding connection end nodes, thermal sensing nodes and power control nodes are mapped into a heterogeneous node set;

[0044] An event-driven edge weight update strategy is adopted to establish three types of multimodal edges between heterogeneous node sets;

[0045] On each edge, the interference intensity score of the edge is dynamically calculated based on the perception data characteristics, current state change rate and historical risk sensitivity of the nodes at both ends of the edge, and is attached to the graph structure as an edge weight to form a local heterogeneous graph;

[0046] The core interference source nodes and key coupling paths in the local heterogeneous graph are identified through the joint analysis of spectral clustering and node centrality, and the local load priority interference map is output.

[0047] Preferably, the process of generating a charging safety response instruction is:

[0048] Based on the historical load data and current operating status of all sensing nodes, the state entropy of each main variable in the multi-dimensional time series feature tensor is counted, and a global risk entropy fluctuation trend sequence representing the amplitude and direction of state entropy change is constructed;

[0049] The global risk entropy fluctuation trend sequence is integrated with the local load priority interference map to identify the system risk evolution section where the critical path in the map is located, and the cumulative risk value of each critical path is calculated;

[0050] Dynamically adjust the data sampling parameters of the corresponding nodes based on the risk accumulation value of each path and the priority labels of the nodes involved;

[0051] Integrate all perception parameter adjustments and control execution actions to output charging safety response instructions.

[0052] The charging pile load test data analysis system based on intelligent perception includes:

[0053] Multi-source mapping module: Dynamically maps and semantically segments charging pile groups and generates data priority sequences based on risk sensitivity;

[0054] Feature construction module: Cross-domain correlation of handshake, thermal image and harmonic data in a multi-scale sliding window, combined with historical and environmental information to form a multi-dimensional time series feature tensor;

[0055] Trajectory analysis module: This module extracts load evolution trajectories based on a priority adaptive model and detects cluster density mutations and entropy drift inflection points to identify candidate safety hazard events.

[0056] Interference map module: This module maps candidate safety hazard events to local areas and integrates connection terminals, thermal sensing, and power control nodes to generate a load priority interference map.

[0057] Safety response module: Dynamically adjusts the perception weight and sampling frame rate based on the interference map and global risk entropy fluctuations to generate charging safety response instructions.

[0058] Compared with the prior art, the present invention provides a charging pile load test data analysis method and system based on intelligent perception, which has the following beneficial effects:

[0059] The present invention achieves high-precision capture of transient load characteristics of charging piles by deeply integrating multi-source heterogeneous perception, semantic segmentation and misalignment labeling with risk sensitivity grading; cross-domain sliding window correlation and tensor modeling can extract hidden abnormal patterns under complex noise backgrounds, significantly improving the sensitivity and reliability of abnormality identification; the priority adaptive mechanism can dynamically focus on high-risk data streams, avoiding low-value information from occupying bandwidth and ensuring real-time monitoring of key nodes in the risk germination stage; the sampling and reconstruction strategy driven by the collaborative efforts of local interference maps and global risk entropy ensures continuous optimization of the matching degree between monitoring resources and system risks, reducing redundant transmission and storage overhead; the final generated safety response instructions can trigger automatic load reduction, power outage or warning push on the edge side, realizing the closed-loop linkage of data perception, risk assessment and active intervention, and comprehensively improving the safety, operational stability and operation and maintenance economy of public parking lot charging facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of the method of the present invention;

[0061] Figure 2 Schematic diagram of the system of the present invention. DETAILED DESCRIPTION

[0062] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0063] Example 1: Please refer to Figure 1 As shown, the charging pile load test data analysis method based on intelligent perception according to an embodiment of the present invention includes the following steps:

[0064] S1: Dynamically map charging pile clusters in public parking lots through multi-source heterogeneous perception components, and perform staggered annotation of handshake waveforms, thermal image hotspots, and harmonic pulses based on a scene semantic segmentation network, generating a data priority sequence based on risk sensitivity.

[0065] The process of generating data priority sequence according to risk sensitivity in S1 is as follows:

[0066] The perception data streams collected by multi-source heterogeneous perception components include: handshake waveform data stream, thermal image stream, and harmonic pulse spectrum stream. Each type of perception data stream is assigned an initial weight based on the sampling frequency, data quality stability, and historical validity of the perception node.

[0067] Handshake waveform data stream: This is continuously output by the charging pile communication control board during the charging start-up, maintenance, and termination processes. The sampling frequency is generally 1–5 kHz, which can reflect the plug-in contact status and protocol negotiation integrity in real time.

