Vehicle fault prediction method based on machine learning

Abstract

The invention relates to the technical field of vehicle intelligent operation and maintenance, and discloses a vehicle fault prediction method based on machine learning, and the method comprises the following steps: S1, data collection and preprocessing: obtaining sensor data, environment data and historical fault record data of a vehicle operation state, and carrying out the cleaning, normalization and time sequence slicing of the data; s2, introducing a physical constraint, constructing a physical model based on vehicle dynamics, generating a virtual sample used for expanding training data, and embedding the physical constraint into a loss function in a model training process; and S3, constructing a deep learning model based on the time sequence. According to the method, the fault prediction precision and the physical consistency of the model are improved through virtual sample generation and multi-objective optimization; an online monitoring and dynamic adjustment mechanism is introduced to enhance the long-term adaptability of the model; collaborative optimization of real-time performance and low computing resource consumption is realized, and the method is suitable for vehicle fault prediction in a complex scene.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle intelligent operation and maintenance, and specifically to a vehicle fault prediction method based on machine learning. Background Art

[0002] With the continuous development of intelligent and networked technologies, vehicle fault prediction has become an important technical means to ensure vehicle safety and reduce operation and maintenance costs. At present, the widely used technologies are mainly based on machine learning, which builds models to predict the possibility of failure by analyzing the vehicle's sensor data and historical fault records. This type of method uses large-scale data to learn the characteristic patterns of vehicle operating conditions, thereby discovering potential fault hazards in advance. With the support of vehicle networking and sensor technology, these technologies have achieved certain results in practical applications and have gradually become an important part of the vehicle intelligent operation and maintenance system.

[0003] However, there are still significant problems with the fault prediction schemes based on machine learning in the existing technology. First, the problem of data scarcity is prominent, and insufficient data on rare fault categories makes it difficult for the model to effectively identify and predict. Second, existing methods often rely entirely on data-driven, ignoring the physical laws of vehicle dynamics and lacking physical interpretability, resulting in reduced credibility and safety of the model. Third, the existing models are not able to adapt to the diversity in the actual operating environment. For example, under different road conditions, climate environments, and load conditions, the model performance is prone to significant decline. Finally, fixed prediction models lack dynamic adjustment capabilities and are difficult to adapt to the impact of sensor aging or environmental changes during long-term operation. These problems severely limit the application effect of existing technologies in complex vehicle operation scenarios. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a vehicle fault prediction method based on machine learning, which solves the problems of data scarcity, insufficient physical interpretation, poor scenario adaptability and weak model dynamic adjustment capability in the prior art.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: A vehicle fault prediction method based on machine learning, comprising the following steps:

[0006] S1. Data collection and preprocessing: obtaining sensor data, environmental data, and historical fault record data of vehicle operation status, and cleaning, normalizing, and time series slicing the above data;

[0007] S2, introduce physical constraints, build a physical model based on vehicle dynamics, generate virtual samples for expanding training data, and embed physical constraints into the loss function during model training;

[0008] S3. Build a deep learning model based on time series, extract the time series features of sensor data, and predict vehicle fault categories;

[0009] S4. Train the model through a multi-objective optimization loss function to jointly optimize fault prediction accuracy, physical consistency, and model robustness;

[0010] S5. Verify model performance and evaluate the model’s fault prediction capability and generalization performance through real-world scenario testing.

[0011] Preferably, the sensor data includes speed, acceleration, engine speed, torque, fuel consumption rate and coolant temperature, the environmental data includes air temperature, humidity, road slope and vehicle load, and the fault record data includes fault type and fault occurrence time.

[0012] Preferably, the physical constraint is constructed through the following vehicle dynamics relationship: the engine output force is equal to the sum of air resistance, ground friction and vehicle inertia, the air resistance is determined by the square of the vehicle speed and the air density, the drag coefficient and the frontal area, and the inertia is determined by the vehicle acceleration and mass.

[0013] Preferably, after the generated virtual samples calculate the unobserved physical variables based on the physical model, they are combined with the real collected data to form extended samples, and the data distribution of the virtual samples is adjusted by a distribution alignment method.

[0014] Preferably, the deep learning model adopts a combination of a long short-term memory network and a Transformer structure, uses LSTM to extract short-term dependency features of the time series, and uses the Transformer's self-attention mechanism to extract long-term dependency features and global features.

