Physical-attention fusion prediction method for vehicle-mounted ECU micro-disturbance impact failure

Abstract

This invention discloses a physical-attention fusion prediction method for micro-disturbance impact failure of vehicle ECUs, comprising: 1. Based on the operating conditions of the vehicle ECU, real-time determination of transient impact events is achieved through temperature, humidity, and instantaneous vibration impact intensity; the impact energy index and cumulative damage index are calculated; and a physical quantification model of micro-disturbance cumulative damage and transient impact is established; 2. In the prediction stage, the encoder hidden state is weighted and modulated using physical perception attention weights, combined with a threshold control screening attention mechanism, to realize the context vector calculation of physical-data dual-channel residual fusion; 3. Based on the synchronous output of a multi-task classification head, the failure type probability prediction of each key subsystem of the vehicle ECU at future time steps is predicted, and a dynamic weighted joint loss function is used to optimize model training. This invention can solve the problem of decreased prediction accuracy of traditional data-driven models under complex operating conditions and small sample conditions, providing support for the reliability assurance of vehicle ECUs.

Description

Technical Field

[0001] This invention relates to the field of reliability prediction technology for automotive electronic control units (ECUs), specifically to a physical-attention fusion prediction method for micro-disturbance impact failure of automotive ECUs. Background Technology

[0002] With the development of intelligent connected vehicles, the Electronic Control Unit (ECU), as the core controller of the vehicle's electronic system, is widely used in key modules such as powertrain, chassis control, and vehicle communication. In modern high-end models, the number and functions of ECUs are becoming increasingly complex. Their operating environment is often subjected to various stress coupling effects such as high temperature, humidity, and vibration. The cumulative damage caused by micro-disturbances and transient shocks gradually degrade the performance of their internal semiconductor devices, solder joints, and circuits, ultimately leading to multimodal failures of multiple MCUs, power supplies, communication, and other subsystems. These failures often manifest as complex nonlinear and multimodal coupling phenomena, posing a significant challenge to failure prediction of vehicle ECUs.

[0003] Traditional data-driven failure prediction methods based on LSTM, Transformer, and other similar technologies rely on large-scale failure data for training. However, in real-world automotive scenarios, extreme environments lead to scarce failure samples and high acquisition costs, severely limiting the model's generalization ability. Furthermore, existing models lack a physical quantitative representation of transient impacts and cumulative damage caused by micro-perturbations, making it difficult to accurately capture failure mechanisms driven by key physical processes such as temperature gradient changes and second-order difference of vibration acceleration. In addition, ECU failures exhibit complex characteristics of multiple subsystems and multiple modes, making it difficult for traditional single-task prediction models to simultaneously predict multiple types of failures. Moreover, the imbalance between normal and failed state samples in the data further weakens the identification effect on a few key failure modes. Some studies have attempted to integrate physical models with data-driven techniques, but these often focus on single stress scenarios and are ill-suited to complex nonlinear failure processes under multi-stress coupling conditions. While the introduced attention mechanisms enhance the model's focus on key time and features, the lack of saliency guidance based on physical damage makes the model susceptible to noise, reducing the stability and reliability of predictions.

[0004] Given the current situation, how to build a lightweight failure prediction model that combines physical interpretability and data-driven adaptability in the vehicle environment where edge computing resources are limited has become a technical bottleneck that urgently needs to be overcome in the field of vehicle ECU reliability assurance. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention proposes a physics-attention fusion prediction method for micro-disturbance impact failure of vehicle ECUs. This method combines the physical quantification mechanisms of cumulative damage from micro-disturbances and transient impact damage to effectively extract key features of the environment and electronic control status. Furthermore, it integrates an attention mechanism based on physical saliency to achieve multi-task, multi-label joint prediction. This improves prediction accuracy and reliability under complex stress coupling and meets the real-time online monitoring needs of environments with limited edge computing resources. Ultimately, it enables early warning and accurate diagnosis of vehicle ECU failures, ensuring the safe and stable operation of vehicle electronic systems.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] The present invention provides a physics-attention fusion prediction method for micro-disturbance impact failure of automotive ECUs, characterized by the following steps:

[0008] Step 1: Collect multi-dimensional monitoring data and construct an input sequence of environmental impact parameters. and time series key state matrix and failure type set ;

[0009] Step 2: Determine the transient impact event, calculate the impact energy index, and normalize the cumulative damage index to generate the physical attention weight distribution at future time t. ;

