A new energy truck quality estimation method, system and device based on multi-source data fusion and a storage medium

Abstract

The application discloses a new energy truck quality estimation method, system and device based on multi-source data fusion and a storage medium, relates to the technical field of vehicle engineering, and comprises the following steps: mapping and aligning dynamic time sequence characteristics and road space environment characteristics, and constructing a multivariate time sequence characteristic vector; calculating a dynamics enhancement characteristic as a mechanism constraint; combining and processing the multivariate time sequence characteristic vector and the dynamics enhancement characteristic to generate an enhanced multivariate time sequence characteristic matrix; and taking the enhanced multivariate time sequence characteristic matrix as input, extracting time sequence long dependence characteristics and multi-scale local detail characteristics of the enhanced multivariate time sequence characteristic matrix; fusing the time sequence long dependence characteristics and the multi-scale local detail characteristics, and solving through regression mapping to output a continuous estimation value of the total mass of the vehicle. The method disclosed by the application does not need to additionally install hardware sensors, can realize high-precision and strong-robust real-time estimation of the total mass of the vehicle under all working conditions, and can be adapted to large-scale commercial operation scenarios of new energy trucks.

Description

Technical Field

[0001] This invention relates to the field of vehicle engineering technology, specifically to a method, system, device, and storage medium for estimating the quality of new energy trucks based on multi-source data fusion. Background Technology

[0002] With the increasing severity of the global energy crisis and environmental pollution, new energy vehicles, especially new energy trucks, are being used more and more widely in the logistics and transportation sector. Compared with traditional fuel trucks, new energy trucks have significant advantages such as zero emissions, low noise, and low operating costs, making them an important choice for urban delivery and short-distance transportation. However, the energy management system of new energy trucks faces unique challenges. The vehicle's load condition is a key factor affecting energy consumption and power performance. Under different load conditions, the optimal control strategy for the vehicle differs significantly. For example, the energy management strategies for empty and fully loaded vehicles should be completely different under conditions such as start-up acceleration, hill climbing, and regenerative braking. Currently, new energy truck mass estimation technologies are mainly divided into three major technical routes.

[0003] First, there's the direct measurement method based on added hardware sensors: This method directly measures load by installing weighing and displacement sensors on the suspension and axles. While this method offers high accuracy, it suffers from drawbacks such as high hardware installation costs, sensor susceptibility to damage from bumpy road conditions, high maintenance costs, and inability to adapt to large-scale commercial vehicle deployment. Second, there's the recursive estimation method based on vehicle dynamics models: This method relies on traditional vehicle longitudinal dynamics equations and uses algorithms such as Kalman filtering and recursive least squares to recursively estimate mass. This method doesn't require additional hardware, but it heavily depends on the accurate calibration of inherent vehicle parameters and ideal driving conditions. In real-world operating scenarios such as complex terrain, gear shifting, acceleration / deceleration, and variable loads, it is greatly affected by external disturbances such as slope, rolling resistance, and air resistance, leading to a significant decrease in estimation accuracy and robustness. Third, there's the data-driven deep learning estimation method: This method uses onboard operating data to learn the mapping relationship between operating condition characteristics and vehicle mass through neural networks. It doesn't require complex physical model calibration, has stronger adaptability to operating conditions, and is currently the mainstream research direction in the industry.

[0004] Previous research indicates that current studies face three major technical bottlenecks: a lack of standardized preprocessing systems for multi-source heterogeneous data, insufficient spatiotemporal matching accuracy, and weak effective sample selection capabilities. Most existing data-driven solutions rely solely on a single data source, the vehicle's CAN bus, neglecting the crucial impact of spatial environmental information such as road elevation, slope, and terrain on quality estimation. Furthermore, for multi-source heterogeneous sensor data from CAN bus, GPS, and other sources, there is a lack of standardized preprocessing procedures encompassing time alignment, downsampling, missing value imputation, noise smoothing, and outlier removal, failing to address core issues such as inconsistent sampling frequencies, asynchronous timestamps, and mismatched spatial coordinates among multi-sensor data. Additionally, existing solutions do not specifically remove strong noise samples that disrupt quality mapping relationships, such as those related to gear shifting, braking, low speed, and acceleration anomalies, leading to models learning incorrect operating condition characteristics. This severely reduces the accuracy and generalization ability of quality estimation; the heterogeneous modal data fusion method is rigid and cannot achieve dynamic weighted fusion that adapts to operating conditions, and the mining of complementary information between modes is insufficient. Existing multi-source data fusion schemes generally adopt fixed splicing and constant weighting fusion methods, which cannot dynamically adjust the weight coefficients of vehicle dynamic temporal features and road spatial environment features according to the real-time driving conditions of new energy trucks such as flat road cruising, climbing, acceleration and deceleration, and cannot mine complementary information between two types of heterogeneous modes. In complex terrain and variable operating condition scenarios, the fixed fusion mode is easily affected by external interference and cannot specifically strengthen the contribution of core effective features, resulting in insufficient feature fusion effectiveness and poor model adaptability to operating conditions; the physical causal relationship between input features and vehicle quality is weak, and there is a lack of feature enhancement mechanism guided by dynamic priors, making model learning difficult and lacking robustness.

[0005] Most existing data-driven methods directly use raw sensor data as model input without combining the longitudinal dynamic equations of new energy trucks to uncover core features that are strongly physically related to the overall vehicle mass. This results in a lack of clear causal relationship between the input features and the estimation target, making neural network learning difficult and prone to overfitting. At the same time, existing solutions cannot enhance the quality representation capability of features through lightweight feature derivation without adding additional sensors. The estimation accuracy and robustness under complex operating conditions cannot meet the needs of commercial deployment. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is that existing methods for estimating the quality of new energy trucks have insufficient spatiotemporal matching accuracy, weak ability to screen effective samples, insufficient mining of complementary information between modes, lack of feature enhancement mechanism guided by prior dynamics, and the problem of how to enhance the model's learning ability and robustness.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a method for estimating the mass of new energy trucks based on multi-source data fusion, comprising: aligning the dynamic temporal features characterizing the vehicle's driving state with the road spatial environment features characterizing the vehicle's current geographical location through temporal and spatial dimension mapping to construct a multivariate temporal feature vector; utilizing vehicle driving state information and inherent vehicle parameters to calculate dynamic enhancement features with strong physical causal correlation to vehicle commands, serving as mechanistic constraints to guide neural network learning; combining and standardizing the multivariate temporal feature vector and dynamic enhancement features to generate an enhanced multivariate temporal feature matrix with strong physical mechanism constraints; using the enhanced multivariate temporal feature matrix as input, employing a regression algorithm with a multi-scale feature extraction mechanism to extract the temporal long-term dependency features of the enhanced multivariate temporal feature matrix in the continuous time dimension and the multi-scale local detail features within the local time window; fusing the temporal long-term dependency features and the multi-scale local detail features, and solving through regression mapping to output a continuous estimate of the vehicle's total mass.

