A vehicle load estimation method and system based on a parallel hybrid architecture

Abstract

The application provides a vehicle load estimation method and system based on a parallel hybrid architecture, and belongs to the technical field of vehicle load regulation and control; the method comprises the following steps: acquiring vehicle composite data in real time based on a multi-modal sensor and performing pretreatment; inputting the pretreated vehicle composite data into a parallel hybrid architecture to predict a load motion state and a load compensation value; adopting an improved EEMD algorithm to separate the predicted value of the parallel hybrid architecture; fusing the separation results based on a federal Kalman filter to obtain an estimated value of the vehicle load; verifying the estimated value based on a three-mode redundant architecture to determine whether to trigger a vehicle load fault recovery process; and the application can realize high-precision estimation under static conditions and adaptive optimization under dynamic conditions while reducing costs and hardware dependence by means of multi-source data fusion and parallel hybrid architecture modeling, and by combining physical equations with machine learning.

Description

Technical Field

[0001] This invention belongs to the field of vehicle load control technology, and particularly relates to a vehicle load estimation method and system based on a parallel hybrid architecture. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Virtual sensor technology has moved from theoretical research to commercial application, particularly in high-end passenger vehicles equipped with air suspension. However, it is still in a period of optimization and expansion. In commercial vehicle air suspension scenarios, existing virtual load estimation technologies still face some unresolved technical challenges, such as:

[0004] (1) Insufficient adaptability under heavy load conditions: Most existing commercial vehicle solutions are based on pressure sensors or height sensors. In this way, if the load is large (such as a passenger car with a load of more than 10 tons), the static error will increase sharply when going over speed bumps or bumpy roads, and the error in estimating the center of gravity shift will also increase significantly when cornering dynamically. This cannot meet the requirements of commercial vehicle load monitoring and suspension stability.

[0005] (2) Difficulty in decoupling multi-axle loads: The calculation of axle load distribution for commercial vehicles with three or more axles depends on complex mechanical models. Existing algorithms based on two-degree-of-freedom suspension models are difficult to accurately decouple the dynamic loads of each axle, resulting in excessive deviations in the estimation of axle load transfer during braking.

[0006] (3) Poor robustness under harsh working conditions: Commercial vehicles often travel on unpaved roads, and the existing dynamic compensation algorithms (such as wavelet transform) have high static load separation errors under high frequency vibration, which makes it impossible to effectively suppress interference in scenarios such as gravel roads and washboard roads. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides a vehicle load estimation method and system based on a parallel hybrid architecture. By combining multi-source data fusion and parallel hybrid architecture modeling, physical equations and machine learning are combined, which can reduce costs and hardware dependence while achieving high-precision estimation under static conditions and adaptive optimization under dynamic conditions.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0009] The first aspect of this invention provides a vehicle load estimation method based on a parallel hybrid architecture.

[0010] A vehicle load estimation method based on a parallel hybrid architecture includes:

[0011] Real-time acquisition of vehicle composite data based on multimodal sensors;

[0012] Perform preprocessing operations on the obtained vehicle composite data;

[0013] The preprocessed vehicle composite data is input into a parallel hybrid architecture to predict the load motion state and load compensation value; wherein, the parallel hybrid architecture includes a 7-DOF vehicle dynamics model with temperature compensation term and a data-driven model using a lightweight LSTM network;

[0014] An improved EEMD algorithm is used to dynamically separate the predicted values ​​of the parallel hybrid architecture, and the static load and dynamic load are separated by adjusting the amplitude of adaptive white noise.

[0015] The vehicle load is estimated by fusing the separation results of static and dynamic loads using federated Kalman filtering. The estimated value is then verified using a three-modal redundancy architecture, and the vehicle load fault recovery process is determined based on the verification results.

[0016] Furthermore, an improved EEMD algorithm is used to dynamically separate the predicted values ​​of the parallel hybrid architecture. The static and dynamic loads are separated by adaptive white noise amplitude adjustment, including: first, injecting adaptive amplitude white noise into the original signal; then, extracting the intrinsic mode function through Hilbert transform, and using kurtosis index to screen out effective IMF components to reconstruct the static load, so as to achieve dynamic load separation.

[0017] Furthermore, preprocessing operations are performed on the vehicle composite data, including sequentially performing moving mean filtering, temperature compensation, wavelet denoising, and time alignment operations on the vehicle composite data.

[0018] Furthermore, the 7-DOF vehicle dynamics model includes: the vertical, pitch, and roll motions of the sprung mass, the vertical motions of four unsprung masses, and the tire stiffness and damping characteristic equations; wherein, the model parameters of the 7-DOF vehicle dynamics model are updated through a real-time parameter identification algorithm.

