Electronic guiding vehicle steering system fault prediction method based on time series data analysis

Abstract

The invention relates to the technical field of intelligent monitoring, in particular to an electronic guide vehicle steering system fault prediction method based on time sequence data analysis, which comprises the following steps: mapping a steering system into perception, decision and execution virtual computing nodes to construct a digital twin model; extracting inter-node time lag correlation through multivariable Granger causality operation, and constructing a virtual causality matrix; fault parameters such as current, angle and bias are injected into the model for numerical simulation, and a response sequence is generated through time domain iteration; using a dynamic time warping algorithm to calculate the morphological difference between the time sequence and the reference sequence so as to generate a time sequence feature; and constructing a digital twinning dynamic graph, inputting the dynamic graph attention network, and extracting spatio-temporal evolution characteristics to generate a fault prediction result. According to the method, accurate deduction of the full-life-cycle fault evolution process of the steering system is realized through digital twinning.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, specifically to a method for predicting faults in electronic steering systems based on time-series data analysis. Background Technology

[0002] Currently, the R&D model is undergoing a profound transformation from physical verification to model-based systems engineering and fully virtual simulation verification. In this computer-aided design process, building a high-fidelity digital twin model that can accurately reproduce the behavior of the system throughout its entire lifecycle is the core foundation for realizing design parameter iteration and reliability assessment. By running high-precision numerical models in a virtual simulation environment, R&D personnel aim to predict the performance degradation trend of the system in the long-term operation of the future through computational deduction during the design phase.

[0003] In existing virtual simulation modeling technologies for steering systems, there is a common contradiction between insufficient physical mechanism model representation capabilities and difficulty in computational convergence. Traditional simulation methods mainly rely on deterministic differential equations to construct multibody dynamics or finite element models. However, when dealing with the complex nonlinear electromechanical coupling characteristics of steering systems, explicit physical equations are difficult to construct accurate mathematical mapping relationships, resulting in significant truncation errors and cumulative deviations in simulation models when simulating long-term wear and aging behavior, which cannot meet the requirements of high-precision design verification. To compensate for the shortcomings of mechanistic models, current technologies attempt to introduce data-driven algorithms to enhance the prediction accuracy of simulation models. However, building robust time-series prediction models requires massive and evenly distributed training samples. In actual engineering design phases, measured data from physical prototypes are extremely scarce, especially data covering extreme operating conditions where the system is on the verge of failure. This small sample size makes trained neural network models or regression models prone to overfitting when extrapolating in virtual space, resulting in a significant decrease in the computational confidence of the simulation system under extreme boundary conditions. In summary, due to the limitations of mathematical representation in physical modeling and the scarcity of data-driven modeling samples, existing virtual simulation systems generally suffer from low simulation accuracy, poor model generalization ability, and huge computational overhead during the design optimization process when performing full life-cycle fault evolution calculations. Therefore, how to utilize limited time-series data to construct an efficient computing architecture that integrates mechanisms and data characteristics to achieve accurate fault evolution simulation and state prediction in a virtual environment is a key problem that urgently needs to be solved in the field of computer-aided simulation and modeling.

[0004] Therefore, this invention proposes a fault prediction method for electronic steering systems based on time-series data analysis. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting faults in electronic steering systems based on time-series data analysis, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A fault prediction method for electronic steering systems based on time-series data analysis includes: The target system is mapped as virtual computing nodes to construct a digital twin model; historical observation time series data is acquired, multivariate Granger causality operation is performed on the historical observation time series data, the time lag correlation between the virtual computing nodes is extracted, and a digital twin virtual causality matrix is ​​constructed. Configure the operating parameters of the digital twin model, and perform numerical simulation of multi-condition fault parameter injection in the digital twin model; perform time-domain iterative calculation in the digital twin simulation environment to generate a digital twin response sequence; Multi-scale sliding window segmentation is performed on the digital twin response sequence, and the temporal morphological difference between the digital twin response sequence and the fault-free baseline sequence is calculated using a dynamic time warping algorithm to generate digital twin temporal morphological features. A digital twin dynamic graph is constructed using the temporal morphological features of the digital twin and the virtual causal matrix of the digital twin. The graph is then input into a dynamic graph attention network. Based on the virtual causal matrix of the digital twin, spatiotemporal evolution features of the virtual computing nodes with the temporal morphological features are aggregated to generate fault prediction data.

[0007] Preferably, the specific process of constructing the digital twin model includes: According to the signal transmission path of the electronic steering system, the physical entity is deconstructed into multiple functionally independent units and mapped as virtual computing nodes; the virtual computing nodes include perception nodes, decision nodes, and execution nodes. A multi-dimensional attribute vector is configured for each virtual computing node, and the real-time collected current, torque, angle and vibration timing signals are used as the dynamic input features of the corresponding node; according to the physical connection relationship of the steering system and the bus communication protocol, directed logical links between each virtual computing node are defined to construct a virtual topology network; The sensor time-series data under historical fault-free operating conditions are injected into the virtual topology network, and the distribution boundary of each virtual computing node under normal conditions is extracted using a clustering algorithm to construct a digital twin model.

[0008] Preferably, the multivariate Granger causality operation involves introducing historical observations of all virtual computing nodes as predictors, performing statistical regression on the current state of any target virtual computing node to obtain a first prediction bias under full information conditions, and then sequentially removing historical observations of specific virtual computing nodes from the predictors and recalculating the regression on the current state of the target virtual computing node to obtain a second prediction bias after missing information about specific nodes. By comparing the second prediction deviation with the first prediction deviation, the time lag correlation between one virtual computing node and another is obtained, and a digital twin virtual causal matrix is ​​constructed.