[0015] Preferably, the loss function of the multi-objective optimization includes fault prediction loss, physical consistency loss and robustness loss, wherein the fault prediction loss is calculated by the error between the actual fault label and the predicted value, the physical consistency loss is calculated by the degree of deviation between the predicted value and the physical model, and the robustness loss is calculated by the stability of the model output change after simulating the input data disturbance.

[0016] Preferably, in S1, in order to enhance the robustness of the model, an adversarial training strategy is introduced in the training process, and the stability and accuracy of the model in complex and noisy environments are improved by adding disturbances to the input data to simulate sensor anomalies.

[0017] Preferably, the optimization of the loss function uses a Bayesian optimization method to dynamically adjust the weight coefficients to balance the fault prediction accuracy, physical consistency and robustness objectives.

[0018] Preferably, the verification of the model includes performance indicator testing and scenario testing, wherein the performance indicators include the accuracy, recall rate and F1 value of fault prediction, and the scenario testing includes generalization performance evaluation under different loads, road conditions and climate environments.

[0019] Preferably, the scenario test also includes testing under extreme conditions, targeting scenarios where multiple sensor failures occur simultaneously, to verify the reliability and predictive ability of the model in complex scenarios.

[0020] The present invention provides a vehicle fault prediction method based on machine learning. It has the following beneficial effects:

[0021] 1. The present invention adopts a technical solution that combines physical constraints with deep learning, generates virtual samples by introducing a vehicle dynamics model, and embeds physical consistency constraints in model training, thereby achieving efficient prediction of rare fault categories. Compared with the prediction method in the prior art that completely relies on data-driven, it solves the problem of poor training effect caused by data scarcity and significantly improves the generalization ability and prediction accuracy of the model.

[0022] 2. The present invention uses a multi-objective optimization loss function to take fault prediction accuracy, physical consistency and robustness as joint training objectives, and achieves balance by dynamically adjusting weights. Compared with the model training method in the prior art that only optimizes a single objective, it solves the problem that it is difficult to balance physical constraints and prediction accuracy, and achieves the effect of improving model reliability and practical applicability.

[0023] 3. The present invention deploys models collaboratively on vehicle hardware and in the cloud, so that the fault prediction module can strike a balance between real-time performance and in-depth analysis capabilities. Compared with the technical solutions in the prior art that simply rely on vehicle computing or cloud analysis, this solves the problems of insufficient real-time performance or excessive consumption of computing resources, and ensures efficient operation and low-cost deployment of the model.

[0024] 4. The present invention ensures the long-term stability of the model by continuously monitoring the model performance and performing incremental learning based on actual operating data. Compared with the prediction method of fixed models in the prior art, it solves the problem of gradual failure of the model due to changes in the vehicle operating environment, and achieves long-term adaptability and accuracy in diversified scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] Please see attached Figure 1 The embodiment of the present invention provides a vehicle fault prediction method based on machine learning, comprising the following steps:

[0028] S1, data collection and preprocessing, obtaining sensor data, environmental data and historical fault record data of vehicle operation status, and cleaning, normalizing and time series slicing the above data;

[0029] In this embodiment, the sources of data collection include the following parts:

[0030] The multi-dimensional data of vehicle operation status mainly comes from sensor data, environmental data and historical fault record data.

[0031] Generally speaking, vehicle sensor data is the core data source for fault prediction, which directly reflects the operating status of the vehicle. Specifically, sensor data includes but is not limited to the following:

[0032] Speed, usually measured in kilometers per hour (km / h), is used to reflect the vehicle's driving conditions;

[0033] Acceleration, in meters per second squared (m / s 2 ) is used as a unit to reflect the dynamic changes of the vehicle;

[0034] Engine speed, measured in revolutions per minute (RPM), describes the operating state of the engine;

[0035] Torque, measured in Newton meters (Nm), reflects the power output of the engine;

[0036] Fuel consumption rate, measured in liters per 100 kilometers (L / 100km), reflects the energy efficiency of the vehicle;

[0037] Coolant temperature, measured in degrees Celsius (°C), is used to assess the health of the engine thermal management system.

