Fault diagnosis method for new energy electric vehicle power system

Abstract

The invention discloses a fault diagnosis method for a new energy electric vehicle power system, and relates to the technical field of fault diagnosis of the new energy electric vehicle power system, and the method comprises the steps: mapping a power parameter of a power system assembly into an energy flow tensor, and obtaining the inter-assembly energy transmission efficiency through tensor decomposition; performing wavelet transform on the transmission efficiency to extract degradation characteristics, and constructing a component health degree index; constructing a probability graph network based on the health degree index to calculate a fault state transition probability; fault coupling distribution is obtained through variational inference, and an early warning threshold value is adjusted in a self-adaptive mode; and when the health degree index is lower than a threshold value, determining a fault source and an affected component based on fault coupling distribution, and generating a fault propagation path and residual life early warning. The method can accurately identify the fault source and predict the fault propagation path, and improves the fault diagnosis accuracy.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for power systems of new energy electric vehicles, and more specifically, to a fault diagnosis method for power systems of new energy electric vehicles. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the reliability and safety of the power system have received increasing attention. The power system of new energy electric vehicles mainly includes core components such as power batteries, motors, and electronic controls. There are complex energy transfer relationships and fault coupling effects among these components.

[0003] Currently, power system fault diagnosis mainly relies on the condition monitoring and threshold judgment of individual components, making it difficult to effectively identify fault propagation paths between components.

[0004] Existing fault diagnosis methods typically employ rule-based expert systems or single data-driven models. These methods neglect the energy transfer relationships between components of the power system, leading to inaccurate fault location; they set fixed fault warning thresholds without considering the dynamic characteristics of component degradation; and they lack analysis of fault coupling effects, making it difficult to predict fault propagation paths.

[0005] Traditional data-driven fault diagnosis methods, such as support vector machines and neural networks, can classify faults using historical data, but they often treat the power system as an independent assembly of components, ignoring the energy flow relationships between them. This can lead to the system misdiagnosing fault symptoms as root causes, affecting the accuracy and timeliness of fault diagnosis.

[0006] Furthermore, existing technologies generally employ fixed fault warning thresholds, failing to adapt to the performance degradation of components during use. Common battery management systems on the market use fixed voltage and temperature thresholds for fault diagnosis, making it difficult to accurately reflect the actual health status of the battery under different operating conditions and degrees of degradation. This method is prone to false alarms or missed alarms, affecting the reliability of the system. Summary of the Invention

[0007] This invention provides a fault diagnosis method for the power system of new energy electric vehicles, which can solve the problems in the prior art.

[0008] This invention provides a fault diagnosis method for the power system of a new energy electric vehicle, comprising the following steps: Construct an energy flow tensor network between components to map the power parameters of each component in the power system of new energy electric vehicles into energy flow tensors; The energy transfer efficiency between components is obtained through tensor decomposition. Wavelet transform is used to extract component degradation features from energy transfer efficiency, and a component health index is constructed. A probabilistic graphical network is constructed based on the component health index to calculate the probability of component failure state transition. The fault coupling distribution among components is obtained by variational inference, and the fault early warning threshold is adaptively adjusted according to the fault coupling distribution. When the component health index is lower than the fault warning threshold, the fault source component and the affected component are determined based on the fault coupling distribution, and fault propagation path and remaining lifetime warning information are generated.

[0009] Furthermore, an energy flow tensor network is constructed between components to map the power parameters of each component in the new energy electric vehicle power system into energy flow tensors, including: Collect power parameters of the power system components and construct an initial mapping matrix; The initial mapping matrix is ​​scaled to obtain a multi-scale energy flow feature sequence, and an energy flow tensor network is constructed. The path weights are calculated based on the energy flow tensor network, and a mapping function is constructed based on the path weights to map the power parameters of each component to the energy flow tensor.

[0010] Furthermore, the energy transfer efficiency between components obtained through tensor decomposition includes: The energy flow tensor is decomposed into nonnegative components to obtain the initial core tensor. Construct the energy balance constraints of the initial core tensor and calculate the recursive core tensor; The efficiency difference and compensation coefficient between different transmission paths between components are calculated based on the recursive core tensor. An efficiency compensation optimization equation is constructed, and the energy transfer efficiency between components is obtained by solving the efficiency compensation optimization equation.

[0011] Furthermore, wavelet transform is used to extract component degradation features from energy transfer efficiency, and a component health index is constructed, including: Wavelet transform is performed on the energy transfer efficiency of the component to obtain the wavelet coefficient sequence, which is then divided into continuous degradation component and instantaneous response component according to the frequency range. Operating condition compensation coefficient is constructed, and the wavelet coefficient sequence is compensated using the operating condition compensation coefficient. The component health index is calculated based on the compensated degradation characteristic sequence.

[0012] Furthermore, constructing a probabilistic graphical network based on the component health index includes: The transition correlation degree is calculated based on the differences between adjacent states in the component health index sequence, and a probabilistic graphical network is constructed based on the transition correlation degree.

[0013] Furthermore, calculating the component failure state transition probability includes: A bidirectional propagation analysis is performed on the probabilistic graphical network to extract the forward influence strength and backward coupling strength of the state transition, and to determine the state transition weights. The state transition weights and transition correlations are weighted and combined to obtain the component failure state transition probability.

[0014] Furthermore, variational inference is used to obtain the fault coupling distribution between components, and the fault early warning threshold is adaptively adjusted based on the fault coupling distribution, including: A variational distribution of component fault coupling is constructed using conditional random fields, and the variational distribution is optimized by expectation maximization iteration to obtain the fault coupling distribution between components. The correlation between component faults is calculated based on the fault coupling distribution, and the fault early warning threshold is adaptively adjusted by combining the fault propagation risk coefficient.

[0015] Furthermore, adaptively adjusting the fault early warning threshold based on the fault coupling distribution includes: The amplitude and phase features of the fault coupling distribution are extracted, the rate of change of the amplitude features and the drift of the phase features are calculated, a correction function for the fault warning threshold is constructed, the adjustment coefficient of the warning threshold is calculated based on the correction function, and the fault warning threshold is dynamically updated.

[0016] Furthermore, based on the fault coupling distribution, the fault source component and affected components are identified, and fault propagation path and remaining lifetime early warning information are generated, including: Maximum entropy time series analysis is used to calculate the component failure propagation probability; The fault coupling distribution is prioritized based on the fault propagation probability to identify the fault source component, and the affected components whose coupling strength with the fault source component exceeds a preset strength threshold are extracted. Construct a propagation relationship graph between the fault source component and the affected components to obtain the fault propagation path; The component coupling accumulation value is calculated based on the fault propagation path, and the remaining lifespan warning information is generated by combining the component health index.

[0017] Furthermore, the probability of component fault propagation is calculated using maximum entropy time series analysis, including: Extract the temporal change characteristics of component states, calculate the state transition entropy values ​​between components, and obtain the fault propagation probability based on the state transition entropy values.

[0018] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method described in any of the foregoing embodiments.

[0019] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps in the method described in any of the preceding claims.

[0020] This invention constructs an energy flow tensor network, which can comprehensively reflect the energy transfer relationship and fault propagation characteristics between components, thereby improving the accuracy of fault diagnosis. The component degradation feature extraction method based on wavelet transform and variational inference can effectively identify the dynamic change trend of component health status. By using a probabilistic graphical network to calculate the fault state transition probability, the invention achieves accurate characterization of the component fault evolution process. By adaptively adjusting the fault warning threshold and combining fault source component identification and propagation path prediction, the invention can promptly detect potential faults and warn of fault propagation risks, providing a reliable basis for preventive maintenance of the power system of new energy electric vehicles. Attached Figure Description

[0021] Figure 1 A flowchart of a fault diagnosis method for a power system of a new energy electric vehicle provided in an embodiment of the present invention; Figure 2 A comparison chart of fault prediction lead time provided in an embodiment of the present invention. Detailed Implementation

[0022] like Figure 1 As shown, Figure 1 A flowchart of a fault diagnosis method for a power system of a new energy electric vehicle provided by an embodiment of the present invention, the method comprising the following steps: Construct an energy flow tensor network between components to map the power parameters of each component in the power system of new energy electric vehicles into energy flow tensors; The energy transfer efficiency between components is obtained through tensor decomposition. Wavelet transform is used to extract component degradation features from energy transfer efficiency, and a component health index is constructed. A probabilistic graphical network is constructed based on the component health index to calculate the probability of component failure state transition. The fault coupling distribution among components is obtained by variational inference, and the fault early warning threshold is adaptively adjusted according to the fault coupling distribution. When the component health index is lower than the fault warning threshold, the fault source component and the affected component are determined based on the fault coupling distribution, and fault propagation path and remaining lifetime warning information are generated.