[0068] Thermal image stream: captured by an infrared thermal imager placed near the gun-line interface and at key cable bends, with a frame rate of 10–60 fps, used to monitor temperature rise rate and hotspot distribution.

[0069] Harmonic pulse spectrum stream: collected by the harmonic analysis module connected in parallel to the AC output side, with a frequency band covering 0–10 kHz and a refresh period of 200 ms, reflecting harmonic content and transient pulses;

[0070] Based on the scene semantic segmentation network, various sensory data streams are processed in time and space synchronously. Dislocation annotation is used to label the phase characteristics of handshake waveforms, temperature gradient changes of thermal image hot spots, and spectral anomalies of harmonic pulses.

[0071] Based on the risk contribution matrix constructed from the historical accident database, the correlation between each data stream and three typical accidents: overload, poor contact, and harmonic impact was calculated. The characteristic manifestations of similar accidents were divided into several prototype patterns through clustering algorithms. Decision trees, random forests, and other methods were used to evaluate the splitting gain of each type of feature for identifying target accidents. The feature contribution ranking was obtained, and the ranking results were normalized to form a three-row, three-column matrix. The rows corresponded to the three types of data streams: handshake waveform, thermal image hotspot, and harmonic pulse, and the columns corresponded to the three types of accidents. Each value in the matrix represents the proportion of importance of the data stream in determining the accident.

[0072] The correlation degree is coupled with the real-time parking lot load density to obtain a risk sensitivity coefficient that reflects the characteristics of the scenario. The current number of concurrent charging, average power utilization, temperature, and humidity environmental indicators of the parking lot are counted once a minute and mapped into three scenario parameters: "overload excitation quantity", "contact abnormality excitation quantity", and "harmonic excitation quantity". Each column of the risk contribution matrix is ​​multiplied element by element with the corresponding scenario parameter and then summarized to obtain the contribution score of the three data streams of handshake waveform, thermal image hotspot, and harmonic pulse to the comprehensive risk in the current scenario. This score is recorded as the risk sensitivity coefficient. The higher the score, the more critical the data stream is to accident warning.

[0073] Sort the data streams according to the risk sensitivity coefficient to form a data priority candidate list, set dynamic thresholds, eliminate low-correlation data, and output the final data priority sequence.

[0074] S2: Perform multi-scale sliding window cross-domain correlation tracking on the data priority sequence, combine historical waveform fingerprints, power factor drift trajectory and environmental electromagnetic interference matrix, and construct a multi-dimensional time series feature tensor of charging load behavior.

[0075] The multi-scale sliding window cross-domain correlation tracking process for the data priority sequence in S2 is:

[0076] Prioritize the perceived data streams based on the risk sensitivity coefficient, defining data streams above the set threshold as high-weight data and data streams below the set threshold as low-weight data;

[0077] Short windows are assigned to high-weight data, and long windows are assigned to low-weight data, constructing a multi-scale sliding window set. High-weight data uses short windows, typically 100ms to 500ms, to capture fine-grained transients, while low-weight data uses long windows, typically 1s to 5s, to observe slowly changing trends. The sliding step size of high-weight windows is set to 50% of their length (ensuring overlap) to avoid missing spikes, while the sliding step size of low-weight windows is equal to their length (no overlap) to save bandwidth and computation. Each data stream is automatically bound to a corresponding window and step size based on its own weight and written to a sliding window queue. Window parameters can be inserted, deleted, or changed during operation.

[0078] The handshake waveform, thermal imaging hotspot, and harmonic pulse fragments are synchronously read within each sliding window, and cross-domain data delays are eliminated through timestamp alignment and phase correction methods. When the sliding window is triggered, the scheduler synchronously pulls the handshake waveform, thermal imaging frame, and harmonic fragments. If any stream is temporarily unavailable, it waits for no more than one sampling cycle. All three types of data use the unified time base calibrated in the previous stage. If the alignment error is higher than 2ms, a synchronous correction will be triggered. The phase offset of the main frequency component of the handshake waveform is detected and corrected to the reference phase. The harmonic spectrum is linearly shifted according to the fundamental frequency synchronization to eliminate the spectral line displacement caused by time base drift, and generate multi-domain data fragments with consistent time and phase.