[0009] Preferably, the numerical simulation of multi-condition fault parameter injection specifically involves determining the characteristic variables to be simulated, including execution current, feedback angle, and link transmission delay, from the attribute vector of the virtual computing node. Different offsets are configured for the feature variables, and the dynamic input values ​​of the corresponding virtual computing nodes are rewritten to deviate from the healthy distribution boundary during simulation; a digital twin response sequence is generated by performing continuous time-domain computation in the virtual topology network.

[0010] Preferably, the continuous computation in the time domain is performed to determine the time advance increment of the digital twin simulation environment and initialize the digital twin response sequence with the physical state observation value at the current moment; within each simulation step, each virtual computing node updates the output attribute vector of the current step according to the received dynamic input value containing perturbation. Based on the directed logical links defined by the virtual topology network, the updated node output attribute vector is passed to the downstream associated virtual computing nodes as the input for the next simulation step of the downstream nodes; the node state update and link information transmission are repeatedly executed to record the output changes of each virtual computing node over time, forming a digital twin response sequence.

[0011] Preferably, the specific process of calculating the temporal morphological difference between the digital twin response sequence and the fault-free reference sequence includes extracting the digital twin response sequence and the pre-stored fault-free reference sequence, projecting the two sets of sequences onto a unified digital twin virtual feature space using a standardization function, and performing mean centering on the projected sequences to obtain a digital twin standard sequence pair. Within the digital twin simulation step, the Euclidean distance between each time sampling point in the disturbed response sequence and the corresponding sampling point in the search window of the fault-free baseline sequence is calculated to construct a dynamic distance matrix; a dynamic programming algorithm is applied to traverse the dynamic distance matrix to determine the minimum cumulative cost path; and a time series morphological difference degree is generated based on the minimum cumulative cost path.

[0012] Preferably, the digital twin dynamic graph uses the temporal morphological features of the digital twin as the dynamic components of the virtual computing nodes, and establishes directed weighted edges in the virtual topology network using the time lag correlation determined by the virtual causal matrix of the digital twin as the weight. Based on the iteration frequency of the digital twin simulation environment, the virtual topology network is represented in a time-based manner to generate a digital twin map slice sequence consisting of N discrete simulation step sizes; The slice sequence is input into a dynamic graph attention network to extract spatiotemporal evolution feature vectors.

[0013] Preferably, the dynamic graph attention network performs spatial dimension aggregation of the digital twin, combines the topological constraints of the digital twin virtual causal matrix within the digital twin graph slice, and calculates the state influence weights between virtual computing nodes through attention operators; Perform digital twin time dimension evolution, extract node state transition components between adjacent digital twin map slices, and capture the state evolution trajectory of virtual computing nodes within continuous discrete simulation steps; The state influence weights and state evolution trajectories are fused to generate a spatiotemporal evolution feature vector.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention deconstructs the steering system into virtual nodes for perception, decision-making, and execution, and introduces multivariate Granger causal operations to extract the time lag correlation between nodes. This approach breaks through the mathematical limitations of traditional physical differential equations in dealing with complex electromechanical couplings. It uses virtual causal matrices to construct a topological network that reflects real communication protocols and logical links. Compared with explicit physical equations that are prone to truncation errors, it can more accurately characterize the hidden dynamic interaction relationships inside the steering system, significantly improving the digital twin model's accuracy in reproducing complex nonlinear behaviors.

[0015] 2. This invention can autonomously generate a large number of digital twin response sequences by performing multi-condition fault parameter injection and time-domain iterative calculation in the digital twin model. This data augmentation method based on virtual simulation effectively makes up for the problem of scarce measured data of physical prototypes under extreme boundary conditions such as the system being on the verge of failure. By rewriting the feature variable offset and using clustering algorithms to determine the healthy distribution boundary, this scheme can build a high-confidence training dataset in the virtual space without relying on a large number of measured failure samples, effectively avoiding the overfitting phenomenon caused by the lack of samples in traditional data-driven models.

[0016] 3. This invention calculates temporal morphological differences using a dynamic time warping algorithm and extracts spatiotemporal evolution features using a dynamic graph attention network. Compared to traditional threshold monitoring or static prediction models, this method uses graph slice sequences to capture the state evolution trajectory of virtual computing nodes within continuous simulation steps and the cross-node influence weights. This computational architecture, which integrates causal logic constraints and morphological features, can more sensitively capture subtle signs of performance degradation in steering systems, thereby achieving accurate extrapolation and prediction of the entire lifecycle fault evolution process while maintaining low computational overhead. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a fault prediction method for electronic steering systems based on time-series data analysis. Figure 2 This is a data flow diagram of a fault prediction method for electronic steering systems based on time-series data analysis. Figure 3 The flowchart for injecting numerical simulation and time-domain iteration logic into the fault parameters of this invention is shown. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1:

[0020] Please see Figure 1 and Figure 2 This invention provides a fault prediction method for electronic steering systems based on time-series data analysis, the technical solution of which is as follows: A fault prediction method for electronic steering systems based on time-series data analysis includes: The target system is mapped as virtual computing nodes to construct a digital twin model; historical observation time series data is acquired, multivariate Granger causality operation is performed on the historical observation time series data, the time lag correlation between the virtual computing nodes is extracted, and a digital twin virtual causality matrix is ​​constructed. Configure the operating parameters of the digital twin model, and perform numerical simulation of multi-condition fault parameter injection in the digital twin model; perform time-domain iterative calculation in the digital twin simulation environment to generate a digital twin response sequence; Multi-scale sliding window segmentation is performed on the digital twin response sequence, and the temporal morphological difference between the digital twin response sequence and the fault-free baseline sequence is calculated using a dynamic time warping algorithm to generate digital twin temporal morphological features. A digital twin dynamic graph is constructed using the temporal morphological features of the digital twin and the virtual causal matrix of the digital twin. The graph is then input into a dynamic graph attention network. Based on the virtual causal matrix of the digital twin, spatiotemporal evolution features of the virtual computing nodes with the temporal morphological features are aggregated to generate fault prediction data.

