A mobile power supply fault intelligent detection method based on electrical performance data

Abstract

The present application relates to a kind of mobile power supply fault intelligent detection method based on electrical performance data.First, four-dimensional time series data such as voltage, current, temperature and estimated internal resistance are collected, denoised, resampled and windowed, and phase space reconstruction is realized by delay coordinate embedding, generating multiple sets of trajectory point clouds reflecting the evolution of operating point.Subsequently, according to the physical signal flow direction, each parameter point cloud is fused to form a topological graph with energy cycle closure, the topological index of nodes and edges is calculated to extract the system health vector, and the normal operating condition distribution model is used to identify abnormal patterns.After detecting the abnormality, the key fault parameters and time period are locked by reverse mapping, and the stage semantic judgment is carried out combined with the operating condition finite state machine model.The method can realize high-sensitivity monitoring of mobile power supply health status, automatic identification of abnormality and fault positioning, and improve system operation reliability.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnosis and fault detection technology, and in particular to an intelligent fault detection method for mobile power supplies based on electrical performance data. Background Technology

[0002] With the widespread adoption of portable electronic devices, power banks, as core components providing backup power for smart terminals and wearable devices, have become a focus of industry attention due to their safety, reliability, and health management capabilities. Currently, fault detection and health status assessment of power banks mainly rely on the real-time acquisition and analysis of electrical performance parameters (such as voltage, current, temperature, and internal resistance). However, in the low-power, high-constraint hardware environment of embedded devices, achieving accurate, early, and interpretable fault identification still faces many challenges.

[0003] Current mainstream detection and diagnosis solutions fall into two main categories: First, shallow machine learning methods based on artificial feature engineering, where experts typically extract customized indicators such as mean, variance, and frequency domain features from electrical performance data based on experience, and combine these with models such as support vector machines and decision trees to achieve anomaly detection. Second, end-to-end deep learning models that have emerged in recent years, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers, automatically learn complex fault modes through training on large-scale labeled samples. The former has significant advantages in terms of computing power and interpretability, but it is highly dependent on artificial feature design, struggles to capture high-order hidden patterns in time-series data, and has poor robustness to data disturbances or novel fault types. The latter, while capable of discovering complex patterns, is limited by embedded MCU resources, making direct deployment at the edge difficult, and requires a large number of labeled samples and long-term iterative training. Furthermore, its "black box" structure weakens the physical interpretability of the diagnostic results.

[0004] In summary, the current field of mobile power bank fault detection urgently needs a method that does not require large-scale labeled samples, does not rely on complex manual feature engineering, and does not require additional sensor hardware or high-computing platforms. This method should be able to efficiently, with low power consumption and interpretability, capture latent fault modes under complex operating conditions and achieve early warning and classification. Summary of the Invention

[0005] This application provides a mobile power bank fault intelligent detection method based on electrical performance data, which aims to solve the problems mentioned in the background art, such as the difficulty in direct deployment at the edge due to the limitation of embedded MCU resources, and the need to rely on a large number of labeled samples and long-term iterative training.

[0006] This application provides a method for intelligent fault detection of mobile power banks based on electrical performance data, specifically including: S1: Obtain the four-dimensional electrical performance timing data stream of voltage, current, temperature and estimated internal resistance in the embedded MCU of the mobile power supply, and perform window truncation processing on the four-dimensional electrical performance timing data stream with a fixed sampling interval to obtain the original timing data window; S2: Based on the electrical performance time-series data stream of each dimension in the original time-series data window, select the optimal time delay and embedding dimension to perform phase space reconstruction processing to generate four sets of independent electrical parameter phase space trajectory point clouds; S3: Based on the physical signal flow constraints from the charging IC inside the power bank to the battery cell and then to the protection board, the topology map of the combined operation trajectory of electric-thermal-resistance is obtained based on the point cloud of the phase space trajectory of four independent electrical parameters. S4: Calculate three types of low-dimensional topological invariants based on the combined electric-thermal-resistance operating trajectory topology map to generate a topological health vector; S5: Input the topology health vectors generated by multiple consecutive sliding windows into the normal operating condition distribution model, identify and output abnormal pattern clusters; S6: When the abnormal pattern cluster is detected, the corresponding original phase space point is locked and mapped back to the original four-dimensional time series coordinates. By statistically analyzing the fluctuation intensity and phase shift consistency of each dimension during the abnormal period, a fault feature tracing result including fault sensitivity parameter sorting and key period marking is generated. S7: Combine the preset finite state machine model to align the key time period markers in the fault feature tracing results to the specific working stages of constant current, constant voltage, or trickle flow, so as to output the preliminary judgment result of the mobile power supply fault type.

[0007] This application provides a mobile power bank fault intelligent detection method based on electrical performance data, which has the following beneficial effects: By reconstructing four-dimensional electrical performance time-series data (voltage, current, temperature, and estimated internal resistance) into a phase space trajectory point cloud, and combining this with cross-graph edge constraint fusion of the mobile power supply's internal circuit topology, a "joint electrical-thermal-resistance operating trajectory topology map" is constructed, enabling high-order nonlinear dynamic modeling of the equipment's operating state. Compared to traditional fault detection methods based on threshold comparison or statistical feature classification, this scheme eliminates the reliance on explicit fault labels and manual feature engineering, enabling the capture of early degradation signs hidden in dynamic responses, such as slow SEI film thickening or gradual deterioration of protection board contact resistance—potential risks that are difficult to judge through instantaneous values. This topology map uses nodes to represent transient operating points, edges to describe state transition paths, and loops to reflect energy cycle characteristics, elevating system behavior from static parameter monitoring to dynamic evolution analysis. This significantly improves the sensitivity and foresight of anomaly identification, maintaining stable discrimination capabilities, especially under complex operating condition switching. Attached Figure Description

[0008] Figure 1This is the main flowchart of a mobile power bank fault intelligent detection method based on electrical performance data.

[0009] Figure 2 This is a sub-flowchart of a mobile power bank fault intelligent detection method based on electrical performance data.

[0010] Figure 3 This is another sub-flowchart of a mobile power bank fault intelligent detection method based on electrical performance data. Detailed Implementation

[0011] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0012] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0013] like Figure 1 As shown, this application provides a method for intelligent fault detection of mobile power banks based on electrical performance data, specifically including: S1: Obtain the four-dimensional electrical performance timing data stream of voltage, current, temperature and estimated internal resistance in the embedded MCU of the mobile power supply, and perform window truncation processing on the four-dimensional electrical performance timing data stream with a fixed sampling interval to obtain the original timing data window; S2: Based on the electrical performance time-series data stream of each dimension in the original time-series data window, select the optimal time delay and embedding dimension to perform phase space reconstruction processing to generate four sets of independent electrical parameter phase space trajectory point clouds; S3: Based on the physical signal flow constraints from the charging IC inside the power bank to the battery cell and then to the protection board, the topology map of the combined operation trajectory of electric-thermal-resistance is obtained based on the point cloud of the phase space trajectory of four independent electrical parameters. S4: Calculate three types of low-dimensional topological invariants based on the combined electric-thermal-resistance operating trajectory topology map to generate a topological health vector; S5: Input the topology health vectors generated by multiple consecutive sliding windows into the normal operating condition distribution model, identify and output abnormal pattern clusters; S6: When the abnormal pattern cluster is detected, the corresponding original phase space point is locked and mapped back to the original four-dimensional time series coordinates. By statistically analyzing the fluctuation intensity and phase shift consistency of each dimension during the abnormal period, a fault feature tracing result including fault sensitivity parameter sorting and key period marking is generated. S7: Combine the preset finite state machine model to align the key time period markers in the fault feature tracing results to the specific working stages of constant current, constant voltage, or trickle flow, so as to output the preliminary judgment result of the mobile power supply fault type.

[0014] Step S1: Acquire the four-dimensional electrical performance timing data stream of voltage, current, temperature, and estimated internal resistance from the embedded MCU of the mobile power supply, and perform window truncation processing on the four-dimensional electrical performance timing data stream at fixed sampling intervals to obtain the original timing data window. Specifically, this includes: S1.1: Acquire the analog electrical signal output in real time from the analog-to-digital conversion interface inside the embedded microcontroller of the power bank, and use multi-channel synchronous sampling technology to perform parallel acquisition and processing of voltage signal, current signal, temperature sensing signal and internal resistance estimation signal to generate a four-dimensional raw electrical performance data stream containing timestamps.