[0079] A recursive cross-domain correlation matching algorithm is used to track the load waveform morphology, thermal image temperature rise gradient, and harmonic amplitude changes between adjacent sliding windows to generate a cross-domain matching vector. The peak and valley vectors of the handshake waveform, the temperature rise gradient vector of the thermal image, and the spectral power vector of the harmonics are extracted from the aligned segments. Within the current sliding window, the initial correlation coefficient between the vectors is first calculated. Then, the recursive phase is entered, and the vectors are extended backward frame by frame, and a secondary comparison is performed with the corresponding vector of the next sliding window. If the correlation coefficient continues to increase, the next sliding window is merged into the current matching segment and recursively repeated. If it drops below a threshold, the recursion is terminated. A cross-domain matching vector is recorded for each successful recursive segment. The vector contains the coordinated change amplitude, duration, and maximum correlation gradient of the three types of features.

[0080] Trajectory continuity constraints and abnormal sparseness penalty terms are introduced to smooth and compensate cross-domain matching vectors to address short-term distortion or frame drops. The system maintains the three most recent sliding window matching results for each type of data stream and calculates a continuity score. When the continuity score drops significantly, it indicates that the trajectory may be broken or distorted for a short time, triggering the compensation logic. If a sliding window has an invalid value for a long time, the matching confidence is deducted by the window length multiplied by the penalty coefficient. Linear interpolation or spline interpolation is used to restore continuity in the broken segments, and exponential smoothing is used for segments with decreased confidence to suppress the sharp fluctuations caused by single-frame outliers. The integrated multi-scale cross-domain correlation results are output.

[0081] The process of combining historical waveform fingerprints, power factor drift trajectory and environmental electromagnetic interference matrix in S2 is as follows:

[0082] Retrieve the waveform fingerprint template group that matches the current charging pile device model from the historical database, match the template group with the real-time handshake waveform collected in the sliding window, and output a similarity score vector; the operation and maintenance platform uses the pile model-firmware version-power segment as the first-level index to store the handshake waveform fingerprint template. Each template contains a typical handshake segment of 1 second in length and the corresponding event label. The template library is automatically clustered and de-redundant every three months to maintain representativeness and compressed storage. The handshake waveform in the sliding window is first low-pass filtered to remove high-frequency noise, and then amplitude normalization and baseline shift correction are performed to ensure the same benchmark as the template. The sequence morphology matching algorithm is called to compare with the template one by one. To avoid alignment errors of segments of different lengths, the matching algorithm only retains the top five templates with the smallest alignment distance and outputs a similarity score vector with the same length.

[0083] Exponentially weighted recursive smoothing is performed on real-time power factor data to extract short-term oscillation characteristics reflecting instantaneous fluctuations and long-term drift characteristics describing trend changes, forming a dual-channel feature vector. The smart energy meter uploads the power factor sequence at a 100-millisecond granularity. A real-time smoothing curve is maintained for this sequence on the edge to reduce single-point fluctuations and extract long-term trends. The short-term oscillation channel is the difference between the original sequence and the smoothing curve, representing instantaneous fluctuations. The long-term oscillation channel uses the smoothing curve value to represent trend changes. The two channels are spliced ​​together to form a fixed-length power factor feature vector.

[0084] An environmental electromagnetic interference matrix was constructed based on data collected by the sensor array. The degree of interference of noise in each frequency band on the handshake waveform and harmonic pulses was calculated using frequency band analysis methods, and the frequency domain interference impact weight vector was obtained. Multiple magnetic field probes were deployed on the charging pile partitions, the side walls of the cable trench, and the edges of the parking spaces. Reference probes were also deployed outside the field. The 50Hz-10kHz frequency band was uniformly divided into 16 sub-segments on a logarithmic scale. The root mean square noise within one minute of each probe in each sub-segment was calculated. The reference probe value was subtracted from the probe measured value to obtain the interference amplitude between each probe and the frequency band.

[0085] The waveform similarity score vector, dual-channel power factor feature vector, and frequency domain interference weight vector are uniformly mapped to the multidimensional feature embedding space through a tensor embedding mechanism to form an embedding result tensor. The waveform similarity vector, power factor dual-channel vector, and interference weight vector are truncated to a uniform length using the feature group ID + fixed-length value range. The three types of vectors are sequentially written into different depth channels of the three-dimensional data block, and the source labels are recorded to form an embedding result tensor.