[0009] As a preferred embodiment of the new energy truck quality estimation method based on multi-source data fusion described in this invention, the time and space dimension mapping alignment includes using the timestamp of dynamic time series features as the unique sample benchmark, performing linear interpolation and downsampling processing on the road spatial environment features, so as to perform one-to-one corresponding time and space dimension standardized matching mapping on continuous time steps with the same sampling frequency.

[0010] As a preferred embodiment of the new energy truck quality estimation method based on multi-source data fusion described in this invention, the method further includes: before constructing the multivariate time-series feature vector, dynamic cleaning of non-stationary operating condition samples; real-time acquisition of the vehicle's current gear status, target gear status, brake opening status, current vehicle speed, and longitudinal acceleration; when the current gear status is not equal to the target gear status, or the brake opening status is greater than zero, or the current vehicle speed is less than or equal to a set speed threshold, or the longitudinal acceleration exceeds any one or more of the preset normal driving physical range, the current time step is determined to be a non-stationary noise operating condition, and the corresponding time-series data samples are stripped and removed online.

[0011] As a preferred embodiment of the new energy truck quality estimation method based on multi-source data fusion described in this invention, the road spatial environment features include kinematically corrected road slope and terrain feature vectors. The kinematically corrected road slope is calculated by forward difference based on road elevation data from adjacent sampling points to obtain a basic differential slope value. Real-time longitudinal acceleration data of the vehicle is introduced to dynamically compensate and correct the basic differential slope value, eliminating the slope perception error caused by the vehicle's non-uniform speed. The terrain feature vector is constructed by setting several mutually exclusive absolute slope value intervals to define plain terrain, hilly terrain, and hilly terrain respectively. Based on the interval into which the corrected road slope falls, the terrain category of the current vehicle is determined, and the terrain category is converted into a geometric vector without category numerical weight bias as the terrain feature vector.

[0012] As a preferred embodiment of the new energy truck quality estimation method based on multi-source data fusion described in this invention, the construction of a multivariate temporal feature vector includes: taking dynamic temporal features as a vehicle dynamic feature branch and road spatial environment features as a road environment feature branch; using a nonlinear gated logic unit, predicting in real time step by step the first adaptive weight of the vehicle dynamic feature branch and the second adaptive weight of the road environment feature branch according to the current driving conditions, wherein the algebraic sum of the first adaptive weight and the second adaptive weight is 1; and using the corresponding weights to perform a weighted summation of the features of the two branches step by step to generate an adaptively fused temporal feature vector.

[0013] As a preferred embodiment of the new energy truck mass estimation method based on multi-source data fusion described in this invention, the dynamic enhancement features include: based on the vehicle's longitudinal dynamic equilibrium equation, using real-time motor torque, gear ratio, and final drive ratio parameters to back-calculate the vehicle's temporal driving force characteristics; within time steps where the vehicle's longitudinal acceleration is greater than zero, calculating the ratio of real-time motor torque to longitudinal acceleration to construct torque-acceleration ratio features; calculating the first-order difference of key dynamic temporal parameters between adjacent time steps to construct temporal dynamic difference features; introducing a linear correlation coefficient to perform statistical validity verification on the derived features, eliminating weakly correlated redundant features with absolute values ​​below a set threshold, and forming dynamic enhancement features strongly correlated with the overall vehicle mass.

[0014] As a preferred embodiment of the new energy truck mass estimation method based on multi-source data fusion described in this invention, the extraction of the enhanced multivariate temporal feature matrix in terms of long-term temporal dependency features in the continuous time dimension and multi-scale local detail features in the local time window includes: parallelly connecting the enhanced multivariate temporal feature matrix to a long-term hidden layer feature extraction channel and a multi-scale local window feature extraction channel; using the long-term hidden layer feature extraction channel to capture the macroscopic temporal dynamic dependency trend of the vehicle during the entire process of starting, cruising, or continuous climbing, generating long-term temporal dependency features; simultaneously using the multi-scale local window feature extraction channel, through multiple parallel time-domain sliding windows of different sizes, capturing the microscopic local significant operating condition features caused by rapid terrain switching or sudden changes in driving force, generating multi-scale local detail features; performing global average pooling to eliminate time axis size differences between the long-term temporal dependency features and the multi-scale local detail features, then performing channel aggregation, and directly outputting a continuous estimate of the vehicle's total mass through a regression mapping layer, and using this estimate as a control parameter to input into the truck's vehicle controller in real time to dynamically adjust the vehicle's energy management control strategy.

[0015] Another objective of this invention is to provide a new energy truck quality estimation system based on multi-source data fusion. This system can generate an enhanced multivariate time-series feature matrix with strong physical mechanism constraints by combining and standardizing the multivariate time-series feature vector with the dynamic enhancement features. This solves the problem that current new energy truck quality estimation methods lack a feature enhancement mechanism guided by prior dynamics.

