Vehicle-mounted power assembly remote monitoring system and method

Abstract

The invention relates to the technical field of vehicle engineering and intelligent monitoring, and discloses a remote monitoring system and method for a vehicle-mounted power assembly, and the method comprises the steps: constructing a multi-modal data collection frame for collecting various types of data; designing a federated learning driven anomaly detection model, deploying a global model at the cloud, and processing data and updating the global model by edge nodes; implementing a control strategy of reinforcement learning optimization, and generating a self-adaptive control instruction according to an anomaly detection result; a distributed feedback calibration is performed to adjust the model parameters and the reward function. The system comprises a vehicle-mounted terminal, a cloud server, a distributed database and a control execution module, and interacts data by means of a 5G-V2X communication protocol. According to the invention, comprehensive and intelligent monitoring and control of the vehicle-mounted power assembly are realized, the anomaly detection accuracy and the power assembly efficiency are improved, the system adaptability is enhanced, and reliable operation of the vehicle is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical fields of vehicle engineering and intelligent monitoring, and particularly to an on-vehicle powertrain remote monitoring system and method. Background Art

[0002] With the booming development of the automotive industry, the degree of vehicle intelligence and networking is constantly deepening. As the core component of a vehicle, the stable and efficient operation of the on-vehicle powertrain is crucial for the overall performance, safety, and reliability of the vehicle. However, there are many bottlenecks in the current monitoring technology of on-vehicle powertrains, making it difficult to meet the growing market demand.

[0003] In terms of data collection, the data types collected by traditional methods are single, mainly focusing on basic parameters such as engine speed and vehicle speed, ignoring the complex environmental factors and other key data during the operation of the powertrain. For example, changes in environmental temperature and humidity can affect the performance of engine coolant and lubricating oil, thereby affecting the working efficiency and wear degree of the powertrain; altitude and road slope directly affect the load of the engine and the shifting logic of the transmission. However, existing monitoring systems fail to comprehensively collect this environmental data, resulting in incomplete analysis of the powertrain working conditions, inability to accurately evaluate its operating state, and difficulty in detecting potential fault hazards in advance.

[0004] Anomaly detection technology also faces challenges. Most existing anomaly detection models are based on fixed rules or simple algorithms and lack the adaptive ability to abnormal patterns of the powertrain under different working conditions. Driving habits and road conditions vary greatly in different regions, making the operating conditions of the powertrain complex and changeable. In urban congested road conditions, the vehicle starts and stops frequently, and the powertrain bears frequent impacts and load changes; while on the highway, the vehicle travels at a high speed for a long time, and the powertrain is in a high-load and stable operating state. Traditional detection models are difficult to accurately identify anomalies under these complex working conditions and are prone to false alarms or missed alarms. In addition, the centralized data processing method uploads a large amount of vehicle data to the central server, which not only brings high data transmission costs and privacy risks, but also greatly increases the server processing pressure and makes it difficult to achieve real-time detection.

[0005] The control strategy of the powertrain also has deficiencies. Traditional control methods lack intelligence and refinement and cannot be adaptively adjusted according to the real-time state of the powertrain and the driving requirements of the vehicle. In terms of engine power control, it cannot accurately match the actual power demand of the vehicle, resulting in fuel waste or insufficient power; the shifting logic of the transmission also often fails to fully consider the driver's intention and the vehicle driving conditions, affecting driving comfort and the service life of the powertrain. For example, when climbing a slope or overtaking, it cannot timely adjust the engine power and transmission gear, resulting in slow vehicle power response.

[0006] The imperfect data interaction and feedback calibration mechanism is also a major problem. The data synchronization between the vehicle terminal and the server is not timely, which makes the vehicle data obtained by the server lag and cannot reflect the latest status of the powertrain. Moreover, due to the lack of an effective feedback calibration mechanism, it is difficult to dynamically optimize the monitoring model and control strategy according to the actual operation results of the vehicle, resulting in the system gradually deviating from the optimal performance state during long-term operation, and unable to continuously ensure the stable operation of the powertrain. With the deepening of the development of automobile intelligence, users have higher and higher requirements for vehicle performance and safety, and traditional vehicle powertrain monitoring technology can no longer meet market demand. Therefore, the development of a vehicle powertrain remote monitoring system and method that can comprehensively collect multimodal data, realize efficient anomaly detection, implement intelligent and precise control, and have real-time feedback calibration capabilities has become an important issue that needs to be urgently solved in the current field of automotive technology. Summary of the invention