[0010] Step 3: Construct a physics-attention fusion prediction network, including: a temporally dependent encoding layer, a physics-aware threshold-controlled joint attention layer, a context-aware decoding layer, and a multi-task output layer, and then... The process is performed to obtain the predicted probability vector of the vehicle ECU at a future time t. ;

[0011] Step 4: Train and optimize the parameters of the physics-attention fusion prediction network using a dynamically weighted joint loss function to obtain the optimal physics-attention fusion prediction model, which is then used to predict the input sequence. Make predictions and output future time steps. The prediction probability vector of the kth key subsystem ,in, This represents the total time step for prediction.

[0012] The physical-attention fusion prediction method for micro-disturbance impact failure of vehicle ECUs described in this invention is also characterized in that step 1 includes the following steps:

[0013] Step 1.1: Based on the actual operating conditions of the vehicle ECU, collect the ambient temperature at the i-th time step. ,humidity and three-dimensional vibration acceleration The data is used to calculate the instantaneous impact intensity of temperature using equation (1). Instantaneous impact intensity of humidity Instantaneous impact intensity of vibration acceleration Thus, the environmental micro-perturbation impact vector at the i-th time step is constructed. And form an input sequence of environmental impact parameters. N is the total number of time steps;

[0014] (1)

[0015] In equation (1), For a fixed sampling time interval, Let i be the ambient temperature at the (i-1)th time step. Let be the change in ambient temperature between the (i-1)th time step and the ith time step. The humidity at the (i-1)th time step. The change in humidity between the (i-1)th time step and the ith time step. Let be the three-dimensional vibration acceleration at the i-th time step. Let be the three-dimensional vibration acceleration at the (i-1)th time step. Let be the three-dimensional vibration acceleration at the (i-2)th time step; This represents the change in three-dimensional vibration acceleration between the (i-1)th time step and the ith time step;

[0016] Step 1.2: Denote the key state parameter vector of the k-th key subsystem in the vehicle ECU at the i-th time step as... ,in, Let be the s-th critical state parameter of the k-th critical subsystem at the i-th time step; thus, obtain the s-th critical state parameter of the k-th critical subsystem at the i-th time step. The key state parameter vector at each time step is: ; and then the key state matrix of the time series is obtained. ,in, This indicates concatenation, and T indicates transpose;

[0017] Step 1.3: Define the set of failure types for K types of critical subsystems. ,in, Let represent the set of failure types of the k-th critical subsystem, and ,in, This represents the f-th failure type in the k-th critical subsystem; This represents the total number of failure types of the k-th critical subsystem.

[0018] Furthermore, step 2 includes the following steps:

[0019] Step 2.1: Determine whether equation (2) is true. If it is true, it means that a transient impact event has occurred at the i-th time step. Then, use equation (3) to calculate the impact energy index at the i-th time step. Otherwise, it indicates that there is no transient impact event, and let the impact energy exponent at the i-th time step be... :

[0020] (2)

[0021] (3)

[0022] In equations (2) and (3), These represent the historical standard deviations of the temperature gradient, humidity gradient, and vibration acceleration, respectively. These are the sensitivity coefficients for temperature gradient, humidity gradient, and vibration acceleration, respectively. These are the relaxation time constants for temperature, humidity, and vibration / shock, respectively. In order to be in The three-dimensional vibration acceleration at time t, where L is the length of the transient impact backtracking window;

[0023] Step 2.2: Use equation (4) to obtain the cumulative damage at the i-th time step. Thus, the cumulative damage sequence is obtained. ;

[0024] (4)

[0025] In equation (4), For standard temperature parameters, Where k is the activation energy for semiconductor oxidation, k is the Boltzmann constant, and n is the fatigue crack propagation index. The impact-damage coupling coefficient is... This is a humidity reference parameter. is the humidity variation scale factor, and m is the wet stress coefficient;

[0026] Step 2.3, Normalized to a probability distribution, this generates the physical attention weight vector for future time t at the i-th time step. Thus, the physical attention weight distribution at future time t is obtained. .