[0019] Furthermore, the structure of the data-driven model using the lightweight LSTM network is as follows: 10 nodes in the input layer, 8 nodes in the single hidden layer, and 1 node in the output layer. The weights are solidified into DSRAM after INT8 quantization. Among them, the 10 nodes in the input layer include pressure gradient, height change rate, and frequency domain features.

[0020] Furthermore, based on the fusion of static and dynamic load separation results using federated Kalman filtering, the following steps are taken: First, local filters are used to process the pressure sensor subgroup, IMU subgroup, and altitude sensor subgroup respectively; then, a main filter is used to fuse the local results, and a sensor confidence weight factor is introduced into the covariance matrix update formula of the main filter.

[0021] Furthermore, the three-mode redundancy architecture includes a main channel, a monitoring channel, and a protection channel; wherein, the main channel is used to perform vehicle load estimation, the monitoring channel is used to run a 7-DOF vehicle dynamics model, and the protection channel is used to implement rule checker logic. When the deviation of the results of the three channels exceeds the set threshold, the fault recovery process is triggered.

[0022] A second aspect of the present invention provides a vehicle load estimation system based on a parallel hybrid architecture.

[0023] A vehicle load estimation system based on a parallel hybrid architecture includes:

[0024] The data acquisition module is configured to acquire vehicle composite data in real time based on multimodal sensors;

[0025] The preprocessing module is configured to perform preprocessing operations on the obtained vehicle composite data;

[0026] The parallel hybrid architecture module is configured to input preprocessed vehicle composite data into the parallel hybrid architecture to predict load motion state and load compensation value; wherein, the parallel hybrid architecture includes a 7-DOF vehicle dynamics model with temperature compensation term and a data-driven model using a lightweight LSTM network;

[0027] The dynamic load separation module is configured to: use an improved EEMD algorithm to perform dynamic load separation on the predicted values ​​of the parallel hybrid architecture, and separate static load and dynamic load by adjusting the amplitude of adaptive white noise;

[0028] The vehicle load estimation module is configured to: obtain an estimated value of the vehicle load by fusing the separation results of static load and dynamic load based on federated Kalman filtering; verify the obtained estimated value based on a three-modal redundancy architecture; and determine whether to trigger the vehicle load fault recovery process based on the verification results.

[0029] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a vehicle load estimation method based on a parallel hybrid architecture as described in the first aspect of the present invention.

[0030] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a vehicle load estimation method based on a parallel hybrid architecture as described in the first aspect of the present invention.

[0031] The above one or more technical solutions have the following beneficial effects:

[0032] (1) This invention inputs preprocessed vehicle composite data into a parallel hybrid architecture to predict load motion state and load compensation value. The parallel hybrid architecture includes a 7-DOF vehicle dynamics model with temperature compensation term and a data-driven model using a lightweight LSTM network. The 7-DOF model introduces a real-time identification algorithm for axle load distribution parameters, which can automatically adjust the stiffness matrix of each axle according to the load, solving the mechanical imbalance problem of the traditional 2-DOF model under heavy load. In the parallel hybrid architecture, the physical model and the lightweight LSTM network are operated in parallel. The LSTM can learn the nonlinear load distribution characteristics under heavy load to compensate for the error of the physical model under extreme load. At the same time, the multi-sensor fusion algorithm can automatically increase the weight of the pressure sensor under heavy load and suppress the sensor noise amplification effect caused by heavy load through federated Kalman filtering. Therefore, compared with the prior art, even under heavy load, whether facing speed bumps, bumpy roads, or sharp bends, it can maintain smaller static error and dynamic offset error, thereby meeting the requirements of commercial vehicle load supervision and suspension stability.

[0033] (2) This invention calculates the vertical displacement of sprung mass and the motion state of unsprung mass based on a 7-DOF vehicle dynamics model. The model has a built-in multi-axis dynamic coupling term. By solving the cross-influence of acceleration and suspension displacement of each axle in real time, accurate dynamic axle load distribution can be achieved. At the same time, federated Kalman filtering is introduced into the axle load decoupling state vector, and the load transfer of each axle is corrected in real time using the three-axis angular velocity data of the IMU, which can reduce the delay in axle load transfer estimation during braking. In this way, the problem of excessive deviation in axle load transfer estimation during braking due to the difficulty in accurately decoupling the dynamic load of each axle can be avoided.