[0007] The object of the present invention is to provide a vehicle-mounted powertrain remote monitoring system and method to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solution: a vehicle-mounted powertrain remote monitoring method, the method comprising:

[0009] Construct a multimodal data collection framework, including real-time collection of powertrain operation data, environmental data and vehicle status data through the vehicle terminal; the operation data includes engine speed, torque, vibration spectrum and gearbox gear timing; the environmental data includes ambient temperature and humidity, altitude and road slope; the vehicle status data includes vehicle speed, acceleration and battery voltage;

[0010] Designing a federated learning-driven anomaly detection model, including deploying a global model on a cloud server and preprocessing and extracting features from local data based on edge computing nodes; the federated learning updates the global model by aggregating local model parameters of multiple vehicles to dynamically adapt to powertrain anomaly patterns under different operating conditions;

[0011] Implementing a control strategy optimized by reinforcement learning, including generating adaptive control instructions based on the output of the anomaly detection model; the control instructions include engine power adjustment, transmission shift logic correction, and drive shaft torque distribution optimization;

[0012] Perform distributed feedback calibration, including periodically synchronizing data between the vehicle terminal and the cloud server, and adjusting the weight parameters of the federated learning model and the reward function of the reinforcement learning strategy based on the actual operation results.

[0013] Preferably, the multimodal data acquisition framework includes:

[0014] Synchronously collect multi-source heterogeneous data of the powertrain through the in-vehicle sensor network; perform time-frequency domain decomposition on the vibration spectrum using wavelet transform to extract high-frequency abnormal features; encapsulate the gearshift timing data of the gearbox through the CAN bus protocol; use the Kalman filter to suppress noise in the vehicle speed and acceleration data.

[0015] Preferably, the specific construction of the federated learning-driven anomaly detection model includes:

[0016] Deploy a lightweight convolutional neural network in the edge computing node to extract spatio-temporal features of the local vibration spectrum;

[0017] Design a differential privacy mechanism to add Gaussian noise to the local model parameters to ensure data privacy;

[0018] Adopt an adaptive weighted aggregation algorithm to update the global model, and its weight allocation formula is:

[0019]

[0020] where w g is the global model parameter, w i is the local model parameter of the i-th vehicle, and α i is the dynamic weight coefficient calculated according to the data quality and anomaly detection confidence.

[0021] Preferably, the reinforcement learning-optimized control strategy includes:

[0022] Construct a Markov decision process model, define the state space as the set of real-time operating parameters of the powertrain, the action space as the set of control instructions, and the reward function is based on the powertrain efficiency improvement rate and anomaly suppression effect;

[0023] Adopt the deep deterministic policy gradient algorithm to optimize the control strategy and output continuous control instructions through the Actor-Critic network.

[0024] Preferably, the distributed feedback calibration includes:

[0025] Deploy an incremental learning module in the in-vehicle terminal to update the local model parameters according to the real-time control effect;

[0026] Design a sliding window mechanism to screen historical data, and only retain the data with a correlation higher than the threshold with the current working condition for model calibration.

[0027] Preferably, the feature extraction further includes:

[0028] Construct a Gram angular field for the engine speed and torque data to convert the time series data into two-dimensional image features;

[0029] Dimensionality reduction is performed on multi-dimensional environmental data using principal component analysis to extract key influencing factors.

[0030] Preferably, the calculation formula for the dynamic weight coefficient is:

[0031]

[0032] where C i is the anomaly detection confidence of the i-th vehicle, and C j is the anomaly detection confidence of the j-th vehicle, and β is the temperature coefficient used to adjust the smoothness of the weight distribution.

[0033] Preferably, the design of the reward function includes:

[0034] R(t) = λ1·η(t) - λ2·ΔE(t) + λ3·S(t)

[0035] where R(t) is the reward value, η(t) is the powertrain efficiency, ΔE(t) is the change in the anomaly index, S(t) is the smoothness of the control command, and λ1, λ2, and λ3 are balance coefficients.

[0036] Preferably, the method further includes:

[0037] Construct a virtual simulation environment on the cloud server to simulate the response of the powertrain under extreme working conditions;

[0038] Fuse the simulation data with the real data through digital twin technology for the adversarial training of the federated learning model.