[0023] In one optional embodiment, constructing an inter-component energy flow tensor network to map the power parameters of each component in the new energy electric vehicle power system to energy flow tensors includes: Collect power parameters of the power system components and construct an initial mapping matrix; The initial mapping matrix is ​​scaled to obtain a multi-scale energy flow feature sequence, and an energy flow tensor network is constructed. The path weights are calculated based on the energy flow tensor network, and a mapping function is constructed based on the path weights to map the power parameters of each component to the energy flow tensor.

[0024] First, power parameters were collected for each component of the trolley's power system, including power data for key nodes such as battery pack output power, motor input power, motor output power, and transmission system output power. High-precision power sensors were used to monitor each node in real time during the data collection process, with a sampling frequency of 100Hz to ensure the timeliness and accuracy of the data.

[0025] After acquiring the power parameters, an initial mapping matrix is ​​constructed. This matrix uses time as the row and power parameters as the column, recording the power distribution of the system at different times. For a four-wheel drive electric vehicle, the matrix includes power data for components such as the battery pack, four drive motors, and the transmission system. The initial mapping matrix has an n×m dimension, where n represents the number of sampling time points and m represents the number of power parameters. Power data within 300 seconds can be collected to form a 30000×10 initial mapping matrix.

[0026] The initial mapping matrix is ​​transformed into a multi-scale energy flow feature sequence. A sliding window method is used to scale the initial matrix, with window sizes ranging from 1 s to 10 s and a step size of 0.5 s, generating energy flow features at different time scales. For the data within each window, feature parameters such as power change rate, power fluctuation amplitude, and power transfer efficiency between adjacent components are calculated. For energy transfer from battery to motor, the transfer efficiency is calculated as the ratio of motor input power to battery output power; for energy transfer from motor to drive system, the transfer efficiency is calculated as the ratio of drive system output power to motor output power. The initial mapping matrix is ​​then converted into a set of sequences containing features at multiple scales.

[0027] An energy flow tensor network is constructed based on multi-scale energy flow feature sequences. This network is represented by a graph structure, where nodes represent components of the dynamic system, and edges represent the energy flow relationships between components. The weight of each edge is determined by its energy transfer efficiency; edges with higher efficiency have larger weights, and vice versa. The network structure uses a three-dimensional tensor representation: the first dimension represents the relationships between components, the second dimension represents the time series, and the third dimension represents features at different scales. For example, the energy flow tensor network has dimensions of 10×300×20, where 10 represents the number of components, 300 represents the number of time points, and 20 represents the feature dimension.

[0028] Using a constructed energy flow tensor network, the weights of each energy transfer path in the system are calculated. The path weights are calculated using the cumulative energy method, which calculates the proportion of cumulative energy transferred along a specific path to the total energy. For a complete energy transfer path from the battery to the wheel, the product of the transfer efficiencies of each node along the path is calculated as the total path efficiency, which is then multiplied by the total energy output from the battery to obtain the effective energy transferred along that path. By comparing the energy transfer efficiencies of different paths, the main energy transfer paths and bottleneck nodes are identified. Under normal operating conditions, the energy transfer efficiency of the main path is typically above 85%; when the efficiency drops below 75%, it may indicate a fault in that path.

[0029] Based on the calculated path weights, a mapping function is constructed to map the power parameters of each component to an energy flow tensor. The mapping function uses a weighted summation method, combining the power parameters of each component according to their path weights to generate a tensor representing the energy flow state of the system. For motor faults, the energy flow tensor will show a significant reduction in energy transfer efficiency along the path from the battery to the motor; for transmission system faults, it will manifest as abnormal energy transfer along the path from the motor to the wheels. The mathematical expression of the mapping function is the sum of the products of the power of each component and its corresponding path weight, and the generated energy flow tensor has the same dimension as the number of nodes in the energy flow tensor network.

[0030] This invention achieves precise location and diagnosis of power system faults through energy flow mapping analysis technology. Fault identification can be achieved without disassembling equipment, significantly reducing maintenance time and costs. Multi-scale energy flow characteristic analysis captures the dynamic characteristics of the system at different time scales, improving the accuracy and sensitivity of fault diagnosis. The path-weighted mapping function transforms complex power parameters into intuitive energy flow tensors, making fault location clearer and more precise. This greatly improves the maintenance efficiency and reliability of new energy electric vehicles, extends vehicle lifespan, and reduces operating costs.

[0031] In one optional embodiment, obtaining the inter-component energy transfer efficiency through tensor decomposition includes: The energy flow tensor is decomposed into nonnegative components to obtain the initial core tensor. Construct the energy balance constraints of the initial core tensor and calculate the recursive core tensor; The efficiency difference and compensation coefficient between different transmission paths between components are calculated based on the recursive core tensor. An efficiency compensation optimization equation is constructed, and the energy transfer efficiency between components is obtained by solving the efficiency compensation optimization equation.

[0032] First, an iterative nonnegative matrix factorization technique is employed to decompose the three-dimensional energy flow tensor into a combination of several low-dimensional core tensors and factor matrices. For the energy flow tensor containing 10 components, 300 time points, and 20 feature dimensions, alternating least squares decomposition is used, with 500 iterations and a convergence threshold of 0.001. The resulting initial core tensor has a dimension of 5×5×5, retaining over 95% of the information of the original tensor, effectively reducing data dimensionality while preserving key features.

[0033] After nonnegative decomposition, the initial core tensor can be used to characterize the energy transfer relationships between components in the power system of new energy electric vehicles, but it does not directly satisfy the principle of energy conservation. Therefore, energy balance constraints need to be constructed for the initial core tensor to ensure that the analysis results conform to physical laws. The energy balance constraints are based on two fundamental principles: the total energy inflow to each node equals the total energy outflow plus the node's own losses; and the total energy input of the system equals the effective output plus the losses of each link in the system. Based on these two constraints, the element values ​​of the initial core tensor are adjusted to satisfy the energy balance requirements while maintaining the main characteristics. The constraint adjustment uses the Lagrange multiplier method, and the constraint strength is controlled by setting weight coefficients. The weight value is set to 0.8, indicating that the emphasis on energy balance constraints is higher than the preservation of original data features.

[0034] Calculating the recursive core tensor is an iterative optimization process of the initial core tensor under energy balance constraints. The backpropagation algorithm is used; in each iteration, the error between the current tensor and the ideal tensor satisfying the energy balance condition is calculated, and the tensor element values ​​are adjusted along the error gradient direction. During iteration, the error threshold is set to 0.01, and the maximum number of iterations is 200. For a battery output power of 120kW, the energy flow from the battery to the motor in the iteratively optimized recursive core tensor is 118kW, the internal loss of the motor is 3kW, the energy flow from the motor to the transmission system is 115kW, and the energy flow from the transmission system to the wheels is 112kW.

[0035] Based on the recursive core tensor, the efficiency difference and compensation coefficient for different transmission paths between components can be calculated. The efficiency difference represents the gap between the actual transmission efficiency and the theoretical efficiency of a specific path, calculated by ratios of the corresponding elements in the recursive core tensor. For normally operating new energy electric vehicles, the transmission efficiency from battery to motor is typically 98%, from motor to transmission system is 97%, and from transmission system to wheels is 96%. When the actual efficiency of a transmission path is significantly lower than the theoretical value, it indicates a potential fault in that path. The introduction of compensation coefficients aims to adjust the model's sensitivity to different types of faults. For motor faults, a higher motor transmission efficiency compensation coefficient of 1.2 is set; for transmission system faults, a higher transmission efficiency compensation coefficient of 1.15 is set. By adjusting the compensation coefficients for different paths, the model's ability to detect specific types of faults is improved.

[0036] Constructing an efficiency compensation optimization equation combines efficiency difference values ​​and compensation coefficients to form a mathematical model for fault diagnosis. The objective is to minimize the difference between the compensated efficiency and the standard efficiency, while also considering the mutual influence between components. The optimization equation contains three terms: an efficiency difference term, representing the gap between the actual efficiency and the standard efficiency; a compensation term, introducing compensation coefficients to adjust the detection sensitivity for different fault types; and a smoothing term, ensuring the continuity of efficiency changes between adjacent components. The weight ratio of each term is set to 3:2:1, reflecting the higher importance placed on efficiency difference detection than on compensation adjustment and smoothing.

[0037] The efficiency compensation optimization equation is solved using gradient descent, with a learning rate of 0.01, a maximum number of iterations of 1000, and a convergence threshold of 0.001. During the optimization process, calculation stops when the iteration error is less than the threshold or when the maximum number of iterations is reached. The solution result represents the energy transfer efficiency between components and can be directly used for fault diagnosis.

[0038] This invention achieves a precise characterization of the energy transfer state within a system through nonnegative decomposition of the energy flow tensor and recursive core tensor calculation, overcoming the limitations of traditional methods in depicting complex coupling relationships among multiple components. The introduction of efficiency difference values ​​and compensation coefficients enhances the method's ability to identify different types of faults, particularly its sensitivity to detecting minor initial faults. The construction and solution of the efficiency compensation optimization equation enables a direct mapping from energy flow data to fault location, achieving high-accuracy diagnosis without requiring extensive historical data or expert experience.