[0015] Sub-step S1.1 is located at the beginning of the data acquisition process and is responsible for converting the analog electrical signals of the physical world inside the power bank into a timing data stream that can be processed by the digital system. This step directly affects the analog-to-digital converter (ADC) interface and the front-end signal conditioning circuit of the embedded microcontroller unit (MCU).

[0016] Configure the multi-channel ADC module inside the MCU, enable synchronous sampling mode or hardware triggering mechanism, set the sampling priority and gain parameters of the four channels for voltage, current, temperature and internal resistance estimation, and ensure that each channel completes signal locking within a single clock cycle to eliminate time between channels.

[0017] The battery terminal voltage signal is acquired by a voltage divider resistor network, the current voltage drop signal is obtained by a precision sampling resistor connected in series in the circuit, the battery surface temperature signal is read by an NTC thermistor voltage divider circuit, and the battery internal resistance is estimated in real time by combining the open circuit voltage method or AC injection method, forming a four-channel parallel analog input source.

[0018] The ADC conversion sequence is started to quantize the four analog signals, mapping the continuously changing voltage amplitude to discrete digital coded values, and generating a four-dimensional digital signal sample containing the original quantization noise.

[0019] A high-precision system timestamp is attached to each sampling point, with a timestamp accuracy down to the microsecond level, for subsequent time alignment and phase analysis, forming a four-dimensional raw data tuple with a time index.

[0020] DMA (Direct Memory Access) technology is used to transfer ADC conversion results to the SRAM buffer in batches, reducing CPU interrupt load and ensuring data integrity and real-time performance under high-frequency sampling.

[0021] By using multi-channel synchronous sampling and timestamp marking processing, the analog electrical signal from the previous step is transformed into a four-dimensional raw electrical performance data stream containing timestamps, realizing the spatiotemporal alignment of multi-source heterogeneous data and providing a high-fidelity data foundation for subsequent denoising and reconstruction.

[0022] S1.2: Based on the high-frequency noise components in the original four-dimensional electrical performance data stream, a moving average filtering method is used to perform digital smoothing and denoising processing to eliminate the influence of sensor quantization error and electromagnetic interference, thereby generating a pure four-dimensional electrical performance data stream with high signal-to-noise ratio characteristics.

[0023] S1.3: For the pure four-dimensional electrical performance data stream, perform equal-interval resampling operation according to the preset 10 Hz sampling frequency standard to unify the time base of the data in each dimension and eliminate transmission jitter, and generate a standard four-dimensional electrical performance time series with constant time resolution.

[0024] Receive a clean four-dimensional electrical performance data stream after it has been processed by moving average filtering. This data stream contains time series data of four dimensions: voltage, current, temperature, and estimated internal resistance. The timestamps of each dimension may have microsecond-level jitter.

[0025] A unified time reference grid is constructed, based on a preset 10 Hz sampling frequency, to generate a series of equally spaced standard time points covering the entire duration of the current data window. The time interval between adjacent standard time points is strictly fixed at 100 milliseconds.

[0026] For each dimension of clean time-series data, a linear interpolation method is used to calculate the signal amplitude at non-standard time points. By finding two adjacent original sampling points and their corresponding timestamps, the interpolation result at the standard time point is determined using the proportional relationship.

[0027] The above linear interpolation operation is performed on the four dimensions of voltage, current, temperature and internal resistance respectively to ensure that all dimensions have corresponding values ​​on the same standard time point sequence, thus eliminating time alignment errors caused by multi-channel ADC acquisition delay or transmission jitter.

[0028] The interpolated four-dimensional data are reorganized in chronological order to form a standard four-dimensional electrical performance time series with a matrix structure. Each row represents a standard time point, and each column represents an electrical performance dimension, ensuring strict synchronization and constant resolution of the data on the time axis.

[0029] By using an equal-interval resampling process, the non-uniform sampling data from the previous step is transformed into a standard four-dimensional electrical performance time series with constant time resolution, thereby unifying the time base of multi-source heterogeneous data and providing high-quality time-aligned input data for subsequent phase space reconstruction.

[0030] S1.4: Apply a sliding window truncation strategy with a fixed duration of thirty seconds to the standard four-dimensional electrical performance time series, extract continuous data segments according to the current system clock cycle, so as to form a set of candidate time series data windows with a unified time span constraint.

[0031] S1.5: Based on the data integrity verification results in the candidate time series data window set, invalid windows with packet loss or out-of-bounds anomalies are removed, and the original time series data window with complete structure and time alignment is finally output as the direct input object for phase space reconstruction processing. The phase space reconstruction processing is the core step of converting a one-dimensional time series into a multi-dimensional phase space trajectory using the delayed coordinate embedding method to recover the dynamic structure of the chaotic system.

[0032] Perform multi-dimensional integrity checks on the candidate time-series data window set and construct a joint judgment logic based on timestamp continuity, data volume consistency and physical boundary constraints of each dimension.

[0033] Iterate through the timestamp sequence within each candidate window and calculate the time interval difference between adjacent sampling points. The uniformity of the time base can be verified using the following formula: in, Let be the timestamp of the i-th sampling point. If there exists any If the deviation from the preset sampling period threshold is 0.05, the window is marked as having timing jitter abnormality.

[0034] The system counts the number of valid data points in four dimensions—voltage, current, temperature, and estimated internal resistance—within the statistical window and compares them with the preset standard sampling number. If the missing data point rate in any dimension exceeds the tolerance limit, it is judged as an abnormal communication packet loss and is removed.

[0035] Extreme value out-of-bounds detection is performed on data from various dimensions. A safe range is set based on the chemical characteristics of the power bank battery, and illegal values ​​are filtered out using inequality constraints. in, For the i-th voltage sample value, and These are the battery cutoff voltage and charging limit voltage, respectively. Similarly, boundary checks are performed on the current, temperature, and internal resistance. Outliers that exceed the physically reasonable range are marked as sensor malfunctions.

[0036] Based on the combined results of the timing, quantity, and boundary checks, a logical AND operation is used to generate the final validity flags, retaining only windows where all flags are normal.

[0037] By using a multi-dimensional joint integrity verification process, the candidate time series data window generated in the previous step is transformed into a structurally complete, time-aligned, and physically meaningful original time series data window, thereby ensuring high reliability of the input data for phase space reconstruction.

[0038] Step S2: Based on the electrical performance time-series data stream of each dimension in the original time-series data window, select the optimal time delay and embedding dimension for phase space reconstruction processing to generate four sets of independent electrical parameter phase space trajectory point clouds. Specifically, this includes: S2.1: Perform detrending and normalization preprocessing on the voltage time series, current time series, temperature time series and estimated internal resistance time series in the original time series data window to eliminate dimensional differences and remove linear drift interference, and generate a standardized four-dimensional electrical performance time series data stream.

[0039] S2.2: Based on the standardized four-dimensional electrical performance time-series data stream, the autocorrelation decay characteristics of each dimension of the data are calculated using the mutual information method to determine the optimal time delay parameter. At the same time, the nearest neighbor method is used to analyze the phase space expansion degree to determine the optimal embedding dimension, and a phase space reconstruction configuration parameter set containing the optimal time delay parameter and the optimal embedding dimension is generated.

[0040] It receives a standardized four-dimensional electrical performance time-series data stream and performs adaptive optimization processing of phase space reconstruction parameters for each of the four independent dimensional sequences: voltage, current, temperature, and estimated internal resistance.

[0041] The mutual information method is used to calculate the statistical dependence between time series of each dimension and their delayed replicas, and a mutual information function curve is constructed to quantify the degree of nonlinear correlation. The preset time delay search interval is traversed, and the mutual information value under different delay step sizes is calculated. The time step corresponding to the first occurrence of a local minimum in the mutual information function is selected as the optimal time delay parameter for that dimension, ensuring that the reconstructed coordinate components maintain both independence and dynamic correlation.