[0039] Preferably, the present invention further includes an on-vehicle powertrain remote monitoring system for implementing the above-mentioned on-vehicle powertrain remote monitoring method, including:

[0040] An on-vehicle terminal integrating a multi-modal sensor module, an edge computing unit, and a communication module; the sensor module is used to collect powertrain operation data, environmental data, and vehicle status data; the edge computing unit is used for data preprocessing and local model inference;

[0041] A cloud server deploying a federated learning global model, a reinforcement learning policy library, and a virtual simulation platform;

[0042] A distributed database storing historical operation data and model parameters of multiple vehicles;

[0043] A control execution module embedded in the vehicle ECU for receiving and executing adaptive control commands;

[0044] The on-vehicle terminal, the cloud server, and the control execution module achieve real-time data interaction through the 5G-V2X communication protocol.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] In the data acquisition stage, the multi-modal data acquisition framework realizes the comprehensive acquisition of powertrain operation data, environmental data, and vehicle status data. By synchronously collecting multi-source heterogeneous data through an in-vehicle sensor network and using wavelet transform to perform time-frequency domain decomposition on the vibration spectrum, high-frequency abnormal features that are difficult to discover by traditional methods can be extracted, which helps to detect potential faults of the powertrain in advance. Encapsulating the gearshift timing data of the gearbox using the CAN bus protocol ensures the accuracy and real-time performance of data transmission; the Kalman filter suppresses the noise of vehicle speed and acceleration data, improves the data quality, and provides a reliable basis for subsequent analysis and decision-making. This enables the monitoring system to more comprehensively and accurately grasp the operating state of the powertrain, laying a solid foundation for precise diagnosis and control.

[0047] The anomaly detection model driven by federated learning greatly improves the accuracy and adaptability of anomaly detection. Deploying a lightweight convolutional neural network at the edge computing node not only reduces the amount of data transmission but also can quickly extract the spatio-temporal features of the local vibration spectrum, effectively capturing the abnormal signals of the powertrain. The introduction of the differential privacy mechanism provides strong protection for data privacy and protects the sensitive information of users during the data sharing process. The adaptive weighted aggregation algorithm dynamically assigns weights to update the global model according to the data quality and anomaly detection confidence, enabling the model to quickly adapt to the abnormal patterns under different working conditions. Whether it is complex urban road conditions or harsh field environments, it can accurately detect anomalies, greatly improving the reliability and stability of the system.

[0048] The control strategy optimized by reinforcement learning realizes the intelligent and refined control of the powertrain. By constructing a Markov decision process model, reasonably defining the state space, action space, and reward function, the control problem of the powertrain is transformed into an optimizable decision-making process. The reward function designed based on the powertrain efficiency improvement rate and anomaly suppression effect effectively balances the relationship between performance optimization and fault prevention. The application of the deep deterministic policy gradient algorithm enables the system to output continuous and accurate control instructions, realizing the optimal distribution of engine power, gearbox shift logic, and drive shaft torque. During vehicle driving, the system can automatically adjust the powertrain parameters according to the real-time road conditions and vehicle status, improving fuel economy and power performance, while reducing component wear and extending the service life of the powertrain.

[0049] The distributed feedback calibration mechanism ensures the continuous optimization and adaptability of the system. The incremental learning module deployed in the vehicle terminal updates the local model parameters according to the real-time control effect, enabling the system to quickly adapt to vehicle individual differences and working conditions changes. The sliding window mechanism filters historical data and only retains the data highly relevant to the current working condition for model calibration, improving the data processing efficiency and avoiding the interference of invalid data on the model. By periodically synchronizing the data between the vehicle terminal and the cloud server and adjusting the weight parameters of the federated learning model and the reward function of the reinforcement learning strategy, the system can continuously learn and optimize, and always maintain the best performance state during long-term operation, continuously providing guarantee for the stable operation of the vehicle powertrain.

[0050] The application of the virtual simulation environment constructed by the cloud server and digital twin technology further expands the functions of the system. The virtual simulation environment simulates the response of the powertrain under extreme working conditions, providing rich training data for the federated learning model and enhancing the model's adaptability to complex working conditions. Digital twin technology fuses simulation data with real data for adversarial training, improving the robustness and anomaly detection ability of the model. This enables the system to still work stably and accurately in the face of various extreme situations, effectively improving the safety and reliability of the vehicle.