[0027] Furthermore, step 3 includes the following steps:

[0028] Step 3.1, the time-dependent coding layer... Perform temporal dependency capture processing and output the hidden state vector at the i-th time step step by step. Thus, the encoded hidden state sequence is obtained. ;

[0029] Step 3.2, the physical perception-guided threshold screening combined with the attention layer pair and Perform joint attention calculation to obtain the context vector at future time t. ;

[0030] Step 3.3, the context-aware decoding layer will and After concatenation, the input vector for the predicted time t is obtained. Thus, the hidden state is decoded based on the context of the previous prediction time t-1. and The context-aware decoded hidden state at prediction time t is obtained. ;

[0031] Step 3.4, the multi-task output layer will and After concatenation, the fused feature vector at future time t is obtained. Thus, K independent fully connected sub-classification heads are used to classify the subclasses respectively. The process is performed to output the predicted probability vector of the failure type of the vehicle ECU at a future time t. ,in, This represents the k-th fully connected subclass head pair. The predicted probability vector at future time t:

[0032] Calculate the k-th fully connected subclass head pair using equation (11) Predicted probability vector at future time t ,in, Let represent the predicted probability that the k-th fully connected subclass head is in a normal state at future time t. The predicted probability of the vehicle ECU being in the f-th failure type at future time t is given by the k-th fully connected subclassifier head.

[0033] (11)

[0034] In equation (11), Let be the parameter matrix of the k-th fully connected subclass header, and Softmax represent the activation function.

[0035] Furthermore, step 3.2 includes the following steps:

[0036] Step 3.2.1: Calculate the bond vector after physical driving at the i-th time step using equation (5). Sum value vector :

[0037] (5)

[0038] In equation (5), There are two linear matrices to be learned;

[0039] Step 3.2.2: Calculate the data-driven attention score for future time t at time step i using equation (6). ;

[0040] (6)

[0041] In equation (6), There are three linear matrices to be learned. Let i be the value vector driven by the data at the i-th time step. The query vector driven by data at future time t. Let d be the key vector after data-driven operation at the i-th time step, and d be the dimension of the hidden state. Let represent the context-aware decoding hidden state at the previous prediction time t-1. When t=1, let ; Let be the hidden state vector at the Nth time step;

[0042] Step 3.2.3: Use equation (7) to obtain the threshold-controlled screening attention score modulation term at the i-th time step. :

[0043] (7)

[0044] In equation (8), Let be the physical attention weight vector for future time t at the j-th time step;

[0045] Step 3.2.4: Calculate the hybrid attention score for future time t at time step i using equation (8). :

[0046] (8)

[0047] In equation (9), This is the physical guidance intensity hyperparameter. It is a positive number;

[0048] Step 3.2.5: Use equation (9) to obtain the fault-sensitive attention weights for the i-th time step in the data-driven process to the future time t. :

[0049] (9)

[0050] In equation (9), Indicates selecting only The time step number corresponding to the time;

[0051] Step 3.2.6: Calculate the context vector at future time t using equation (10). :

[0052] (10).

[0053] Furthermore, step 4 includes the following steps:

[0054] Step 4.1: Construct the k-th type of key subsystem using equation (13) in future time steps. Multi-class cross-entropy loss function :

[0055] (12)

[0056] In equation (12), This indicates that the k-th key subsystem is in the future time step. The function indicating the true label of type f, if a failure of type f occurs, lets Otherwise, let , This indicates that the k-th key subsystem is in the future time step. The predicted probability of the f-th type, where f=0 represents the normal state. The serial number indicating the failure type;

[0057] Step 4.2: Construct the future time step using equation (14) joint loss function :

[0058] (13)

[0059] In equation (13), This indicates that the k-th key subsystem is in the future time step. The weighting coefficients are:

[0060] (14)

[0061] In equation (14), It is a hyperparameter used to adjust the weights of physical attention assistance; It is a step into the future. The average physical attention weights are:

[0062] (15)

[0063] In equation (15), This represents the view of future moments at the i-th time step. The physical attention weight vector;

[0064] Step 4.3: Minimize the joint loss function using the backpropagation algorithm and gradient descent optimizer. This allows for the training and parameter optimization of the physics-attention fusion prediction network, resulting in a physics-attention fusion prediction model with optimal parameters.

[0065] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the physical-attention fusion prediction method, and the processor is configured to execute the program stored in the memory.

[0066] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the physical-attention fusion prediction method.

[0067] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0068] 1. Based on key environmental impact parameters such as temperature gradient, humidity gradient and second-order difference of vibration acceleration, this invention establishes a physical quantitative model of cumulative damage from micro-disturbances and transient impact damage, reveals the key failure mechanism under multi-physics coupling, provides physical guidance for subsequent data-driven prediction, and improves the physical interpretation capability of the model.