[0039] In one optional embodiment, wavelet transform is performed on the energy transfer efficiency to extract component degradation features, and a component health index is constructed, including: Wavelet transform is performed on the energy transfer efficiency of the component to obtain the wavelet coefficient sequence, which is then divided into continuous degradation component and instantaneous response component according to the frequency range. Operating condition compensation coefficient is constructed, and the wavelet coefficient sequence is compensated using the operating condition compensation coefficient. The component health index is calculated based on the compensated degradation characteristic sequence.

[0040] Wavelet transform was performed on the energy transfer efficiency of the components to obtain wavelet coefficient sequences reflecting different frequency characteristics. The time series of energy transfer efficiency between components was decomposed into five levels using the db4 wavelet basis, including efficiency data for key transfer paths such as battery to motor, motor to transmission system, and transmission system to wheels. For efficiency data collected at a sampling frequency of 500Hz, wavelet coefficient sequences with frequency ranges of 0-7.8Hz, 7.8-15.6Hz, 15.6-31.2Hz, 31.2-62.5Hz, 62.5-125Hz, and 125-250Hz were obtained after five-level wavelet decomposition.

[0041] After obtaining the wavelet coefficient sequence, it needs to be divided into continuous degradation components and transient response components according to frequency range. The continuous degradation component mainly reflects the long-term performance degradation trend of the module, corresponding to low-frequency wavelet coefficients, typically including coefficients in the frequency range of 0-15.6Hz. The transient response component mainly reflects the transient response of the module to changes in operating conditions, corresponding to high-frequency wavelet coefficients, typically including coefficients in the frequency range of 15.6-250Hz. The basis for this division is that the module degradation process usually manifests as a slow and continuous performance decline, while efficiency fluctuations caused by changes in operating conditions manifest as a rapid transient response. The frequency division boundary can be appropriately adjusted according to the module type and operating characteristics; for example, 0-7.8Hz is typically used as the degradation component for battery packs, while 0-15.6Hz can be used for motors.

[0042] A compensation coefficient for operating conditions is constructed, determined based on the current operating status of the new energy electric vehicle, primarily considering factors such as vehicle speed, acceleration, load, and ambient temperature. For vehicle speed, a mapping relationship between vehicle speed and efficiency fluctuations is established: the compensation coefficient is 1.15 within the speed range of 0-20 km / h, 1.0 within the range of 20-60 km / h, and 0.9 within the range of 60-100 km / h. For acceleration, a compensation coefficient is applied under rapid acceleration conditions (acceleration greater than 2 m / s²). 2 The compensation coefficient is 1.2, ensuring smooth acceleration (acceleration between 0.5-2 m / s²). 2The compensation coefficient for the following conditions is 1.05 for constant speed or deceleration conditions and 1.0 for constant speed or deceleration conditions. For load factors, the compensation coefficient is 1.1 for full load, 1.05 for half load, and 1.0 for light load. For ambient temperature factors, the compensation coefficient is 1.15 for low temperature environments (below 0℃), 1.0 for normal temperature environments (0-30℃), and 1.08 for high temperature environments (above 30℃). The compensation coefficients for each factor are combined using a weighted average to form the final operating condition compensation coefficient, with weights of 0.3 for vehicle speed, 0.3 for acceleration, 0.25 for load, and 0.15 for ambient temperature.

[0043] The wavelet coefficient sequence is compensated using operating condition compensation coefficients, primarily adjusting the instantaneous response component to reduce the interference of operating condition changes on fault diagnosis. The compensation process employs a multiplicative model, multiplying the wavelet coefficients of the instantaneous response component by the corresponding operating condition compensation coefficient to obtain the compensated wavelet coefficients. The wavelet coefficients of continuously deteriorating components remain unchanged, preserving their original degradation trend characteristics. For the motor assembly, if the current operating condition is a vehicle speed of 75 km / h and an acceleration of 0.8 m / s²,... 2 With half load and an ambient temperature of 25℃, the calculated compensation coefficient is 0.985. Multiplying this compensation coefficient by the wavelet coefficients of the motor's instantaneous response component (frequency range of 15.6-250Hz) yields the compensated instantaneous response characteristics.

[0044] The component health index is calculated based on the compensated degradation feature sequence to achieve a quantitative assessment of the component's current state and future trends. The health index calculation considers three aspects: degradation level, degradation rate, and degradation acceleration. The degradation level represents the degree of deviation between the component's current efficiency and the standard efficiency, calculated using the mean of the wavelet coefficients of the compensated continuous degradation component. The degradation rate represents the speed at which the component's efficiency decreases, calculated using the slope of the wavelet coefficients of the compensated continuous degradation component. The degradation acceleration represents the acceleration or deceleration trend of the component's degradation process, calculated using the second difference of the wavelet coefficients of the compensated continuous degradation component. The weights of the three factors are set to 0.5, 0.3, and 0.2, respectively, and the weighted sum is used to obtain the final health index. The health index ranges from 0 to 1, where 1 represents a fully healthy state and 0 represents a completely failed state. A health index below 0.7 triggers a warning, below 0.5 triggers an alarm, and below 0.3 recommends immediate repair.

[0045] For example, a four-wheel drive electric vehicle has a battery pack output power of 80kW, a front-drive motor input power of 40kW, and a rear-drive motor input power of 38kW. Calculations show that the energy transfer efficiency from the battery to the front-drive motor is 96%, and the energy transfer efficiency from the battery to the rear-drive motor is 92%. Wavelet transform is applied to these two sets of efficiency data to obtain the wavelet coefficient sequence, which is then divided into a continuous degradation component and an instantaneous response component. The current operating conditions are a vehicle speed of 50km / h and an acceleration of 1.2m / s².2 Under conditions of half-load and ambient temperature of 18℃, the calculated operating condition compensation coefficient was 1.02. After applying operating condition compensation to the wavelet coefficient sequence, the calculated health index of the front drive motor was 0.92, and that of the rear drive motor was 0.65. Based on the health index, the rear drive motor was determined to be in an early stage of deterioration and required preventative maintenance. Inspection by maintenance personnel revealed that the rear drive motor's cooling system was blocked, causing excessive temperature and affecting efficiency. After timely cleaning, normal operation was restored.

[0046] This invention achieves time-frequency domain analysis of component energy transfer efficiency through wavelet transform, effectively separating low-frequency characteristics reflecting intrinsic component degradation from high-frequency characteristics reflecting operating condition response. The introduction of an operating condition compensation coefficient enables precise correction of efficiency fluctuations under different operating conditions, eliminating interference from external factors such as vehicle speed and load on diagnostic results, and significantly improving the accuracy of fault identification. This enhances the accuracy, stability, and predictive power of fault diagnosis in new energy electric vehicle power systems, providing reliable technical support for predictive maintenance, reducing unnecessary downtime, lowering operating costs, and extending system lifespan.

[0047] In one alternative embodiment, constructing a probabilistic graphical network based on a component health index includes: The transition correlation degree is calculated based on the differences between adjacent states in the component health index sequence, and a probabilistic graphical network is constructed based on the transition correlation degree.

[0048] By analyzing the interrelationships between component health index sequences, a probabilistic graphical network reflecting the internal state transition patterns of the system is constructed to achieve accurate fault identification and propagation path tracing. The transition correlation degree is calculated based on the differences between adjacent states in the component health index sequences. The component health index sequences are time-series data obtained through wavelet transform and operating condition compensation, reflecting the health status of components at different times. For each key component in the power system of new energy electric vehicles, such as the battery pack, power converter, motor controller, motor, and transmission system, there is a corresponding health index sequence. Each sequence contains health index values ​​at several time points, typically sampled at certain time intervals, such as once per hour or once per day.

[0049] The calculation of transfer correlation considers two aspects: the changing trend of the health status of the same component at different times, and the mutual influence of the health status between different components. For the transfer correlation of the same component, it is calculated by analyzing the differences between adjacent time points in its health index sequence. The difference in health index is calculated; the absolute value of the difference represents the magnitude of the state change, and the sign of the difference indicates the direction of the change. A time window method is used to process the health index sequence, with a window length of 24 hours and a window sliding step of 1 hour. For the data within each window, the mean, standard deviation, and coefficient of variation of the differences in health index at adjacent time points are calculated, and combined with the assigned weights to obtain the self-transfer correlation of the component. The weights are set as follows: mean weight 0.5, standard deviation weight 0.3, and coefficient of variation weight 0.2. This weight allocation considers both the average degree of state change and the volatility and relative intensity of change.

[0050] The transfer correlation between different components is calculated by analyzing the interaction relationships of their health index sequences. A delayed correlation analysis method is used to detect the lagged impact of a change in the health status of one component on other components. Specifically, different time delays are set for the health index sequences of components A and B, typically from 0 to 24 hours with a step size of 1 hour. The correlation coefficients are calculated for each delay, and the delay time with the strongest correlation and its corresponding correlation coefficient are identified. For example, the maximum correlation coefficient between a decrease in battery pack health and a change in motor controller health is 0.78, corresponding to a delay time of 5 hours, indicating that a change in battery pack status will have a significant impact on the motor controller after 5 hours, and the transfer correlation between the two is 0.78. In addition, conditional transfer correlation is also considered, i.e., the degree of mutual influence between components under specific conditions, such as the potentially more significant impact of the battery pack on the motor under high-temperature conditions. Conditional transfer correlation is obtained by calculating the correlation coefficients under different operating conditions in segments.