[0016] As a preferred embodiment of the new energy truck quality estimation system based on multi-source data fusion described in this invention, the system includes: a multi-source data spatiotemporal matching module, a non-stationary operating condition sample cleaning module, a gated modal adaptive fusion module, a dynamic physical mechanism enhancement module, and a multi-scale feature regression estimation module. The multi-source data spatiotemporal matching module uses the timestamps of dynamic temporal features as the sole sample benchmark to perform time alignment and downsampling processing on road spatial environment features, enabling a one-to-one spatiotemporal dimension standardized matching mapping between the road spatial environment features and dynamic temporal features at continuous time steps. The non-stationary operating condition sample cleaning module acquires the vehicle's gear shifting status, braking signals, real-time vehicle speed, and longitudinal acceleration in real time. When the real-time status meets the set non-stationary operating condition boundary conditions, it strips and removes the corresponding temporal data samples online to retain effective samples of stable driving that satisfy the causal relationship of the vehicle's longitudinal dynamics. The gated modal adaptive fusion module splits the effective samples into vehicle dynamic feature branches and road environment feature branches, utilizing nonlinear... The gating logic unit dynamically predicts the adaptive weight coefficients of the two branches in a time-step manner, and performs context-aware feature weighted fusion in a time-step manner to generate an adaptively fused multivariate temporal feature vector. The dynamic physics mechanism enhancement module is used to derive driving force features, torque-acceleration ratio features, and temporal dynamic difference features with strong physical causal relationship with the vehicle mass based on the vehicle's longitudinal dynamic equilibrium equation and using the vehicle's inherent physical parameters and effective samples. Redundant features are eliminated through linear correlation analysis to construct an enhanced multivariate temporal feature matrix. The multi-scale feature regression estimation module is used to input the enhanced multivariate temporal feature matrix in parallel into the long-time hidden layer feature extraction channel and the multi-scale local window feature extraction channel. It captures the long-time dependent features during the continuous driving process of the vehicle and the multi-scale local detail features within the local time window, respectively. The two features are then subjected to high-level deep aggregation and regression mapping to solve, and the continuous estimate of the total mass of the vehicle is output to dynamically adjust the energy management control strategy of the vehicle controller online.

[0017] Another object of the present invention is to provide a new energy truck quality estimation device based on multi-source data fusion, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a new energy truck quality estimation method based on multi-source data fusion.

[0018] Another object of the present invention is to provide a storage medium for quality estimation of new energy trucks based on multi-source data fusion, wherein a computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method for quality estimation of new energy trucks based on multi-source data fusion are implemented.

[0019] The beneficial effects of this invention are: This invention provides a new energy truck quality estimation method based on multi-source data fusion, which simultaneously collects onboard CAN bus operation data, GPS positioning data, and DEM digital elevation model data. It synchronously covers two core dimensions: vehicle dynamic time-series data and road spatial environment data. This solves the industry problems of inconsistent sampling frequencies and timestamp asynchronization among multiple sensors. It constructs multi-dimensional data filtering rules to accurately remove invalid samples that disrupt the quality mapping relationship, such as those related to gear shifting, braking, low speed, and acceleration anomalies, retaining only valid data from stable driving conditions, thus improving the quality of input data from the source. A lightweight gated modal adaptive fusion unit is designed, which decomposes temporal features into two independent branches: vehicle dynamic features and road environment features. Through the fully connected layer of the gated network and the sigmoid activation function, it learns the adaptive weights of the two branches under different driving conditions, achieving context-aware feature weighted fusion step-by-step, significantly improving the quality of the vehicle. The model's adaptability and anti-interference capability under complex terrain and varying operating conditions were improved. Based on the core equation of longitudinal dynamics of new energy trucks, a lightweight core feature with a strong physical causal relationship with the vehicle mass was derived without adding any additional hardware sensors. This significantly reduced the learning difficulty of the neural network, effectively suppressed the model overfitting problem, and significantly improved the estimation accuracy and robustness of the mass estimation model under complex operating conditions. A dual-branch parallel fusion end-to-end regression model was built, forming a complete technical loop with data preprocessing and feature engineering. This model accurately captures the long-term dynamic dependencies related to the vehicle mass during vehicle driving, adapts to the feature time-series change patterns under continuous driving conditions, and completes end-to-end estimation of the vehicle mass through a fully connected layer. A two-level fusion system was constructed to fully meet the high-precision estimation requirements of new energy trucks in all terrain conditions, from empty to fully loaded and from plains to hilly areas. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is the main flowchart of multi-source time-series feature fusion and quality estimation of the new energy truck quality estimation method based on multi-source data fusion in this invention.

[0022] Figure 2 This is an overall flowchart of the new energy truck quality estimation method based on multi-source data fusion in this invention.

[0023] Figure 3 This is a schematic diagram of the gated modal adaptive fusion unit structure of the new energy truck quality estimation method based on multi-source data fusion in this invention.

[0024] Figure 4 This is a map showing the coverage of the actual vehicle load conditions for the new energy truck quality estimation method based on multi-source data fusion in this invention.

[0025] Figure 5 This is a schematic diagram illustrating the dynamic temporal characteristics of vehicles under typical operating conditions in the new energy truck quality estimation method based on multi-source data fusion according to the present invention.

[0026] Figure 6 This is a schematic diagram of the overall system for estimating the quality of new energy trucks based on multi-source data fusion, which is the method for estimating the quality of new energy trucks based on multi-source data fusion in this invention. Detailed Implementation

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0029] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0030] The “dynamic timing feature F1” mentioned in this application can be understood as a set of values ​​that characterize the basic physical motion state and power output of the vehicle, which are collected in real time through the vehicle’s CAN bus or sensors during continuous driving and have a strict temporal sequence correlation in continuous time steps. Specifically, it includes the smoothed drive motor torque, current gear status, target gear status, brake opening status, current vehicle speed, and longitudinal acceleration.

[0031] The “road spatial environment feature F2” mentioned in this application can be understood as a collection of external macroscopic geometric terrain and microscopic road slope spatial information of the truck’s current driving position, and a macroscopic road geometric feature obtained by spatial matching of latitude and longitude obtained by the vehicle’s Global Positioning System (GPS) with the Digital Elevation Model (DEM).

[0032] Specifically, it includes two key features: one is the real-time road slope after dynamic inertial compensation correction; the other is the geometric vector obtained by dividing the terrain using a set of mutually exclusive absolute slope value intervals and converting it through one-hot encoding.

[0033] The "dynamic enhancement feature D" mentioned in this application can be understood as a derived feature that is explicitly derived based on the vehicle's longitudinal dynamic balance equation, using the vehicle's inherent physical parameters and basic driving state information, and has a strong physical causal relationship with the overall vehicle mass.