[0051] The vehicle powertrain remote monitoring system and method of the present invention, through the collaborative application of a series of innovative technologies, realizes the all-round and intelligent remote monitoring and control of the powertrain, improves the overall performance and safety of the vehicle, reduces the maintenance cost, and promotes the development of the automotive industry towards the direction of intelligence and high efficiency, with significant economic and social benefits. Brief Description of the Drawings

[0052] Figure 1 It is the working principle diagram of the vehicle powertrain remote monitoring method described in the present invention;

[0053] Figure 2 It is the working principle diagram of the construction diagram of the federated learning-driven anomaly detection model;

[0054] Figure 3 It is the working principle diagram of the reinforcement learning to optimize the control strategy. Detailed Embodiments

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0056] Please refer to Figures 1-3, the present invention provides a technical solution: a method for remote monitoring of vehicle powertrain, the method comprising:

[0057] Construct a multi-modal data acquisition framework: With the help of in-vehicle terminals, real-time collect powertrain operation data, environmental data, and vehicle status data. The operation data includes engine speed, torque, vibration spectrum, and gearbox shift sequence, etc., which are the core indicators reflecting the working state of the powertrain; the environmental data includes environmental temperature and humidity, altitude, and road surface gradient, etc., which have an important impact on the performance of the powertrain; the vehicle status data includes vehicle speed, acceleration, and battery voltage, etc., which helps to comprehensively understand the driving condition of the vehicle.

[0058] Design an anomaly detection model driven by federated learning: Deploy a global model on the cloud server, and use edge computing nodes to preprocess local data and extract features. Federated learning updates the global model by aggregating the local model parameters of multiple vehicles, so as to dynamically adapt to the anomaly patterns of the powertrain under different working conditions, and improve the accuracy and generality of anomaly detection.

[0059] Implement a control strategy optimized by reinforcement learning: Generate adaptive control instructions based on the output results of the anomaly detection model. These control instructions include engine power adjustment, gearbox shift logic correction, and drive shaft torque distribution optimization, etc., aiming to make reasonable control adjustments in time when anomalies are detected or the performance of the powertrain needs to be optimized.

[0060] Execute distributed feedback calibration: Periodically synchronize the data between the in-vehicle terminal and the cloud server, and adjust the weight parameters of the federated learning model and the reward function of the reinforcement learning strategy according to the actual operation results. This step can continuously optimize the model and control strategy to better adapt to the actual operation of the vehicle.

[0061] The present invention will be further described below in conjunction with Embodiments 1 to 5:

[0062] Embodiment 1:

[0063] This embodiment aims to elaborate in detail the specific working mode of the multi-modal data acquisition framework to ensure that the collected data is accurate and effective, providing a reliable basis for subsequent analysis and decision-making.

[0064] Inside the vehicle, construct an in-vehicle sensor network. This network consists of various types of sensors, each responsible for collecting different types of data. For example, use a speed sensor to collect the engine speed, a torque sensor to obtain the engine torque, and a vibration sensor to measure the vibration spectrum. Through the collaborative work of the sensors, synchronous acquisition of multi-source heterogeneous data of the powertrain is realized.

[0065] For vibration spectrum data, due to the complexity of the information it contains, wavelet transform is used for time-frequency domain decomposition. Wavelet transform is a signal processing technique that can analyze the vibration spectrum in both the time and frequency dimensions. Through wavelet transform, high-frequency abnormal features can be extracted, and these features are often important indicator signals for early faults in the powertrain. For example, when a component of the powertrain becomes loose or worn, the high-frequency components in the vibration spectrum will change, and the high-frequency abnormal features extracted by wavelet transform can detect this change in a timely manner.

[0066] The gear shift timing data of the transmission is encapsulated through the CAN bus protocol. The CAN (Controller Area Network) bus is a serial communication protocol widely used in the automotive electronics field, with advantages such as high reliability and strong real-time performance. During vehicle operation, the gear shift change information of the transmission is encoded and transmitted according to the CAN bus protocol to ensure the accurate and rapid transmission of data for subsequent analysis of the working state of the transmission.