[0042] Based on the determined optimal time delay parameter, the nearest neighbor method is applied to evaluate the sufficiency of phase space unfolding. The embedding dimension is gradually increased, and the proportion of points incorrectly identified as neighbors at the current dimension is calculated. When the proportion of false nearest neighbors drops below a preset threshold, it is determined that the phase space trajectory has been fully unfolded and has no self-intersection or overlap. This critical dimension is then determined as the optimal embedding dimension.

[0043] The above optimization process is repeated for each of the four dimensions to generate a set of configuration parameters containing the optimal time delay and optimal embedding dimension for each of voltage, current, temperature and internal resistance, providing an accurate geometric mapping reference for subsequent delay coordinate embedding.

[0044] By using mutual information minimum location and false nearest neighbor ratio convergence criterion, the original one-dimensional time series is transformed into a phase space reconstruction configuration parameter set with optimal topological expansion characteristics, realizing efficient and lossless mapping from scalar observation to multi-dimensional state space, and significantly improving the geometric resolution of fault feature extraction.

[0045] For example, for a standardized voltage sequence, the maximum delay search range is set to 50 sampling points. The mutual information curve is calculated to have its first local minimum at a delay step size of 8, thus determining the optimal voltage time delay to be 8. Subsequently, the nearest neighbor method is applied, with a threshold of 0.05. When the embedding dimension increases from 2 to 3, the proportion of false nearest neighbors decreases from 0.12 to 0.03, determining the optimal embedding dimension for voltage to be 3. Similarly, the optimal time delays for current, temperature, and internal resistance are calculated to be 6, 12, and 10, respectively, with the optimal embedding dimension for each being 3. Finally, a configuration parameter set {voltage: (8, 3), current: (6, 3), temperature: (12, 3), internal resistance: (10, 3)} is generated, ensuring that the subsequently generated phase space point cloud has a clear manifold structure and effectively avoiding feature aliasing caused by trajectory folding.

[0046] S2.3: Based on the optimal time delay parameter and optimal embedding dimension in the phase space reconstruction configuration parameter set, perform a delay coordinate embedding transformation operation on each dimension of the standardized four-dimensional electrical performance time series data stream, map the one-dimensional scalar time series into a multi-dimensional state vector sequence, and generate voltage phase space trajectory point clouds, current phase space trajectory point clouds, temperature phase space trajectory point clouds and internal resistance phase space trajectory point clouds that characterize the instantaneous state distribution of the system.

[0047] The system receives a standardized four-dimensional electrical performance time-series data stream, along with the optimal time delay parameters and optimal embedding dimension determined by the preceding steps, as the input basis for phase space reconstruction.

[0048] For the four independent dimensions of voltage, current, temperature, and estimated internal resistance, delay coordinate mapping matrices are constructed respectively to transform the one-dimensional scalar sequence into a multi-dimensional state vector.

[0049] Using the Takens embedding theorem, a sliding window slicing operation is performed on each dimension of standardized time series data. The slice step size is determined based on the optimal time delay parameter, and the vector length is determined based on the optimal embedding dimension.

[0050] Traverse the entire time window to generate a point cloud set consisting of continuous state vectors, where each vector represents a transient operating point of the system in phase space.

[0051] The mapping process is repeated for the four dimensions of voltage, current, temperature and internal resistance respectively, generating four sets of independent phase space trajectory point clouds.

[0052] The generated point cloud is aligned to ensure that the indices of the point cloud in each dimension correspond strictly on the time axis, thus preserving the dynamic evolution characteristics of the original time series.

[0053] By using delayed coordinate embedding transformation, the one-dimensional standardized time-series data from the previous step is transformed into four sets of independent electrical parameter phase space trajectory point clouds that characterize the instantaneous state distribution of the system, thereby achieving a dimensionality upgrade from the time domain to the phase space and revealing the hidden dynamic structural features.

[0054] S2.4: Perform local neighborhood density estimation and noise point filtering on the generated voltage phase space trajectory point cloud, current phase space trajectory point cloud, temperature phase space trajectory point cloud and internal resistance phase space trajectory point cloud respectively, in order to remove outlier trajectory points caused by sensor quantization error and generate four sets of independent electrical parameter phase space trajectory point clouds with clear geometric topology.

[0055] Step S3: Based on the physical signal flow constraints from the charging IC inside the power bank to the battery cell and then to the protection board, a topology map of the combined electrical-thermal-resistance operating trajectory is obtained based on the point cloud of the phase space trajectory of four sets of independent electrical parameters. Specifically, this includes: S3.1: Obtain the four sets of independent electrical parameter phase space trajectory point clouds generated in the previous steps. Based on the physical signal flow constraints inside the mobile power supply from the charging IC to the battery cell and then to the protection board, perform semantic mapping processing on the nodes in the four sets of independent electrical parameter phase space trajectory point clouds to generate a constrained phase space node set with circuit cascade attribute tags.

[0056] The system receives four independent phase space trajectory point clouds generated in the previous steps, representing voltage, current, temperature, and estimated internal resistance. It extracts the timestamp index and three-dimensional coordinate vector of each node in each point cloud to construct an original node dataset containing spatiotemporal attributes.

[0057] The internal circuit topology of the power bank is analyzed, the physical signal cascade path from the output of the charging IC through the battery cell to the MOSFET of the protection board is defined, and a hierarchical mapping rule based on causal flow is established. Voltage nodes are marked as driving layers, current nodes are marked as response layers, and temperature and internal resistance nodes are marked as state feedback layers.

[0058] Hierarchical attribute injection is performed on each node in the original node dataset. Based on the timestamp synchronization principle, the voltage phase space point, current phase space point, temperature phase space point and internal resistance phase space point at the same time are logically bound to generate a multi-dimensional node tuple with a unified time base.

[0059] Circuit cascade constraint labels are introduced into the multidimensional node tuples. The dependency weights between nodes are defined according to the direction of physical signal flow. The unidirectional influence coefficient of voltage node on current node is set to 1.0, the Joule thermal coupling coefficient of current node on temperature node is the reciprocal of the resistance value, and the Arrhenius relation coefficient of temperature node on internal resistance node is the activation energy parameter, forming a weighted directed correlation matrix.

[0060] The semantic enhancement of multidimensional node tuples is performed using a weighted directed correlation matrix. Each node is appended with structured metadata, including its circuit level, upstream driving source identifier, and downstream feedback target identifier, to generate a constrained phase space node set with circuit cascade attribute tags.

[0061] Through the hierarchical mapping and semantic enhancement processing described above, the geometric discrete point cloud generated in the previous step is transformed into structured graph node data containing physical causal relationships, realizing semantic alignment from pure mathematical space to physically interpretable space, and providing a topological foundation with clear electrical meaning for subsequent cross-graph edge constraint fusion.

[0062] S3.2: Based on the constrained phase space node set with circuit cascade attribute markers, perform k-nearest neighbor connection strategy calculation within each set of independent electrical parameter phase space trajectory point clouds to generate four sets of local adjacency relationship subgraphs representing the transient evolution path of a single parameter.

[0063] Receive a constrained set of phase space nodes with circuit cascade attribute tags, and extract the node coordinate vectors from the phase space trajectory point cloud of four independent electrical parameters: voltage, current, temperature, and internal resistance.

[0064] For each set of phase space trajectory point clouds, a KD tree index structure is constructed to accelerate distance query operations in high-dimensional space and reduce computational complexity.

[0065] The nearest neighbor parameter k is set to 6. Based on the Euclidean distance metric, a k-nearest neighbor search operation is performed on each node in the point cloud to retrieve the k candidate neighbor nodes that are geometrically closest to it.

[0066] An adaptive distance threshold filtering mechanism is introduced to remove abnormal connecting edges whose distance exceeds three standard deviations of the local density mean, thus ensuring the physical tightness of adjacency relationships.

[0067] Each selected adjacent edge is assigned a weight, which is defined as the reciprocal of the Euclidean distance between nodes, reflecting the probability strength of state transitions.

[0068] Based on the calculated edge weights, a sparse adjacency matrix is ​​constructed. The matrix elements represent the connection strength between nodes, and non-zero elements correspond to existing physical connections.

[0069] The adjacency matrix is ​​symmetricized to ensure the undirectedness of the graph structure, forming a local adjacency subgraph that represents the transient evolution path of a single parameter.