[0038] As an option, environmental data can also be included in the collection scope to supplement and explain the external conditions of the vehicle's operating status. Specifically, environmental data includes temperature, humidity, road slope and vehicle load. Among them, temperature and humidity reflect external weather conditions, road slope is closely related to the load of the vehicle's power system, and load data can directly affect vehicle acceleration and energy consumption.

[0039] In addition, this embodiment also combines the fault label data in the vehicle's historical maintenance records. Specifically, the fault label data should include the time when the fault occurred, the fault category, and the scope of impact. These data are used to annotate the collected operating status data to form a complete training sample.

[0040] In this embodiment, in order to ensure data quality, the collected raw data is preprocessed as follows:

[0041] Generally speaking, vehicle operation data may contain outliers or missing values, which need to be cleaned to remove or repair problematic data.

[0042] Specifically, the processing of outliers can be combined with statistical analysis methods and physical constraints. For example, in speed sensor data, samples with negative speed or speed mutation exceeding the physical limit of the vehicle are considered outliers and are removed. In acceleration data, samples exceeding the design acceleration limit of the vehicle (such as ±15m / s 2 ) is also considered abnormal.

[0043] In one possible implementation, the missing data can be supplemented by interpolation methods or time series models. For example, when a small number of time periods are missing in the coolant temperature record, linear interpolation can be performed based on the temperature trend of adjacent time points. If the missing time period is long, a time series regression model based on the dynamic characteristics of the sensor can be used for prediction and supplementation.

[0044] In this embodiment, in order to ensure the standardization of the model input data, the cleaned data is normalized:

[0045] As an option, the Min-Max normalization method is used to map the numerical features to the [0,1] interval. Specifically, the normalization process is performed according to the following formula:

[0046]

[0047] Among them, x is the original data, x min and x max They represent the minimum and maximum values ​​of the data features respectively, and x′ is the normalized value.

[0048] In another implementation, the Z-Score standardization method can also be used to adjust the data to a distribution with a mean of zero and a standard deviation of one. The calculation formula is:

[0049]

[0050] Among them, μ is the sample mean, σ is the sample standard deviation, and z is the standardized value.

[0051] This normalization process can avoid the uneven impact of different features on model training due to dimensional differences, such as the bias of speed and acceleration features on model weights.

[0052] In this embodiment, in order to capture the dynamic characteristics of the time series, the preprocessed data is divided into time windows:

[0053] Generally speaking, vehicle failure is a dynamic process, and the data at a single time point may not be able to reflect the complete failure mode. Therefore, a fixed time window is used to slice the continuous time series into samples of fixed length.

[0054] Specifically, in one possible implementation, every 10 seconds is used as a time window. Assuming that 10 data points are collected per second, a sample contains 10×10=100 data points. In this way, the dynamic changes of sensor features can be captured in a short time, providing input support for the time series feature extraction of the subsequent model.

[0055] As an extension, the length of the time window can be adjusted according to the specific dynamic characteristics of the vehicle fault. For example, for situations where the engine speed changes rapidly, the window time can be shortened to 5 seconds to improve the accuracy of the dynamic response; while for characteristics that change slowly, such as fuel consumption rate, the window time can be appropriately extended to 30 seconds to capture long-term trends.

[0056] Through the processing of this embodiment, a complete and high-quality input data set is generated. These data cover the core information of the vehicle's operating status in terms of content, and the standardization of the data is improved through the operations of cleaning, normalization and time series slicing, providing a solid foundation for the training and optimization of subsequent models. At the same time, through the multiple selective implementations in the specific embodiments, the method of the present invention has good adaptability and can be adjusted according to different vehicles or application scenarios to meet diverse needs.

[0057] S2, introduce physical constraints, build a physical model based on vehicle dynamics, generate virtual samples for expanding training data, and embed physical constraints into the loss function during model training;

[0058] In this embodiment, a physical constraint model is constructed based on vehicle dynamics theory:

[0059] Generally speaking, the running state of a vehicle is determined by the balance relationship of multiple forces. Specifically, when a vehicle is driving, its engine output force needs to overcome air resistance, ground friction and vehicle inertia. This dynamic relationship can be described as:

[0060] F engine =F drag +F friction +F inertia

[0061] in,

[0062] F engine is the engine output force;

[0063] F drag is the air resistance, which can be calculated by the following formula:

[0064]

[0065] Where ρ is the air density, C d is the air resistance coefficient, A is the frontal area of ​​the vehicle, and v is the speed of the vehicle;

[0066] F friction is the ground friction, and the calculation formula is:

[0067] F friction =μmg

[0068] Among them, μ is the ground friction coefficient, m is the vehicle mass, and g is the acceleration due to gravity;

[0069] F inertia is the inertia force, which can be expressed as:

[0070] F inertia =m·a

[0071] Where a is the vehicle acceleration.