[0021] The specific process of constructing the digital twin model includes: According to the signal transmission path of the electronic steering system, the physical entity is deconstructed into multiple functionally independent units and mapped as virtual computing nodes; the virtual computing nodes include perception nodes, decision nodes, and execution nodes. A multi-dimensional attribute vector is configured for each virtual computing node, and the real-time collected current, torque, angle and vibration timing signals are used as the dynamic input features of the corresponding node; according to the physical connection relationship of the steering system and the bus communication protocol, directed logical links between each virtual computing node are defined to construct a virtual topology network; The sensor time-series data under historical fault-free operating conditions are injected into the virtual topology network, and the distribution boundary of each virtual computing node under normal conditions is extracted using a clustering algorithm to construct a digital twin model.

[0022] The mapping process to virtual computing nodes involves decomposing the physical steering mechanism, sensors, controllers, and actuators into multiple functionally independent units, and mapping each of these units to a virtual computing node in the digital space. These virtual computing nodes are divided into three core categories: sensing nodes responsible for collecting raw physical signals, decision nodes responsible for processing data and issuing instructions, and execution nodes responsible for performing specific steering actions. The sensing nodes include a high-precision angle sensor deployed on the steering column, a torque sensor that monitors the driver, and displacement sensors and vibration accelerometers installed at the steering motor and tie rod, and convert analog quantities into digital signals; The decision node corresponds to the electronic steering control unit and its internal core algorithm module, integrating a microprocessor, bus communication interface, and logic operation unit; the decision node is responsible for receiving signals from the sensing node and calculating the target steering current and torque command according to the preset control path tracking algorithm; The control path tracking algorithm is based on the principle of geometric tracking. By adjusting the front wheel steering angle of the vehicle, the rear axle center of the vehicle can reach a predefined aiming point along an arc. The process includes: The decision node obtains the vehicle's current horizontal position coordinates, vertical position coordinates, and heading angle in the global coordinate system in real time through the bus communication interface. Simultaneously, it retrieves a preset reference path point sequence composed of a series of coordinate points from the storage unit. The microprocessor obtains the current vehicle speed and calculates the dynamic pre-aiming distance by multiplying the current vehicle speed by the pre-aiming time (1 to 2 seconds in this embodiment). The microprocessor traverses the reference path point sequence, calculates the Euclidean distance between each path point and the current rear axle center position of the vehicle, and retrieves the point whose distance is closest to the aforementioned dynamic pre-aiming distance, determining it as the current pre-aiming target point. The logic unit performs geometric relationship calculations. Through a coordinate system rotation transformation matrix, the global coordinates of the target point are transformed into a local vehicle coordinate system with the rear axle center as the origin and the vehicle's heading as the vertical axis, thus obtaining the lateral offset distance of the target point relative to the vehicle's centerline. Based on the vehicle's fixed wheelbase and the geometric triangle relationship formed by the target point and the rear axle center, the logic unit derives the radius of curvature of the vehicle's driving arc using the properties of isosceles triangles, which is the square of the target distance divided by twice the lateral offset distance. The desired steering angle required for the front wheels is calculated by dividing the wheelbase length by this radius of curvature and obtaining the arctangent value. The control algorithm based on the geometric tracking principle, through deep coupling of coordinate transformation and the properties of isosceles triangles, significantly improves the motion control performance of the electronically guided vehicle.

[0023] The decision node converts the desired steering angle into a specific execution control quantity. The microprocessor maps the desired steering angle of the front wheels to the target rotation angle of the steering motor according to the mechanical transmission ratio of the steering system. Combined with the preset torque compensation coefficient and current mapping table, it calculates the target current command and target torque required to maintain the desired steering angle. The execution node consists of a steering motor, a power drive circuit, and a mechanical reduction transmission mechanism. Specifically, it includes a current monitoring unit for the brushless DC motor, a motor position resolver, and a force sensing unit for the end gear rack. The execution node directly responds to the instructions of the decision node, converting electrical energy into mechanical energy to drive the vehicle to steer.

[0024] After the node is defined, a multi-dimensional attribute vector is configured for each virtual computing node; sensor data collected in real time from physical entities is accessed, and current timing signals, torque timing signals, angle timing signals, and vibration timing signals are used as dynamic input features for the corresponding virtual computing nodes; for example, the current and torque signals of the steering motor are mapped to the execution node, and the steering wheel angle signal is mapped to the sensing node. Based on the physical connection relationship and bus communication protocol of the steering system, directed logical links between each virtual computing node are defined, and a virtual topology network reflecting the internal logic of the system is constructed. The virtual topology network includes not only the constraints of physical connections, but also the communication logic of data interaction. Sensor time-series data under historical fault-free operating conditions are injected into the constructed virtual topology network. These normal operating condition data are processed using a clustering algorithm (K-means clustering in this embodiment) to extract the feature distribution boundaries of each virtual computing node under fault-free conditions. These feature distribution boundaries define the threshold range for normal system operation. By combining the topology, dynamic features, and normal distribution boundaries, the digital twin model is constructed.