[0034] (3) The present invention uses an improved EEMD algorithm to dynamically separate the predicted values ​​of the parallel hybrid architecture. The EEMD algorithm can effectively separate high-frequency vibration noise from static load through multiple integrated empirical mode decompositions, and then select the low-frequency components of IMF3-IMF4 to reconstruct the static load to suppress the interference of unpaved roads. At the same time, the sensor priority strategy can automatically increase the weight of the IMU in harsh working conditions and combine the vehicle acceleration signal to compensate for the load fluctuation caused by road excitation in real time. In addition, the lightweight LSTM network input contains 10-dimensional vibration features (such as kurtosis and root mean square value), which can effectively learn the load fluctuation patterns corresponding to different road textures. Compared with the existing technology, it can reduce the dynamic error in extreme scenarios such as washboard roads.

[0035] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0036] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0037] Figure 1 This is a flowchart of a vehicle load estimation method based on a parallel hybrid architecture according to Embodiment 1 of the present invention.

[0038] Figure 2 This is a flowchart of the preprocessing operation performed in Embodiment 1 of the present invention.

[0039] Figure 3 This is a schematic diagram of the calibration process between the calibration platform, ECU, and hydraulic press in Embodiment 1 of the present invention.

[0040] Figure 4 This is a fault recovery flowchart for triggering vehicle load fault recovery in Embodiment 1 of the present invention. Detailed Implementation

[0041] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0042] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0043] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0044] Example 1

[0045] This embodiment discloses a vehicle load estimation method based on a parallel hybrid architecture.

[0046] like Figure 1 As shown, a vehicle load estimation method based on a parallel hybrid architecture includes:

[0047] Step S1: Acquire vehicle composite data in real time based on multimodal sensors;

[0048] Step S2: Perform preprocessing operations on the obtained vehicle composite data;

[0049] Step S3: Input the preprocessed vehicle composite data into the parallel hybrid architecture to predict the load motion state and load compensation value; wherein, the parallel hybrid architecture includes a 7-DOF vehicle dynamics model with temperature compensation term and a data-driven model using a lightweight LSTM network.

[0050] Step S4: The improved EEMD algorithm is used to perform dynamic load separation on the predicted values ​​of the parallel hybrid architecture, and the static load and dynamic load are separated by adjusting the amplitude of adaptive white noise.

[0051] Step S5: Based on the separation results of static load and dynamic load fusion using federated Kalman filtering, obtain an estimated value of vehicle load; verify the obtained estimated value based on a three-modal redundancy architecture, and determine whether to trigger the vehicle load fault recovery process based on the verification results.

[0052] Based on the above process, this invention combines physical equations with machine learning through multi-source data fusion and parallel hybrid architecture modeling. This enables high-precision estimation under static conditions and adaptive optimization under dynamic conditions while reducing costs and hardware dependence. To facilitate understanding of the technical solution of this invention, the specific implementation methods are further explained and described below.

[0053] In step S1, vehicle composite data is acquired in real time based on multimodal sensors.

[0054] Real-time data such as air spring pressure, suspension height, acceleration, and vehicle speed are collected in real time by multimodal sensors to form composite vehicle data. Specifically, the multimodal sensors include a basic sensor network and extended signal sources. The basic sensor network includes a high-linearity pressure sensor, a non-contact magnetoresistive height sensor, and a 6-axis MEMS IMU. The high-linearity pressure sensor can monitor the internal pressure of the air spring in real time and calculate the vehicle body load. It can be directly integrated into the five-way valve or at the air spring chamber interface. The non-contact magnetoresistive height sensor can measure the relative height between the vehicle body and the wheels with high precision. It can be installed between the suspension control arms and the vehicle body and connected by a linkage mechanism. The 6-axis MEMS IMU can monitor the vehicle body's three-axis acceleration (X / Y / Z) and three-axis angular velocity (pitch / yaw / roll) to dynamically identify the vehicle's motion state and provide instantaneous feedback for suspension adjustment. It can be installed near the vehicle's center of gravity to minimize measurement errors. The extended signal sources are used to collect and reuse CDC shock absorber current signals and analyze CAN bus vehicle speed and steering angle data.

[0055] In step S2, preprocessing operations are performed on the obtained vehicle composite data.

[0056] like Figure 2 As shown, preprocessing operations are performed on the vehicle composite data, including: sequentially performing moving mean filtering, temperature compensation, wavelet denoising, and time alignment operations on the vehicle composite data. Specifically:

[0057] 1) Apply moving mean filtering to the composite vehicle data to eliminate high-frequency noise. Specifically:

[0058] First, missing values ​​are processed by aligning the timestamps of multiple sensors; and data scales are normalized (e.g., pressure in bars, height in mm, acceleration in g) for pre-filtering. Then, a sliding window design is implemented, and the mean is calculated point-by-point by iteratively updating the window data to achieve filtering and eliminate high-frequency noise. It should be noted that the sliding mean filtering method used is sufficient to eliminate high-frequency noise; this embodiment does not impose specific limitations on the choice of implementation method.