[0034] The "nonlinear gated logic unit" described in this application can be understood as a feedforward neural network layer architecture used to dynamically balance the contribution of bimodal heterogeneous features on the time axis. It consists of two fully connected layers and a Sigmoid activation function cascaded in sequence. It performs a fully connected nonlinear mapping between the vehicle dynamic features and road environment features input at the current sampling step, outputs a scalar in the range (0, 1) as the first adaptive weight, and naturally derives the second adaptive weight through algebraic complement logic, thereby realizing feature-level context-aware fusion step by step.

[0035] Example 1

[0036] Reference Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a new energy truck quality estimation method based on multi-source data fusion, and solves the problems of spatiotemporal mismatch of cross-modal data sources, weak physical causal correlation of pure data-driven models, and the inability of feature extraction networks to take into account both long-term macroscopic laws and local transient changes in traditional quality estimation methods.

[0037] Specifically, the dynamic temporal features F1, representing the vehicle's driving state, and the road spatial environment features F2, representing the vehicle's current geographical location, are mapped and aligned in both time and space dimensions to construct a multivariate temporal feature vector V. The vehicle's dynamic temporal features F1 and the road spatial environment features F2 achieve a one-to-one standardized matching mapping across consecutive time steps through a specific spatiotemporal alignment mechanism, thus forming a multivariate temporal feature vector V that simultaneously reflects both the vehicle's own state and the external spatial environment. Subsequently, the system utilizes vehicle driving state information and inherent vehicle parameters to calculate dynamic enhancement features D, which have a strong physical causal relationship with the vehicle's commands, and uses these features as mechanistic constraints to guide the neural network's learning. These dynamic enhancement features D directly rely on the vehicle's longitudinal dynamics system and are obtained through explicit physical mechanism derivation, which can strengthen the causal relationship between the input data and the final estimation target from a white-box mechanism level.

[0038] Next, the multivariate temporal feature vector V is combined and standardized with the dynamic enhancement feature D to generate an enhanced multivariate temporal feature matrix M with strong physical mechanism constraints. During the combination and standardization process, the system eliminates the influence of different dimensions on model training through normalization or standardization algorithms. The temporal feature vector generated by pure data-driven approach is deeply concatenated with the dynamic enhancement feature D containing physical priors to construct a high-dimensional enhanced multivariate temporal feature matrix M with strong physical mechanism constraints, thus defining clear feature boundaries for subsequent networks.

[0039] Subsequently, the enhanced multivariate temporal feature matrix M is used as input. A regression algorithm based on a multi-scale feature extraction mechanism is employed to extract the temporal long-term dependency feature M1 in the continuous time dimension and the multi-scale local detail feature M2 within the local time window. The multi-scale feature extraction mechanism adopts a parallel dual-channel processing architecture. One channel focuses on capturing the macroscopic temporal dynamic dependency trend of the vehicle during long-term driving to generate the temporal long-term dependency feature M1. The other channel dynamically captures the microscopic local salient features caused by sudden changes in operating conditions through multiple different local time domain sliding windows, thereby generating the multi-scale local detail feature M2. Finally, the temporal long-term dependency feature M1 and the multi-scale local detail feature M2 are fused and solved through a regression mapping layer. The final end-to-end continuous estimate of the total vehicle mass is output. This continuous estimate can then be used as a key control parameter to be fed back to the vehicle controller in real time for online dynamic adjustment of the vehicle's industrial control strategy.

[0040] Example 2

[0041] Reference Figure 1 and Figure 2 This is the second embodiment of the present invention. Unlike the previous embodiment, this embodiment provides a precise spatiotemporal mapping of bimodal heterogeneous data based on linear interpolation and downsampling alignment. This solves the problem of missing spatiotemporal matching caused by different sampling frequencies and low refresh rates in cross-modal data. It also eliminates the problem of non-steady-state mechanical noise such as gear shifting, braking, low speed, and abnormal acceleration disrupting the physical strong causal chain of quality estimation, thus laying the foundation for training the quality assessment model and derivation of high-value features.

[0042] Specifically, when performing time and space dimension mapping alignment, the system uses the timestamp of the acquired vehicle CAN bus dynamic timing feature F1 as the unique sample benchmark to standardize and align the road spatial environment feature F2 obtained through GPS and DEM digital elevation model. Since the sampling frequency of vehicle CAN bus data typically varies between 10Hz and 100Hz, while the refresh rate of spatial environment data is lower, this embodiment uses linear interpolation to fill in missing values ​​on the time axis of GPS latitude and longitude data and DEM digital elevation model data, and uniformly performs downsampling processing, uniformly defining its reference frequency as 10Hz. The linear interpolation method is as follows: , in, For missing values, This is the data set preceding the missing values. This is the set of data following the missing values.

[0043] This mechanism enables the road spatial environment features F2, such as road elevation and slope, to achieve a fully aligned and one-to-one standardized spatiotemporal dimension matching mapping with the dynamic temporal features F1 at continuous time steps with the same sampling frequency.

[0044] Before constructing the multivariate time-series feature vector V, the system performs dynamic cleaning of non-stationary operating condition samples to eliminate the disruption of the strong causal chain of quality estimation by non-steady-state mechanical noise. During this process, the system acquires the vehicle's current gear status (curgear), target gear status (selgear), brake opening status (brake), current vehicle speed (v), and longitudinal acceleration (a) in real time through onboard sensors and the CAN bus. This includes speed unit conversion, speed data smoothing, gear data type conversion, and acceleration calculation. Speed ​​unit conversion converts the original speed data from kilometers per hour to meters per second. Speed ​​data smoothing reduces potential noise or fluctuations in the original speed data using a sliding window averaging method with a window size of 40 data points. This means that for each time point's speed value, the average of the 40 data points before and after it is taken as the smoothed value. This method effectively reduces high-frequency noise in the data while preserving the overall trend of speed change. With the smoothed speed data, the vehicle's acceleration can be calculated. Acceleration is the rate of change of velocity over time. It is calculated by dividing the change in velocity between two adjacent time points by the time interval. For each time point (starting from the second data point), the change in velocity is obtained by subtracting the speed of the vehicle at the previous time point from the current speed. This change in velocity is then divided by the time interval (in seconds) between the two time points to obtain the acceleration value.