[0069] 2. This invention introduces a threshold-controlled attention mechanism based on physical perception attention weights, which uses physical saliency as a weight to guide the model to focus on micro-perturbations and transient impact features, enhances the sensitivity to key time nodes and feature dimensions, effectively filters out irrelevant noise, and improves the model's discrimination ability and robustness in complex environments.

[0070] 3. This invention designs a multi-task learning framework for various failure characteristics of the three major subsystems of vehicle ECU: MCU, power supply and communication. It realizes multi-label joint prediction, can simultaneously identify multiple failure modes of multiple subsystems, enhances the model's coverage and recognition accuracy of complex multimodal failure states, and improves generalization performance to adapt to the complex and diverse failure requirements of vehicle ECU. Attached Figure Description

[0071] Figure 1 This is a flowchart of the impact quantification model for micro-disturbances of an on-board ECU based on the present invention;

[0072] Figure 2This is a structural diagram of the physics-attention fusion prediction model of the present invention;

[0073] Figure 3 This is a structural diagram of the joint attention layer for threshold control screening guided by physical perception in the prediction model of this invention. Detailed Implementation

[0074] In this embodiment, refer to Figures 1-3 This paper presents a physics-attention fusion prediction model method for micro-disturbance impact failure of automotive ECUs. The method primarily enhances the detection and prediction capabilities of micro-disturbance impact events by combining a dynamic environmental disturbance judgment mechanism with multi-scale, multi-modal feature fusion. Specifically, the method first dynamically calculates and judges environmental impact parameters and their changes. Combining the energy and cumulative damage indicators of the environmental disturbance, a physics attention adjustment mechanism is established to guide the model to focus on key environmental and structural features. Then, using a physics-based attention mechanism, the model's attention to information at different time scales is dynamically adjusted to achieve accurate capture of micro-disturbance impacts. Finally, multi-task, multi-scale feature fusion improves the model's prediction accuracy and robustness. The specific steps include:

[0075] Step 1: Collect multi-dimensional monitoring data and construct an input sequence of environmental impact parameters. and time series key state matrix and failure type set ,like Figure 1 As shown:

[0076] Step 1.1: Based on the actual operating conditions of the vehicle ECU, collect the ambient temperature at the i-th time step. ,humidity and three-dimensional vibration acceleration The data is used to calculate the instantaneous impact intensity of temperature using equation (1). Instantaneous impact intensity of humidity Instantaneous impact intensity of vibration acceleration The vibration acceleration is expressed in a second-order difference form to construct the environmental micro-perturbation impact vector at the i-th time step. And form an input sequence of environmental impact parameters. N is the total number of time steps;

[0077] Step 1.2: Denote the key state parameter vector of the k-th key subsystem in the vehicle ECU at the i-th time step as... ,in, Let be the s-th critical state parameter of the k-th critical subsystem at the i-th time step; thus, obtain the s-th critical state parameter of the k-th critical subsystem at the i-th time step. The key state parameter vector at each time step is: ; and then the key state matrix of the time series is obtained. ,in, The symbol indicates splicing, and T indicates transpose. In this embodiment, K=3 key subsystems representing the main failures of the vehicle ECU are selected, including the MCU, power supply, and communication subsystem. The key state parameters that contribute the most to the failure diagnosis and prediction in each subsystem are selected to form a set of parameters, specifically including: key features of the MCU subsystem: chip temperature, pin impedance abnormality index, firmware verification error flag; key features of the power supply subsystem: main power supply voltage, current load, voltage transient noise; key features of the communication subsystem: CAN bus error count, retransmission rate, packet loss rate.

[0078] Step 1.3: Define the set of failure types for K types of critical subsystems. ,in, Let represent the set of failure types of the k-th critical subsystem, and ,in, This represents the f-th failure type in the k-th critical subsystem; This represents the total number of failure types in the k-th critical subsystem. In this embodiment, specific common major failure types are defined for the MCU, power supply, and communication subsystems. Specifically, MCU subsystem failure types include: system crash, pin desoldering, firmware anomaly, and overheating; power supply subsystem failure types include: abnormal voltage, excessive ripple, voltage regulator failure, and overcurrent; and communication subsystem failure types include: transmission failure, reception error, packet loss, and bus error. This provides a basis for directly predicting the probability of specific fault categories in subsequent modeling steps, and the model output can cover all specific fault categories.