[0051] Constructing a probabilistic graphical network based on transfer correlation is the process of mapping each component and its interrelationships to a graph structure. A probabilistic graphical network consists of nodes and edges, where nodes represent individual components in the system, and edges represent the mutual influence relationships between components. The direction of an edge indicates the direction of influence, and the weight of an edge represents the magnitude of the transfer correlation. A transfer correlation threshold is set, typically 0.3, and a connection is established between two components only when their transfer correlation exceeds the threshold. This filters out weak correlations, making the network structure clearer. For self-transfer correlation, it is represented as a self-loop of a node, with the self-loop weight reflecting the autocorrelation of the component's state. Node attributes include component type, current health index, and historical health trend; edge attributes include the magnitude of transfer correlation, response latency, and direction of influence. To further enhance the network, virtual nodes are introduced to represent external environmental factors such as temperature, humidity, and road conditions. While these factors are not internal components of the system, they can still affect component performance.

[0052] After the probabilistic graphical network is constructed, it needs to be trained and its parameters optimized to accurately reflect the system's state transition patterns. Training data includes normal operation data and known fault data. Normal operation data is used to establish the baseline network structure, while known fault data is used to verify the network's ability to identify fault modes. Training employs the expectation-maximization algorithm with 100 iterations and a convergence threshold of 0.01. During training, the transition correlation is continuously updated based on newly acquired health index data, allowing the network structure to dynamically adapt to changes in system state. Network parameter optimization includes adjusting the transition correlation threshold, optimizing node attribute weights, and refining edge attribute descriptions, with the goal of maximizing the network's accuracy in fault identification.

[0053] This invention precisely quantifies the interrelationships of state changes among components within a system by calculating the transition correlation degree obtained from the differences between adjacent states of a component health index sequence. This overcomes the limitations of traditional diagnostic methods that focus only on a single component and ignore the interactions between components. The probabilistic graphical network constructed based on the transition correlation degree intuitively displays the complex state transition paths within the system, making the fault propagation mechanism visible and effectively supporting fault root cause tracing and propagation prediction. The dynamic update mechanism of the probabilistic graphical network enables the diagnostic system to continuously learn, adapt to characteristic changes brought about by equipment aging and changes in the usage environment, and maintain long-term diagnostic accuracy.

[0054] In one optional embodiment, calculating the component failure state transition probability includes: A bidirectional propagation analysis is performed on the probabilistic graphical network to extract the forward influence strength and backward coupling strength of the state transition, and to determine the state transition weights. The state transition weights and transition correlations are weighted and combined to obtain the component failure state transition probability.

[0055] Bidirectional propagation analysis of probabilistic graphical networks allows for the accurate calculation of component failure state transition probabilities, enabling precise fault diagnosis and prediction. This analysis is a crucial tool for gaining a deep understanding of the state transition patterns of components within a system. Bidirectional propagation analysis comprises two directions: forward propagation and backward propagation, revealing the interrelationships between components from causal and consequential perspectives, respectively. Forward propagation analysis focuses on the impact of a component's state change on downstream components, while backward propagation analysis focuses on the correlation between a component's state change and upstream components.

[0056] Forward propagation analysis starts from the source node and propagates downstream along the network edges, calculating the intensity of the impact of the source node's state change on other nodes. The calculation of forward impact intensity uses a decay propagation model, meaning the impact gradually decreases as it propagates along the path. Starting from the source node, propagation proceeds step-by-step along directed edges. For each edge encountered, the impact intensity is multiplied by the transition correlation degree of that edge, while also considering the decay factor due to the path length. The decay factor is set to the power of the baseline decay rate, with the default baseline decay rate being 0.85. During propagation, if multiple paths exist from the source node to the target node, the principle of maximum impact is applied, selecting the path with the strongest impact intensity as the final forward impact intensity. For circular paths, a maximum number of loops is set to 3; propagation stops after exceeding this number to avoid infinite loops.

[0057] Backpropagation analysis starts from the target node and traces upstream against the network edges to calculate the coupling strength between the target node's state changes and the upstream nodes. The backward coupling strength calculation uses a reverse attribution model, meaning that the target node's state changes are proportionally attributed to upstream nodes. Starting from the target node, propagation proceeds upstream against the directed edges. For each edge encountered, the coupling strength is multiplied by the edge's transition correlation degree and the node's contribution rate. The node contribution rate represents the proportion of contribution a specific upstream node makes to the target node's state changes under the influence of multiple upstream nodes; it is obtained by normalizing the transition correlation degrees of each upstream node. Similarly, for multiple paths from the target node to the source node, the path with the highest coupling strength is selected as the final backward coupling strength. In complex networks, the node's self-state maintenance capability, i.e., self-transition correlation degree, must also be considered. A higher value indicates a more stable node state and less susceptibility to external influences.

[0058] After extracting the forward influence strength and backward coupling strength of state transitions, it is necessary to determine the state transition weights, which comprehensively reflect the probability of state transitions between components. The state transition weights are determined by a weighted combination of the forward influence strength and the backward coupling strength, calculated as the forward influence strength multiplied by the forward weight coefficient plus the backward coupling strength multiplied by the backward weight coefficient. The sum of the forward and backward weight coefficients is 1, with default settings of 0.6 and 0.4, indicating that the forward influence accounts for a larger proportion in state transitions. In specific application scenarios, the weight coefficients can be adjusted according to system characteristics. For example, for rapidly responding power electronic components, the forward weight can be appropriately increased; for cooling systems with significant thermodynamic characteristics, the backward weight can be appropriately increased. After calculating the state transition weights, normalization processing is required to ensure that the sum of all possible state transition weights is 1, facilitating subsequent probability calculations.

[0059] The component failure state transition probability is obtained by weighting and combining the state transition weights and the transition correlation. The component failure state transition probability represents the likelihood that, under given conditions, the failure state of one component will transition to that of another component; it is a core indicator for fault diagnosis and prediction. The weighted combination uses a linear weighting method, i.e., the state transition weight is multiplied by a weight coefficient α, plus the transition correlation multiplied by a weight coefficient β, where the sum of α and β is 1, with default settings of 0.7 and 0.3. The choice of weight coefficients reflects the relative importance of the state transition weights and transition correlation in the final probability calculation; a larger weighting of the state transition weights indicates that the bidirectional propagation analysis results are more decisive for predicting failure state transitions. Different weight coefficients can be used for different types of component pairs in the system to adapt to their specific fault propagation characteristics.

[0060] To improve the accuracy of state transition probabilities, both the component's intrinsic characteristics and external environmental factors must be considered. Component characteristics include reliability level, lifespan, and maintenance cycle, while external environmental factors include temperature, humidity, and vibration intensity. These factors are incorporated into the state transition probability calculation through correction factors, which are determined based on historical data and expert experience. The corrected component failure state transition probability serves as the final basis for fault diagnosis and prediction. A high transition probability indicates a greater risk of fault propagation, requiring priority monitoring and prevention.

[0061] like Figure 2 As shown, Figure 2This chart compares the lead time for fault prediction in this embodiment, demonstrating the predictive capabilities of different methods for various component faults. This technical solution achieves a lead time of 27.5 hours for battery component fault prediction, 4.7 hours longer than the closest time-series analysis model. For control system faults, this solution achieves a lead time of 24.8 hours, while other methods are all below 20 hours. For transmission system faults, this solution achieves a lead time of 20.3 hours, while traditional methods average only 13.2 hours. For cooling system fault prediction, this solution achieves a lead time of 33.6 hours, significantly higher than the 25.8 hours of other methods. For sensor faults, this solution detects anomalies 18.5 hours earlier, while the threshold monitoring method only achieves 8.7 hours. Overall, the average lead time for fault prediction in this technical solution is 24.9 hours, 9.6 hours longer than traditional methods on average.

[0062] This invention overcomes the limitations of traditional unidirectional causal analysis by performing bidirectional propagation analysis on probabilistic graphical networks. It considers both the intensity of the fault's propagation from its source downstream and the coupling relationship between the result and the upstream cause, achieving comprehensive and multi-faceted fault correlation analysis. The extraction and combination of forward influence strength and backward coupling strength makes fault propagation path identification more accurate, especially for systems with complex interactions, revealing hidden fault transmission mechanisms. The weighted combination of state transition weights and transition correlation degrees considers both static network structure and dynamic propagation characteristics, overcoming the inability of a single indicator to fully reflect system complexity.