[0045] Subsequently, the system performs online logical discrimination based on the driving boundary conditions of the vehicle's longitudinal dynamics. When the real-time sampling state meets any one or more of the following rejection conditions, the sampling record corresponding to the current time step is determined to be a non-stationary noise condition. The core issue with shifting and braking data is that it disrupts the causal relationship between vehicle driving characteristics and overall vehicle mass, causing the model to learn incorrect patterns. This type of data belongs to strong noise and outlier samples, easily leading to overfitting of the model to invalid patterns and reducing generalization ability. Only by retaining valid samples of stable driving without additional external forces or mechanical intervention can the neural network learn stable mapping relationships, thereby ensuring the accuracy of mass estimation. When the two are different, i.e. This indicates that the vehicle is shifting gears, and the data at this time cannot be used for load identification; when the brake signal is greater than zero, i.e. This indicates that the vehicle is braking and the data at this time cannot be used for load identification.

[0046] Under low-speed conditions, vehicle speed, acceleration, and motor torque signals are prone to drastic fluctuations and noise interference, representing non-steady-state abnormal samples. This can lead to overfitting of neural networks to invalid features and reduce estimation generalization ability. To eliminate data from stopped conditions and reduce calculation errors, a speed threshold is set. The preferred value is 1 m / s. This applies only when the vehicle speed... Load identification is only performed at certain speeds. This speed threshold is determined based on the vehicle's operating characteristics during start-up, stopping, and low-speed driving phases. During these phases, vehicle speed, acceleration, and motor torque signals fluctuate significantly, easily introducing unsteady-state noise and increasing calculation errors. Therefore, data below 1 m / s is discarded to improve the stability and accuracy of load identification. Load identification based on higher acceleration can improve accuracy, hence the setting of a lower acceleration limit. Furthermore, because there are outliers in the acceleration data, which are much larger than the actual values, an acceleration upper limit is set. Outlier removal. To eliminate data during deceleration and to identify load based on significant acceleration, a lower limit for acceleration is set. , To ensure the data is valid, an acceleration upper limit is set to remove outlier acceleration data after velocity derivative. , This constitutes valid data. Once any of the above judgments is triggered, the system will perform an online process to completely strip and remove the time-series data sample containing destructive noise, thereby laying a highly reliable and clean data foundation for subsequent model training and feature derivation.

[0047] The cleaned and matched road spatial environment features F2 are further refined to include kinematically corrected road slope and terrain feature vectors. Specifically, the kinematically corrected road slope is calculated by first using forward differencing based on road elevation data from adjacent sampling points to obtain the basic differential slope value, as shown below: , in, The slope is the ratio of the height difference to the horizontal distance. For at any time The height value, For at any time The height value, This represents the straight-line distance (i.e., the horizontal distance) between two local extreme points. In order to be in Changes in distance in direction, In order to be in Changes in distance along a direction.

[0048] To eliminate the physical inertial disturbances to elevation differential perception caused by non-uniform speed driving states such as high torque acceleration or deceleration while climbing hills during actual operation, the system explicitly incorporates real-time longitudinal acceleration data of the vehicle to dynamically compensate and correct the basic differential slope value. This effectively eliminates the slope perception error caused by non-uniform vehicle speed driving. The corrected slope is expressed as: , in, To correct the road slope angle, For elevation difference, In order to be in The speed of the car at any given moment For time intervals, For vehicles at any time The acceleration.

[0049] Furthermore, to extract macroscopic spatial features and eliminate ordinal bias, the system sets several mutually exclusive intervals of absolute slope values ​​when constructing the terrain feature vector. These intervals are used to define plain terrain, hilly terrain, and hilly terrain, respectively. Specifically, the slope of the plain terrain falls into... The slope of the hilly terrain falls into The slope of the hilly terrain falls into The system accurately determines the macro-terrain category of the current vehicle based on the specific interval where the corrected real-time road slope falls. It traverses each slope interval, updates the cumulative distance for each terrain type based on the slope value category, and calculates its proportion in the total distance. The system uses smoothed elevation data (including elevation information at each location), cumulative mileage data recording the vehicle's cumulative travel distance, and interpolated elevation data to fill in missing values ​​as input data to ensure data continuity. It traverses the sorted extreme value index, calculates the slope value based on adjacent extreme points, and classifies the slope value into plains, gentle hills, or hilly areas. After each slope calculation, the cumulative distance for the corresponding terrain type is updated based on its slope value. For each terrain type, its cumulative distance is calculated, and the proportion of each terrain type is calculated based on the ratio of cumulative distance to total mileage. For the three terrain classification results (plains, gentle hills, and hilly areas), one-hot encoding is used for vectorization to avoid the ordinal weight bias introduced by numerical encoding. The conversion rule is as follows: Plain (−2%≤i≤2%): terrain=[1,0,0], Minimal hillock (−5%≤i<−2% or 2%) <i≤5%):terrain=[0,1,0], Heavy hills (i<−5% or i>5%): terrain=[0,0,1].

[0050] This avoids introducing unreasonable ordinal weight bias due to directly inputting numerical values ​​in subsequent deep learning regression modeling.

[0051] Example 3

[0052] Reference Figures 1-3 This is the third embodiment of the present invention. Unlike the previous embodiment, this embodiment provides a lightweight gating unit for dual-modal adaptive weighted fusion and a lightweight feature enhancement and screening mechanism based on the longitudinal dynamic equation of new energy trucks. This solves the defects of fixed splicing fusion method, which cannot adaptively adjust feature weights according to driving conditions, is easily affected by complex conditions, and cannot mine complementary information between heterogeneous modes. It realizes accurate spatiotemporal matching and adaptive, dynamic feature weighting of dual modes, and significantly improves the estimation accuracy and robustness of the model under complex and stable conditions.

[0053] Specifically, after precise spatiotemporal matching and dynamic cleaning of non-stationary operating condition samples, the system constructs a multivariate temporal feature vector V, which divides the secondary multi-source data into two heterogeneous modes: vehicle dynamic temporal mode, which comes from the vehicle CAN bus and is preprocessed to a unified 10Hz sampling rate; and road spatial environment mode, which comes from the GPS+DEM digital elevation model and is synchronized with the CAN data timestamp. The parameter definitions of the two models are shown in Table 1.