[0086] A multi-core kernel function fusion algorithm is used to perform nonlinear feature superposition on the embedded tensor, generating a comprehensive set of weight coefficients for time series modeling and risk assessment. The embedded tensor is fed into each kernel function pipeline, yielding four sets of nonlinear mapping results. A weight update path is established at the pipeline exit to automatically adjust the kernel weights based on the historical contribution of the mapping results to risk assessment accuracy. Finally, the four mapping sets are linearly fused according to the latest kernel weights to generate a comprehensive set of weight coefficients for time series modeling and risk assessment. This set includes: key waveform pattern weights, power factor drift weights, frequency domain interference suppression weights, and kernel fusion confidence.

[0087] The process of constructing the multi-dimensional time series feature tensor of the charging load behavior in S2 is:

[0088] The handshake waveform, thermal image gradient, harmonic amplitude and comprehensive weight coefficient set in the cross-domain correlation results are stored in the tensor draft according to the three-dimensional index of time axis, data domain and risk weight;

[0089] Time axis T: uses the same unified millisecond time base as the multi-scale sliding window;

[0090] Data domain D: four subdomains including handshake waveform, thermal image gradient, harmonic amplitude and additional imported comprehensive weight;

[0091] Risk weight R: In the comprehensive weight coefficient set output by the kernel function fusion step, each risk-related factor corresponds to a layer of weight;

[0092] For each time slice, first write the handshake waveform vector, then write the corresponding thermal image gradient vector and harmonic amplitude vector, and finally write the weight vector corresponding to the three.

[0093] Double threshold correction for missing anomalies is applied to the draft tensor, short-term missing samples are repaired using local interpolation, and extreme outliers are removed using random forest discrimination. Within the same subdomain, continuous missing samples less than 3 sampling cycles are considered short-term missing samples and repaired using local interpolation. Continuous missing samples ≥ 3 sampling cycles are marked as severe missing samples. Local interpolation includes linear interpolation and two-dimensional spline interpolation. Linear interpolation: handshake waveforms and harmonic amplitudes are filled using linear interpolation. Two-dimensional spline interpolation: thermal image gradients are interpolated using spline interpolation due to their spatial distribution characteristics. The repaired tensor fragments are fed into the outlier detection pipeline. The system maintains a trained random forest model. Input features include adjacent time slice differences, cross-domain correlation, sliding window variance, etc. The random forest outputs a 0 / 1 label, 1 represents an extreme outlier. Extreme outliers are directly replaced by the median of the past five frames of the subdomain to prevent pulse spikes from contaminating the spectrum.

[0094] Wavelet decomposition and reconstruction operators are introduced to perform multi-resolution decomposition of the tensor, separating low-frequency trends from high-frequency pulse components while retaining complete spectral information. Each subdomain is decomposed along the time axis using the Daubechies-4 wavelet, with the number of layers automatically selected based on the sampling frequency. The decomposed approximate coefficients represent low-frequency trends, while the detail coefficients represent high-frequency pulses. In the low-frequency channel, the complete approximate coefficients are retained to highlight slow-changing trends. In the high-frequency channel, the detail coefficients are subjected to energy threshold screening, retaining only the pulse segments with the top 30% energy, removing weak high-frequency noise.

[0095] The platform reduces redundant dimensions through tensor orthogonal projection mapping while retaining 95% of the cumulative explained variance to form a multidimensional time series feature tensor; all subdomains are z-score normalized according to their respective dimensions to eliminate scale differences; the platform runs an improved Tucker decomposition on the tensor dimension: first, the principal components are searched in the time axis direction; secondly, a secondary orthogonalization is performed on the joint dimension of the data domain and the risk weight; the cumulative explained variance is calculated after each decomposition step; when the cumulative explained variance reaches 95%, the decomposition is stopped and soft threshold truncation is performed on the remaining redundant dimensions. This not only removes duplicate information but also retains the high-order features that are most discriminative of load behavior.

[0096] S3: Based on the multi-dimensional time series feature tensor, a priority adaptive model is used to extract the load evolution trajectory. By identifying the cluster density mutation and entropy drift inflection points, candidate safety hazard events are detected.

[0097] The process of extracting the load evolution trajectory using the priority adaptive model in S3 is as follows:

[0098] An adaptive network combining gated recurrent units and an attention mechanism is used to sequentially encode multidimensional temporal feature tensors. The encoder uses a two-layer stack of gated recurrent units (GRUs), with the bottom layer processing local temporal details and the upper layer capturing long-range dependencies. The output of each GRU layer is followed by a 1×1 channel attention module, which redistributes the importance of each feature domain within the same time step, splitting the output multidimensional temporal feature tensor into consecutive frames along the time axis, while maintaining the reduced data domain and risk weight structure within the frame. Each frame is accompanied by a timestamp and tensor metadata to ensure that the original sampling position can be traced within the network. After the double-layer GRU+attention process, a hidden state vector is generated at each time step, which already contains cross-domain coupling features, risk weight references, and temporal context information.