[0079] Step 2: Identify transient impact events, calculate the impact energy index, and normalize the cumulative damage index to generate a physical attention weight distribution, such as... Figure 1 As shown:

[0080] In this step, the established physical model, based on the gradient changes of input environmental parameters, can realistically simulate the dynamic response of the system under micro-disturbances and transient impacts. Utilizing the gradient information of environmental changes, the occurrence of transient impact events is determined, ensuring timely identification of energy transfer characteristics in transient pulses induced by micro-disturbances. Through exponential calculation of the impact energy, the model can quantify the intensity and duration of the impact, reflecting the energy accumulation process of the system under the influence of micro-disturbances. This step, by establishing a model that conforms to actual physical laws, effectively integrates the dynamic characteristics of micro-disturbances and impacts, improving the sensitivity to microscopic changes in the system.

[0081] Step 2.1: Determine whether equation (2) is true. If it is true, it means that a transient impact event has occurred at the i-th time step. Then, use equation (3) to calculate the impact energy index at the i-th time step. By amplifying the impact energy index, the intensity of micro-disturbance impact is quantified, enhancing the response capability to micro-damage caused by transient impacts; otherwise, it indicates no transient impact event, and the impact energy index at the i-th time step is set to... Considering only steady-state damage, we characterize the asymptotic damage caused by "micro-perturbations";

[0082] (2)

[0083] (3)

[0084] In equations (2) and (3), These represent the historical standard deviations of the temperature gradient, humidity gradient, and vibration acceleration, respectively. These are the sensitivity coefficients for temperature gradient, humidity gradient, and vibration acceleration, respectively. These are the relaxation time constants for temperature, humidity, and vibration / shock, respectively. In order to be in The three-dimensional vibration acceleration at time t, where L is the length of the transient impact backtracking window;

[0085] Step 2.2: Use equation (4) to obtain the cumulative damage at the i-th time step. Thus, the cumulative damage sequence is obtained. This characterizes the total degree of damage suffered by the vehicle ECU system under multi-stress coupling at the current stage:

[0086] (4)

[0087] In equation (4), For standard temperature parameters, Where k is the activation energy for semiconductor oxidation, k is the Boltzmann constant, and n is the fatigue crack propagation index. The impact-damage coupling coefficient was calibrated through a three-dimensional acceleration test of the vehicle ECU. This is a humidity reference parameter. is the humidity variation scale factor, and m is the wet stress coefficient;

[0088] Step 2.3, Normalized to a probability distribution, this generates the physical attention weight vector for future time t at the i-th time step. Thus, the physical attention weight distribution at future time t is obtained. ;

[0089] Step 3: Construct a physics-attention fusion prediction network, including: a temporally dependent encoding layer, a physics-aware threshold-controlled joint attention layer, a context-aware decoding layer, and a multi-task output layer, and then... The process is performed to obtain the predicted failure type probability of the vehicle ECU at a future time t, such as... Figure 2 As shown:

[0090] Step 3.1, the time-dependent coding layer... Perform temporal dependency capture processing and output the hidden state vector at the i-th time step step by step. Thus, the encoded hidden state sequence is obtained. ;

[0091] Step 3.2, the physical perception-guided threshold screening combined with the attention layer pair and Perform joint attention calculation to obtain the context vector at future time t. Compared to traditional purely data-driven methods, the introduction of physical information reduces reliance on large amounts of labeled data, helping to maintain good performance in environments with insufficient data or noise. By utilizing physical parameters, potential anomalies and fault signals in the system can be captured earlier and more accurately, providing early warnings. Figure 3 As shown:

[0092] Step 3.2.1: Calculate the bond vector after physical driving at the i-th time step using equation (5). Sum value vector Strengthen the characteristics of critical moments of injury:

[0093] (5)

[0094] In equation (5), There are two linear matrices to be learned;

[0095] Step 3.2.2: Calculate the value vector after data-driven operation at the i-th time step using equation (6). This allows us to obtain the data-driven attention score for future time t at i time steps. ;

[0096] (6)

[0097] In equation (6), There are three linear matrices to be learned. Let i be the value vector driven by the data at the i-th time step. The query vector driven by data at future time t. Let d be the key vector after data-driven operation at the i-th time step, and d be the dimension of the hidden state. Let represent the context-aware decoding hidden state at the previous prediction time t-1. When t=1, let ; Let be the hidden state vector at the Nth time step;

[0098] Step 3.2.3: Use equation (7) to obtain the threshold-controlled screening attention score modulation term at the i-th time step. This dynamic adjustment mechanism focuses the model on different regions, filtering out noise and irrelevant information. Simultaneously, it enhances the model's adaptability under varying operating conditions, enabling it to automatically adjust the filtering threshold based on the actual environment.