[0063] In one optional embodiment, variational inference is used to obtain the fault coupling distribution between components, and the fault warning threshold is adaptively adjusted according to the fault coupling distribution, including: A variational distribution of component fault coupling is constructed using conditional random fields, and the variational distribution is optimized by expectation maximization iteration to obtain the fault coupling distribution between components. The correlation between component faults is calculated based on the fault coupling distribution, and the fault early warning threshold is adaptively adjusted by combining the fault propagation risk coefficient.

[0064] First, a Conditional Random Field (CRF) model is constructed to analyze the fault coupling relationships between components. Based on the fault coupling distribution, the warning threshold is adaptively adjusted to achieve accurate prediction and early warning of system faults. Constructing the variational distribution of component fault coupling using CRF is a key step in understanding the internal fault propagation mechanism of the system. CRF is an undirected graphical model suitable for describing the interdependencies of fault states among components in a system. In the power system of new energy electric vehicles, various components such as battery packs, battery management systems, power electronic converters, motor controllers, and drive motors are all susceptible to failure, and the fault states of these components are often interconnected.

[0065] When constructing a Conditional Random Field (CRF) model, each component in the system is represented as a node in a graph, and the fault coupling relationships between components are represented as edges between nodes. Node features include the component's basic attributes, historical fault records, and current health status. Edge features describe the physical connections, functional dependencies, and signal transmission characteristics between components. To accurately represent node and edge features, feature functions are introduced, which can be divided into node feature functions and edge feature functions. Node feature functions measure the degree of matching between the fault state of a single component and its attributes, while edge feature functions measure the degree of coordination between the fault states of connected components. Each feature function also corresponds to a weight parameter, which reflects the importance of the feature in determining the fault state.

[0066] Node feature functions may include: the correspondence between component health index and fault state, the relationship between component lifespan and fault probability, and the impact of environmental factors on fault probability. Edge feature functions may include: the consistency of fault states of adjacent components, the temporal characteristics of fault propagation between components, and the impact of signal interference on fault propagation. To enhance the expressive power of the model, different forms of feature functions can be introduced, such as linear functions, threshold functions, and Gaussian functions, with the appropriate function form selected based on the specific scenario.

[0067] The variational distribution of component failure coupling represents the possible combinations of failure states of each component in a system and their probability distribution. Variational distributions are typically constructed using methods such as the mean-field approximation or the structured mean-field approximation to reduce computational complexity. In the mean-field approximation, it is assumed that the failure states of each component are independent, and the overall joint distribution can be represented as the product of the marginal distributions. The structured mean-field approximation, on the other hand, preserves some of the dependencies between components, making it closer to the actual system. The parameters of the variational distribution include the marginal probabilities of each component's failure state and the mutual information between component pairs; these parameters need to be obtained through data learning.

[0068] The core process for obtaining the fault coupling distribution between components is to optimize the variational distribution using the expectation-maximization (EM) iterative method. The E-step algorithm consists of two alternating steps: the E-step and the M-step. The E-step fixes the model parameters and estimates the posterior distribution of the latent variables; the M-step fixes the distribution of the latent variables and optimizes the model parameters. In the fault diagnosis scenario, the E-step estimates the fault state distribution of each component, and the M-step updates the weight parameters of the feature function. The E-step calculates the approximate posterior distribution using variational inference methods, including the edge failure probability of components and the joint failure probability of component pairs; the M-step updates the feature function weights using gradient descent to make the distribution generated by the model as close as possible to the distribution of the observed data.

[0069] In iterative optimization, setting appropriate convergence conditions is crucial. Common convergence conditions include the magnitude of parameter changes, the change in log-likelihood value, and an upper limit on the number of iterations. The threshold for the magnitude of parameter changes is typically set to 0.001, the threshold for the change in log-likelihood value is set to 0.01, and the upper limit on the number of iterations is set to 100. When any of these conditions is met, the iteration stops, and the final variational distribution parameters are output. To prevent overfitting, a regularization term can be introduced. Common regularization methods include L1 regularization and L2 regularization. The former promotes sparsity of the weight parameters, while the latter prevents the weight parameters from becoming too large.

[0070] After obtaining the inter-component fault coupling distribution using the expectation-maximization algorithm, it is necessary to calculate the component fault correlation, which quantifies the degree of interrelation between the fault states of different components. Component fault correlation can be calculated using methods such as mutual information, conditional probability, or correlation coefficient. Mutual information represents the amount of information about the fault state of another component contained in the fault state of one component; conditional probability represents the probability that another component is in a fault state given that the fault state of one component is known; and the correlation coefficient represents the degree of linear correlation between the fault states of two components. These indicators can be considered comprehensively and weighted to obtain the final component fault correlation.

[0071] The fault propagation risk coefficient is a modified index that incorporates timing and importance factors based on the correlation of component faults. The timing factor considers the time delay of fault propagation, while the importance factor considers the functional criticality of the component within the system. The timing factor can be determined by analyzing historical data to determine the sequence and time intervals of faults among components, while the importance factor can be determined based on the functional classification of the component and the scope of its impact. The fault propagation risk coefficient is calculated by multiplying the component fault correlation by the timing factor and then by the importance factor. The larger the coefficient, the higher the probability and severity of fault propagation.

[0072] Adaptive adjustment of fault warning thresholds based on fault propagation risk coefficients is a key technology for achieving accurate early warning. Traditional early warning methods typically use fixed thresholds, which are difficult to adapt to changes in system state and environment. Adaptive early warning thresholds are dynamically adjusted based on the fault propagation risk coefficient; components with higher risk coefficients correspond to lower warning thresholds, thereby detecting potential faults earlier. The adaptive adjustment formula is the baseline threshold minus the product of the fault propagation risk coefficient and an adjustment coefficient. The adjustment coefficient can be set according to the early warning sensitivity requirements, typically between 0.05 and 0.2. A time factor can also be introduced, allowing the warning threshold to change over time; for example, a more lenient threshold can be used for new components, while a more stringent threshold can be used for aging components.

[0073] For example, the system comprises five core components: a battery module, a battery management system, a DC-DC converter, an inverter, and a drive motor. State parameters of these components during operation are collected, including voltage, current, temperature, and vibration, while historical fault data is recorded, including fault time, fault type, and faulty component. Based on this data, a conditional random field model is constructed, designing 18 node characteristic functions and 12 edge characteristic functions, with initial weights set based on expert experience. A variational distribution is constructed using the mean-field variational approximation, and iterative optimization is performed using the expectation-maximization algorithm, reaching convergence after 32 iterations. The obtained fault coupling distribution shows that the mutual information value between the battery management system and the DC-DC converter is 0.68, indicating a high correlation between their faults. The component fault correlation matrix is ​​calculated; the correlation between battery management system faults and DC-DC converter faults is 0.72. Considering a time series factor of 1.2 and an importance factor of 1.1, the fault propagation risk coefficient is 0.95. Based on this risk coefficient, the voltage fluctuation warning threshold for the DC-DC converter is adaptively adjusted from the standard value of 5% to 3.4%. During actual operation, a 4.2% fluctuation in the output voltage of the DC-DC converter was detected, which was below the standard threshold but above the adaptive threshold, triggering a system warning. Inspection revealed an intermittent fault in the battery management system's communication module, causing abnormal control signals and consequently unstable DC-DC converter output. Replacing the communication module promptly restored normal system operation.

[0074] This invention utilizes conditional random fields to construct a variational distribution of component fault coupling, overcoming the limitations of traditional models that assume independent faults. It accurately captures the complex fault dependencies within the system, significantly improving the accuracy of fault diagnosis. Combined with an innovative mechanism that adaptively adjusts the warning threshold based on the fault propagation risk coefficient, it achieves a shift from a "one-size-fits-all" approach to personalized warnings, increasing sensitivity to critical faults while reducing false alarm rates. This method is particularly suitable for new energy electric vehicle power systems with complex fault mechanisms and frequent component interactions, significantly improving the foresight, accuracy, and reliability of system fault prediction.

[0075] In one optional embodiment, adaptively adjusting the fault warning threshold based on the fault coupling distribution includes: The amplitude and phase features of the fault coupling distribution are extracted, the rate of change of the amplitude features and the drift of the phase features are calculated, a correction function for the fault warning threshold is constructed, the adjustment coefficient of the warning threshold is calculated based on the correction function, and the fault warning threshold is dynamically updated.

[0076] The amplitude and phase characteristics of fault coupling distribution are extracted, and the early warning threshold is dynamically adjusted based on the changes in these characteristics to achieve early prediction and accurate warning of faults. The amplitude characteristics of fault coupling distribution characterize the strength of the mutual influence between faults in components and are a key indicator for quantifying the possibility of fault propagation. In the power system of new energy electric vehicles, the fault coupling strength between different components often exhibits different amplitude characteristics, reflecting the tightness of the functional and physical connections between the components.