[0072] As an option, under different road conditions, the friction coefficient μ and the air resistance coefficient C d It can be adjusted according to the vehicle type and road conditions. For example, on a slippery road, the μ value may be significantly reduced, while for a vehicle traveling at high speed, C d The impact is more significant.

[0073] In this embodiment, a virtual sample is generated through a physical model:

[0074] In practical applications, it is difficult to collect data on certain vehicle states or fault categories, resulting in imbalanced training data. To alleviate this problem, virtual samples are generated based on the above dynamic model to supplement the lack of real data.

[0075] In general, the process of generating virtual samples includes the following steps:

[0076] Use the speed v and acceleration a data collected by the sensor to substitute into the dynamics formula to calculate the air resistance F in turn drag , ground friction F friction and inertial force F inertia .

[0077] In a possible implementation, the physical quantity obtained by the above calculation is combined with other state variables (such as engine speed, fuel consumption rate, etc.) collected by the sensor to form a complete virtual sample.

[0078] The generated virtual samples are distributed and adjusted. Specifically, the distribution differences between real data and virtual data are counted, and the virtual data is corrected using distribution alignment technology to make it closer to the distribution range of real data.

[0079] As an extension, the generation of virtual samples can also be adjusted according to specific fault scenarios. For example, for an engine overheating fault under high load conditions, the value range of mmm can be increased to generate virtual data that is more in line with the actual working conditions.

[0080] In this example, physical constraints are introduced to guide the training of the model:

[0081] During the model training process, in order to ensure that the prediction results are consistent with the laws of physics, a physical constraint regularization term is introduced into the loss function. Specifically, this constraint term is used to measure the degree of deviation between the model output results and the physical model. For example, when the model predicts the vehicle acceleration Penalties are imposed on the predictions when they are inconsistent with the theoretical values ​​calculated from the kinetic formula.

[0082] As an implementation approach, the embedding of physical constraints can be based on the following rules:

[0083] For the predicted vehicle state variables (such as acceleration, fuel consumption rate, etc.), calculate the deviation between them and the values ​​derived from the physical model;

[0084] The larger the deviation, the higher the penalty of the regularization term, thus guiding the prediction results in the model training to be closer to the physical laws.

[0085] Generally speaking, the introduction of this physical constraint will not significantly increase the complexity of model training, because the physical quantities required to calculate the deviation (such as air resistance and inertial force) can be quickly solved using simple formulas.

[0086] In some embodiments, the extension of this embodiment may include the following:

[0087] As a possibility, virtual sample generation can be combined with the boundary conditions of the actual collected data. For example, in mountainous road conditions, since the slope significantly affects the vehicle's operating state, the slope parameter can be incorporated into the physical model to generate virtual data with regional characteristics.

[0088] In addition, the design of physical constraints can also be extended to multi-objective optimization. For example, in addition to constraining the acceleration to be consistent with the dynamic model, the output of the model can be further optimized by combining the fuel consumption and energy efficiency models.

[0089] Through this embodiment, the dynamics of the vehicle is successfully combined with the machine learning method. The generation of virtual samples expands the data coverage, and the introduction of physical constraints improves the physical interpretation and robustness of the model. This combination enables the model to have good fault prediction capabilities even when data is scarce, and the prediction of the operating status is more realistic.

[0090] S3. Build a deep learning model based on time series, extract the time series features of sensor data, and predict vehicle fault categories;

[0091] In this embodiment, the input of the deep learning model is the processed time series feature matrix:

[0092] Generally, sensor data is collected continuously in the form of time series and converted into standardized samples by time window division in step 1. Specifically, each sample is a two-dimensional feature matrix of fixed length, where rows represent time steps and columns represent data features of different sensors.