[0054] Table 1. Parameter Definition Table for Dual-Modal Multi-Source Data

[0055] For each CAN data timestamp Ti, the latitude and longitude coordinates output by the GPS module at that time are matched. Through UTM coordinate transformation and DEM elevation matching, the road elevation value and slope calculation result corresponding to the coordinates are obtained. The vehicle dynamic features and road environment features under the same timestamp Ti are aligned to form a 10-dimensional original feature vector under that time step. The feature vectors of the continuous time series are concatenated to form a multivariate time series feature matrix M of size [N,10], where N is the total length of the continuous time steps.

[0056] To address the shortcomings of traditional fixed-joining fusion methods, such as their inability to dynamically adjust feature weights based on driving conditions, susceptibility to interference from complex operating conditions, and inability to mine complementary information between modes, this invention introduces a lightweight gated modal adaptive fusion unit to achieve context-aware feature weighted fusion, such as... Figure 3 As shown, the gated network structure consists of two fully connected (FC) layers and a sigmoid activation function, dividing the temporal feature matrix into two independent branches, each weighted separately, namely the vehicle dynamic feature branch: X dyn ∈R N×6 Corresponding to 6-dimensional vehicle dynamic features; Road environment feature branch: X env ∈R N×4 This corresponds to 4D road environment features. The gated network learns the adaptive weights of the two branches at each time step through two fully connected layers and a sigmoid activation function, as shown below: , , in, for Weighting coefficients of vehicle dynamic features at time steps. for Weighting coefficients for road environment features at each time step. For the learnable weights of the gating network, for Time step vehicle dynamics feature branch. for Time step road environment feature branch. These are the bias parameters for the gated network. The function is a sigmoid activation function with an output range of (0,1), ensuring that the sum of the weights of the two branches is 1.

[0057] The features of the two branches are weighted step-by-step and then concatenated to form a fused feature vector, which is represented as follows: , in, This is the weighted vehicle dynamic feature vector. This is the weighted feature vector of the road environment.

[0058] Through this mechanism, the model can dynamically adjust the weights according to the working conditions: automatically increasing the weight of slope features during uphill driving and strengthening the contribution of motor torque and acceleration features during flat road cruising, thus achieving adaptive fusion.

[0059] To eliminate the dimensional differences between different features and improve the convergence speed and stability of the neural network, Z-Score normalization is performed on all feature dimensions after fusion, as follows: , in, for Time step The standardized values ​​of the 3D features For the first The complete feature vector after gated adaptive weighted fusion at each time step For the training set The mean of the dimensional features, For the training set The mean of the dimensional features.

[0060] After standardization, all feature dimensions satisfy a distribution with a mean of 0 and a variance of 1.

[0061] Finally, a standardized multivariate temporal feature matrix M of size [N, 10] is generated, which serves as the basis for subsequent temporal sample construction.

[0062] To reduce the learning difficulty of neural networks, strengthen the physical correlation between input features and vehicle mass, and improve the estimation accuracy and robustness of the model under complex working conditions, this invention, based on the longitudinal dynamic equation of new energy trucks, derives three types of lightweight features that are strongly correlated with vehicle mass and integrates them into a standardized time-series feature matrix without adding additional sensors.

[0063] During vehicle operation, the driving force must overcome rolling resistance, air resistance, gradient resistance, and acceleration resistance, which can be expressed as: , in, The driving force for the vehicle is calculated from the torque output of the drive motor. For rolling resistance, , For air resistance, , For slope resistance, , To increase resistance, , The total mass of the vehicle (including load, which is the estimation target of this invention) is the total mass of the vehicle. It is the acceleration due to gravity. The rolling resistance coefficient, The road slope angle, The air drag coefficient, For windward area, This is the rotational mass conversion factor.

[0064] The driving force characteristics are calculated based on the drive motor torque, gearbox transmission ratio, final drive ratio, and tire radius, and are directly related to the vehicle mass, expressed as: , in, For the torque of the drive motor, The gear ratio of the current gear. Main reduction ratio, For the mechanical efficiency of the transmission system, This refers to the tire rolling radius. All of the above parameters are inherent to the vehicle and can be obtained from the vehicle manual.

[0065] The torque-acceleration ratio directly reflects the motor torque output per unit acceleration and is positively correlated with the vehicle mass, expressed as: , in, Torque-acceleration ratio, for Longitudinal acceleration at the time step.

[0066] The time-series dynamic differential feature captures the changing trend of vehicle dynamic response under different loads, including vehicle speed and motor torque. acceleration Three core features are used to calculate the first-order difference features between adjacent time steps, which are expressed as follows: , in, This is the first-order difference feature of the characteristic, reflecting the instantaneous rate of change of the feature, which can effectively distinguish the differences in the dynamic response of vehicles under different loads. for The core feature values ​​of the time step for The core feature values ​​of the time step.

[0067] After constructing the dynamic enhancement feature D, to avoid feature dimensional redundancy, Pearson correlation coefficient analysis is used to analyze the linear correlation between all features and the true value of the vehicle mass, expressed as: , in, The Pearson correlation coefficient is used. Features With the true value of quality covariance, Features With the true value of quality The variance.

[0068] Strongly correlated features with an absolute correlation coefficient ≥ 0.2 are retained, while weakly correlated redundant features are removed, resulting in an enhanced multivariate temporal feature matrix M of size [N, 13], which is then directly input into the subsequent LSTM-1D-CNN quality estimation model.

[0069] Example 4

[0070] Reference Figure 1 and Figure 2 This is the fourth embodiment of the present invention. Unlike the previous embodiment, this embodiment provides a deep learning network architecture with dual-branch parallel fusion and regression. This solves the problem that traditional methods cannot take into account long-term time-series dependencies, sudden changes in local operating conditions, and insufficient deep fusion of heterogeneous features at high levels. It enables high-precision and robust continuous estimation of the total mass of new energy trucks from raw multi-source data.