[0077] When extracting amplitude features, the fault coupling distribution data needs to be preprocessed, including filtering, denoising, and normalization. Filtering uses a low-pass filter with a cutoff frequency set to 1 / 5 of the sampling frequency to effectively remove high-frequency noise interference. Denoising employs wavelet transform, selecting the db4 wavelet basis function, with a decomposition level of 4, and using soft thresholding to threshold the wavelet coefficients. Normalization maps the amplitude features to the 0-1 interval, facilitating comparisons between different components. Amplitude feature extraction primarily focuses on four statistics: peak value, mean, variance, and skewness. The peak value reflects the maximum strength of fault coupling, the mean represents the overall coupling level, the variance describes the fluctuation of coupling strength, and the skewness reflects the degree of asymmetry in the coupling distribution. These statistics together constitute the amplitude feature vector, comprehensively characterizing the strength characteristics of fault coupling.

[0078] Extracting phase features from fault coupling distribution is an effective way to analyze the timing relationships of faults in system components. Phase features characterize the temporal correlation of fault occurrence between components, that is, the time difference between the fault state changes of one component and the fault state changes of another component. Phase feature extraction first requires converting time-domain data into frequency-domain data; the commonly used transformation method is Fourier transform. For non-stationary signals, short-time Fourier transform or wavelet transform can be used to obtain time-varying phase information. In the frequency domain, phase features mainly focus on indicators such as phase angle, phase difference, group delay, and phase consistency. The phase angle represents the initial phase of each frequency component, the phase difference reflects the phase offset between the two component signals, the group delay describes the rate of change of the phase angle with respect to frequency, and phase consistency quantifies the stability of the phase relationship.

[0079] Phase feature extraction needs to consider the multi-scale characteristics of the signal. Multi-resolution decomposition of the fault coupling distribution and analysis of phase features at different scales can capture the temporal patterns of fault propagation. Multi-scale analysis employs wavelet packet transform with a decomposition scale of 3, obtaining phase information from 8 frequency bands. For fault modes with obvious periodicity, such as communication faults between the battery management system and the motor controller, Hilbert transform can be used to extract the instantaneous phase, intuitively reflecting the fault propagation delay. For fault modes with strong randomness, such as performance degradation caused by temperature fluctuations, cross-correlation functions are used to analyze the phase relationship and determine the time delay corresponding to the maximum correlation.

[0080] Calculating the rate of change of amplitude characteristics is a crucial step in identifying fault development trends. The rate of change of amplitude characteristics represents the magnitude of change in fault coupling strength per unit time, directly reflecting the speed of fault development. The rate of change is calculated using the central difference method, which approximates the amplitude characteristics at continuous time points through differentiation. To avoid the influence of instantaneous fluctuations on the rate of change estimation, a sliding window averaging is introduced, with a window length of 10 time units. The rate of change can be divided into short-term and long-term rates. The short-term rate of change focuses on the trend within the last 30 minutes, while the long-term rate of change focuses on the overall trend within 24 hours. The sign of the rate of change indicates the direction of increase or decrease in fault coupling strength; a positive value indicates an increase in fault risk, and a negative value indicates a decrease in fault risk. The absolute value of the rate of change indicates the degree of drastic change; the larger the absolute value, the faster the fault develops, requiring more timely early warning.

[0081] Calculating the drift of phase characteristics is an important means of analyzing changes in fault propagation modes. The drift of phase characteristics represents the stability of the fault timing relationship between components; a large drift indicates an unstable fault propagation path, while a small drift indicates a fixed fault propagation pattern. The drift is calculated using the standard deviation of the phase difference, reflecting the degree of fluctuation in the phase relationship. To enhance the timeliness of the drift, an exponentially weighted moving average is introduced, giving higher weight to recent phase changes. The drift calculation also needs to consider the periodicity of the phase; for phase difference changes close to 2π, phase expansion processing is required to avoid calculation errors. In multi-component systems, a phase drift matrix can be constructed to comprehensively describe the phase relationship changes of each component pair within the system.

[0082] Constructing a correction function for the fault warning threshold is the core step in achieving dynamic threshold adjustment. The correction function maps the amplitude characteristic change rate and phase characteristic drift to threshold adjustment coefficients, enabling the warning threshold to adaptively change according to the fault development characteristics. The correction function is designed as a piecewise function, setting different mapping relationships for different amplitude change rate and phase drift ranges. When the amplitude change rate is positive and the phase drift is small, it indicates that the fault is accelerating according to a fixed pattern; the correction function outputs a large negative adjustment coefficient to lower the warning threshold and issue an early warning. When the amplitude change rate is negative and the phase drift is large, it indicates that the fault fluctuates randomly but generally tends to ease; the correction function outputs a moderate positive adjustment coefficient to raise the warning threshold and avoid false alarms. When both the amplitude change rate and phase drift are at a moderate level, the correction function outputs an adjustment coefficient close to zero, maintaining a relatively stable warning threshold.

[0083] The calculation of the early warning threshold adjustment coefficient based on the correction function serves as a bridge connecting feature analysis and threshold adjustment. The early warning threshold adjustment coefficient is the output of the correction function and is directly used for the dynamic updating of the early warning threshold. The calculation of the adjustment coefficient requires comprehensive consideration of multiple statistics of the amplitude characteristic change rate and multiple scales of the phase characteristic drift. Weighting factors are introduced during the calculation process; the weights of different statistics and scales can be obtained through training with historical data or set based on expert experience. To improve the stability of the adjustment coefficient, a limiting process is introduced, restricting the adjustment coefficient to between -0.5 and 0.5 to prevent excessive adjustment from causing drastic threshold fluctuations. An adjustment step size constraint is set, with each adjustment not exceeding 10% of the baseline threshold, ensuring smooth and controllable threshold changes.

[0084] The fault warning threshold is dynamically updated by combining the baseline warning threshold with an adjustment coefficient to obtain a real-time warning threshold that adapts to the current fault development characteristics. The baseline warning threshold is set based on historical data statistics or expert experience, representing the fault judgment standard under normal conditions. Dynamic updates employ either an additive or multiplicative model; the additive model is suitable for absolute thresholds, while the multiplicative model is suitable for relative thresholds. The update frequency is determined based on system characteristics; critical components such as the battery management system can use a higher frequency, such as once per minute, while auxiliary components such as the cooling system can use a lower frequency, such as once every 10 minutes. Upper and lower limits must also be set for threshold updates to prevent the threshold from becoming too high or too low under extreme conditions.

[0085] This invention accurately captures changes in fault development trends and propagation patterns by calculating the rate of change of amplitude characteristics and the drift of phase characteristics, thus enabling fault prediction to be timely and forward-looking. A correction function for the fault warning threshold is constructed, establishing a mapping relationship between characteristic changes and threshold adjustments, solving the problem that fixed thresholds are difficult to adapt to dynamic system changes. Based on the correction function, the warning threshold adjustment coefficient is calculated, quantifying the characteristic analysis results into specific adjustment parameters, realizing the transformation from qualitative analysis to quantitative adjustment. Dynamic updates to the fault warning threshold enable the warning mechanism to adaptively adjust according to the fault development characteristics, effectively reducing false alarm and missed alarm rates while ensuring timely warnings.

[0086] In one optional embodiment, determining the fault source component and affected components based on the fault coupling distribution, and generating fault propagation path and remaining lifetime warning information includes: Maximum entropy time series analysis is used to calculate the component failure propagation probability; The fault coupling distribution is prioritized based on the fault propagation probability to identify the fault source component, and the affected components whose coupling strength with the fault source component exceeds a preset strength threshold are extracted. Construct a propagation relationship graph between the fault source component and the affected components to obtain the fault propagation path; The component coupling accumulation value is calculated based on the fault propagation path, and the remaining lifespan warning information is generated by combining the component health index.

[0087] First, a time-series state space needs to be constructed, dividing the historical operating data of each component of the trolley power system into several states, such as normal, slightly abnormal, moderately abnormal, and severely abnormal. State division can be based on key performance indicators of the components, such as the voltage, temperature, and internal resistance of the battery components; the current, speed, and torque of the motor components; and the communication quality and response time of the electronic control unit. For continuous indicators, uniform binning or thresholds set based on expert experience can be used for discretization. Time-series data collection should cover a sufficiently long time period, generally recommended to be no less than 3 months, to ensure that system behavior under various operating conditions and environmental conditions is captured.

[0088] After the state space is constructed, the Lagrange multiplier method is introduced to solve the maximum entropy model. The constraint condition is set to match the expected value of each component's state transition with the expected value of historical observation data. During the iterative solution process, an improved generating function method is used to accelerate convergence. The iteration termination condition is that the entropy change between two consecutive iterations is less than a preset threshold, typically 0.001. After the maximum entropy model is trained, the state transition probability matrix between the system's components can be obtained. This matrix represents the probability distribution of other components transitioning to their respective states at the next time step, given that the current component is in a certain state. The fault propagation probability is the conditional probability in the state transition probability matrix of transitioning from an abnormal state of one component to an abnormal state of another component.