[0059] 2) Perform temperature compensation on the vehicle composite data to correct the inherent temperature drift error of the sensors, i.e.:

[0060] ;

[0061] in, This indicates the pressure value after temperature compensation. Indicates the temperature coefficient; Indicates the original pressure. Indicates the current temperature. Indicates the reference temperature.

[0062] Based on this, wavelet denoising is performed on the vehicle composite data to preserve effective signal features; and time alignment is also performed on the vehicle composite data. It should be noted that wavelet denoising and time alignment are only for providing more accurate and analyzable data for the parallel hybrid architecture. Existing similar technologies can also be used for denoising and time alignment operations. Therefore, this embodiment does not impose specific limitations on the actual implementation method.

[0063] In step S3, the preprocessed vehicle composite data is input into a parallel hybrid architecture to predict the load motion state and load compensation value. The parallel hybrid architecture includes a 7-DOF vehicle dynamics model with temperature compensation and a data-driven model using a lightweight LSTM network. Specifically, this can be achieved through the following methods:

[0064] 1) 7-DOF vehicle dynamics model.

[0065] A 7-DOF vehicle dynamics model is used for vehicle attitude processing, and a real-time parameter identification algorithm is added to improve the accuracy and precision of vehicle processing data. Specifically, the 7-DOF vehicle dynamics model includes: vertical, pitch, and roll motions of sprung mass, vertical motions of four unsprung mass, and tire stiffness and damping characteristic equations; the model parameters of the 7-DOF vehicle dynamics model are updated through a real-time parameter identification algorithm.

[0066] Furthermore, the 7-DOF vehicle dynamics model employed includes: vertical motion of the vehicle body. Body roll movement Vehicle body pitch motion Longitudinal movement Lateral movement Lateral motion and wheel rotation movement When processing vehicle attitude using this model, the specific steps include: establishing the equations of motion for each degree of freedom based on Newton's laws of motion and the theorem of angular momentum, such as the vertical dynamic equation of the vehicle body:

[0067] ;

[0068] in, For vehicle body weight, The vertical forces of each suspension are calculated; combined with real-time data such as wheel speed, acceleration, and steering angle collected by sensors, the motion equations are solved to obtain vehicle attitude parameters (such as roll angle, pitch angle, vertical displacement, etc.), providing an attitude reference for load estimation.

[0069] The Recursive Least Squares (RLS) algorithm is used for real-time parameter identification, and key vehicle parameters (such as suspension stiffness, damping coefficient, and tire stiffness) are updated online. The deviation between the attitude parameters predicted by the model and the attitude parameters measured by the sensors is used as input. The parameter values ​​are iteratively corrected through the RLS algorithm to minimize the deviation. The parameter update period is set (e.g., 10ms) to ensure that the parameters can be adapted in real time when the vehicle load changes (e.g., passengers getting on and off, cargo being added or removed) or the working conditions are switched (e.g., asphalt road / gravel road).

[0070] Compared to low-DOF models (such as 3DOF models), 7DOF models can more comprehensively reflect the dynamic characteristics of vehicles and reduce attitude calculation errors; real-time parameter identification can eliminate parameter drift caused by vehicle aging and load changes, ensuring that the model always matches the actual vehicle state and improving the accuracy and precision of subsequent load estimation.

[0071] 2) A data-driven model using a lightweight LSTM network is adopted.

[0072] The micro LSTM network used in this invention is a lightweight model, employing a single-layer LSTM architecture. The input feature dimensions include sensor data such as wheel height and acceleration, and the projected output dimension (directly corresponding to the predicted wheel load values). The input data dimension is represented as [10, 1, 8], where 10 is the time step (10 frames of historical data), 1 is the batch size (batch_size=1, suitable for real-time processing), and 8 represents the number of features per frame. Forward propagation only outputs the result of the last time step, discarding hidden states. The output shape is represented as [1, 4], corresponding to the real-time load estimation for the four wheels. This lightweight characteristic results in only about 500 parameters, optimized for embedded deployment, and provides direct end-to-end output.