[0025] By deconstructing the steering system into three types of virtual computing nodes—perception, decision-making, and execution—and combining them with a dynamic pre-aiming geometric tracking algorithm, an efficient mapping between physical entities and digital twin models is achieved. This not only improves the smoothness of path tracking and steering accuracy at different vehicle speeds, but also deeply restores the internal logical constraints of the system through a virtual topology network.

[0026] The multivariate Granger causality operation involves introducing historical observations of all virtual computing nodes as predictors, performing statistical regression on the current state of any target virtual computing node to obtain the first prediction bias under full information conditions, and then sequentially removing historical observations of specific virtual computing nodes from the predictors and recalculating the regression on the current state of the target virtual computing node to obtain the second prediction bias after missing information about specific nodes. By comparing the second prediction deviation with the first prediction deviation, the time lag correlation between one virtual computing node and another is obtained, and a digital twin virtual causal matrix is ​​constructed.

[0027] Specifically, the historical operational observation sequence collected and stored by physical entity sensors is extracted, including sensor values ​​of sensing nodes, logical operation status of decision nodes, and key parameters of execution nodes. The historical observation data of all these nodes are used as global predictive factors and introduced into a multivariate statistical regression model. The current state value of any selected target virtual computing node in the system is fitted and predicted through the multivariate regression model, and the prediction results are compared with the actual observation values ​​to calculate the first prediction deviation. The multivariate regression model is implemented by establishing a linear lag regression equation. The value of the target node at the current moment is used as the dependent variable, and the historical observations of all nodes in the system over a continuous period of time are used as independent variables. The influence weights of each historical time step on the current state are obtained by fitting the model using the least squares method. The sum of squared residuals between the model's predicted values ​​and the actual sampled values ​​is calculated and defined as the first prediction bias. Perform item-by-item elimination and comparison operations on predictor factors. While keeping the data of other nodes unchanged, remove the historical observation sequence of a specific virtual computing node from the entire set of predictor factors in turn; use the incomplete set of predictor factors to recalculate the current state of the target virtual computing node, thereby obtaining the second prediction bias under the interference of missing specific node information. This elimination and regression process is executed cyclically among all virtual computing nodes in the system to traverse all possible node association combinations. The second prediction bias generated by each pair of nodes is quantitatively compared with the first prediction bias. If it is found that the second prediction bias significantly increases compared to the first prediction bias after removing the historical data of a specific node, it is logically determined that there is a statistically significant guiding relationship between the specific node and the target node. The judgment of significant increase is based on constructing a test statistic that follows a distribution. The value is determined by the ratio of the difference between the first prediction bias and the second prediction bias. The probability value of this statistic under the current sample size is calculated by analyzing the significance level of the deviation change. When the probability value is less than the preset significance threshold (set to 0.05 in this embodiment), it is determined to be a significant increase. By analyzing the time order parameter in the regression model, the time lag correlation value is extracted. The microprocessor retrieves the coefficients of each delay term corresponding to a specific node in the regression equation, identifies the delay with the largest absolute value of the weight coefficient, and multiplies the delay order by the fixed signal sampling period of the system to extract the time lag correlation value. The time lag correlation values ​​and causal influence strengths calculated between all pairs of virtual computing nodes are structured and mapped into a preset matrix model to construct a digital twin virtual causal matrix. In the digital twin virtual causal matrix, the intersection value of each row and each column represents the depth and time lag of the logical influence between the corresponding nodes.

[0028] By constructing a virtual causal matrix through multivariate Granger causality operations, the implicit logical guiding relationships between computing nodes are quantified, which can effectively eliminate complex feedback interference within the system. At the same time, by using the mapping between delay order and sampling period, the abstract correlation is transformed into specific physical signal transmission delay, thus providing a precise time-series compensation basis for the digital twin model.

[0029] The numerical simulation of multi-condition fault parameter injection specifically involves determining the characteristic variables to be simulated, including execution current, feedback angle, and link transmission delay, from the attribute vector of the virtual computing node. Different offsets are configured for the feature variables, and the dynamic input values ​​of the corresponding virtual computing nodes are rewritten to deviate from the healthy distribution boundary during simulation; a digital twin response sequence is generated by performing time-domain iterative calculation in the virtual topology network.