[0099] (7)

[0100] In equation (8), Let be the physical attention weight vector for future time t at the j-th time step;

[0101] Step 3.2.4: Calculate the hybrid attention score for future time t at time step i using equation (8). :

[0102] (8)

[0103] In equation (9), This is the physical guidance intensity hyperparameter, which can be initially set to 0.5 and dynamically adjusted through training. It should be a positive number to prevent ;

[0104] Step 3.2.5: Use equation (9) to obtain the fault-sensitive attention weights for the i-th time step in the data-driven process to the future time t. :

[0105] (9)

[0106] In equation (9), Indicates selecting only The corresponding time step number is used to ensure that the physical channel independently retains the integrity of physical information and avoids being overwhelmed by data features.

[0107] Step 3.2.6: Calculate the context vector at future time t using equation (10). :

[0108] (10)

[0109] Step 3.3, the context-aware decoding layer will and After concatenation, the input vector for the predicted time t is obtained. Thus, the hidden state is decoded based on the context of the previous prediction time t-1. and The context-aware decoded hidden state at prediction time t is obtained. ;

[0110] Step 3.4, the multi-task output layer will and After concatenation, the fused feature vector at future time t is obtained. Thus, K independent fully connected sub-classification heads are used to classify the subclasses respectively. The process involves assigning an independent fully connected subclassification head to each selected key subsystem, outputting a predicted failure type probability vector for the vehicle ECU at future time t. ,in, This represents the k-th fully connected subclass head pair. The output prediction probability vector:

[0111] Step 3.4.1: Calculate the k-th fully connected sub-classification head pair using equation (11). Output probability vector at future time t ,in, This represents the probability that the vehicle ECU will be in a normal state at future time t, as predicted by the k-th fully connected subclassifier. The probability that the on-board ECU is in the f-th failure type at future time t, as predicted by the k-th fully connected subclassifier head;

[0112] (11)

[0113] In equation (11), Let be the parameter matrix of the k-th fully connected subclass header, and Softmax represent the activation function;

[0114] Step 3.4.2: Based on the multi-task, multi-label probabilistic model constructed in Step 3.4.1, generate future time windows. The prediction results are output synchronously for each future time step. and each specific failure type probability of occurrence If the probability of any failure type f is... Exceeding the preset threshold within a consecutive Q=5 time window Then determine the k-th type of subsystem in time. A failure of type f occurs, and the failure type that first meets the condition is taken as the final failure diagnosis result.

[0115] Step 4: Train and optimize the parameters of the physics-attention fusion prediction network using a dynamically weighted joint loss function to obtain the optimal physics-attention fusion prediction model, which is then used for... Make predictions and output future time steps. Failure probability vector of the kth critical subsystem ,in, This represents the total time step for prediction.

[0116] Step 4.1: Construct the k-th type of key subsystem using equation (13) in future time steps. Multi-class cross-entropy loss function :

[0117] (12)

[0118] In equation (12), This indicates that the k-th key subsystem is in the future time step. The function indicating the true label of type f is one-hot encoded. If a failure of type f occurs, let... For the other categories, let , This indicates that the k-th key subsystem is in the future time step. The model predicts the probability for the f-th type, where index f=0 represents the normal state. The serial number indicating the failure type;

[0119] Step 4.2: Construct the future time step using equation (14) joint loss function :

[0120] (13)

[0121] In equation (13), This indicates that the k-th key subsystem is in the future time step. The weighting coefficients are:

[0122] (14)

[0123] In equation (14), This is a hyperparameter used to adjust the weights for physical attention assistance; the default value is 0.5. It is a step into the future. The average physical attention weights are:

[0124] (15)

[0125] In equation (15), This represents the view of future moments at the i-th time step. The physical attention weight vector;

[0126] Step 4.3: In this embodiment, the maximum number of iterations (epoch_number) is set to 200. The backpropagation algorithm and the learning rate are then used. The Adam gradient descent optimizer minimizes the joint loss function. This allows for the training and parameter optimization of the physics-attention fusion prediction network. Training stops when either the number of iterations reaches `epoch_number` or the validation set loss does not decrease for P=10 consecutive epochs. The model parameters with the minimum validation set loss are saved, resulting in the optimal physics-attention fusion prediction model, which is then used for... Make predictions and output the future of the vehicle ECU. The failure type probability set for each time step.