[0073] The data-driven model based on a lightweight LSTM network has the following structure: a 10-node input layer, an 8-node single hidden layer, and a 1-node output layer. The weights are quantized using INT8 and then stored in DSRAM. The 10 nodes in the input layer contain pressure gradient, height change rate, and frequency domain features. Furthermore, the lightweight model's micro-LSTM network is an optimization of the traditional LSTM structure, specifically including: 1) Simplified network structure: reducing the number of hidden layers (to only 1 layer), reducing the number of neurons in the hidden layer (from the usual 64-128 to 16-32), and removing redundant gating connections; 2) Quantization compression: using 8-bit integer quantization instead of 32-bit floating-point operations to reduce computational complexity; 3) Input feature filtering: retaining only sensor data strongly correlated with the load (such as pressure sensor output, vertical acceleration, and vehicle height) to reduce input dimensionality. Other aspects remain unchanged compared to the traditional LSTM structure.

[0074] This invention employs a lightweight LSTM network-based data-driven model, reducing the demand for computing resources and enabling real-time operation on automotive embedded chips (such as MCUs), thus meeting the real-time requirements for vehicle dynamic load estimation. Notably, while the data-driven model is lightweight, it retains the advantages of LSTM in modeling time-series data (such as dynamic load changes on continuously bumpy roads), making it more suitable for handling time-varying load characteristics compared to traditional neural networks (such as CNNs). Furthermore, it is adaptable to the low-power environment of automotive systems, avoiding overheating issues caused by excessive computing power.

[0075] In step S4, the improved EEMD algorithm is used to dynamically separate the predicted values ​​of the parallel hybrid architecture, and the static load and dynamic load are separated by adjusting the amplitude of adaptive white noise.

[0076] To better achieve dynamic load separation, an improved EEMD algorithm is used to perform dynamic load separation on the predicted values ​​of the parallel hybrid architecture, that is, to separate static load and dynamic load by adjusting the amplitude of adaptive white noise.

[0077] The improved EEMD algorithm, used to separate static and dynamic components from vehicle acceleration signals, can be summarized in its basic implementation process as follows:

[0078] First, adaptive amplitude white noise is injected into the original signal, that is:

[0079] ;

[0080] in, This indicates the amplitude of the added white noise. Indicates the adaptive coefficient for operating conditions. It represents the standard deviation of the signal.

[0081] Subsequently, the intrinsic mode functions are extracted by Hilbert transform, and the effective IMF components are screened using the kurtosis index to reconstruct the static load, thereby achieving dynamic load separation.

[0082] As an optional embodiment, in the specific implementation process, the original acceleration signal can first be subjected to 100 Gaussian noise perturbations (noise standard deviation 0.1); and after each perturbation, EMD decomposition is performed to obtain multiple sets of IMF components; then, the 3rd and 4th order IMF components are selected and weighted in a 7:3 ratio to form the static components; finally, the static components are subtracted from the original signal to obtain the dynamic components. Thus, this invention improves stability through multiple noise-added decompositions (traditional EMD performs a single decomposition), intelligently selects specific IMF combinations (3+4 order) instead of simply using all IMFs; and directly separates the components using arithmetic operations, avoiding complex post-processing; in the obtained output results, the static components can reflect low-frequency information such as vehicle weight distribution, while the dynamic components contain high-frequency information such as road vibration.

[0083] In step S5, the vehicle load is estimated by fusing the separation results of static and dynamic loads using a federated Kalman filter. The estimated value is then verified using a three-modal redundancy architecture, and the vehicle load fault recovery process is determined based on the verification results. This can be achieved through the following methods:

[0084] Step S5-1: Based on the separation results of static load and dynamic load through federated Kalman filtering, obtain the estimated value of vehicle load.

[0085] Federated Kalman filtering is used as the multi-sensor fusion algorithm. Specifically, it takes the local state estimates x_local and covariance matrix p_local of each sensor as input, and outputs the fused global state estimate x_est and covariance matrix p_fused. The intermediate fusion process uses a weighted fusion based on the information matrix (the inverse of the covariance matrix). First, the inverses of the local covariance matrices are calculated and summed, then the sum is inverted again to obtain the fused covariance matrix. Simultaneously, the state estimate fusion employs an information-weighted average, weighting the local state estimates using the information matrices of each sensor, and finally normalizing using the fused information matrix.

[0086] Based on the fusion of static and dynamic load separation results using federated Kalman filtering, the process includes: first, using local filters to process the pressure sensor subgroup, IMU subgroup, and altitude sensor subgroup separately; then, using a master filter to fuse the local results, and introducing a sensor confidence weight factor into the covariance matrix update formula of the master filter. Specifically:

[0087] 1) Local filters process each sensor subgroup.