[0067] Vehicle speed and acceleration data are prone to noise interference during the acquisition process, and a Kalman filter is used to suppress the noise. The Kalman filter is an optimal linear recursive filter that can effectively remove noise and obtain more accurate vehicle speed and acceleration data through the prediction of the system state and the fusion of measurement data. For example, during vehicle driving, due to reasons such as road bumps, the data collected by the acceleration sensor will have noise, and the Kalman filter can perform an optimal estimate of the acceleration based on the previous state and the current measurement data to improve the reliability of the data.

[0068] Embodiment 2:

[0069] In the specific construction process of the anomaly detection model driven by federated learning, various technical means are used to improve the performance and data security of the model. A lightweight convolutional neural network (CNN) is deployed at the edge computing node. The lightweight CNN is a neural network architecture designed specifically for resource-constrained edge computing devices, which has the characteristics of small computational complexity and few model parameters, and can quickly process local vibration spectrum data without affecting the accuracy of anomaly detection. For example, lightweight CNN models such as MobileNet and ShuffleNet reduce the computational complexity and memory occupancy by optimizing the network structure and parameters, and are suitable for running at the edge computing node.

[0070] To ensure data privacy, a differential privacy mechanism is designed. Gaussian noise is added to the local model parameters before they are uploaded to the cloud server. Gaussian noise is a common type of random noise whose probability distribution conforms to the Gaussian distribution. By adding Gaussian noise, even if an attacker obtains the uploaded model parameters, it is difficult to accurately reconstruct the original data from the noise, thus protecting the data privacy of vehicle users.

[0071] The adaptive weighted aggregation algorithm is used to update the global model. Its weight assignment formula is:

[0072]

[0073] where w g is the global model parameter, w i is the local model parameter of the i-th vehicle, and α i is the dynamic weight coefficient calculated based on data quality and anomaly detection confidence. The calculation formula for the dynamic weight coefficient is:

[0074]

[0075] where C i is the anomaly detection confidence of the i-th vehicle, C j is the anomaly detection confidence of the j-th vehicle, and β is the temperature coefficient used to adjust the smoothness of the weight distribution. For example, when a vehicle has a high anomaly detection confidence and good data quality, its corresponding dynamic weight coefficient α i will be relatively large, and when updating the global model, the local model parameters of this vehicle will contribute more to the global model.

[0076] In terms of feature extraction, in addition to processing the vibration spectrum, a Gramian angular field is also constructed for engine speed and torque data. The Gramian angular field is a method of converting time series data into two-dimensional image features. By converting engine speed and torque data into two-dimensional images, image processing techniques can be used for feature extraction and analysis. For example, using a convolutional neural network to process the Gramian angular field image can extract richer features and improve the accuracy of anomaly detection.

[0077] For multi-dimensional environmental data, principal component analysis (PCA) is used for dimensionality reduction. PCA is a commonly used data analysis method that can convert multiple related environmental data (such as environmental temperature and humidity, altitude, road surface slope, etc.) into a few uncorrelated principal components, which retain most of the information of the original data. Through PCA dimensionality reduction, not only can the complexity of data processing be reduced, but also the key influencing factors that have a greater impact on the working state of the powertrain can be extracted, improving the efficiency and performance of the model.

[0078] Example 3:

[0079] This embodiment details the specific implementation of the control strategy optimized by reinforcement learning. By constructing a reasonable model and algorithm, precise control of the powertrain is achieved. A Markov Decision Process (MDP) model is constructed. In this model, the state space is defined as the set of real-time operating parameters of the powertrain, including engine speed, torque, vibration spectrum, transmission gear, vehicle speed, acceleration, etc. These parameters comprehensively reflect the current working state of the powertrain and are the basis for decision-making. The action space is the set of control commands, i.e., control commands such as engine power adjustment, transmission shift logic correction, and drive shaft torque distribution optimization. The reward function is designed based on the powertrain efficiency improvement rate and the abnormal suppression effect, and the formula is:

[0080] R(t) = λ1·η(t) - λ2·ΔE(t) + λ3·S(t)

[0081] where R(t) is the reward value, η(t) is the powertrain efficiency, ΔE(t) is the change in the abnormal index, S(t) is the smoothness of the control command, and λ1, λ2, λ3 are balance coefficients. For example, when the powertrain efficiency increases, η(t) increases, and the reward value R(t) will increase accordingly; when the change in the abnormal index decreases, the value of -λ2·ΔE(t) increases, which will also increase the reward value R(t); at the same time, the higher the smoothness S(t) of the control command, the higher the reward value R(t). By adjusting the balance coefficients λ1, λ2, λ3, the relationship between the powertrain efficiency, abnormal suppression effect, and control command smoothness can be balanced according to actual needs.