[0127] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0128] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A physics-attention fusion prediction method for micro-disturbance impact failure of vehicle ECUs, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional monitoring data and construct an input sequence of environmental impact parameters. and time series key state matrix and failure type set ; Step 2: Determine the transient impact event, calculate the impact energy index, and normalize the cumulative damage index to generate the physical attention weight distribution at future time t. ; Step 3: Construct a physics-attention fusion prediction network, including: a temporally dependent encoding layer, a physics-aware threshold-controlled joint attention layer, a context-aware decoding layer, and a multi-task output layer, and then... The process is performed to obtain the predicted probability vector of the vehicle ECU at a future time t. ; Step 3.1, the time-dependent coding layer... Perform temporal dependency capture processing and output the hidden state vector at the i-th time step step by step. Thus, the encoded hidden state sequence is obtained. ; Step 3.2, the physical perception-guided threshold screening combined with the attention layer pair and Perform joint attention calculation to obtain the context vector at future time t. ; Step 3.2.1: Calculate the bond vector after physical driving at the i-th time step using equation (5). Sum value vector : (5) In equation (5), There are two linear matrices to be learned; Step 3.2.2: Calculate the data-driven attention score for future time t at time step i using equation (6). ; (6) In equation (6), There are three linear matrices to be learned. Let i be the value vector driven by the data at the i-th time step. The query vector driven by data at future time t. Let d be the key vector after data-driven operation at the i-th time step, and d be the dimension of the hidden state. Let represent the context-aware decoding hidden state at the previous prediction time t-1. When t=1, let ; Let be the hidden state vector at the Nth time step; Step 3.2.3: Use equation (7) to obtain the threshold-controlled screening attention score modulation term at the i-th time step. : (7) In equation (8), Let be the physical attention weight vector for future time t at the j-th time step; Step 3.2.4: Calculate the hybrid attention score for future time t at time step i using equation (8). : (8) In equation (9), This is the physical guidance intensity hyperparameter. It is a positive number; Step 3.2.5: Use equation (9) to obtain the fault-sensitive attention weights for the i-th time step in the data-driven process to the future time t. : (9) In equation (9), Indicates selecting only The time step number corresponding to the time; Step 3.2.6: Calculate the context vector at future time t using equation (10). : (10) Step 3.3, the context-aware decoding layer will and After concatenation, the input vector for the predicted time t is obtained. Thus, the hidden state is decoded based on the context of the previous prediction time t-1. and The context-aware decoded hidden state at prediction time t is obtained. ; Step 3.4, the multi-task output layer will and After concatenation, the fused feature vector at future time t is obtained. Thus, K independent fully connected sub-classification heads are used to classify the subclasses respectively. The process is performed to output the predicted probability vector of the failure type of the vehicle ECU at a future time t. ,in, This represents the k-th fully connected subclass head pair. The predicted probability vector at future time t: Calculate the k-th fully connected subclass head pair using equation (11) Predicted probability vector at future time t ,in, Let represent the predicted probability that the k-th fully connected subclass head is in a normal state at future time t. The predicted probability of the vehicle ECU being in the f-th failure type at future time t is given by the k-th fully connected subclassifier head. (11) In equation (11), Let be the parameter matrix of the k-th fully connected subclass header, and Softmax represent the activation function; Step 4: Train and optimize the parameters of the physics-attention fusion prediction network using a dynamically weighted joint loss function to obtain the optimal physics-attention fusion prediction model, which is then used to predict the input sequence. Make predictions and output future time steps. The prediction probability vector of the kth key subsystem ,in, This represents the total time step for prediction.