[0070] By using the k-nearest neighbor connection strategy, the discrete phase space point cloud from the previous step is transformed into a local adjacency subgraph with topological connectivity, realizing the mapping from point set to graph structure and providing basic units for cross-graph fusion.

[0071] S3.3: Based on the cascade logic defined by the physical signal flow direction constraint, perform cross-graph edge constraint fusion operation on the four sets of local adjacency relationship subgraphs to establish state transition association edges between point clouds of different electrical parameters phase space trajectories, so as to generate an initial fusion topology graph structure containing multi-dimensional coupling information.

[0072] It receives four sets of local adjacency subgraphs and a set of phase space nodes marked with circuit cascading attributes, and extracts the nodes in each subgraph that correspond to the same sampling time index as cross-graph mapping reference points.

[0073] Based on the physical signal flow from the charging IC to the battery cell and then to the protection board inside the power bank, a causal coupling weight coefficient matrix is ​​defined between voltage-current, current-internal resistance, and internal resistance-temperature to establish the priority logic for cross-graph edge connections.

[0074] For any two reference node pairs in quantum graphs with different electrical parameters, calculate their normalized Euclidean distance and time synchronization deviation in their respective phase spaces to construct a multidimensional similarity metric.

[0075] The confidence weights of cross-graph connection edges are calculated using the following formula. : in, For normalized Euclidean distance, For scale parameters, This is the phase synchronization deviation angle.

[0076] Node pairs with confidence weights exceeding a preset threshold of 0.7 are selected, and directed connections across the graph are established to form an initial fusion topology that represents the multi-parameter collaborative evolution.

[0077] By using a cross-graph edge constraint fusion processing method, the independent subgraphs generated in the previous step are transformed into an initial fused topology graph structure containing multi-dimensional coupling information, thereby achieving the expected technical effect of explicit modeling of the dynamic relationship between electricity, heat and resistance.

[0078] S3.4: Verify and optimize the loop closure characteristics in the initial fusion topology structure, remove redundant connection edges that violate the law of energy conservation and complete key energy cycle paths to generate an electrical-thermal-resistance joint operation trajectory topology with strict energy cycle closure characteristics.

[0079] The system receives the initial fused topology graph structure containing multidimensional coupling information generated in the preceding steps. This structure is formed by fusing four sets of phase space subgraphs of voltage, current, temperature and internal resistance through cross-graph edge constraints. Nodes represent transient operating points and edges represent state transition paths.

[0080] Traverse all closed loops in the initial fusion topology graph, and calculate the algebraic sum of the weights of each edge in each loop based on Kirchhoff's voltage law and the principle of energy conservation to construct a loop energy balance verification index.

[0081] Calculate the energy residual, set an energy residual threshold, compare the calculated energy residual with the threshold, and identify redundant connection edges whose residuals exceed the threshold. These edges are usually caused by sensor noise or nonlinear interference, which violate the energy conservation constraints of the physical system.

[0082] Perform pruning operations to remove redundant connection edges from the initial merged topology graph, disconnect non-physical spurious topology connections, and clean up the physical consistency of the graph structure.

[0083] To address topological breaks or missing critical energy cycle paths caused by pruning operations, a dynamic time warping algorithm is used to analyze the phase synchronization of adjacent nodes in the time series.

[0084] Based on the phase synchronization score, high-confidence virtual connection edges are added between the broken critical nodes to ensure that the signal flow from the charging IC to the battery cell and then to the protection board forms a complete closed loop on the topology graph.

[0085] The completed topology graph is tested for connectivity to confirm that all critical working nodes are in at least one valid closed loop and that the loop direction conforms to the energy flow logic of the charging and discharging process.

[0086] Through the above loop closure characteristic verification and optimization process, the initial fusion topology map of the previous step is transformed into an electrical-thermal-resistance joint operation trajectory topology map with strict energy cycle closure characteristics, thereby achieving the expected technical effect of eliminating physical paradoxical connections and strengthening the topological expression of fault characteristics.

[0087] Step S4: Calculate three types of low-dimensional topological invariants based on the combined electric-thermal-resistance operating trajectory topology map to generate a topological health vector. Specifically, this includes: S4.1: Based on the node set and adjacency matrix in the joint operation trajectory topology of electric-thermal-resistance, perform temporal difference calculation on the degree centrality index of each node within the sliding time window to generate a node degree centrality decay rate sequence that characterizes the instability trend of key operating points.

[0088] The system receives the topology map of the combined electric-thermal-resistance operating trajectory generated in the previous steps, and extracts the node set and corresponding adjacency matrix from the map as input for calculation. It iterates through each node in the node set, counts the number of valid connections it establishes with other nodes within the current sliding time window, and calculates the degree centrality index of that node. Based on the temporal continuity of the power bank's charging and discharging process, it obtains the historical degree centrality index of the corresponding node in the previous time window. A degree centrality decay rate calculation model is constructed, and the node degree centrality decay rate is calculated using the following formula: in, The degree centrality decay rate is the rate at which nodes decay. Degree centrality of the current window node The degree centrality of the nodes in the previous window is determined. The above difference operation is performed on all critical operating point nodes in the topology graph to generate a degree centrality decay rate sequence containing instability trend information for each node. This sequence quantifies the connectivity stability changes of phase space trajectory points within their local neighborhoods, reflecting the smoothness of electrochemical state transitions within the battery. Through temporal difference processing, the static topological connectivity structure is transformed into a dynamic instability trend indicator, enabling real-time capture of the microstructural evolution of critical operating points.

[0089] S4.2: Using the transient operating point distribution state identified by the node degree centrality decay rate sequence, perform Shannon entropy calculation on the probability density function of the weight distribution of the connecting edges in the topology graph of the electric-thermal-resistance joint operation trajectory to generate an edge weight distribution entropy scalar that characterizes the degree of nonlinear distortion of the current and voltage response.

[0090] The node degree centrality decay rate sequence generated in the previous steps is received and used as a dynamic weight adjustment factor to correct the original physical impedance weights of each connecting edge in the topology diagram of the combined electric-thermal-resistance operating trajectory. Based on the nonlinear energy transfer characteristics of the internal circuitry of the mobile power supply, the weight values ​​of each connecting edge in the topology diagram are normalized to construct an edge weight distribution set reflecting the probability of transient energy flow. Based on this edge weight distribution set, the continuous weight values ​​are discretized into several equal-width intervals using histogram statistics, and the proportion of edges in each interval to the total number of edges is calculated to generate an edge weight probability density function. The Shannon entropy formula is used to perform uncertainty measurement calculations on the edge weight probability density function to quantify the degree of nonlinear distortion of the current and voltage responses in the phase space trajectory.

[0091] S4.3: Based on the non-uniformity of energy transfer reflected by the edge weight distribution entropy scalar, area integral and time-series ratio analysis are performed on the smallest closed loop composed of three or more edges in the topology diagram of the combined electric-thermal-resistance operation trajectory to generate the minimum loop area shrinkage rate parameter that reflects the degradation of the battery polarization recovery capability.

[0092] Receive the topology map of the combined electrical-thermal-resistance trajectory and its adjacency matrix, and identify all minimal closed loops consisting of three or more edges in the map. For each minimal closed loop, extract the three-dimensional coordinate sequence of the nodes in phase space and construct a closed geometric path.

[0093] The geometric area of ​​each minimum closed loop projected onto the phase space plane is calculated using Green's formula or the shoelace formula.

[0094] Establish an area time series within a sliding time window, and select corresponding loops in the current window and the previous window for matching. Calculate the area ratio of loops with the same physical meaning at continuous time steps, and define it as the instantaneous shrinkage factor.

[0095] An exponentially weighted moving average was applied to the instantaneous shrinkage factor over M consecutive sliding windows to eliminate transient noise interference and generate a smooth minimum loop area shrinkage rate parameter. This parameter quantifies the degradation trend of the battery's polarization recovery capability; continuous area shrinkage indicates that increased internal resistance leads to decreased energy cycle efficiency.

[0096] By using the above-mentioned area integral and time-series ratio analysis method, the topological characteristics of the previous step are transformed into the minimum loop area shrinkage rate parameter, which reflects the degradation of the battery polarization recovery capability, thus realizing the quantitative capture of early weak fault characteristics.