[0082] The Deep Deterministic Policy Gradient (DDPG) algorithm is used to optimize the control strategy. The DDPG algorithm is a reinforcement learning algorithm based on the Actor-Critic network that can output continuous control commands. The Actor network is responsible for generating control commands, and the Critic network is used to evaluate the value of the control commands generated by the Actor network. During the training process, the Actor network generates control commands based on the current state, the Critic network calculates the reward value according to the actual response of the powertrain and feedbacks it to the Actor network, and the Actor network adjusts its own parameters according to the reward value to generate better control commands. Through continuous training, the DDPG algorithm can find the optimal control strategy to achieve efficient control of the powertrain.

[0083] Example 4:

[0084] Deploy an incremental learning module in the in-vehicle terminal. The incremental learning module can update the local model parameters according to the real-time control effect. When a control instruction is executed during vehicle operation, the incremental learning module will collect the actual response data of the powertrain and compare it with the expected result. If there are differences, the incremental learning module will adjust the local model parameters based on these differences so that the local model can better reflect the actual situation of the current vehicle. For example, when the engine power is adjusted, the incremental learning module will monitor parameters such as the actual output torque and speed of the engine, and adjust the engine-related parameters in the local model according to the changes in these parameters to improve the prediction accuracy of the model for engine performance.

[0085] Design a sliding window mechanism to screen historical data. The sliding window mechanism will slide in the historical data and only retain the data with a correlation higher than the threshold with the current working condition for model calibration. For example, set a time window, and the data within the window is the vehicle operation data in the recent period. At each time step, the sliding window moves forward by one time unit, incorporates new data into the window, and deletes the earliest data in the window. By calculating the correlation between the data in the window and the current working condition, such as the similarity measure based on the powertrain operation parameters, environmental data, and vehicle state data, the data with a correlation higher than the threshold is screened out. These screened data are used to update the weight parameters of the federated learning model and the reward function of the reinforcement learning strategy to ensure that the model and strategy can be accurately calibrated according to the current working condition, improving the adaptability and accuracy of the system.

[0086] Embodiment 5:

[0087] Build a virtual simulation environment on the cloud server. The virtual simulation environment can simulate the response of the powertrain under extreme working conditions. For example, simulate the working state of the powertrain under harsh environments such as high temperature, high cold, high altitude, and extreme driving conditions such as rapid acceleration, rapid deceleration, and frequent gear shifting. By conducting a large number of simulation experiments in the virtual simulation environment, data that is difficult to collect during actual operation can be obtained, enriching the training data of the federated learning model and improving the model's adaptability to extreme working conditions.

[0088] The simulation data is fused with the real data by using digital twin technology. Digital twin technology is a technology that creates a digital mapping for physical objects by using data such as physical models, sensor updates, and operation history. In the present invention, the simulation data generated in the virtual simulation environment is fused with the real data collected by the vehicle terminal for the adversarial training of the federated learning model. For example, the powertrain data under extreme working conditions obtained by simulation is combined with the normal working condition data during actual operation, enabling the federated learning model to learn the differences and laws between different working conditions. During the adversarial training process, through technologies such as the generative adversarial network (GAN), the model is continuously optimized to improve its detection ability for abnormal working conditions and its adaptability to complex working conditions, thereby enhancing the performance of the entire vehicle powertrain remote monitoring system.

[0089] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0090] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for remote monitoring of a vehicle powertrain, characterized in that, The method includes: Constructing a multi-modal data acquisition framework, including real-time acquisition of the operating data, environmental data, and vehicle status data of the powertrain through an in-vehicle terminal; the operating data includes engine speed, torque, vibration spectrum, and gearbox gear shift sequence; the environmental data includes environmental temperature and humidity, altitude, and road surface gradient; the vehicle status data includes vehicle speed, acceleration, and battery voltage; Designing an anomaly detection model driven by federated learning, including deploying a global model on a cloud server and preprocessing and feature extraction of local data based on edge computing nodes; the federated learning updates the global model by aggregating the local model parameters of multiple vehicles, and is used to dynamically adapt to the anomaly patterns of the powertrain under different working conditions; Implementing a control strategy optimized by reinforcement learning, including generating an adaptive control instruction according to the output result of the anomaly detection model; the control instruction includes engine power adjustment, gearbox shift logic correction, and drive shaft torque distribution optimization; Performing distributed feedback calibration, including periodically synchronizing the data between the in-vehicle terminal and the cloud server, and adjusting the weight parameters of the federated learning model and the reward function of the reinforcement learning strategy according to the actual operation results.