2. The physics-attention fusion prediction method for micro-disturbance impact failure of vehicle ECUs according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Based on the actual operating conditions of the vehicle ECU, collect the ambient temperature at the i-th time step. ,humidity and three-dimensional vibration acceleration The data is used to calculate the instantaneous impact intensity of temperature using equation (1). Instantaneous impact intensity of humidity Instantaneous impact intensity of vibration acceleration Thus, the environmental micro-perturbation impact vector at the i-th time step is constructed. And form an input sequence of environmental impact parameters. N is the total number of time steps; (1) In equation (1), For a fixed sampling time interval, Let i be the ambient temperature at the (i-1)th time step. Let be the change in ambient temperature between the (i-1)th time step and the ith time step. The humidity at the (i-1)th time step. The change in humidity between the (i-1)th time step and the ith time step. Let be the three-dimensional vibration acceleration at the i-th time step. Let be the three-dimensional vibration acceleration at the (i-1)th time step. Let be the three-dimensional vibration acceleration at the (i-2)th time step; This represents the change in three-dimensional vibration acceleration between the (i-1)th time step and the ith time step; Step 1.2: Denote the key state parameter vector of the k-th key subsystem in the vehicle ECU at the i-th time step as... ,in, Let be the s-th critical state parameter of the k-th critical subsystem at the i-th time step; thus, obtain the s-th critical state parameter of the k-th critical subsystem at the i-th time step. The key state parameter vector at each time step is: ; and then the key state matrix of the time series is obtained. ,in, This indicates concatenation, and T indicates transpose; Step 1.3: Define the set of failure types for K types of critical subsystems. ,in, Let represent the set of failure types of the k-th critical subsystem, and ,in, This represents the f-th failure type in the k-th critical subsystem; This represents the total number of failure types of the k-th critical subsystem.

3. The physics-attention fusion prediction method for micro-disturbance impact failure of vehicle ECUs according to claim 2, characterized in that, Step 2 includes the following steps: Step 2.1: Determine whether equation (2) is true. If it is true, it means that a transient impact event has occurred at the i-th time step. Then, use equation (3) to calculate the impact energy index at the i-th time step. Otherwise, it indicates that there is no transient impact event, and let the impact energy exponent at the i-th time step be... : (2) (3) In equations (2) and (3), These represent the historical standard deviations of the temperature gradient, humidity gradient, and vibration acceleration, respectively. These are the sensitivity coefficients for temperature gradient, humidity gradient, and vibration acceleration, respectively. These are the relaxation time constants for temperature, humidity, and vibration / shock, respectively. In order to be in The three-dimensional vibration acceleration at time t, where L is the length of the transient impact backtracking window; Step 2.2: Use equation (4) to obtain the cumulative damage at the i-th time step. Thus, the cumulative damage sequence is obtained. ; (4) In equation (4), For standard temperature parameters, Where k is the activation energy for semiconductor oxidation, k is the Boltzmann constant, and n is the fatigue crack propagation index. The impact-damage coupling coefficient is... This is a humidity reference parameter. is the humidity variation scale factor, and m is the wet stress coefficient; Step 2.3, Normalized to a probability distribution, this generates the physical attention weight vector for future time t at the i-th time step. Thus, the physical attention weight distribution at future time t is obtained. .

4. The physics-attention fusion prediction method for micro-disturbance impact failure of vehicle ECUs according to claim 3, characterized in that, Step 4 includes the following steps: Step 4.1: Construct the k-th type of key subsystem using equation (13) in future time steps. Multi-class cross-entropy loss function : (12) In equation (12), This indicates that the k-th key subsystem is in the future time step. The function indicating the true label of type f, if a failure of type f occurs, lets Otherwise, let , This indicates that the k-th key subsystem is in the future time step. The predicted probability of the f-th type, where f=0 represents the normal state. The serial number indicating the failure type; Step 4.2: Construct the future time step using equation (14) joint loss function : (13) In equation (13), This indicates that the k-th key subsystem is in the future time step. The weighting coefficients are: (14) In equation (14), It is a hyperparameter used to adjust the weights of physical attention assistance; It is a step into the future. The average physical attention weights are: (15) In equation (15), This represents the view of future moments at the i-th time step. The physical attention weight vector; Step 4.3: Minimize the joint loss function using the backpropagation algorithm and gradient descent optimizer. This allows for the training and parameter optimization of the physics-attention fusion prediction network, resulting in a physics-attention fusion prediction model with optimal parameters.

5. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the physics-attention fusion prediction method according to any one of claims 1-4, the processor being configured to execute the program stored in the memory.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the physics-attention fusion prediction method according to any one of claims 1-4.

Images