[0088] Pressure sensor subgroup: Static loads are processed using extended Kalman filter (EKF). The tire ground pressure collected by the pressure sensor is used as input, and combined with the vehicle's attitude parameters (roll angle, pitch angle) when stationary, the static load (such as the static distribution of occupants / cargo) is estimated.

[0089] IMU Subgroup: The dynamic load is processed using unscented Kalman filtering (UKF). Based on the triaxial acceleration and angular velocity collected by the IMU, the dynamic load transfer during vehicle acceleration, braking and steering is calculated.

[0090] Height sensor subgroup: Kalman filter (KF) is used to assist in dynamic load estimation. The suspension deformation is inferred from the change in vehicle height to correct the amplitude of dynamic load.

[0091] 2) The main filter fuses the local results. It receives the outputs of three local filters (static load estimate, dynamic load estimate, and correction value) to construct the fused state equation. The covariance matrix update formula introduces the sensor confidence weight factor; the specific formula is as follows:

[0092] ;

[0093] in, The covariance matrix updated by the master filter. Let be the covariance matrix of the local filter; This represents the sensor confidence weighting factor. These correspond to three subgroups respectively.

[0094] Confidence weighting factor Calculations: Based on consistency checks of sensor data (e.g., residual analysis) and environmental adaptability (e.g., IMU on bumpy roads). Lower, pressure sensor on smooth road surface (Raise), dynamically adjust and .

[0095] Based on the above design, separating local filtering from main filtering reduces the impact of individual sensor failures on the overall results and improves system fault tolerance. The confidence weighting factor makes the fusion process more "intelligent," prioritizing the trust of data from highly reliable sensors to reduce errors caused by noise or interference (such as reduced weight of pressure sensors due to temperature effects), thus improving fusion accuracy. Simultaneously, it achieves effective separation and fusion of static and dynamic loads, solving the problem that a single sensor cannot accurately capture both static (e.g., load during parking) and dynamic (e.g., load transfer during rapid acceleration) characteristics simultaneously.

[0096] Step S5-2: Verify the obtained estimated value based on the three-mode redundancy architecture, and determine whether to trigger the vehicle load fault recovery process based on the verification results.

[0097] The architecture principle of the triple redundancy architecture is as follows: 1) Triple redundancy design: The main execution channel undertakes the main control functions; the verification channel runs the same algorithm independently; and the arbitration channel makes the final output decision. 2) Security monitoring mechanism: Faults are detected by comparing the outputs of the main and backup channels, and a "2-out-of-2" voting logic is adopted.

[0098] The key component functions of the triple redundancy architecture are: 1) Deviation detection unit, which calculates the absolute difference between the primary and backup channels in real time; 2) Safety monitoring mechanism: determines whether to start fault response by judging threshold, activates safety mode, and records fault codes and actual deviation values ​​simultaneously; 3) Implicit arbitration logic, which trusts the output of the primary channel by default and switches to safety state when a fault occurs.

[0099] Furthermore, the three-mode redundancy architecture includes a main channel, a monitoring channel, and a protection channel. The main channel is used to estimate the vehicle load. The monitoring channel uses a simplified 2-DOF suspension model to run a 7-DOF vehicle dynamics model. The protection channel is used to preset the load change rate threshold and implement rule checker logic. When the deviation of the results from the three channels exceeds the preset threshold (load change rate threshold), the fault recovery process is triggered.

[0100] The fault recovery process includes: switching to a backup model when a sensor fails, and initiating a memory double-buffering strategy when computing resources are abnormal. Specifically, for example... Figure 4 As shown, during the fault monitoring phase, the main algorithm output value is temporarily stored in a buffer, while the monitoring module simultaneously requests a verification value from the simplified model and compares the main output with the simplified model's verification value. During the fault determination phase, when a deviation is detected, the fault type (numerical out-of-bounds, logical anomalies, etc.) is analyzed, and the specific fault mode is identified through type matching. During the recovery execution phase, the safety module initiates a three-level recovery, including: 1) soft reset: clearing the algorithm's abnormal state; 2) data rollback: restoring to the previous valid state; 3) mode switching, downgrading to a safe operating mode. During the recovery confirmation phase, continuous monitoring is performed for 3-5 cycles to confirm normal recovery, and a completion signal is sent via the CAN bus. The entire process can be completed within 10ms, achieving an instantaneous fault recovery rate of over 99% through output buffering, dual verification mechanisms, and tiered recovery design. Specifically:

[0101] 1) Implementation of Tri-Mode Redundancy Architecture. The tri-mode redundancy architecture calculates vehicle load estimates through three independent paths, specifically including:

[0102] Path 1: Analytical calculation based on the 7-DOF model and physical formulas;

[0103] Path 2: Data-driven prediction based on micro LSTM networks;

[0104] Path 3: Fusion results based on federated Kalman filtering (i.e., the output of step S3 above).