2. The on-vehicle power-train remote monitoring method according to claim 1, wherein The multi-modal data acquisition framework includes: Synchronously acquiring multi-source heterogeneous data of the powertrain through an in-vehicle sensor network; performing time-frequency domain decomposition on the vibration spectrum by using wavelet transform to extract high-frequency anomaly features; encapsulating the gearbox gear shift sequence data through the CAN bus protocol; using a Kalman filter to suppress the noise of the vehicle speed and acceleration data.

3. The on-vehicle powertrain remote monitoring method according to claim 1, characterized in that The specific construction of the anomaly detection model driven by federated learning includes: Deploying a lightweight convolutional neural network in the edge computing node to extract the spatio-temporal features of the local vibration spectrum; Designing a differential privacy mechanism to add Gaussian noise to the local model parameters to ensure data privacy; Updating the global model by using an adaptive weighted aggregation algorithm, and its weight allocation formula is: Among them, w g is the global model parameter, w i is the local model parameter of the i-th vehicle, and α i is the dynamic weight coefficient calculated according to data quality and anomaly detection confidence.

4. A vehicle power-train remote monitoring method according to claim 1, characterized in that, The control strategy optimized by reinforcement learning includes: Constructing a Markov decision process model, defining the state space as the set of real-time operating parameters of the powertrain, the action space as the set of control instructions, and the reward function based on the powertrain efficiency improvement rate and anomaly suppression effect; Using the deep deterministic policy gradient algorithm to optimize the control strategy and outputting continuous control instructions through the Actor-Critic network.

5. A method for remote monitoring of a vehicle powertrain according to claim 1, characterized in that, The distributed feedback calibration includes: Deploying an incremental learning module in the in-vehicle terminal to update the local model parameters according to the real-time control effect; Designing a sliding window mechanism to screen historical data, and only retaining the data with a correlation higher than the threshold with the current working condition for model calibration.

6. The on-vehicle powertrain remote monitoring method according to claim 1, wherein The feature extraction also includes: Constructing a Gramian angular field for the engine speed and torque data to convert the time series data into two-dimensional image features; Using principal component analysis to reduce the dimension of the multi-dimensional environmental data and extract the key influencing factors.

7. A vehicle powertrain remote monitoring method according to claim 3, characterized in that The calculation formula of the dynamic weight coefficient is: Among them, C i is the anomaly detection confidence of the i-th vehicle, and C j is the anomaly detection confidence of the j-th vehicle. β is the temperature coefficient, which is used to adjust the smoothness of the weight distribution.

8. A method for remote monitoring of a vehicle powertrain according to claim 4, characterized in that The design of the reward function includes: R(t) = λ1·η(t) - λ2·ΔE(t) + λ3·S(t) Wherein, R(t) is the reward value, η(t) is the powertrain efficiency, ΔE(t) is the change in the anomaly index, S(t) is the smoothness of the control command, and λ1, λ2, and λ3 are balance coefficients.

9. A method for remote monitoring of a vehicle powertrain according to claim 1, characterized in that, The method further includes: Constructing a virtual simulation environment in the cloud server to simulate the response of the powertrain under extreme conditions; Fusing simulation data and real data through digital twin technology for adversarial training of the federated learning model.

10. A vehicle powertrain remote monitoring system for implementing the method according to any one of claims 1-9, characterized in that, It includes: An in-vehicle terminal integrating a multi-modal sensor module, an edge computing unit, and a communication module; the sensor module is used to collect powertrain operation data, environmental data, and vehicle status data; The edge computing unit is used for data preprocessing and local model inference; A cloud server deploying a federated learning global model, a reinforcement learning policy library, and a virtual simulation platform; A distributed database storing historical operation data and model parameters of multiple vehicles; A control execution module embedded in the vehicle ECU for receiving and executing adaptive control commands; The in-vehicle terminal, the cloud server, and the control execution module achieve real-time data interaction through the 5G-V2X communication protocol.

Images