[0093] As an implementation method, for samples with a window length of 10 seconds, assuming that the samples are collected 10 times per second, each sample contains 10 time steps, and each step has n characteristic values ​​(such as speed, acceleration, fuel consumption rate, etc.). Finally, a 10×n characteristic matrix is ​​formed. In this matrix, each column reflects the time change trend of a certain feature, and the rows reflect the temporal correlation of multi-sensor data.

[0094] In this embodiment, the feature extraction module of the deep learning model adopts an architecture that combines LSTM and Transformer:

[0095] In general, the long short-term memory network (LSTM) is suitable for extracting short-term dependencies in time series, especially for modeling dynamically changing vehicle operation states. Specifically, LSTM captures local features in the sequence through memory units and forget gate mechanisms, avoiding the gradient vanishing problem of traditional recurrent neural networks (RNNs).

[0096] In one possible implementation, the input matrix of the model is gradually input into the LSTM layer. The features of each time step are extracted by the LSTM unit into hidden states, which carry the time dependency information of the sequence and eventually form a new time series feature vector.

[0097] To further extract long-term dependencies and global features, this embodiment introduces the self-attention mechanism in the Transformer structure. Specifically, the self-attention mechanism captures the dependencies within the global time range by calculating the relative importance between features. Compared with LSTM, Transformer can process long time series more efficiently and identify complex interactions between features.

[0098] As an extension, the Transformer structure can use a multi-head attention mechanism to enhance feature capture capabilities. Each attention head calculates correlations on different feature dimensions, and ultimately achieves multi-angle modeling of features by merging the outputs of different heads.

[0099] In this embodiment, the output module of the deep learning model is designed for fault prediction of multi-classification problems:

[0100] Generally, the output module is used to convert the feature representation generated by the feature extraction module into a specific prediction result. Specifically, the global features extracted by LSTM and Transformer are passed to the fully connected layer to map the high-dimensional features to the fault category space.

[0101] As an implementation method, assume that the fault categories include k categories such as engine failure, brake system abnormality, and sensor failure, and the number of neurons in the output layer is k. The prediction probability of each category is calculated through the Softmax activation function, and the category with the highest probability is taken as the final prediction result.

[0102] In another possible implementation, for the vehicle health status assessment scenario, the output layer can be designed as a scalar to quantify the current health status of the vehicle. For example, the output health score can be between 0 and 100, with lower values ​​indicating higher risk of failure.

[0103] In this embodiment, the model training data includes a combination of real data and virtual data:

[0104] In general, to improve the generalization ability of the model, the training data should cover as many fault categories and operating scenarios as possible. The virtual samples generated in step 2 expand the scope of the training data set, especially in terms of the number of samples of rare faults.

[0105] As an option, real data and virtual data in the training data can be mixed in a certain proportion. For example, for some scenarios where categories are scarce, the proportion of virtual samples can be appropriately increased to improve the model's prediction performance on that category.

[0106] In this embodiment, in order to improve the robustness of the model, an adversarial training method is used:

[0107] Specifically, adversarial training simulates the actual situation of sensor noise or abnormal input by adding disturbances to the input data. For example, in the vehicle speed feature, random noise within a certain range can be added to generate samples with a certain degree of uncertainty.

[0108] In one possible implementation, the adversarial perturbation is generated by the following formula:

[0109]

[0110] in,

[0111] x is the original sample,

[0112] ∈ is the disturbance intensity,

[0113] is the gradient of the loss function with respect to the input data. Generated adversarial examples

[0114] x adv It can be used for model training to improve the model's adaptability in abnormal scenarios.

[0115] Through the model design and training of this embodiment, high-precision prediction of vehicle failures was successfully achieved. The model captures short-term changes in time series through LSTM and extracts long-term dependent features in combination with the Transformer structure. The addition of virtual data expands the coverage of training samples, while adversarial training improves the robustness of the model to noise and abnormal inputs. This combination of deep learning architecture and optimization method provides an efficient solution for vehicle operation health monitoring and fault prediction.

[0116] S4. Train the model through a multi-objective optimization loss function to jointly optimize fault prediction accuracy, physical consistency, and model robustness;

[0117] In this embodiment, the loss function design of multi-objective optimization includes the following core parts:

[0118] Generally, the loss function of the model needs to comprehensively consider multiple optimization objectives. Specifically, the loss function in the present invention is composed of fault prediction loss, physical consistency loss and robustness loss.