[0030] Based on the physical characteristics and common failure modes of the steering system, the core characteristic variables to be simulated are determined. These variables specifically include the execution current of the actuator, the feedback angle of the sensing node, and the link transmission delay in the virtual topology network; the execution current reflects the load and power state of the steering motor, the feedback angle represents the actual position output of the steering mechanical structure, and the link transmission delay simulates the lag of bus communication under high load or electromagnetic interference. For each type of characteristic variable, a step-wise offset parameter is set. For the execution current variable, the average current value in historical fault-free data is first calculated, and the offset is set based on the average current value. Specifically, the offset interval is divided into 10 levels, and the step amplitude is set to 5%. When simulating the power steering attenuation condition, the real-time current input value is multiplied by a gain coefficient of 1.05, 1.10 up to 1.50 using a multiplication operator. When the gain coefficient is set in the range of 1.05 to 1.20, it is used to simulate slight power steering resistance; when the coefficient increases to the range of 1.30 to 1.50, it is used to characterize the overload state caused by a partial short circuit in the motor winding by simulating an abnormal increase in current. For the feedback angle variable, fault simulation is achieved by superimposing two types of quantization operators. First, fixed angle offset quantization: an offset value is set in 0.5-degree increments between 0.5 and 5.0 degrees, and directly added to the original angle sequence to simulate sensor installation displacement or mechanical transmission clearance. Second, random pulse offset quantization: instantaneous pulse noise with an amplitude of 10% of the original signal peak-to-peak value is injected along the time axis at a frequency randomly triggered once every 100 sampling points to simulate signal drift of the angle sensor when subjected to electromagnetic interference. To address link transmission latency in the virtual topology network, a time delay operator is set based on the 5-millisecond standard period specified by the bus communication protocol. The quantization process is divided into three gradients: the first gradient is 10 to 50 milliseconds, in 10-millisecond increments, simulating data queuing latency in the controller area network under mild congestion; the second gradient is 60 to 120 milliseconds, simulating retransmission latency caused by damage to the physical bus link; and the third gradient is 130 to 200 milliseconds, simulating severe lag when communication nodes are on the verge of offline. By adding a corresponding time counter offset to the logical link function of the virtual topology, the synchronization time of data between computing nodes is changed. Since the increase in communication latency will cause changes in the arrival order of data packets, a first-in-first-out virtual buffer queue is established. When simulating retransmission or severe lag, the logical link function should be able to handle the discarding or overwriting logic of old data to prevent the digital twin model from having operational logic conflicts due to receiving data with future timestamps. The quantized offset is input into the digital twin model, which automatically compares the real-time data value after the offset with the Euclidean distance of the normal state distribution boundary. When the current gain exceeds 15% or the angle offset exceeds 2 degrees, it is determined that the current feature value has crossed the healthy distribution boundary. Before performing the distance comparison, the fluctuation range of each variable under normal operating conditions is extracted in advance to establish a standardized feature reference vector. When calculating the Euclidean distance, the current, angle, and time delay need to be normalized to eliminate the influence of different unit dimensions on the distance calculation, thereby ensuring the accuracy of the judgment threshold. At this time, the digital twin model will automatically trigger continuous time-domain calculation, record the dynamic response of each node in the deviation state, and generate a digital twin response sequence containing fault evolution characteristics. By establishing a multi-dimensional step quantization injection model with current, angle, and time delay, and combining a virtual cache queue with a physical dynamics coupling mechanism, the problem of inaccurate judgment caused by physical logic disconnection, timing disorder, and inconsistent dimensions commonly encountered in digital twin fault simulation is effectively solved. It can not only accurately restore the complete evolution chain from performance degradation to severe damage, but also ensure the sensitivity of the judgment benchmark through a normalized distance algorithm.

[0031] See Figure 3 The execution of continuous time-domain computation involves determining the time progression increment of the digital twin simulation environment and initializing the digital twin response sequence to the physical state observation value at the current moment. In this embodiment, the digital twin simulation step size is set to 5ms, the total simulation duration is set to 2000ms, and a total of 400 time-series sampling points are generated. Within each simulation step size, each virtual computing node updates the output attribute vector of the current step size based on the received dynamic input value containing perturbations. Based on the directed logical links defined by the virtual topology network, the updated node output attribute vector is passed to the downstream associated virtual computing nodes as the input for the next simulation step of the downstream nodes; the node state update and link information transmission are repeatedly executed to record the output changes of each virtual computing node over time, forming a digital twin response sequence.

[0032] The time advance increment of the digital twin simulation environment is determined, and the response sequence is initialized with the physical state observations at the current moment. To ensure the logical rigor of the communication simulation, a time counter offset is added to the logic link function, and a first-in-first-out virtual buffer queue is established. When the simulation is retransmitted or severely delayed, the logic link function handles the logic of discarding or overwriting old data to prevent data packet out-of-order due to time delay or logical conflicts caused by receiving future timestamp data. Within each simulation step, each virtual computing node receives dynamic input values ​​containing the aforementioned disturbances and updates the output attribute vector of the current step. Before performing distance comparison, the fluctuation range of each variable under normal operating conditions is extracted in advance, and a standardized feature reference vector is established. When calculating the Euclidean distance between the real-time data after the injected offset and the healthy boundary, the current, angle, and time delay are normalized to eliminate the influence of dimensions. When the current gain exceeds 15% or the angle offset exceeds 2 degrees, the feature value is determined to be out of bounds and the continuous calculation in the time domain is triggered. The updated output vector of each node is passed to the downstream node according to the directed logical link defined by the virtual topology, serving as the input for its next simulation step. By repeatedly executing node state updates and link information transmission, the dynamic response of each node over time is fully recorded, ultimately generating a digital twin response sequence containing fault evolution characteristics. By introducing a virtual buffer queue and time counter offset in the time-domain iteration, the problem of communication disorder and timing conflict caused by high latency in digital twin simulation is solved from the underlying logic, which significantly improves the stability of simulation operation. At the same time, the triggering mechanism built by using standardized feature benchmarks and normalized Euclidean distance eliminates the dimensional differences between different physical quantities.

[0033] The specific process of calculating the temporal morphological difference between the digital twin response sequence and the fault-free reference sequence includes extracting the digital twin response sequence and the pre-stored fault-free reference sequence, projecting the two sets of sequences onto a unified digital twin virtual feature space using a normalization function, and performing mean centering on the projected sequences to obtain a digital twin standard sequence pair. Within the digital twin simulation step, the Euclidean distance between each time sampling point in the disturbed response sequence and the corresponding sampling point in the search window of the fault-free baseline sequence is calculated to construct a dynamic distance matrix; a dynamic programming algorithm is applied to traverse the dynamic distance matrix to determine the minimum cumulative cost path; and a time series morphological difference degree is generated based on the minimum cumulative cost path.