[0099] A data priority sequence is embedded in the encoding process, and the hidden state update rate of different perception data streams is dynamically adjusted using a priority gating unit. The data priority sequence obtained from the risk sensitivity stage is embedded into the GRU gating function in a one-to-one correspondence with the time frame. High-priority data streams are given a greater probability of opening the gate in the gating, which means that the corresponding hidden state is updated more frequently. The update rate of low-priority data streams is suppressed, and the hidden state is refreshed only when the trend changes significantly, reducing noise interference. The gating weight is refreshed in real time with the scenario risk and recalculated every 30 seconds, so that the model remains highly sensitive to sudden high-risk signals.

[0100] Aggregate and analyze continuous hidden states to generate several candidate evolution trajectory segments. Continuous hidden state vectors are merged into 1-second windows, and summary metrics such as maximum value, mean value, and rate of change are extracted using a dual-path pooling and statistical approach. Window summaries that meet the two rules of consistent trend and consistent direction can be concatenated to form candidate evolution trajectory segments. If either rule is not met, a new trajectory segment is created to prevent unrelated patterns from being mixed into the same path. Each candidate segment records the start and end time, average risk weight, main feature domain identifier, and the number of internal windows.

[0101] The joint discrimination criterion of density peak recognition and entropy drift detection is used to evaluate the similarity and risk mutation between candidate evolution trajectory fragments, and highly correlated fragments are spliced ​​to form a complete evolution trajectory; the density of the hidden state distribution of the candidate fragments is estimated to find local density peaks; if the height difference of the density peaks of two fragments is within the tolerance range, it indicates that the two behavior patterns are similar; the entropy of the risk weight distribution within the fragment is calculated; when the entropy value drifts slowly over time rather than suddenly, it indicates that the risk characteristics transition smoothly and can be spliced ​​with adjacent fragments; the system only splices fragments when the density peaks are similar and the entropy drift is smooth, so as to prevent the high-risk mutation segments from being merged with the normal segments; the spliced ​​segments are connected into a complete trajectory on the time axis, and the overall risk volatility, duration and other indicators are recalculated; the load evolution trajectory with time annotations and priority labels is output.

[0102] The process of detecting a candidate safety hazard event in S3 is as follows:

[0103] The load evolution trajectory is dynamically time-warped and compared with the standard safe operation trajectory library to calculate the trajectory shape deviation and time synchronization error. Candidate reference trajectories are screened based on the main feature domain + average risk weight level. Then, the trajectory length is fine-screened within ±10% to ensure similarity in the comparison time series. Improved DTW is performed on the selected reference trajectory and the evolution trajectory to be tested: coarse-grained linear alignment is performed first, followed by iterative alignment with local constraints to reduce computational costs. Trajectory shape deviation: DTW cumulative distance divided by the number of comparison points reflects shape differences. Time synchronization error: the average time difference between the two trajectories at key inflection points after alignment.

[0104] A dual-threshold anomaly scoring function is constructed, where the first threshold is based on cluster density mutations, and the second threshold is based on the magnitude of the entropy drift inflection point. Sliding window clustering is performed on the target trajectory hidden state stream, observing the sample density distribution within the window. A mutation is identified when the density increase or decrease ratio of adjacent windows exceeds 2.5 standard deviations of the mean over the past 30 days. The time derivative of the risk weight distribution entropy is calculated. Any reverse jump with a magnitude exceeding the historical 95th percentile is considered an entropy drift inflection point. Two scores are generated for each trajectory subsequence: a density mutation score and an entropy drift score. These scores are then weighted with a ratio of 0.6:0.4, and the result is written into the anomaly score field.

[0105] If both the deviation and the anomaly score of a trajectory subsequence exceed the corresponding thresholds, the trajectory subsequence is marked as a potential safety hazard segment, and the start and end times of the trigger window are recorded;

[0106] Adjacent potential hidden danger segments are merged in time and space to generate a continuous risk segment. If the time interval between adjacent hidden danger segments is ≤ half a minute, or the physical location difference is ≤ one pile distance, the two segments are determined to belong to the same risk chain and enter the merging process. When merging, the earliest start time and the latest end time of the two segments are used as the boundaries of the new segment, and the risk count is accumulated.