[0119] Fault prediction loss is the main optimization objective of the model, which is used to evaluate the model's ability to classify fault categories. As an option, a cross entropy loss function can be used to quantify the error between the model's predicted value and the true label. Cross entropy loss is suitable for multi-classification problems. The smaller its value, the closer the category distribution predicted by the model is to the true distribution.

[0120] The physical consistency loss is one of the innovative designs of the present invention, which is used to ensure that the model output results conform to the laws of vehicle dynamics. Specifically, the loss measures the degree of deviation between the state variables predicted by the model (such as acceleration and energy consumption) and the theoretical values ​​calculated by the dynamic model. As an implementation method, the physical consistency of acceleration can be constrained, that is, according to the relationship between the air resistance, friction and inertia of the vehicle, the deviation between the model prediction value and the theoretical value is calculated, and the deviation is added to the loss function.

[0121] Robustness loss is a goal designed to enhance model adaptability and improve the stability of the model in the presence of noisy data or abnormal inputs. Generally, adversarial training can be used to generate perturbation data to evaluate the sensitivity of the model's output results when there are slight changes in the input data. Optimizing robustness loss can effectively reduce the dependence of prediction results on noise, thereby improving the reliability of the model.

[0122] In this embodiment, the combination of loss functions is achieved by weighted summation:

[0123] In one possible implementation, the loss function is defined as:

[0124]

[0125] in,

[0126] represents the failure prediction loss;

[0127] Indicates loss of physical consistency;

[0128] represents robustness loss;

[0129] λ1, λ2, and λ3 are weight coefficients used to adjust the impact of each target on the total loss.

[0130] As an option, the initial value of the weight coefficient can be set based on experience or experimental results. For example, in a scenario where physical interpretation needs to be strengthened, the value of λ2 can be appropriately increased so that physical consistency accounts for a higher proportion of the total loss.

[0131] In another possible implementation, the weight coefficients can be dynamically adjusted through an optimization algorithm. Specifically, the Bayesian optimization method is used to search for the optimal weight combination based on the performance of the model on the validation set, thereby achieving a dynamic balance of various optimization objectives.

[0132] In this embodiment, model training combines multi-objective optimization and adversarial training strategies:

[0133] Generally speaking, model training needs to ensure the stability of the training process while optimizing the multi-objective loss function. Specifically, the training process includes the following steps:

[0134] First, the real data is mixed with the virtual samples generated in step 2 in a certain proportion to form a complete training data set. As an implementation method, categories with fewer virtual samples can be preferentially selected for data augmentation to alleviate the category imbalance problem.

[0135] Secondly, adversarial samples are added to the training data to improve the robustness of the model. Adversarial samples can be generated by perturbing the original input data. For example, a small amount of random noise is added to the speed feature to simulate the actual scenario of sensor abnormality. The proportion of adversarial samples can be dynamically adjusted according to actual needs to ensure the adaptability of the model under complex conditions.

[0136] Finally, the model parameters are optimized by gradient descent. As an extension, the AdamW optimizer can be used, which can more accurately control weight decay while dynamically adjusting the learning rate, thereby improving the convergence efficiency of training.

[0137] The optimization method in this embodiment also includes a distribution adaptive strategy:

[0138] In one possible implementation, the distribution of training data may vary due to different operating scenarios. For example, the characteristics of sensor data under different vehicle types or road conditions may vary significantly. In order to adapt to this distribution difference, a distribution adaptive regularization strategy can be introduced in model training to encourage the model to maintain consistent prediction capabilities in multiple scenarios.

[0139] Specifically, the feature distribution of data from different scenarios can be statistically analyzed to calculate the prediction deviation of the model under distribution drift, and the deviation can be added to the loss function as a regularization term. For example, the distribution characteristics of speed and acceleration may be different on slopes and highways. By optimizing the distribution adaptive regularization term, the generalization performance of the model can be improved.

[0140] Through the multi-objective optimization design and model training of this embodiment, a balance between fault prediction accuracy, physical consistency and robustness is successfully achieved. In particular, through dynamic weight adjustment and distribution adaptive strategy, the model can be flexibly optimized according to the needs of different application scenarios, improving the applicability and reliability in practical applications. The above method provides a complete solution for the comprehensive performance optimization of the vehicle fault prediction system.