[0034] The digital twin response sequence and a pre-stored fault-free baseline sequence are extracted. To eliminate the influence of differences in physical dimensions on feature comparison, a normalization function is used to project the two sets of sequences onto a unified digital twin virtual feature space. Within the projection space, mean centering is performed on the two sets of sequences, i.e., the overall mean of the sequence is subtracted from the value of each sampling point in the sequence to eliminate DC component bias, thereby obtaining a standard digital twin sequence pair. For standard sequence pairs, multi-scale sliding window segmentation is performed to divide long sequences into segments with local features. Within a set digital twin simulation step size, dynamic time warping logic is used to calculate the morphological differences between the disturbed response sequence and the fault-free baseline sequence. Specifically, for each time sampling point in the response sequence, a search window is set near the corresponding time axis of the fault-free baseline sequence. The Euclidean distance between the sampling point and all points within the search window is calculated, and the calculation results are filled into the corresponding coordinate positions to construct a dynamic distance matrix. The multi-scale sliding window segmentation includes three sets of sliding windows: short-term (20 steps), medium-term (40 steps), and long-term (80 steps), which respectively capture instantaneous, evolutionary, and degenerative features; Based on this, a dynamic programming algorithm is applied to perform a global optimization traversal of the dynamic distance matrix. The dynamic programming algorithm starts from the starting coordinates of the dynamic distance matrix and iteratively calculates the local minimum cost value of each cell. This cost value is obtained by summing the Euclidean distance of the current coordinate and the minimum cumulative cost value of its adjacent predecessor cells (including the cells to the left, below, and to the lower left). Through this layer-by-layer recursive approach, the algorithm can automatically find a minimum cumulative cost path extending from the starting point to the ending point in the matrix. The minimum cumulative cost path represents the best matching relationship between the two sets of sequences under time-domain stretching or compression. Based on the value of the minimum cumulative cost path, a temporal morphological difference degree representing the degree of deviation of the steering system's operating state is generated. Specifically, the sum of the values ​​of all cells on the minimum cumulative cost path is extracted and normalized and smoothed according to the total length of the path. The resulting value is the temporal morphological difference degree. The magnitude of the temporal morphological difference degree directly reflects the degree of waveform distortion of the disturbed response sequence relative to the fault-free reference sequence.

[0035] By combining standardized projection with dynamic time warping algorithm, the problem of diagnostic failure caused by time axis offset or phase lag in steering system signals under different operating conditions is solved. Through the calculation of minimum cumulative cost path, the nonlinear morphological distortion of digital twin response sequence relative to reference sequence can be accurately captured. Even when the signal frequency fluctuates slightly or there is random pulse interference, the morphological features reflecting the fault evolution trend can still be accurately extracted.

[0036] The digital twin dynamic graph uses the temporal morphological features of the digital twin as the dynamic components of the virtual computing nodes, and establishes directed weighted edges in the virtual topology network using the time lag correlation determined by the virtual causal matrix of the digital twin as the weight. Based on the iteration frequency of the digital twin simulation environment, the virtual topology network is represented in a time-based manner to generate a digital twin map slice sequence consisting of N discrete simulation step sizes; The slice sequence is input into a dynamic graph attention network to extract spatiotemporal evolution feature vectors.

[0037] The calculated temporal morphological difference is used as a dynamic attribute component of the virtual computing nodes to characterize the deviation of each node's state in real time. Simultaneously, using the digital twin virtual causal matrix previously extracted from multivariate Granger causality operations, the temporal lag correlation values ​​between nodes are set as the weights of corresponding edges in the graph. Based on the original virtual topology network, directed weighted edges are established according to causal direction, thereby constructing a digital twin graph structure that includes both physical connection logic and statistical causal strength. The continuous simulation timeline is divided into fixed step sizes, and the dynamic features of nodes and edge weight states at each discrete simulation step size are extracted to generate a sequence of digital twin graph slices consisting of 32 consecutive step sizes. Each slice represents the spatial distribution characteristics of the steering system at a specific instant, while the temporal arrangement between slices records the propagation and evolution path of the fault's impact in the topology network. The generated graph slice sequence is input into the dynamic graph attention network. During each layer of computation, the attention mechanism is used to filter the neighbor node information based on the weights defined by the causal matrix, and the contribution between the current virtual computing node and its associated nodes is calculated. Feature aggregation is performed in the spatial dimension, and the perturbation received by the downstream execution node is nonlinearly mapped to the historical state of its upstream decision or perception node to generate a spatial feature vector. By using a cyclic structure to process slice sequences in the time dimension, spatiotemporal evolution feature vectors are extracted. This integrates the spatial coupling of the physical structure of the steering system with the temporal delay of fault evolution, and can capture the evolution trend of steering assist fade or sensor drift from latent to explosive in the network. The spatiotemporal evolution feature vectors are mapped to a preset fault classification space, and the probability of fault occurrence and the estimated remaining life are output for the next 500 to 2000 millisecond steps, thus generating the final fault prediction data.

[0038] By deeply integrating the Granger causality matrix into the dynamic graph attention network, the unity of physical topological logic and statistical causal strength is achieved, accurately characterizing the propagation lag and evolution path of steering failure. By adopting time-segmented slicing sequences and spatiotemporal aggregation mechanisms, the perception depth of potential risks such as boost fade and signal drift is significantly enhanced.