[0107] The output includes candidate security risk events including risk type, impact data domain and risk level.

[0108] S4: The charging entities and coupling ranges corresponding to the candidate safety hazard events are marked as candidate areas, and the connection end nodes, thermal sensing nodes and power control nodes in the candidate areas are integrated to construct a local load priority interference map.

[0109] The process of constructing the local load priority interference map in S4 is as follows:

[0110] According to the list of candidate safety hazard events, the corresponding connection end nodes, thermal sensing nodes and power control nodes are mapped into a heterogeneous node set;

[0111] An event-driven edge weight update strategy is used to establish three types of multimodal edges between heterogeneous node sets: handshake waveform, temperature rise gradient, and power factor. When a candidate event is confirmed and written to the node pool, an edge establishment task is triggered. The system scans all node pairs in parallel. Handshake waveform edges connect two connecting nodes that have handshake protocol interactions. Temperature rise gradient edges are established between pairs of thermal sensing nodes with superimposed heat diffusion in the thermal imaging area. Power factor edges are established if the power factor curves of the power control nodes show synchronous drift in the same time period.

[0112] On each edge, the interference intensity score of the edge is dynamically calculated based on the perception data characteristics, current state change rate and historical risk sensitivity of the nodes at both ends of the edge, and is attached to the graph structure as an edge weight, forming a local heterogeneous graph with time-varying edge properties. After the edge is successfully established, the system subscribes to the real-time streams of the nodes at both ends: handshake phase change rate, temperature rise rate, power factor slope; for each edge, the weighted average of the current state change rate and historical risk sensitivity is calculated: the state change rate reflects the degree of suddenness from the sliding window average of the last 10 seconds; the historical risk sensitivity comes from the proportion of the risk number list length in the node attributes; if the nodes at both ends fluctuate in the same direction at the same time, the score increases; if they fluctuate in different directions, the score decreases; the increase can be divided into five levels: weak, mild, moderate, severe, and extreme, corresponding to a 0.1-0.5 incremental weight adjustment; the final edge weight is refreshed every 2 seconds;

[0113] Through the joint analysis of spectral clustering and node centrality, the core interference source nodes and critical coupling paths in the local heterogeneous graph are identified; the edge weight matrix at the current moment is extracted, and Laplace matrix decomposition is performed on it. The nodes are mapped to the k-dimensional embedding space using the first k eigenvectors; k-means clustering is performed in the embedding space to group closely coupled nodes into the same subgraph, and the k value is automatically selected according to the graph modularity; for each subgraph in the clustering result, weighted betweenness centrality and weighted degree centrality are calculated; the weight comes from the real-time interference intensity of the edge; the higher the centrality, the more critical the node is in risk propagation; a recursive search is performed from the root node to the edge with the largest weight in the subgraph to form a path set; the cumulative interference intensity is calculated for each path, and the top 10% are selected as the key coupling paths; the local load priority interference map containing node attributes, edge weight dynamics and path hierarchy is output.

[0114] S5: Based on the local load priority interference map, combined with the potential unstable nodes and the global risk entropy fluctuation trend in the monitoring data, the perception weight and sampling frame rate are dynamically reconstructed, and charging safety response instructions are generated.

[0115] The process of generating the charging safety response instruction in S5 is as follows:

[0116] Based on the historical load data and current operating status of all sensing nodes, the state entropy of each main variable in the multi-dimensional time series feature tensor is counted, and a global risk entropy fluctuation trend sequence representing the amplitude and direction of state entropy change is constructed;

[0117] The global risk entropy fluctuation trend sequence is integrated with the local load priority interference map to identify the system risk evolution section where the critical path in the map is located, and the cumulative risk value of each critical path is calculated;

[0118] Dynamically adjust the data sampling parameters of the corresponding nodes based on the risk accumulation value of each path and the priority labels of the nodes involved;

[0119] Generate high-frequency sampling and real-time perception instructions for key nodes in high-risk paths and push them to edge computing units for execution. Reduce the sampling frequency for non-critical nodes in low-risk areas.

[0120] Integrate all perception parameter adjustments and control execution actions, and output charging safety response instructions including timestamp, target node number, control parameter set and feedback receipt mechanism.