[0089] The fault coupling distribution is prioritized based on the fault propagation probability to identify fault source components. The fault coupling distribution represents the strength relationship of the mutual influence of faults among components in the system, and can be represented by an adjacency matrix. For n components in a trolley power system, an n×n fault coupling matrix is ​​constructed, where the matrix element (i,j) represents the degree of fault influence of component i on component j, and the value range is usually normalized to the [0,1] interval. Prioritization is based on the asymmetry of fault propagation, calculating the outgoing and incoming degrees of each component. The outgoing degree represents the total strength of the influence of the component on other components, and the incoming degree represents the total strength of the influence of the component on other components. The difference between the outgoing degree and the incoming degree is the net influence of the component, and the component with the largest net influence is likely to be the fault source component.

[0090] Once the fault source component is identified, it is necessary to extract the affected components whose coupling strength exceeds a preset strength threshold. The strength threshold is set based on system reliability requirements, typically between 0.3 and 0.7. A lower threshold includes more affected components, resulting in a more comprehensive analysis, but also a greater computational burden; a higher threshold retains only the most significant influence relationships, providing a more focused analysis but potentially missing potential risks. The threshold can be dynamically adjusted based on the system's criticality and computational resources. The time-varying characteristics of coupling strength must also be considered during the extraction process. Some components may only exhibit strong coupling relationships under specific operating conditions; therefore, the stability of the coupling strength should be verified under various operating conditions.

[0091] Constructing a propagation graph between the fault source component and affected components is an effective means of visualizing fault propagation paths. The propagation graph adopts a directed weighted graph structure, where nodes represent system components, edges represent fault propagation paths, and edge weights represent propagation probabilities or coupling strengths. The graph is constructed using a sparse representation, retaining only coupling relationships exceeding a threshold to reduce computational complexity. For large dynamic systems, hierarchical clustering methods can be introduced to aggregate functionally similar components into subsystems, reducing the number of nodes while retaining key structural information. After the graph is constructed, an improved depth-first search algorithm is used to identify propagation paths from the fault source to each affected component. The path weight is defined as the product of the weights of all edges along the path, representing the probability of the fault propagating along that path.

[0092] Calculating the cumulative coupling value of components based on the fault propagation path is an important indicator for assessing the severity of a fault. The cumulative coupling value considers the cascading effect of fault propagation, focusing not only on direct impacts but also on the accumulation of indirect effects. The calculation process employs an iterative method. Initially, only the fault source component has a cumulative coupling value of 1, while other components have a value of 0. In each iteration, the cumulative coupling value of a component is updated by adding its current value to the influence transmitted from all predecessor nodes. The influence is calculated by multiplying the cumulative coupling value of the predecessor node by its corresponding edge weight. The iteration terminates when the change in the cumulative coupling value of all nodes is less than a preset convergence threshold, typically 0.01. The final cumulative coupling value reflects the overall impact of the fault source on each component; a larger value indicates a more severe impact of the fault on that component.

[0093] The component health index is a comprehensive indicator that measures the difference between the current state and the ideal state of a component. Health calculation comprehensively considers multi-dimensional information such as component performance parameters, operating time, environmental conditions, and historical maintenance records. For battery components, health can be based on capacity decay rate, internal resistance growth rate, and charge / discharge efficiency; for motor components, health can be based on output power fluctuations, temperature rise characteristics, and noise levels; for electronic control units (ECUs), health can be based on computational load, response delay, and error rate. The overall health score is synthesized using a weighted average method, with weights determined using expert experience or sensitivity analysis based on historical data. The health index is typically normalized to the 0-1 range, where 1 represents complete health and 0 represents complete failure.

[0094] The remaining lifetime warning information is generated by combining the component coupling cumulative value and the health index. The remaining lifetime prediction uses a degradation model approach, treating component health as a process of degradation over time. The lifetime is defined as the time it takes for the health to gradually decrease from its initial value to the failure threshold. The degradation model can be a linear model, an exponential model, or a Weibull model, and the model parameters are obtained by fitting historical data. The coupling cumulative value serves as a moderating factor for the degradation rate; a larger coupling cumulative value indicates a faster degradation rate and a shorter predicted lifetime. Warning information generation is based on a comparison of the remaining lifetime with a preset warning time threshold. A warning is triggered when the remaining lifetime falls below the threshold. Warning levels are divided into three levels: alert, warning, and emergency, each corresponding to different remaining lifetime ranges.

[0095] For example, a fault diagnosis analysis was performed on the powertrain system of an electric vehicle. The system comprises five main components: a battery management unit (BMU), a motor controller, a power conversion unit, a thermal management unit, and an on-board diagnostic unit (OBD). Three months of operational data were collected, and maximum entropy time-series analysis was used to calculate the fault propagation probability matrix between components. The analysis results showed that the fault propagation probability from the BMU to the motor controller was 0.72, to the power conversion unit 0.65, to the thermal management unit 0.43, and to the OOD unit 0.38; while the fault propagation probability from other components to the BMU did not exceed 0.25. Based on the propagation probabilities, the net influence of each component was calculated, and the BMU had the largest net influence, thus identifying it as the fault source component.

[0096] A coupling strength threshold of 0.4 was set, and the motor controller and power conversion unit were identified as the main affected components. A fault propagation graph was constructed, identifying three main propagation paths: Battery Management Unit → Motor Controller (0.72); Battery Management Unit → Power Conversion Unit (0.65); Battery Management Unit → Motor Controller → Thermal Management Unit (0.31). The cumulative coupling values ​​for the components were calculated: Battery Management Unit = 1, Motor Controller = 0.72, Power Conversion Unit = 0.65, Thermal Management Unit = 0.22, and On-board Diagnostics Unit = 0.38. Based on the current health index (Battery Management Unit 0.78, Motor Controller 0.85, Power Conversion Unit 0.82), the predicted remaining lifespan was 8 months for the Battery Management Unit, 10 months for the Motor Controller, and 11 months for the Power Conversion Unit, generating a medium-level warning and recommending priority repair of the Battery Management Unit. Inspection by maintenance personnel revealed that the temperature sensor in the Battery Management Unit exhibited drift, leading to improper battery charge / discharge control strategies, which in turn affected the performance of the Motor Controller and Power Conversion Unit. After replacing the temperature sensor, the system returned to normal operation.

[0097] This invention employs maximum entropy time-series analysis to calculate the probability of component fault propagation, overcoming the limitations of traditional methods in handling uncertain data and improving the accuracy of fault propagation pattern recognition. Based on the fault propagation probability, priority ranking is performed to determine the fault source component, enabling the tracing of the root cause from surface fault phenomena and avoiding repetitive problems caused by addressing only the symptoms. A fault propagation relationship graph is constructed and the propagation path is analyzed, visually demonstrating the diffusion process of fault impact and providing a basis for targeted fault isolation and control.

[0098] In one optional embodiment, calculating the component failure propagation probability using maximum entropy time series analysis includes: Extract the temporal change characteristics of component states, calculate the state transition entropy values ​​between components, and obtain the fault propagation probability based on the state transition entropy values.

[0099] First, component status refers to the operating condition of each component in the electric vehicle's power system at a specific moment, which can be characterized by performance parameters collected by multiple sensors. For the battery management system, status parameters include cell voltage, battery temperature, state of charge, and internal resistance; for the motor controller, status parameters include phase current, speed, torque, and control signals; for the power converter, status parameters include switching frequency, voltage ripple, and conversion efficiency; and for the thermal management system, status parameters include coolant temperature, flow rate, and radiator efficiency. These raw parameters often suffer from inconsistencies in dimensions, noise interference, and redundant information, requiring data preprocessing and feature extraction.

[0100] Data preprocessing includes outlier detection and handling, missing value imputation, and signal filtering. Outlier detection uses the three-standard-deviation method, marking data points exceeding three standard deviations of the mean as outliers and replacing them with local averages. Missing value imputation uses linear interpolation, supplementing missing segments with historical data under the same operating conditions for larger time intervals. Signal filtering uses a low-pass filter with a cutoff frequency set to 1 / 5 of the main signal frequency to remove high-frequency noise interference. After preprocessing, a sliding window method is used to extract time-series variation features, with a window length of 5 minutes and a sliding step of 1 minute to ensure the capture of dynamic changes in component states.

[0101] Within each window, statistical and trend characteristics are extracted. Statistical characteristics include mean, standard deviation, kurtosis, and skewness, used to describe the distribution characteristics of the parameters; trend characteristics include slope, curvature, and fluctuation frequency, used to describe the dynamic changes of the parameters. For battery systems, particular attention is paid to voltage drop rate, temperature rise rate, and internal resistance growth rate, as these indicators effectively reflect the battery's health status. For motor systems, particular attention is paid to torque fluctuation rate, current imbalance, and control response time, as these indicators reflect the stability of motor control. For power conversion systems, particular attention is paid to efficiency degradation rate, switching loss growth rate, and harmonic content, as these indicators reflect the quality of power conversion.