[0097] S4.4: Combining the node degree centrality decay rate sequence, the edge weight distribution entropy scalar, and the minimum loop area shrinkage rate parameter, perform multi-dimensional feature vector concatenation and normalization processing to generate a topological health vector containing multi-dimensional health evolution information of the system.

[0098] Step S5: Input the topology health vectors generated by multiple consecutive sliding windows into the normal operating condition distribution model, identify and output abnormal pattern clusters. Specifically, this includes: S5.1: Obtain the topology health vectors of multiple consecutive sliding windows output from the previous steps. Construct a reference benchmark dataset based on the set of topology health vectors under historical normal operating conditions of the power bank. Use the kernel density estimation method to fit the probability density distribution of the reference benchmark dataset to generate a normal operating condition distribution model that characterizes the normal operating boundary of the system.

[0099] Receive the continuous sliding window topological health vector sequence generated in the previous steps and construct a reference benchmark dataset based on historical normal operating condition data. Perform decentralization and standardization preprocessing on the multidimensional topological features in the reference benchmark dataset to eliminate dimensional differences and establish a zero-mean distribution benchmark.

[0100] A Gaussian kernel function is used as the core operator for density estimation to construct a non-parametric probability density model to fit the topological state distribution boundary under normal operating conditions. The following formula is used to calculate the probability density of any test point. probability density estimate at : in, For reference sample size, For bandwidth parameters, For Gaussian kernel function, Let a be the a-th reference sample point.

[0101] The bandwidth parameter is dynamically optimized using the Silverman rule of thumb to balance the smoothness of the estimated curve with the ability to preserve local details, ensuring that the density peak accurately corresponds to the center of the normal topological cluster. The eigenvalues ​​of the covariance matrix of the reference dataset are calculated, and the optimal bandwidth is determined based on the principle of maximum likelihood estimation, avoiding the problem of anomaly boundaries being blurred due to excessively large bandwidth or the false alarm rate being increased due to excessively small bandwidth.

[0102] A lookup table index is built in the embedded MCU memory to store key support vectors and their corresponding density gradient information, forming a lightweight normal operating condition distribution model. This model clearly defines the high-density clustering areas of topological health vectors in the three-dimensional feature space, serving as a statistical benchmark for subsequent isolation forest methods to determine the degree of abnormal deviation.

[0103] By using kernel density estimation, discrete historical normal topology vectors are transformed into continuous probability density distribution functions, enabling accurate mathematical modeling of the normal operation boundary of mobile power supplies and providing robust distribution priors for unsupervised anomaly detection.

[0104] S5.2: Receive the normal operating condition distribution model and the current topological health vector input in real time, construct an anomaly detection ensemble structure composed of multiple random partitioning trees based on the isolated forest method, and calculate the isolation depth of the current topological health vector in the feature space using the binary tree recursive partitioning mechanism to generate an original isolation score sequence that characterizes the degree of deviation of data points.

[0105] Receive the normal operating condition distribution model generated by S5.1 and the real-time topology health vector output by S4, and initialize the isolated forest ensemble structure. Set the isolated forest to contain 100 independent random binary trees, with the maximum depth of each tree limited to log2(N), where N is the number of samples within the sliding window, to balance computational complexity and detection accuracy.

[0106] For the current input 3D topological health vector, perform recursive splitting in each random binary tree. Randomly select a feature dimension, and randomly choose a split point between the maximum and minimum values ​​of that dimension to divide the data space into two subspaces.

[0107] Determine whether the current topological health vector falls into the left or right subspace. If it falls into the left subspace, increment the path length by 1 and continue recursively partitioning that subspace; if it falls into the right subspace, process it similarly. Continue the recursive process until the node becomes a leaf node or the maximum depth limit is reached.

[0108] Record the path length of the current topological health vector in each binary tree, which is the number of edges traversed from the root node to the leaf node. Take the arithmetic mean of the path lengths of all 100 binary trees to obtain the average path length.

[0109] Through the above recursive partitioning and score calculation, the topological health vector from the previous step is transformed into a raw isolation score sequence that characterizes the degree of deviation of data points, achieving preliminary quantification of anomalies without labels and providing basic data for subsequent dynamic threshold calibration.

[0110] S5.3: Based on the statistical thresholds of the original isolation score sequence and the normal operating condition distribution model, perform a dynamic threshold adaptive calibration operation to normalize the original isolation score sequence into an anomaly score scalar between zero and one, so as to generate a real-time anomaly confidence index that quantifies the degree to which the current topology deviates from the normal boundary.

[0111] The system receives the original isolation score sequence and normal operating condition distribution model generated in the previous steps, and extracts the statistical feature parameters of the historical normal operating condition topological health vector. It calculates the mean and standard deviation of the isolation scores under normal operating conditions, constructs a dynamic benchmark threshold interval, and eliminates the sensitivity of fixed thresholds to operating condition fluctuations.

[0112] Based on the distribution characteristics of isolated scores within the current sliding window, an exponentially weighted moving average method is used to update local statistics, giving higher weight to recent data to capture short-term drift trends.

[0113] Each score in the original isolated score sequence is compared with a dynamic threshold, and a linear normalization mapping is performed. For scores below the threshold, the mapping is to zero, indicating that they are within the normal safety boundary; for scores above the threshold, a compression mapping is performed based on a nonlinear growth function of the excess amplitude to ensure that the abnormal score scalar strictly converges to the interval between zero and one.

[0114] A time-smoothing factor is introduced to recursively filter outlier scores within a continuous window, suppressing score jumps caused by transient electromagnetic interference and generating a real-time anomaly confidence index with temporal consistency. This index quantifies the degree to which the current topological state deviates from the normal evolutionary boundary, providing a high signal-to-noise ratio input for subsequent clustering.

[0115] By using dynamic threshold adaptive calibration and normalization mapping, the original isolation score from the previous step is transformed into a real-time anomaly confidence index between zero and one, which enables sensitive capture of weak fault precursors and effective suppression of strong interference, significantly improving the robustness of early fault identification.

[0116] S5.4: For the time point when the real-time anomaly confidence index exceeds the preset alarm threshold, extract the topological health vector and its adjacent context vector within the corresponding time window, and use a density-based spatiotemporal clustering algorithm to perform cluster aggregation processing on the topological health vectors corresponding to the high anomaly confidence index, so as to generate a set of candidate anomaly pattern clusters with coherent evolutionary characteristics.

[0117] S5.5: Perform noise filtering and connectivity verification on the candidate abnormal pattern cluster set, remove isolated abnormal points caused by instantaneous interference and merge spatiotemporally adjacent fragmented clusters, and finally output an abnormal pattern cluster that deviates from the normal topological evolution path with complete structure and physical meaning, as the trigger object for the subsequent reverse trajectory projection mechanism.

[0118] like Figure 2 As shown, step S6: When the abnormal pattern cluster is detected, the corresponding original phase space point is locked and mapped back to the original four-dimensional time series coordinates. By statistically analyzing the consistency of fluctuation intensity and phase shift in each dimension during the abnormal period, a fault feature tracing result including fault sensitivity parameter ranking and key period marking is generated. Specifically, this includes: S6.1: Based on the indexes of all abnormal topological nodes contained in the abnormal pattern cluster, retrieve the corresponding original phase space point set from the pre-established phase space point cloud mapping table to generate abnormal phase space point cloud data to be reverse mapped.

[0119] Receive the abnormal pattern cluster output from step S5, extract the unique index identifier set of all abnormal topology nodes contained therein, and use it as the primary key input for reverse retrieval.

[0120] The phase space point cloud mapping table, which is pre-stored in the embedded MCU flash memory, is called. This mapping table is generated synchronously when the phase space trajectory is constructed in the S2 stage. It records the bidirectional binding relationship between each phase space node index and its corresponding original four-dimensional time series data timestamp and relative position within the window.

[0121] Based on the abnormal topology node index set, a parallel hash lookup operation is performed on the phase space point cloud mapping table to quickly locate the specific physical location of each abnormal node in the original data acquisition sequence.

[0122] The original voltage, current, temperature, and estimated internal resistance four-dimensional instantaneous state vector associated with the abnormal node index are read from the mapping table to ensure that the retrieved data retains the original sampling accuracy without being normalized or dimensionality reduced.