[0121] Example 2: Please refer to Figure 2 As shown in FIG, the charging pile load test data analysis system based on intelligent perception includes:

[0122] Multi-source mapping module: Dynamically maps and semantically segments charging pile groups and generates data priority sequences based on risk sensitivity;

[0123] Feature construction module: Cross-domain correlation of handshake, thermal image and harmonic data in a multi-scale sliding window, combined with historical and environmental information to form a multi-dimensional time series feature tensor;

[0124] Trajectory analysis module: This module extracts load evolution trajectories based on a priority adaptive model and detects cluster density mutations and entropy drift inflection points to identify candidate safety hazard events.

[0125] Interference map module: This module maps candidate safety hazard events to local areas and integrates connection terminals, thermal sensing, and power control nodes to generate a load priority interference map.

[0126] Safety response module: Dynamically adjusts the perception weight and sampling frame rate based on the interference map and global risk entropy fluctuations to generate charging safety response instructions.

[0127] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0128] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A charging pile load test data analysis method based on intelligent perception, characterized in that: The following steps are involved: Dynamically map charging pile clusters in public parking lots through multi-source heterogeneous perception components, perform misalignment annotation on perception data streams based on scene semantic segmentation networks, and generate data priority sequences based on risk sensitivity. The process of generating data priority sequence according to risk sensitivity is as follows: The perception data stream is formed by collecting multi-source heterogeneous perception components; Based on the scene semantic segmentation network, various perception data streams are processed in time and space synchronously, and feature labels are annotated on the perception data streams using staggered annotation. Based on the risk contribution matrix constructed from the historical accident database, the correlation degree of typical accidents in each data stream is calculated; The correlation degree and the real-time parking lot load density are coupled and scored to obtain the risk sensitivity coefficient reflecting the scene characteristics, and the final data priority sequence is output; Perform multi-scale sliding window cross-domain correlation tracking on the data priority sequence, combining historical waveform fingerprints, power factor drift trajectory and environmental electromagnetic interference matrix to construct a multi-dimensional time series feature tensor of charging load behavior; Based on the multi-dimensional time series feature tensor, a priority adaptive model is used to extract the load evolution trajectory. By identifying the cluster density mutation and entropy drift inflection points, candidate safety hazard events are detected. The charging entities and coupling ranges corresponding to candidate safety hazard events are marked as candidate areas, and the connection end nodes, thermal sensing nodes, and power control nodes in the candidate areas are integrated to construct a local load priority interference map. Based on the local load priority interference map, combined with the potential unstable nodes and the global risk entropy fluctuation trend in the monitoring data, the perception weight and sampling frame rate are dynamically reconstructed, and charging safety response instructions are generated.

2. The charging pile load test data analysis method based on intelligent perception according to claim 1 is characterized in that: The multi-scale sliding window cross-domain correlation tracking process for the data priority sequence is: Prioritize the perceived data streams based on the risk sensitivity coefficient, defining data streams above the set threshold as high-weight data and data streams below the set threshold as low-weight data; Assign short time windows to high-weight data and long time windows to low-weight data, and build a multi-scale sliding window set; Synchronously read the perception data stream within each sliding window, and eliminate cross-domain data delay through timestamp alignment and phase correction methods; A recursive cross-domain correlation matching algorithm is used to track the load waveform shape, thermal image temperature rise gradient, and harmonic amplitude changes between adjacent sliding windows to generate a cross-domain matching vector. Output the integrated multi-scale cross-domain correlation results.

3. The charging pile load test data analysis method based on intelligent perception according to claim 2 is characterized in that: The process of combining historical waveform fingerprints, power factor drift trajectory and environmental electromagnetic interference matrix is ​​as follows: Retrieve a waveform fingerprint template group that matches the current charging pile device model from the historical database, match the template group with the real-time handshake waveform collected within the sliding window, and output a similarity score vector; Exponentially weighted recursive smoothing is performed on the real-time collected power factor data to extract short-term oscillation features reflecting instantaneous fluctuations and long-term drift features describing trend changes, forming a dual-channel feature vector. An environmental electromagnetic interference matrix based on sensor array data is constructed. The interference degree of each frequency band noise on the handshake waveform and harmonic pulse is calculated through frequency band analysis method, and the frequency domain interference influence weight vector is obtained. The waveform similarity score vector, dual-channel power factor feature vector, and frequency domain interference weight vector are uniformly mapped to the multi-dimensional feature embedding space through the tensor embedding mechanism to form an embedding result tensor; The multi-core kernel function fusion algorithm is used to perform nonlinear feature superposition on the embedded result tensor to generate a set of comprehensive weight coefficients for time series modeling and risk judgment.