[0105] Verification process: Set the deviation threshold to 5% of the estimated load value, and calculate the pairwise deviations of the three path results: , , "Pairwise deviations" refer to the estimation deviations between path 1 and path 2, path 1 and path 3, and path 2 and path 3. , , These are weighting factors corresponding to the pairwise deviations of the three groups, used to quantify the influence of different deviations on the overall verification results. The weight values ​​are based on the inherent accuracy, dynamic response characteristics, and reliability under the current operating conditions of each sensor subgroup: for example, if the pressure sensor has higher accuracy under low-speed static conditions, then the weight corresponding to the deviation it participates in (such as path 1 and path 2) is... A larger value can be set to highlight the impact of this deviation on the verification results; however, under high-speed dynamic conditions, the reliability of the IMU subgroup improves, corresponding to the weight of the deviation (e.g., Adaptability adjustment. In this embodiment, if all deviations are less than or equal to the deviation threshold, the estimated value is deemed valid and the fusion result is output; if any deviation is greater than the deviation threshold, the vehicle load fault recovery process is triggered.

[0106] 2) The fault recovery process includes:

[0107] Level 1 recovery: Activate backup sensors (such as redundant pressure sensors) to re-acquire data and repeat the calculation of the three paths;

[0108] Secondary recovery: If primary recovery fails, switch to a simplified estimation model (such as dual-sensor fusion based only on pressure sensor and IMU) to ensure basic load estimation function;

[0109] Warning message: A fault code is sent to the vehicle controller, and the driver is prompted on the instrument panel that "the load estimation system is operating in a degraded manner," and it is recommended to have it inspected.

[0110] Based on the above design, the three-mode redundancy architecture can significantly reduce the risk of incorrect estimation caused by the failure of a single algorithm or sensor through multi-path verification, thereby improving system reliability. At the same time, the hierarchical fault recovery process ensures that the system can still maintain basic functions when some components fail, avoiding the impact on vehicle safety due to complete failure (such as overload detection, suspension control and other functions that rely on load data).

[0111] The vehicle load estimation method based on a parallel hybrid architecture provided by this invention also includes an online self-calibration mechanism, namely:

[0112] a) The calibration process is triggered every 50,000 kilometers of accumulated driving, such as Figure 3 As shown, during the initialization phase, the ECU sends a zero-point calibration command and confirms that the sensor is fault-free and communication is normal. Subsequently, the hydraulic press applies a 20% reference load (establishing an initial reference point). During the data acquisition phase, the sensor begins to upload real-time data, and the hydraulic press is gradually loaded in five gradients, recording data after each load level remains stable. During the parameter calibration phase, the system automatically identifies feature points, calculates calibration coefficients, and writes them into the ECU's non-volatile memory. During the verification phase, the hydraulic press is unloaded to zero, and the ECU verifies the validity of the calibration parameters.

[0113] b) Under zero-speed conditions on a level road surface, compare the current estimated load value with the baseline value;

[0114] c) Update the spring stiffness parameter matrix using the least squares method. .

[0115] Example 2

[0116] This embodiment discloses a vehicle load estimation system based on a parallel hybrid architecture.

[0117] A vehicle load estimation system based on a parallel hybrid architecture includes:

[0118] The data acquisition module is configured to acquire vehicle composite data in real time based on multimodal sensors;

[0119] The preprocessing module is configured to perform preprocessing operations on the obtained vehicle composite data;

[0120] The parallel hybrid architecture module is configured to input preprocessed vehicle composite data into the parallel hybrid architecture to predict load motion state and load compensation value; wherein, the parallel hybrid architecture includes a 7-DOF vehicle dynamics model with temperature compensation term and a data-driven model using a lightweight LSTM network;

[0121] The dynamic load separation module is configured to: use an improved EEMD algorithm to perform dynamic load separation on the predicted values ​​of the parallel hybrid architecture, and separate static load and dynamic load by adjusting the amplitude of adaptive white noise;

[0122] The vehicle load estimation module is configured to: obtain an estimated value of the vehicle load by fusing the separation results of static load and dynamic load based on federated Kalman filtering; verify the obtained estimated value based on a three-modal redundancy architecture; and determine whether to trigger the vehicle load fault recovery process based on the verification results.

[0123] Example 3

[0124] The purpose of this embodiment is to provide a computer-readable storage medium.