[0123] The retrieved discrete four-dimensional instantaneous state vectors are reassembled according to timestamp order to restore their continuous arrangement structure on the original time axis, forming abnormal phase space point cloud data to be reverse mapped.

[0124] Through the above-mentioned index retrieval and data recombination processing methods, the abstract topological anomaly nodes are transformed into raw electrical performance data fragments carrying complete physical information, realizing accurate backtracking from graph structure features to the original signal domain, and providing a high-fidelity data foundation for subsequent fault feature quantification.

[0125] S6.2: Perform delayed coordinate inverse embedding transformation on the abnormal phase space point cloud data, and use the preset optimal time delay parameter and embedding dimension to restore the three-dimensional phase space trajectory points into a time axis aligned four-dimensional electrical performance time series segment, so as to generate original time series reconstruction data carrying abnormal features.

[0126] S6.3: Based on the four-dimensional sequences of voltage, current, temperature and estimated internal resistance in the original time series reconstruction data, calculate the standard deviation fluctuation coefficient and instantaneous rate of change extreme value of each dimension during the abnormal period to generate a dimension fluctuation intensity vector characterizing the degree of single-parameter instability.

[0127] Receive the original time-series reconstruction data carrying abnormal characteristics output from step S6.2. This data includes time-series segments of four dimensions—voltage, current, temperature, and estimated internal resistance—during the abnormal period.

[0128] The time segments of the four dimensions are subjected to demeaning processing to eliminate the interference of DC component offset on volatility assessment and generate zero-mean signal sequences for each dimension.

[0129] Based on the zero-mean signal sequences of each dimension, the sliding window standard deviation method is used to calculate the local fluctuation amplitude in the neighborhood of each sampling point, and obtain the standard deviation fluctuation coefficient sequence that reflects the degree of signal dispersion.

[0130] First-order difference operations are performed on the time segments of the four dimensions to calculate the numerical changes between adjacent sampling points and generate a sequence of instantaneous rate of change that characterizes the transient response speed.

[0131] Extract the maximum and minimum absolute values ​​from the instantaneous rate of change sequence to determine the most dramatic jump amplitude of signals in each dimension during the abnormal period, and generate an instantaneous rate of change extreme value index.

[0132] By weighting and fusing the mean of the standard deviation volatility coefficient sequence with the instantaneous rate of change extreme value index, a dimensional volatility intensity vector characterizing the degree of single-parameter instability is constructed.

[0133] Through the above processing method, the original time series reconstruction data from the previous step is transformed into a dimensional fluctuation intensity vector that quantifies the instability of each electrical performance dimension, thereby achieving preliminary quantification of the fault contribution of each parameter and providing basic data support for subsequent multi-parameter coupling analysis.

[0134] S6.4: Using the cross-correlation function method, perform phase difference scanning analysis on any two-dimensional electrical performance signal in the original time series reconstruction data, calculate the absolute value of the deviation between the reference phase under normal operating conditions and the actual phase during the current abnormal period, so as to generate a phase offset consistency matrix that characterizes the multi-parameter coupling misalignment.

[0135] Receive the raw time-series reconstruction data carrying abnormal features output from step S6.2. This data includes time-series segments of four dimensions: voltage, current, temperature, and estimated internal resistance, which serve as input for phase difference scanning analysis.

[0136] The four-dimensional electrical performance signals are paired up to construct six sets of bivariate signal pairs, including voltage-current, voltage-internal resistance, current-temperature, etc., to fully cover the internal electrothermal coupling relationship of the power bank.

[0137] For each pair of bivariate signals, the cross-correlation function method is used to calculate their similarity measure under different time lags. By traversing all sampling points within the abnormal time period through a sliding window, a cross-correlation coefficient sequence is generated.

[0138] The optimal time lag between two sets of signals is determined by using the peak position of the cross-correlation coefficient sequence. This lag characterizes the response delay of a physical quantity during circuit transmission.

[0139] A reference phase library is established based on historical normal operating condition data. The reference time lag of each bivariate signal pair in the standard charging and discharging stage is extracted and used as a reference standard for phase deviation calculation.

[0140] Calculate the absolute value of the difference between the actual time lag and the reference time lag during the current abnormal period to obtain the phase offset scalar of a single signal pair, and quantify the degree of misalignment in the multi-parameter coupling relationship.

[0141] The phase offset scalars of six sets of bivariate signal pairs are integrated into a symmetric matrix structure to form a phase offset consistency matrix that characterizes multi-parameter coupling mismatch, where the diagonal elements are zero and the off-diagonal elements reflect the phase mismatch between corresponding dimensions.

[0142] By performing phase difference scanning analysis on any two-dimensional electrical performance signals in the original time-series reconstructed data using the cross-correlation function method, the results of the previous step are transformed into a phase offset consistency matrix that characterizes multi-parameter coupling misalignment, thereby achieving accurate location of the fault root cause at the multi-dimensional signal interaction level.

[0143] For example, for voltage and current signal pairs, the sliding window length is set to 300 sampling points, and the step size is 10 sampling points. When calculating the cross-correlation function, the maximum hysteresis range is set to ±50 sampling points. During the normal constant current charging phase, the baseline hysteresis of voltage and current is 2 sampling points. Within the 30-second window when an anomaly is detected, the calculated actual hysteresis is 8 sampling points. The phase offset scalar is |8-2|=6 sampling points. Similarly, the offsets of the other five signal pairs are calculated, such as the current-internal resistance offset of 4 and the voltage-temperature offset of 12. In the final generated 6x6 phase offset consistency matrix, the off-diagonal elements are significantly larger than the normal threshold, indicating that the internal polarization effect of the battery causes a severe hysteresis in the voltage response, and the thermal management failure exacerbates the decoupling between parameters, providing strong feature support for subsequent fault classification.

[0144] S6.5: Combining the aforementioned dimensional fluctuation intensity vector and phase offset consistency matrix, the fault contribution of each electrical performance dimension is quantitatively evaluated and sorted in descending order using a weighted entropy sorting method. At the same time, the start and end time points of fluctuations exceeding the threshold are extracted to generate fault feature tracing results that include fault-sensitive parameter sorting and key time period marking.

[0145] The dimensional fluctuation intensity vector and phase offset consistency matrix generated in the preceding steps are received as the input data source for the fault contribution quantification assessment.

[0146] Normalization is performed on the four components of voltage, current, temperature and internal resistance in the dimensional fluctuation intensity vector to eliminate the interference of dimensional differences on subsequent entropy value calculation and generate a standardized fluctuation characteristic sequence.

[0147] Based on the standardized fluctuation characteristic sequence, the probability distribution density of data in each dimension is calculated, and the Shannon entropy formula is used to measure the degree of disorder and information uncertainty of each electrical performance parameter during abnormal periods.

[0148] Eigenvalue decomposition is performed on the phase offset consistency matrix to extract the eigenvector corresponding to the largest eigenvalue, which represents the principal direction weight of multi-parameter coupling misalignment.

[0149] By weighting and fusing the fluctuation entropy value with the phase misalignment weight, a comprehensive fault contribution index is constructed to reflect the combined impact of single parameter instability and multi-parameter coupling anomalies.

[0150] Based on the comprehensive fault contribution index, the electrical performance dimensions are sorted in descending order to determine the priority order of fault-sensitive parameters and identify the key physical quantities that dominate fault evolution.

[0151] Traverse the original time-series reconstructed data, set dynamic fluctuation thresholds, and detect the start and end times when signals in each dimension exceed the thresholds.

[0152] By combining the start and end times to form key time period markers, the specific time window in which the fault occurs can be locked, enabling precise location of the fault in the time domain.

[0153] By using weighted entropy sorting and threshold detection, the statistical results of the previous step are transformed into fault feature tracing results that include fault-sensitive parameter sorting and key time period marking. This enables reverse tracing from abstract topological anomalies to specific physical parameters and time periods, providing a high-confidence input basis for semantic mapping in subsequent stages.

[0154] like Figure 3 As shown, step S7 involves aligning the key time period markers in the fault characteristic tracing results to the specific operating stages of constant current, constant voltage, or trickle flow, using a pre-set finite state machine model, to output a preliminary judgment result of the mobile power supply fault type. Specifically, this includes: S7.1: Obtain the key time period markers in the fault feature tracing results and the four-dimensional electrical performance time series data stream collected in real time in the embedded MCU. Based on the voltage change rate threshold judgment logic, perform charge and discharge condition feature extraction processing on the four-dimensional electrical performance time series data stream to generate the original operating condition status identifier sequence representing the current working mode.