[0071] Specifically, a dual-branch parallel fusion + regression architecture is adopted. The final generated enhanced multivariate temporal feature matrix M of size [N,13] is used as a unified input, and the LSTM long temporal feature extraction branch and the 1D-CNN local multi-scale feature extraction branch are simultaneously connected. After the two branches independently complete feature encoding, the complementary information is aggregated by the advanced feature deep fusion module. Finally, the real-time estimate of the total vehicle mass is output through the fully connected regression layer, realizing the end-to-end connection from the original multi-source data to the mass estimation result.

[0072] This study employs an LSTM long-time dynamic dependency capture branch. The core objective is to capture long-time dynamic dependencies strongly correlated with vehicle mass during the vehicle's journey, adapting to the temporal variations of features such as speed, torque, and acceleration during continuous driving of new energy trucks. The branch input is the generated [N,13] enhanced multivariate temporal feature matrix M, where N is the continuous time step length, and the 13-dimensional features include core vehicle dynamic features, road environment features, and dynamic enhancement features D. A two-layer stacked bidirectional LSTM network structure is used, with the first layer being an LSTM... Setting `return_sequences=True` outputs complete temporal hidden layer features, which are then passed to the second LSTM layer for further extraction of higher-order temporal dependencies. Each LSTM layer is followed by a Dropout layer with a dropout rate of 0.2 to suppress overfitting during model training. The hidden layer dimension is set to 64 dimensions based on the input feature dimension and temporal length, controlling the number of parameters while ensuring feature extraction capabilities. The branch ultimately outputs higher-order dynamic features of the vehicle's driving timeline, fully capturing the continuous dynamic response differences under different load conditions such as acceleration, cruising, and hill climbing, providing core temporal feature support for subsequent quality regression.

[0073] The local multi-scale feature extraction branch aims to uncover multi-scale salient features related to vehicle quality within a local time window, focusing on capturing the characteristic patterns of local operating conditions such as abrupt slope changes, instantaneous torque variations, and terrain transitions. This technique echoes the dynamic slope perception algorithm and terrain classification system discussed earlier, and complements LSTM. The branching approach complements long-term temporal dependencies and local detailed features. The multi-scale convolutional kernel employs three parallel one-dimensional convolutional layers with kernel sizes set to 3, 5, and 7 time steps, respectively, corresponding to local feature extraction at ultra-short, short, and medium time scales. This adapts to the time span of feature changes under different operating conditions, addressing the issue that fixed convolutional kernels cannot cover multi-scale operating condition features. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to accelerate model convergence and improve nonlinear fitting capabilities. A max-pooling layer is then applied after the convolutional layers to reduce the dimensionality of the features while preserving salient features within the local window and filtering redundant noise. Finally, the outputs of the three convolutional layers at different scales are concatenated along the channel dimension to form a complete local multi-scale feature vector. The branch's final output integrates local operating condition features across multiple time scales, accurately capturing local feature changes strongly correlated with vehicle quality, such as road slope, terrain changes, and instantaneous driving force, thus compensating for the LSTM branch's insufficient ability to capture local abrupt feature changes.

[0074] The dual-branch advanced feature deep fusion achieves accurate regression of vehicle quality by deeply fusing temporal dynamic features and local multi-scale features. This complements the gated modality adaptive fusion unit, constructing a two-level fusion system of "low-level feature adaptive fusion - advanced feature deep fusion." Global average pooling is performed on the temporal features output from the LSTM branch and the multi-scale features output from the 1D-CNN branch, mapping the two-dimensional temporal features to a one-dimensional global feature vector, eliminating the size difference in the time dimension. The global feature vectors from the two branches are concatenated along the channel dimension to form an advanced fusion feature that integrates long-term temporal dependencies and local multi-scale information. Two fully connected layers are used for non-linear mapping of the advanced fusion feature. The first fully connected layer is set to 32 dimensions and incorporates a batch normalization layer and a ReLU activation function. The second fully connected layer is set to 16 dimensions and incorporates a ReLU activation function and a Dropout layer to further filter redundant features, strengthen the contribution of core features directly related to vehicle quality, and suppress model overfitting. The output layer uses a linear activation function and sets up one output neuron to directly output a continuous estimate of the total vehicle mass (including vehicle weight and load), realizing an end-to-end mapping from multi-source input data to mass estimation results.

[0075] Example 5

[0076] Reference Figure 1 and Figure 2 This is the fifth embodiment of the present invention. Unlike the previous embodiment, this embodiment provides a model training strategy and generalization ability optimization, forming a closed loop with data preprocessing, sample selection, and feature engineering. Through targeted training strategies, the estimation accuracy and robustness of the model in complex commercial scenarios are guaranteed.

[0077] Specifically, a dataset is constructed using valid stable driving samples after full-process standardized preprocessing, invalid sample removal, precise spatiotemporal matching, and feature enhancement. The sample label is the true value of the vehicle mass under the corresponding driving conditions. The dataset is randomly divided into training, validation, and test sets in a ratio of 7:2:1. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and overfit monitoring, and the test set is used for the final verification of the model's generalization ability under all driving conditions. At the same time, stratified sampling is performed for the three types of terrain samples: plains, hills, and hills to ensure a balanced distribution of driving conditions in each dataset.

[0078] The core loss function uses mean squared error (MSE) to accurately measure the deviation between the quality estimate and the true value, aligning with the optimization objective of the regression task. An L2 regularization term is also included to penalize excessively large model weight values, further suppressing overfitting. The optimizer employs the Adam adaptive learning rate optimizer, setting the initial learning rate to 0.001 and incorporating an exponential learning rate decay strategy, reducing the learning rate to 0.9 every 10 epochs to balance convergence speed and accuracy. The training batch size is set to 32, and the maximum number of training epochs is 200. An early stopping mechanism is implemented during training: if the validation set loss does not decrease for 15 consecutive epochs, training is automatically terminated, retaining the model weights with the lowest validation set loss to prevent overfitting to the training set. Furthermore, temporal data augmentation operations are applied to the training set data, increasing the sample size through a time-series sliding window without disrupting physical causality, thereby improving the model's generalization ability. After training, the model is tested using a test set to verify its performance under all working conditions. The core evaluation metrics include mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination R². The focus is on verifying the model's estimation accuracy under all load ranges (no load, half load, full load) and all terrain conditions (plains, hills, and hilly areas) to ensure that the model meets the performance requirements for commercial deployment.