[0125] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a vehicle load estimation method based on a parallel hybrid architecture as described in Embodiment 1 of this disclosure.

[0126] Example 4

[0127] The purpose of this embodiment is to provide an electronic device.

[0128] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a vehicle load estimation method based on a parallel hybrid architecture as described in Embodiment 1 of this disclosure.

[0129] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0130] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0131] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A vehicle load estimation method based on a parallel hybrid architecture, characterized in that, include: Real-time acquisition of vehicle composite data based on multimodal sensors; Perform preprocessing operations on the obtained vehicle composite data; The preprocessed vehicle composite data is input into a parallel hybrid architecture to predict the load motion state and load compensation value; wherein, the parallel hybrid architecture includes a 7-DOF vehicle dynamics model with temperature compensation term and a data-driven model using a lightweight LSTM network; An improved EEMD algorithm is used to dynamically separate the predicted values ​​of the parallel hybrid architecture. The static and dynamic loads are separated by adaptive white noise amplitude adjustment. The process includes: first, injecting adaptive amplitude white noise into the original signal; then, extracting the intrinsic mode function through Hilbert transform, and using kurtosis index to screen out effective IMF components to reconstruct the static load, thereby achieving dynamic load separation. The vehicle load is estimated by fusing the separation results of static and dynamic loads using federated Kalman filtering. The estimated value is then verified using a three-modal redundancy architecture, and the vehicle load fault recovery process is determined based on the verification results. The three-mode redundancy architecture includes a main channel, a monitoring channel, and a protection channel. The main channel is used to estimate the vehicle load, the monitoring channel is used to run a 7-DOF vehicle dynamics model, and the protection channel is used to implement rule checker logic. When the deviation of the results from the three channels exceeds a set threshold, a fault recovery process is triggered.

2. The vehicle load estimation method based on a parallel hybrid architecture as described in claim 1, characterized in that, Preprocessing operations are performed on the vehicle composite data, including: sequentially performing moving mean filtering, temperature compensation, wavelet denoising, and time alignment operations on the vehicle composite data.

3. The vehicle load estimation method based on a parallel hybrid architecture as described in claim 1, characterized in that, The 7-DOF vehicle dynamics model includes: sprung mass vertical, pitch, and roll motions, four unsprung mass vertical motions, and tire stiffness and damping characteristic equations; wherein, the model parameters of the 7-DOF vehicle dynamics model are updated through a real-time parameter identification algorithm.

4. The vehicle load estimation method based on a parallel hybrid architecture as described in claim 1, characterized in that, The structure of the data-driven model using a lightweight LSTM network is as follows: 10 nodes in the input layer, 8 nodes in the single hidden layer, and 1 node in the output layer. The weights are solidified into DSRAM after INT8 quantization. The 10 nodes in the input layer include pressure gradient, height change rate, and frequency domain features.

5. The vehicle load estimation method based on a parallel hybrid architecture as described in claim 1, characterized in that, Based on the fusion of static and dynamic load separation results using federated Kalman filtering, the process includes: first, using local filters to process the pressure sensor subgroup, IMU subgroup, and altitude sensor subgroup respectively; then, using a master filter to fuse the local results, and introducing a sensor confidence weight factor into the covariance matrix update formula of the master filter.

6. A vehicle load estimation system based on a parallel hybrid architecture, employing a vehicle load estimation method based on a parallel hybrid architecture as described in any one of claims 1-5, characterized in that, include: The data acquisition module is configured to acquire vehicle composite data in real time based on multimodal sensors; The preprocessing module is configured to perform preprocessing operations on the obtained vehicle composite data; The parallel hybrid architecture module is configured to input preprocessed vehicle composite data into the parallel hybrid architecture to predict load motion state and load compensation value; wherein, the parallel hybrid architecture includes a 7-DOF vehicle dynamics model with temperature compensation term and a data-driven model using a lightweight LSTM network; The dynamic load separation module is configured to: use an improved EEMD algorithm to perform dynamic load separation on the predicted values ​​of the parallel hybrid architecture, and separate static load and dynamic load by adjusting the amplitude of adaptive white noise; The vehicle load estimation module is configured to: obtain an estimated value of the vehicle load by fusing the separation results of static load and dynamic load based on federated Kalman filtering; verify the obtained estimated value based on a three-modal redundancy architecture; and determine whether to trigger the vehicle load fault recovery process based on the verification results.

7. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps in the vehicle load estimation method based on a parallel hybrid architecture as described in any one of claims 1-5.

8. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the vehicle load estimation method based on a parallel hybrid architecture as described in any one of claims 1-5.

Images