[0141] S5. Verify model performance and evaluate the model’s fault prediction capability and generalization performance through real-world scenario testing;

[0142] In this embodiment, model performance verification includes multiple dimensions:

[0143] Generally, model performance is verified from two aspects: static data testing and dynamic scenario simulation to comprehensively evaluate the accuracy and stability of the model.

[0144] Specifically, in static data testing, the prediction ability of the model is evaluated by the classification performance of the fault category. As an implementation method, indicators such as accuracy, recall, and F1-Score are used to quantify the classification effect of the model. Accuracy reflects the correctness of the overall prediction of the model, recall evaluates the recognition ability of fault samples, and F1-Score takes into account the balance between accuracy and recall.

[0145] In dynamic scenario testing, to verify the physical consistency of the model, the prediction results are compared with the vehicle dynamics. For example, the predicted vehicle acceleration data can be substituted into the dynamics model in step 2 to calculate the corresponding air resistance, friction and other variables. By comparing the changing trends of these physical variables, it is evaluated whether the model meets the physical constraints.

[0146] Alternatively, specific performance deviation metrics can be designed to quantify how well the model performs in terms of physical consistency, for example, whether the tendency of air resistance to increase with the square of speed at higher vehicle speeds is correctly captured by the model.

[0147] In this embodiment, scenario testing mainly verifies the generalization ability of the model through a variety of operating environments:

[0148] Generally, the vehicle operating environment includes different road conditions, climate conditions and vehicle load conditions. In order to verify the performance of the model under these environments, a variety of test scenarios are constructed.

[0149] In one possible implementation, the test scenarios include the following categories:

[0150] Road conditions: flat roads, ramps, high-speed sections, and low-speed congested sections.

[0151] Climate conditions: high temperature environment, low temperature environment and slippery road surface.

[0152] Vehicle load: three states: empty, half-loaded and fully loaded.

[0153] In each scenario, the vehicle's sensor data is collected and input into the model for prediction. By comparing the classification performance indicators in different scenarios, it is evaluated whether the model can adapt to data input with obvious distribution differences.

[0154] As an extension, extreme scenarios can be simulated to verify the robustness of the model. For example, assuming that multiple sensors fail at the same time (such as the data of the velocity sensor and the acceleration sensor are abnormal at the same time), by adding adversarial samples, the fault prediction ability of the model under complex conditions can be tested.

[0155] In this embodiment, the operating efficiency of the model is also evaluated:

[0156] Specifically, the operational efficiency evaluation focuses on the model's reasoning speed and computing resource consumption. In the real-time operation environment of the vehicle, the model needs to complete data processing and fault prediction within a limited time to meet real-time requirements.

[0157] As an implementation method, the real-time performance of the model can be quantified by recording the inference time of the model under different hardware conditions. For example, on an in-vehicle computing device, the single inference time of the model is required to be no more than 100 milliseconds. At the same time, the memory usage of the model is evaluated to ensure its applicability on embedded devices.

[0158] In another possible implementation, the lightweight version of the model can be optimized to reduce the number of parameters and computational complexity to improve operating efficiency. Through the knowledge distillation method, the teacher model with superior performance guides the training of the lightweight student model to ensure a balance between operating efficiency and prediction performance.

[0159] In this example, the verification process also combines the explanatory analysis of the model output results:

[0160] Generally speaking, the black box nature of deep learning models may make the output results difficult to interpret. To this end, the present invention improves the output interpretability of the model by combining physical models with feature importance analysis.

[0161] Specifically, after predicting the fault category, the input features that contribute most to fault prediction can be identified through feature importance evaluation. For example, the SHAP (Shapley Additive Explanations) method is used to calculate the contribution of each feature to the model decision. Through this analysis, it is possible to intuitively show how the model identifies faults based on features such as speed, acceleration, or fuel consumption rate.

[0162] As an option, causal reasoning can be performed on the prediction results based on physical laws. For example, when the prediction results indicate abnormal fuel consumption, the vehicle dynamics model can be combined to infer that the possible cause is insufficient engine output power or abnormal air resistance.