[0155] Receive the key time period markers in the fault feature tracing results, synchronously read the four-dimensional electrical performance time series data stream cached in the embedded MCU, and extract the voltage and current sequence segments aligned with the key time period marker timestamps.

[0156] The extracted voltage sequence segments are subjected to first-order difference operations to calculate the voltage change rate between adjacent sampling points and generate a voltage change rate time vector to quantify the dynamic characteristics of voltage fluctuations.

[0157] Set a threshold range for voltage change rate, and map the values ​​in the voltage change rate time-series vector to discrete state identifiers. When the absolute value of the change rate is lower than the preset lower limit and the duration exceeds the minimum judgment window, it is marked as a candidate state for the constant voltage stage.

[0158] The extracted current sequence fragments are subjected to sliding window mean filtering to eliminate high-frequency noise interference and generate smooth current time series curves to highlight the stable characteristics of the current plateau.

[0159] Calculate the standard deviation of the smoothed current time-series curve within the sliding window. When the standard deviation is lower than the current plateau noise threshold and the mean is maintained within the specified range, it is marked as a candidate state for the constant current stage.

[0160] By combining the state indicators of voltage and current, and using logical AND operations to eliminate contradictory states, when the voltage is in a constant voltage candidate state and the current shows an exponential decay trend, it is confirmed as a trickle charging stage indicator.

[0161] The discrete state identifiers at each time point are arranged in chronological order to generate a sequence of original operating condition state identifiers representing the current working mode, thus realizing the initial transformation from continuous analog signals to discrete operating condition semantics.

[0162] By employing a dual constraint mechanism of voltage change rate threshold determination and current plateau period detection, the key time period data from the previous step is transformed into a sequence of original operating condition status identifiers with physical meaning, thereby achieving coarse-grained identification of charging and discharging conditions and providing basic input for subsequent fine-grained matching of finite state machines.

[0163] For example, the voltage change rate threshold is set to 0.05V / s, and the minimum judgment window is 2 seconds. The voltage data within the critical period is differentially analyzed. If the absolute value of the change rate for 20 consecutive sampling points (10Hz sampling rate) is less than 0.05V / s, it is marked as a constant voltage state. The current data is filtered using a moving average with a window size of 10. If the filtered current standard deviation is less than 0.02A and the mean is greater than 0.5A, it is marked as a constant current state. If the voltage is marked as constant voltage and the current mean is less than 0.1A and shows a decreasing trend, it is marked as a trickle current state. The final output is a sequence of original operating condition identifiers such as [constant current, constant current, constant voltage, trickle current], accurately reflecting the actual working stage of the power bank during abnormal periods.

[0164] S7.2: Receive the original operating condition status identifier sequence, and use the mobile power bank finite state machine model pre-installed in the firmware to perform state transition matching and noise filtering on the original operating condition status identifier sequence to generate a standardized charging stage tag stream with well-defined stage boundaries.

[0165] Receive the original operating condition status identifier sequence generated in the previous steps. This sequence contains discrete operating condition markers obtained based on the detection of instantaneous voltage change rate and current plateau period, which serve as the input data stream for the finite state machine model.

[0166] The charging phase finite state machine model, which is pre-installed in the read-only storage area of ​​the power bank firmware, is invoked. This model defines six standard state nodes and their legal transition paths: idle, precharge, constant current, constant voltage, trickle charge, and fully charged.

[0167] The original operating condition state identifier sequence is traversed in a time sequence. The discrete markers at the current time are matched and verified with the system state at the previous time to identify whether there are illegal state transitions that violate the physical charging logic, such as a direct transition from the constant current stage to the idle state.

[0168] For the identified illegal state transition points, a majority voting filtering mechanism based on time windows is introduced to count the frequency of the state distribution of the N adjacent sampling points before and after the anomaly point, and replace the current anomaly mark with the state value with the highest frequency to eliminate state jitter caused by instantaneous noise of the sensor.

[0169] For the transition period in the state boundary ambiguity region, a hysteresis comparison strategy is adopted, and dual criteria are set for the difference between the upper and lower limits of the voltage threshold and the duration of current stability. The state transition operation is only performed when the real-time electrical parameters continuously meet the target state constraints for more than the preset holding time.

[0170] Through the aforementioned state transition matching and noise filtering, the original operating condition identifiers containing glitches and logical conflicts are transformed into a standardized charging stage label stream that is time-continuous, logically consistent, and has clear boundaries, thereby achieving accurate alignment of operating condition semantics.

[0171] For example, the finite state machine sampling period is set to 100ms, the voltage threshold difference of the hysteresis comparator is 50mV, and the current stabilization time is 2 seconds. When a noise point marked as "constant voltage" appears in the original sequence during the constant current phase, the system detects that this transition does not conform to the physical law that the voltage must gradually increase from constant current to constant voltage, and determines it as an illegal transition. By statistically analyzing 5 sampling points before and after, 9 points are found to be "constant current" and 1 point is "constant voltage". Based on the majority voting principle, the noise point is corrected to "constant current". If the voltage rises from 4.1V to 4.2V, but does not remain above 4.2V for 2 seconds before falling back, the state machine remains in the "constant current" state without transitioning, effectively avoiding frequent state switching caused by grid fluctuations. The output standardized tag stream accurately reflects the actual charging process.

[0172] S7.3: Call the standardized charging stage tag stream and calculate the time axis overlap with the key time period marker. Perform a working condition semantic mapping operation on the key time period marker based on the maximum intersection principle to generate a semantic fault period index containing attributes of constant current stage, constant voltage stage or trickle stage.

[0173] The system receives the standardized charging phase tag stream output by S7.2 and the fault critical period markers generated by S6.5 to construct a time-axis aligned data mapping environment.

[0174] Extract the state transition time points from the standardized charging stage tag stream and convert them into a discrete set of time intervals. Each interval corresponds to a specific operating condition attribute such as constant current, constant voltage, or trickle charging.

[0175] The critical time period markers of the fault are analyzed to determine the start and end timestamps of the abnormal events, forming a fault time window to be judged.

[0176] The overlap duration between the fault time window and the time intervals of all charging stages is calculated, and the intersection operation is used to quantify the coverage of each operating condition stage during the abnormal period.

[0177] By applying the principle of maximum intersection, the numerical values ​​of each overlapping duration are compared, and the charging stage interval with the highest overlap ratio is selected as the dominant working condition background.

[0178] The semantic labels of the dominant working condition phase are assigned to the fault time window to generate a semantic fault time period index containing specific working phase attributes.

[0179] By calculating the overlap of the time axis and processing the maximum intersection mapping, the abstract fault period is accurately anchored to the specific charging and discharging physical process, realizing the semantic association between the fault phenomenon and the working mechanism, and providing a clear stage context for subsequent fault type reasoning based on the mechanism knowledge base.

[0180] For example, the standardized charging stage tag stream shows 0-10s as the constant current stage, 10-25s as the constant voltage stage, and 25-30s as the trickle charging stage. The critical fault period is marked as 8s to 12s. The overlap duration is calculated as follows: 2s overlap with the constant current stage (8s-10s), 2s overlap with the constant voltage stage (10s-12s), and 0s overlap with the trickle charging stage. If the constant current stage takes precedence over the constant voltage stage in the preset priority rules, or if the voltage fluctuation is more significant within 8-10s, then the fault is determined to mainly occur during the transition from constant current to constant voltage or fall into the end of the constant current stage. Assuming another scenario where the fault period is 15s-20s, falling entirely within the 10-25s interval, with an overlap duration of 5s, it is directly mapped to the constant voltage stage. The output fault period index with semantics is "constant voltage stage - 15s-20s", significantly improving the physical interpretability of fault location.

[0181] S7.4: Combining the semantic fault period index with the pre-stored fault mechanism knowledge base, perform fault type reasoning and synthesis processing on the semantic fault period index according to the association rules of stage attributes and topology health vector abnormal patterns to generate a preliminary judgment result of mobile power bank fault type with stage semantic definition.