4. The charging pile load test data analysis method based on intelligent perception according to claim 3 is characterized in that: The process of constructing the multi-dimensional time series feature tensor of charging load behavior is as follows: The perception data stream and comprehensive weight coefficient set in the cross-domain correlation results are stored in the tensor draft according to the three-dimensional index of time axis, data domain and risk weight; Apply double threshold correction for missing anomalies to the tensor draft, use local interpolation to repair short-term missing samples, and use random forest discrimination to remove extreme outliers; The redundant dimensions are reduced by tensor orthogonal projection mapping to form a multi-dimensional time series feature tensor.

5. The charging pile load test data analysis method based on intelligent perception according to claim 4 is characterized in that: The process of extracting load evolution trajectory using the priority adaptive model is as follows: Adopting an adaptive network combining gated recurrent units and attention mechanism to perform sequence encoding on multi-dimensional temporal feature tensors; Embed a data priority sequence during the encoding process and use a priority gating unit to dynamically adjust the hidden state update rate of different perception data streams; Perform aggregate analysis on continuous hidden states to generate several candidate evolution trajectory segments; The similarity and risk mutation between candidate evolution trajectory segments are evaluated using a joint discriminant criterion of density peak identification and entropy drift detection. Output load evolution trajectory with time and priority labels.

6. The charging pile load test data analysis method based on intelligent perception according to claim 5 is characterized in that: The process of detecting candidate safety hazard events is as follows: Perform dynamic time-warping comparison on the load evolution trajectory and the standard safe operation trajectory library to calculate the trajectory shape deviation and time synchronization error; Construct a dual-threshold anomaly scoring function, where the first threshold is based on the cluster density mutation and the second threshold is based on the entropy drift inflection point amplitude; If both the deviation and the anomaly score of a trajectory subsequence exceed the corresponding thresholds, the trajectory subsequence is marked as a potential safety hazard segment; Adjacent potential hidden danger segments are merged in time and space to generate risk continuous segments and output candidate safety hidden danger events.

7. The charging pile load test data analysis method based on intelligent perception according to claim 6 is characterized in that: The process of constructing the local load priority interference map is as follows: According to the list of candidate safety hazard events, the corresponding connection end nodes, thermal sensing nodes and power control nodes are mapped into a heterogeneous node set; An event-driven edge weight update strategy is adopted to establish three types of multimodal edges between heterogeneous node sets; On each edge, the interference intensity score of the edge is dynamically calculated based on the perception data characteristics, current state change rate and historical risk sensitivity of the nodes at both ends of the edge, and is attached to the graph structure as an edge weight to form a local heterogeneous graph; The core interference source nodes and key coupling paths in the local heterogeneous graph are identified through the joint analysis of spectral clustering and node centrality, and the local load priority interference map is output.

8. The charging pile load test data analysis method based on intelligent perception according to claim 7 is characterized in that: The process of generating a charging safety response instruction is as follows: Based on the historical load data and current operating status of all sensing nodes, the state entropy of each main variable in the multi-dimensional time series feature tensor is counted, and a global risk entropy fluctuation trend sequence representing the amplitude and direction of state entropy change is constructed; The global risk entropy fluctuation trend sequence is integrated with the local load priority interference map to identify the system risk evolution section where the critical path in the map is located, and the cumulative risk value of each critical path is calculated; Dynamically adjust the data sampling parameters of the corresponding nodes based on the risk accumulation value of each path and the priority labels of the nodes involved; Integrate all perception parameter adjustments and control execution actions to output charging safety response instructions.

9. A charging pile load test data analysis system based on intelligent perception, applied to the method according to any one of claims 1 to 8, characterized in that: include: Multi-source mapping module: Dynamically maps and semantically segments charging pile groups and generates data priority sequences based on risk sensitivity; Feature construction module: Cross-domain correlation of handshake, thermal image and harmonic data in a multi-scale sliding window, combined with historical and environmental information to form a multi-dimensional time series feature tensor; Trajectory analysis module: This module extracts load evolution trajectories based on a priority adaptive model and detects cluster density mutations and entropy drift inflection points to identify candidate safety hazard events. Interference map module: This module maps candidate safety hazard events to local areas and integrates connection terminals, thermal sensing, and power control nodes to generate a load priority interference map. Safety response module: Dynamically adjusts the perception weight and sampling frame rate based on the interference map and global risk entropy fluctuations to generate charging safety response instructions.

Images