[0039] The dynamic graph attention network performs spatial dimension aggregation of the digital twin, combines the topological constraints of the digital twin virtual causal matrix within the digital twin graph slice, and calculates the state influence weights between virtual computing nodes through attention operators. Perform digital twin time dimension evolution, extract node state transition components between adjacent digital twin map slices, and capture the state evolution trajectory of virtual computing nodes within continuous discrete simulation steps; The state influence weights and state evolution trajectories are fused to generate a spatiotemporal evolution feature vector.

[0040] Perform spatial dimension aggregation of digital twins. Within each discrete digital twin graph slice, define the interaction boundary between nodes by combining the topological constraints determined by the virtual causal matrix of the digital twin. Use the attention operator to weight the node features in the local neighborhood. Specifically, determine the influence weight of each neighbor node on the state of the central node by calculating the similarity between the current virtual computing node and its causally related neighbor nodes. For the generated digital twin map slice sequence, the state transition components of the corresponding virtual computing nodes between adjacent slices are extracted. This process records the fluctuation trajectory of feature values ​​on the time axis by comparing the rate of change of node attribute vectors within continuous simulation steps, thereby effectively identifying state evolution trajectories such as slow rise of assist current or periodic drift of feedback angle; The spatial dimension state influence weights obtained from the above calculations are fused with the temporal dimension state evolution trajectory. During the fusion process, the spatial weights are used as a mask to calibrate the temporal evolution trajectory, enabling the model to identify which temporal fluctuations are caused by real physical causal transmission and which are isolated random noise. Through this nonlinear superposition of multidimensional features, a spatiotemporal evolution feature vector that can highly summarize the complex evolutionary logic of the steering system is generated. The spatiotemporal evolution feature vector is then input to the output layer and mapped to specific fault prediction results.

[0041] By combining the attention operator with the causal matrix topological constraints, a precise modeling of the spatial coupling relationship of the steering system is achieved, which can effectively identify the propagation priority of faults among execution, decision-making, and perception nodes. Simultaneously, by extracting state transition components between adjacent slices, the model's ability to capture subtle trends such as steering assist anomalies or sensor drift is enhanced.

[0042] By introducing multivariate Granger causality operations to construct a virtual causal matrix, the deep time-lag correlations between virtual computing nodes are revealed from a statistical perspective, overcoming the limitation of traditional physical models in describing nonlinear coupling relationships. Combining digital twin fault injection simulation with dynamic time warping algorithms, precise quantification of signal nonlinear distortion under complex operating conditions is achieved, greatly enriching the expressive dimensions of fault features. Finally, by utilizing a dynamic graph attention network to perform spatiotemporal evolution feature aggregation under the constraints of the causal matrix, the ability to capture potential fault evolution trends in steering systems is significantly improved, ensuring prediction lead time and accuracy under multivariate cross-interference environments, providing scientific data support for the safe operation and maintenance of electronically guided vehicles.

[0043] Example 2:

[0044] This implementation applies a fault prediction method for electronic steering systems based on time-series data analysis to the fault prediction process of power assist attenuation caused by partial short circuits in the windings of electronic steering systems.

[0045] On a certain urban rail transit line, an electronic steering system is deconstructed into sensing nodes (displacement and torque sensors), decision nodes (steering ECU), and execution nodes (brushless DC motor). A multi-dimensional attribute vector is configured for each node, and 10A reference current and steering angle signals are collected in real time. The distribution boundary under normal conditions is extracted using the K-means clustering algorithm, and multivariate Granger causality operation is performed to identify a 20ms time lag correlation between the sensing node and the execution node. Based on this, an initial virtual causal matrix is ​​constructed. During the simulation, the execution current is determined as the variable to be simulated in the attribute vector of the execution node; a step offset is set to gradually increase the current gain coefficient from 1.05 to 1.45 to simulate the overload state caused by a partial short circuit in the motor winding; within a simulation step of 5ms, each node updates the output vector according to the input value containing the disturbance, and uses a first-in-first-out virtual buffer queue to process data synchronization to prevent logical conflicts caused by simulation delay, and finally generates a digital twin response sequence containing fault evolution characteristics; The above response sequence and the pre-stored fault-free baseline sequence are extracted, projected onto a unified virtual feature space and mean-centered. The dynamic time warping algorithm is applied to calculate the Euclidean distance between the sampling points of the two sequences within the search window of the simulation step size, and a dynamic distance matrix is ​​constructed. The matrix is ​​traversed by a dynamic programming algorithm to determine the path with the minimum cumulative cost and perform normalization smoothing. The generated time-series morphological difference degree intuitively reflects the degree of waveform distortion caused by the motor short circuit. The calculated morphological difference degree is used as the dynamic component of the node, and combined with the hysteresis weight determined by the causal matrix, a digital twin dynamic graph is constructed. The simulation time axis is divided into a sequence of graph slices consisting of 32 consecutive step sizes and input into the dynamic graph attention network. The network calculates the influence weight between nodes in the spatial dimension through the attention operator, and extracts the state transition components between adjacent slices in the temporal dimension to capture the trajectory of the slow rise of the power steering current. Finally, based on the spatiotemporal evolution feature vector, the fault prediction data is output, showing that the probability of power steering failure within the next 1500ms is 88%, and a warning of the remaining lifespan is given.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting faults in electronic steering systems based on time-series data analysis, characterized in that, include: The target system is mapped as virtual computing nodes to construct a digital twin model; historical observation time series data is acquired, multivariate Granger causality operation is performed on the historical observation time series data, the time lag correlation between the virtual computing nodes is extracted, and a digital twin virtual causality matrix is ​​constructed. Configure the operating parameters of the digital twin model, and perform numerical simulation of multi-condition fault parameter injection in the digital twin model; Perform time-domain iterative calculations in a digital twin simulation environment to generate a digital twin response sequence; Multi-scale sliding window segmentation is performed on the digital twin response sequence, and the temporal morphological difference between the digital twin response sequence and the fault-free baseline sequence is calculated using a dynamic time warping algorithm to generate digital twin temporal morphological features. A digital twin dynamic graph is constructed using the digital twin temporal morphological features and the digital twin virtual causal matrix. This graph is then input into a dynamic graph attention network. Based on the digital twin virtual causal matrix, the spatiotemporal evolution features of the virtual computing nodes representing the digital twin temporal morphological features are aggregated to generate fault prediction results.