[0102] After feature extraction, continuous feature values ​​need to be mapped to discrete state values ​​to facilitate subsequent calculation of state transition entropy. State discretization employs a uniform binning method, dividing the value range of each feature into several equal intervals, each interval corresponding to a state code. State codes are combined to form the component's state vector; for example, the state of a battery management system can be represented as a combination of "voltage state code - temperature state code - internal resistance state code". To reduce the dimensionality of the state space, principal component analysis is used to reduce the dimensionality of the state vector, retaining principal components with a cumulative contribution rate of 95%. Typically, the dimensionality of the reduced state space is between 3 and 5. The state of each component within each time window is mapped to a discrete state code, forming a time series of component states.

[0103] The state transition entropy between components is calculated. State transition entropy is an indicator in information theory that measures the strength of the causal relationship between two time series, effectively capturing nonlinear and asymmetric information transfer processes. For any two components A and B in a trolley power system, the state transition entropy from A to B is calculated to assess the impact of changes in the state of A on changes in the state of B. The calculation of state transition entropy is based on conditional probability, considering the statistical dependence of the current state of component B on its own historical states and the historical states of component A.

[0104] The core of state transition entropy calculation is estimating the joint probability distribution and conditional probability distribution. The probability distribution estimation uses the historical frequency method, which involves counting the frequency of each state combination in the component's state sequence and dividing by the total number of samples to obtain the probability estimate. To improve the robustness of the estimation, especially for sparse state spaces, a Laplace smoothing technique is introduced, adding a small positive number (usually 0.5) to the count of each state combination to avoid the zero-probability problem. The influence length of historical states (i.e., the time lag order) is determined through autocorrelation analysis; for battery systems, it is typically taken as 2-3 orders, for motor systems as 1-2 orders, for power conversion systems as 1 order, and for thermal management systems as 3-4 orders.

[0105] After determining the joint probability distribution and the conditional probability distribution, the state transition entropy is calculated by weighting the logarithmic terms of the conditional probabilities. The weights are the corresponding joint probabilities, ensuring that more frequent state transitions contribute more to the entropy value. The larger the state transition entropy value, the more significant the influence of the source component on the target component, the stronger the information flow, and the higher the potential risk of fault propagation. For a system with n components, all possible component pairs are calculated to obtain an n×n state transition entropy matrix, where the matrix element (i,j) represents the state transition entropy value from component i to component j.

[0106] The fault propagation probability is derived from the state transition entropy. Fault propagation probability is defined as the conditional probability that a failure in one component will cause another component to fail within a specific time window. There is a monotonically increasing relationship between the state transition entropy and the fault propagation probability; the higher the entropy value, the higher the propagation probability. The transition function uses an S-shaped curve to ensure that the propagation probability ranges between 0 and 1, and exhibits high sensitivity in the intermediate entropy range.

[0107] The function parameters are calibrated using historical fault data. The calibration process employs maximum likelihood estimation, selecting the parameter combination that best matches the model's predicted fault propagation probability with historical observations. For newly commissioned tram systems, parameters from similar systems can be used as initial values, and the calibration is gradually updated as operational data accumulates. To improve the model's generalization ability and prevent overfitting, a regularization term is introduced; the regularization strength is determined through cross-validation. After calibration, a mapping function from state transition entropy to fault propagation probability is obtained, leading to the fault propagation probability matrix between components.

[0108] Analysis revealed that the battery management unit (BMU) was the primary source of failure, with its faults most easily propagating to the motor controller and power converter. The thermal management system (TMS) was the most vulnerable component, susceptible to failures in all other components. For example, if the system detected an abnormality in the BMU's temperature control function, fault propagation probability prediction indicated that the motor controller and power converter would be affected in the short term. Targeted investigation revealed that battery temperature sensor drift caused misjudgments in the control algorithm, and overcharging and over-discharging of the battery led to voltage fluctuations, which in turn interfered with the normal operation of the motor controller and power converter. After repairing the temperature sensor, the system returned to normal operation, validating the accuracy of the fault propagation path analysis.

[0109] This invention, by extracting the temporal change characteristics of component states, can comprehensively capture the dynamic operating characteristics of each component in a trolleybus power system, providing a rich source of information for fault propagation analysis. By employing state transition entropy to calculate the information flow between components, it overcomes the limitations of traditional correlation analysis methods, effectively identifying nonlinear and asymmetric causal relationships and improving the accuracy of fault source component identification. Applicable to new energy trolleybus power systems with complex inter-component interactions, it can significantly reduce false diagnosis rates, shorten diagnosis time, lower maintenance costs, improve system reliability, and extend the service life of trolleybuses.

[0110] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method described in any of the foregoing embodiments.

[0111] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps in the method described in any of the preceding claims.

[0112] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A fault diagnosis method for the power system of new energy electric vehicles, characterized in that, Includes the following steps: Construct an energy flow tensor network between components to map the power parameters of each component in the power system of new energy electric vehicles into energy flow tensors; The energy transfer efficiency between components is obtained through tensor decomposition. Wavelet transform is used to extract component degradation features from energy transfer efficiency, and a component health index is constructed. A probabilistic graphical network is constructed based on the component health index to calculate the probability of component failure state transition. The fault coupling distribution among components is obtained by variational inference, and the fault early warning threshold is adaptively adjusted according to the fault coupling distribution. When the component health index is lower than the fault warning threshold, the fault source component and the affected component are determined based on the fault coupling distribution, and fault propagation path and remaining lifetime warning information are generated.

2. The method according to claim 1, characterized in that, Constructing an energy flow tensor network between components to map the power parameters of each component in the power system of a new energy electric vehicle to energy flow tensors includes: Collect power parameters of the power system components and construct an initial mapping matrix; The initial mapping matrix is ​​scaled to obtain a multi-scale energy flow feature sequence, and an energy flow tensor network is constructed. The path weights are calculated based on the energy flow tensor network, and a mapping function is constructed based on the path weights to map the power parameters of each component to the energy flow tensor.

3. The method according to claim 1, characterized in that, The energy transfer efficiency between components obtained through tensor decomposition includes: The energy flow tensor is decomposed into nonnegative components to obtain the initial core tensor. Construct the energy balance constraints of the initial core tensor and calculate the recursive core tensor; The efficiency difference and compensation coefficient between different transmission paths between components are calculated based on the recursive core tensor. An efficiency compensation optimization equation is constructed, and the energy transfer efficiency between components is obtained by solving the efficiency compensation optimization equation.

4. The method according to claim 1, characterized in that, Wavelet transform is used to extract component degradation features from energy transfer efficiency, and a component health index is constructed, including: Wavelet transform is performed on the energy transfer efficiency of the component to obtain the wavelet coefficient sequence, which is then divided into continuous degradation component and instantaneous response component according to the frequency range. Operating condition compensation coefficient is constructed, and the wavelet coefficient sequence is compensated using the operating condition compensation coefficient. The component health index is calculated based on the compensated degradation characteristic sequence.

5. The method according to claim 1, characterized in that, Constructing a probabilistic graphical network based on component health indices includes: The transition correlation degree is calculated based on the differences between adjacent states in the component health index sequence, and a probabilistic graphical network is constructed based on the transition correlation degree.

6. The method according to claim 1, characterized in that, Calculating the component failure state transition probability includes: A bidirectional propagation analysis is performed on the probabilistic graphical network to extract the forward influence strength and backward coupling strength of the state transition, and to determine the state transition weights. The state transition weights and transition correlations are weighted and combined to obtain the component failure state transition probability.

7. The method according to claim 1, characterized in that, Variational inference is used to obtain the fault coupling distribution between components, and the fault early warning threshold is adaptively adjusted based on the fault coupling distribution, including: A variational distribution of component fault coupling is constructed using conditional random fields, and the variational distribution is optimized by expectation maximization iteration to obtain the fault coupling distribution between components. The correlation between component faults is calculated based on the fault coupling distribution, and the fault early warning threshold is adaptively adjusted by combining the fault propagation risk coefficient.

8. The method according to claim 1, characterized in that, Adaptive adjustment of fault early warning thresholds based on fault coupling distribution includes: The amplitude and phase features of the fault coupling distribution are extracted, the rate of change of the amplitude features and the drift of the phase features are calculated, a correction function for the fault warning threshold is constructed, the adjustment coefficient of the warning threshold is calculated based on the correction function, and the fault warning threshold is dynamically updated.

9. The method according to claim 1, characterized in that, Based on the fault coupling distribution, the fault source component and affected components are identified, and fault propagation path and remaining lifetime early warning information are generated, including: Maximum entropy time series analysis is used to calculate the component failure propagation probability; The fault coupling distribution is prioritized based on the fault propagation probability to identify the fault source component, and the affected components whose coupling strength with the fault source component exceeds a preset strength threshold are extracted. Construct a propagation relationship graph between the fault source component and the affected components to obtain the fault propagation path; The component coupling accumulation value is calculated based on the fault propagation path, and the remaining lifespan warning information is generated by combining the component health index.

10. The method according to claim 9, characterized in that, The probability of component failure propagation is calculated using maximum entropy time series analysis, including: Extract the temporal change characteristics of component states, calculate the state transition entropy values ​​between components, and obtain the fault propagation probability based on the state transition entropy values.

Images