[0163] In some embodiments, extended testing of this embodiment includes monitoring of the long-term performance of the model:

[0164] Generally, the vehicle operating environment and usage conditions may change over time. To ensure the long-term applicability of the model, the model can be updated through incremental training or online learning mechanisms.

[0165] Specifically, new operating data can be continuously collected after the model is deployed and used to fine-tune the model. Through online learning, the model can adapt to changes in data distribution, such as seasonal climate change or long-term vehicle performance degradation.

[0166] Through the model verification and scenario testing of this embodiment, the adaptability and prediction performance of the model under different operating conditions are comprehensively evaluated. In particular, through the evaluation of diversified test scenarios and operating efficiency, the reliability and real-time performance of the model in practical applications are ensured. The above verification process provides a solid technical guarantee for the actual implementation of the present invention.

[0167] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A vehicle fault prediction method based on machine learning, characterized in that: The following steps are involved: S1, data collection and preprocessing, obtaining sensor data of vehicle operation status, environmental data and historical fault record data, and cleaning, normalizing and time series slicing the above data; S2, introduce physical constraints, build a physical model based on vehicle dynamics, generate virtual samples for expanding training data, and embed physical constraints into the loss function during model training; S3. Build a deep learning model based on time series, extract the time series features of sensor data, and predict the vehicle fault category; S4. Train the model through a multi-objective optimization loss function to jointly optimize fault prediction accuracy, physical consistency, and model robustness; S5. Verify model performance and evaluate the model’s fault prediction capability and generalization performance through real-world scenario testing.

2. The vehicle fault prediction method based on machine learning according to claim 1, characterized in that: The sensor data includes speed, acceleration, engine speed, torque, fuel consumption rate and coolant temperature, the environmental data includes air temperature, humidity, road slope and vehicle load, and the fault record data includes fault type and fault occurrence time.

3. The vehicle fault prediction method based on machine learning according to claim 1, characterized in that: The physical constraint is constructed through the following vehicle dynamics relationship: the engine output force is equal to the sum of air resistance, ground friction and vehicle inertia, the air resistance is determined by the square of the vehicle speed and the air density, drag coefficient and frontal area, and the inertia is determined by the vehicle acceleration and mass.

4. The vehicle fault prediction method based on machine learning according to claim 1, characterized in that: The generated virtual samples are combined with real collected data to form extended samples after unobserved physical variables are calculated based on the physical model, and the data distribution of the virtual samples is adjusted through a distribution alignment method.

5. The vehicle fault prediction method based on machine learning according to claim 1, characterized in that: The deep learning model adopts a combination of long short-term memory network and Transformer structure, uses LSTM to extract short-term dependency features of time series, and uses the self-attention mechanism of Transformer to extract long-term dependency features and global features.

6. The vehicle fault prediction method based on machine learning according to claim 1, characterized in that: The loss function of the multi-objective optimization includes fault prediction loss, physical consistency loss and robustness loss, wherein the fault prediction loss is calculated by the error between the true fault label and the predicted value, the physical consistency loss is calculated by the degree of deviation between the predicted value and the physical model, and the robustness loss is calculated by the stability of the model output change after simulating the input data disturbance.

7. The vehicle fault prediction method based on machine learning according to claim 1, characterized in that: In S1, in order to enhance the robustness of the model, an adversarial training strategy is introduced in the training process, by adding disturbances to the input data to simulate sensor anomalies, thereby improving the stability and accuracy of the model in complex and noisy environments.

8. The vehicle fault prediction method based on machine learning according to claim 1, characterized in that: The optimization of the loss function adopts the Bayesian optimization method to dynamically adjust the weight coefficients to balance the fault prediction accuracy, physical consistency and robustness objectives.

9. The vehicle fault prediction method based on machine learning according to claim 1, characterized in that: The verification of the model includes performance indicator testing and scenario testing, wherein the performance indicators include the accuracy, recall rate and F1 value of fault prediction, and the scenario testing includes the generalization performance evaluation under different loads, road conditions and climate environments.

10. The vehicle fault prediction method based on machine learning according to claim 1, characterized in that: The scenario testing also includes testing under extreme conditions, targeting scenarios where multiple sensor failures occur simultaneously, to verify the reliability and predictive ability of the model in complex scenarios.

Images