2. The method for predicting faults in an electronic steering system based on time-series data analysis according to claim 1, characterized in that, The specific process of constructing the digital twin model includes: According to the signal transmission path of the electronic steering system, the physical entity is deconstructed into multiple functionally independent units and mapped as virtual computing nodes; the virtual computing nodes include perception nodes, decision nodes, and execution nodes. A multi-dimensional attribute vector is configured for each virtual computing node, and the real-time collected current, torque, angle and vibration timing signals are used as the dynamic input features of the corresponding node; according to the physical connection relationship of the steering system and the bus communication protocol, directed logical links between each virtual computing node are defined to construct a virtual topology network; The sensor time-series data under historical fault-free operating conditions are injected into the virtual topology network, and the distribution boundary of each virtual computing node under normal conditions is extracted using a clustering algorithm to construct a digital twin model.

3. The method for predicting faults in an electronic steering system based on time-series data analysis according to claim 1, characterized in that, The execution of multivariate Granger causality operation includes: using the historical observation time series data of all virtual computing nodes as predictors, performing statistical regression on the current state of any target virtual computing node to obtain the first prediction bias under full information conditions; sequentially removing the historical observation values ​​of virtual computing nodes from the predictors, and re-performing regression calculation on the current state of the target virtual computing node to obtain the second prediction bias after missing node information. By comparing the second prediction deviation with the first prediction deviation, the time lag correlation between one virtual computing node and another is obtained, and a digital twin virtual causal matrix is ​​constructed.

4. The method for predicting faults in an electronic steering system based on time-series data analysis according to claim 1, characterized in that, The specific process of numerical simulation for injecting fault parameters under multiple operating conditions includes determining the characteristic variables to be simulated, including execution current, feedback angle, and link transmission delay, from the attribute vector of the virtual computing node. Different offsets are configured for the feature variables, and the dynamic input values ​​of the corresponding virtual computing nodes are rewritten to deviate from the healthy distribution boundary during simulation; a digital twin response sequence is generated by performing time-domain iterative calculation in the virtual topology network.

5. The method for predicting faults in an electronic steering system based on time-series data analysis according to claim 4, characterized in that, The execution of the continuous time-domain computation process includes determining the time advance increment of the digital twin simulation environment and initializing the digital twin response sequence with the physical state observations at the current moment; Within each simulation step, each virtual computing node updates its output attribute vector for the current step based on the received dynamic input value containing perturbations. Based on the directed logical links defined in the virtual topology network, the updated node output attribute vector is passed to the downstream associated virtual computing node as the input for the next simulation step of the downstream node; Repeatedly execute node state updates and link information transmission, record the output changes of each virtual computing node over time, and form a digital twin response sequence.

6. The method for predicting faults in an electronic steering system based on time-series data analysis according to claim 1, characterized in that, The specific process of calculating the temporal morphological difference between the digital twin response sequence and the fault-free reference sequence includes extracting the digital twin response sequence and the pre-stored fault-free reference sequence, projecting the two sets of sequences onto a unified digital twin virtual feature space using a normalization function, and performing mean centering on the projected sequences to obtain a digital twin standard sequence pair. Within the digital twin simulation step, the Euclidean distance between each time sampling point in the disturbed response sequence and the corresponding sampling point in the search window of the fault-free baseline sequence is calculated, and a dynamic distance matrix is ​​constructed. The dynamic programming algorithm is applied to traverse the dynamic distance matrix to determine the path with minimum cumulative cost; the temporal morphological difference is generated based on the path with minimum cumulative cost.

7. The method for predicting faults in an electronic steering system based on time-series data analysis according to claim 1, characterized in that, The digital twin dynamic graph uses the temporal morphological features of the digital twin as the dynamic components of the virtual computing nodes, and establishes directed weighted edges in the virtual topology network using the time lag correlation determined by the virtual causal matrix of the digital twin as the weight. Based on the iteration frequency of the digital twin simulation environment, the virtual topology network is represented in a time-based manner to generate a digital twin map slice sequence consisting of N discrete simulation step sizes; The slice sequence is input into a dynamic graph attention network to extract spatiotemporal evolution feature vectors.

8. The method for predicting faults in an electronic steering system based on time-series data analysis according to claim 7, characterized in that, The dynamic graph attention network performs spatial dimension aggregation of the digital twin, combines the topological constraints of the digital twin virtual causal matrix within the digital twin graph slice, and calculates the state influence weights between virtual computing nodes through attention operators. Perform digital twin time dimension evolution, extract node state transition components between adjacent digital twin map slices, and capture the state evolution trajectory of virtual computing nodes within continuous discrete simulation steps; The state influence weights and state evolution trajectories are fused to generate a spatiotemporal evolution feature vector.

Images