[0182] Receive the semantic fault period index and extract the charging stage attribute labels and corresponding topology health vector abnormal pattern feature data contained therein.

[0183] The system invokes a pre-stored fault mechanism knowledge base, which contains a mapping rule matrix between the degradation of electrical performance parameters and specific hardware faults at different charging stages.

[0184] Based on stage attribute labels, a subset of faults applicable to the current stage is selected from the mapping rule matrix to limit the inference search space and reduce the false positive rate.

[0185] The three components of the topological health vector—node degree centrality decay rate, edge weight distribution entropy, and minimum loop area shrinkage rate—are matched and compared with the threshold intervals in the subset.

[0186] A weighted voting mechanism is used to calculate the confidence score of each candidate fault type, and the weight coefficient is determined by the co-occurrence frequency of the topological feature and the fault type in historical fault cases.

[0187] If multiple high-confidence candidates exist, a phase offset consistency matrix is ​​introduced as a disambiguation factor to prioritize the fault type that is consistent with the direction of phase change.

[0188] Generate a standardized preliminary judgment result object containing the fault type code, severity level, and recommended handling measures.

[0189] Through the above association rule reasoning and synthesis process, the abstract topological anomaly indicators are transformed into preliminary fault type judgment results with clear physical meaning and maintenance guidance value, thereby realizing the interpretability and engineering applicability of fault diagnosis.

[0190] For example, when the semantic fault period index indicates that the current state is in the constant voltage charging stage, and the topology health vector shows that the minimum loop area shrinkage rate exceeds the 0.85 threshold, while the edge weight distribution entropy exhibits high-frequency oscillation characteristics, the fault mechanism knowledge base retrieves information indicating that the abnormal loop shrinkage rate in the constant voltage stage is mainly associated with SEI film thickening or electrolyte drying. Combined with the phase offset consistency matrix showing an increased phase difference between voltage and current, and ruling out instantaneous impedance changes caused by poor contact, the fault type is identified as an irreversible increase in internal resistance due to battery aging. The system outputs a fault code with a medium severity level, recommending that the user reduce the frequency of fast charging and schedule a battery health check. This process requires no manual intervention, takes less than 5ms within the MCU, and significantly improves the accuracy and response speed of fault location.

[0191] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0192] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0193] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent fault detection of mobile power banks based on electrical performance data, characterized in that, Specifically, it includes: S1: Obtain the four-dimensional electrical performance timing data stream of voltage, current, temperature and estimated internal resistance in the embedded MCU of the mobile power supply, and perform window truncation processing on the four-dimensional electrical performance timing data stream with a fixed sampling interval to obtain the original timing data window; S2: Based on the electrical performance time-series data stream of each dimension in the original time-series data window, select the optimal time delay and embedding dimension to perform phase space reconstruction processing to generate four sets of independent electrical parameter phase space trajectory point clouds; S3: Based on the physical signal flow constraints from the charging IC inside the power bank to the battery cell and then to the protection board, the topology map of the combined operation trajectory of electric-thermal-resistance is obtained based on the point cloud of the phase space trajectory of four independent electrical parameters. S4: Calculate three types of low-dimensional topological invariants based on the combined electric-thermal-resistance operation trajectory topology map to generate a topological health vector; S5: Input the topology health vectors generated by multiple consecutive sliding windows into the normal operating condition distribution model, identify and output abnormal pattern clusters; S6: When the abnormal pattern cluster is detected, the corresponding original phase space point is locked and mapped back to the original four-dimensional time series coordinates. By statistically analyzing the fluctuation intensity and phase shift consistency of each dimension during the abnormal period, a fault feature tracing result including fault sensitivity parameter sorting and key period marking is generated. S7: Combine the preset finite state machine model to align the key time period markers in the fault feature tracing results to the specific working stages of constant current, constant voltage, or trickle flow, so as to output the preliminary judgment result of the mobile power supply fault type.

2. The intelligent fault detection method for mobile power supplies based on electrical performance data according to claim 1, characterized in that, The combined electric-thermal-resistance operating trajectory topology includes node degree centrality, edge weight distribution, and minimum loop area.

3. The intelligent fault detection method for mobile power supplies based on electrical performance data according to claim 2, characterized in that, The three types of low-dimensional topological invariants include the degree centrality decay rate, the edge weight distribution entropy, and the minimum loop area shrinkage rate.

4. The intelligent fault detection method for mobile power supplies based on electrical performance data according to claim 1, characterized in that, The construction of the normal operating condition distribution model is specifically as follows: A reference benchmark dataset is constructed based on the set of topological health vectors obtained from the historical normal operating conditions of power banks. The probability density distribution of the reference benchmark dataset is fitted using the kernel density estimation method to obtain the normal operating condition distribution model.

5. The intelligent fault detection method for mobile power supplies based on electrical performance data according to claim 1, characterized in that, The finite state machine model stores and defines six standard state nodes and their legal transition paths: idle, precharged, constant current, constant voltage, trickle, and full.

6. The intelligent fault detection method for mobile power supplies based on electrical performance data according to claim 1, characterized in that, S1 specifically includes: The analog electrical signal output in real time from the analog-to-digital conversion interface inside the embedded microcontroller of the power bank is acquired. Multi-channel synchronous sampling technology is used to perform parallel acquisition and processing of voltage signal, current signal, temperature sensing signal and internal resistance estimation signal to generate a four-dimensional raw electrical performance data stream containing timestamps. Based on the high-frequency noise components in the original four-dimensional electrical performance data stream, a moving average filtering method is used to perform digital smoothing and denoising processing to generate a clean four-dimensional electrical performance data stream. For the pure four-dimensional electrical performance data stream, perform equal-interval resampling operations to generate a standard four-dimensional electrical performance time series; A sliding window truncation strategy is applied to the standard four-dimensional electrical performance time series to extract continuous data segments according to the current system clock cycle, so as to form a candidate time series data window set; Based on the data integrity verification results in the candidate time series data window set, invalid windows with packet loss or out-of-bounds anomalies are removed, and the original time series data window with complete structure and time alignment is finally output.

7. The intelligent fault detection method for mobile power supplies based on electrical performance data according to claim 1, characterized in that, S2 specifically includes: The voltage time series, current time series, temperature time series and estimated internal resistance time series in the original time series data window are processed to generate a standardized four-dimensional electrical performance time series data stream. Based on the standardized four-dimensional electrical performance time-series data stream, the autocorrelation decay characteristics of each dimension of the data are calculated to determine the optimal time delay parameters. At the same time, the nearest neighbor method is used to analyze the phase space expansion degree to determine the optimal embedding dimension, and a phase space reconstruction configuration parameter set is generated. Based on the optimal time delay parameter and optimal embedding dimension in the phase space reconstruction configuration parameter set, a delay coordinate embedding transformation operation is performed on each dimension of data in the standardized four-dimensional electrical performance time series data stream to map the one-dimensional scalar time series into a multi-dimensional state vector sequence, generating voltage phase space trajectory point cloud, current phase space trajectory point cloud, temperature phase space trajectory point cloud and internal resistance phase space trajectory point cloud. Local neighborhood density estimation and noise point filtering are performed on the generated voltage phase space trajectory point cloud, current phase space trajectory point cloud, temperature phase space trajectory point cloud and internal resistance phase space trajectory point cloud respectively to remove outlier trajectory points caused by sensor quantization error, and generate four sets of independent electrical parameter phase space trajectory point clouds.

8. The intelligent fault detection method for mobile power supplies based on electrical performance data according to claim 7, characterized in that, The voltage time series, current time series, temperature time series, and estimated internal resistance time series in the original time series data window are processed respectively. Specifically, detrending processing and normalization preprocessing are performed on the voltage time series, current time series, temperature time series, and estimated internal resistance time series respectively.

9. The intelligent fault detection method for mobile power supplies based on electrical performance data according to claim 7, characterized in that, The calculation of the autocorrelation decay characteristics of each dimension of data to determine the optimal time delay parameter is performed using the mutual information method.

10. The intelligent fault detection method for mobile power supplies based on electrical performance data according to claim 7, characterized in that, The phase space reconstruction configuration parameter set includes the optimal time delay parameter and the optimal embedding dimension.

Images