[0079] Furthermore, the optimal trained model undergoes lightweight optimization. Redundant neurons and convolutional kernels are removed through structured pruning, and INT8 model quantization technology is employed to significantly compress the number of model parameters and computational load without sacrificing core estimation accuracy. This reduces the model's hardware computing power requirements and adapts it to the operating environment of embedded devices such as the vehicle-mounted T-BOX and vehicle control unit (VCU). The model inference frequency is fully aligned with the 10Hz data sampling frequency of the preprocessing stage, achieving real-time vehicle mass estimation output every 100ms, meeting the online real-time estimation requirements during vehicle operation. The inference link is seamlessly integrated with the data acquisition, preprocessing, and feature calculation stages described above. Vehicle CAN bus data and GPS positioning data are transmitted in real time via the T-BOX and can be directly input into the model to complete mass estimation without additional manual intervention. The output of real-time vehicle mass estimation results can be directly connected to the energy management system and vehicle control system of new energy trucks, providing core input parameters for the optimal control strategies for starting acceleration, climbing, and braking energy recovery under different load conditions. This completely solves the industry pain points of poor adaptability of vehicle control strategies under different loads and insufficient energy consumption optimization in the background technology, realizing a closed-loop technology chain of "real-time acquisition of multi-source data - standardized processing of the whole process - feature enhancement and fusion - online model estimation - vehicle control strategy optimization". It can be fully adapted to the large-scale commercial operation scenarios of new energy trucks without the need to install any additional hardware sensors.

[0080] Example 6

[0081] Reference Figure 4 and Figure 5 This is the sixth embodiment of the present invention. Unlike the previous embodiment, this embodiment uses economic benefit calculations and simulation experiments to scientifically demonstrate the beneficial effects of the present invention.

[0082] First, smoothed vehicle speed, drive motor torque, current gear, target gear, braking signal, and longitudinal acceleration calculated from vehicle speed are collected from the vehicle's CAN bus. GPS trajectory points are simultaneously collected, and road elevation, slope, and terrain type are matched using a DEM (Digital Elevation Model). The actual weighing mass of each experimental condition is recorded as a supervision label. Driving force characteristics, torque-acceleration ratio characteristics, and temporal dynamic difference characteristics are calculated based on the vehicle's basic parameters. All data are standardized to a 10Hz sampling frequency, and timestamp alignment, missing value imputation, noise smoothing, removal of shift / braking / low-speed / abnormal acceleration samples, and Z-score standardization are performed, ultimately forming an enhanced multivariate temporal feature matrix M of size [N, 13].

[0083] Existing real-vehicle samples serve as the basis for validation data, such as Figure 4 As shown, the sample distribution of the three test vehicles within different tonnage ranges is intuitively displayed. The samples cover three new energy trucks: Vehicle 180, Vehicle 868, and Vehicle 989, totaling 26 sets of real load conditions, with a real weighing mass range of 10.56t-52.90t; Figure 5 As shown, time-series line graphs of vehicle dynamic timing characteristics (F1 examples) under typical operating conditions such as hill climbing, flat road cruising, and severe acceleration and deceleration were plotted. For each operating condition, vehicle speed, drive motor torque, current gear, target gear, and braking signal were simultaneously recorded via CAN channels. Based on statistics from real-vehicle samples, the vehicle speed channel accumulated 502,185 sampling records, the drive motor torque channel accumulated 2,510,927 sampling records, the braking channel accumulated 501,960 records, and the current gear and target gear each accumulated 251,094 records, totaling 4,017,260 original records across the five channels, corresponding to a sampling duration of approximately 25,108.1 seconds. These samples cover typical operating states under starting, acceleration, deceleration, constant speed cruising, different loads, and varying road gradients. They can be divided into training, validation, and test sets, and stratified sampling was performed for plains, low hills, and hilly terrain samples to ensure that the evaluation results reflect the vehicle's adaptability across all operating conditions. The actual vehicle samples are shown in Table 2, and the supplementary statistics of the original five-channel samples from the actual vehicle are shown in Table 3.

[0084] Table 2. Coverage of Samples Collected from Actual Vehicles Vehicle 180 9 groups 19.62-50.36 154759 773796 Vehicle 868 8 groups 22.14-52.90 141921 709601 Vehicle 989 9 groups 10.56-50.36 205505 1027530 total 26 groups 10.56-52.90 502185 2510927

[0085] Table 3 Supplementary Statistics of Original Data Collection from Five Channels of Actual Vehicles Vehicle 180 7737.5 154690 77379 77379 1238003 Vehicle 868 7095.7 141857 70961 70961 1135301 Vehicle 989 10274.9 205413 102754 102754 1643956 total 25108.1 501960 251094 251094 4017260

[0086] As shown in the two tables above, the existing samples cover three vehicles, 26 real-world load conditions, and five types of CAN channels, providing a solid data foundation for verifying the method of this invention. To further demonstrate the contributions of each technical unit, the verification experiment can be conducted with two control groups and one experimental group: Control group one uses CAN single-modal input and only uses vehicle dynamic temporal features F1 for LSTM-1D-CNN quality estimation; Control group two uses CAN+GPS / DEM multi-source input, but employs a fixed splicing method to fuse vehicle dynamic features and road environment features; The experimental group adopts the complete scheme of this invention, which introduces gated modal adaptive weighting, dynamic correlation feature enhancement, and an LSTM-1D-CNN dual-branch parallel regression model based on dual-modal spatiotemporal matching. The above control settings can respectively verify the contributions of road environment features, gating fusion mechanism, and dynamic feature enhancement to the estimation accuracy.

[0087] First, vehicle driving data was collected under load conditions and actual weighing was completed and labeled. Then, continuous time window samples were constructed according to the preprocessing procedure of this invention. Next, control group 1, control group 2, and experimental group models were trained respectively. Finally, based on the actual whole vehicle mass of the test set, the mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²) were calculated, and the MAPE of complex terrain samples was calculated separately to evaluate the robustness of the model under slope disturbance. Since the existing data is mainly used to illustrate the sampling coverage and validation organization method, Table 4 is used as an evaluation index recording table and does not pre-set the performance values ​​of the untrained model. After the actual model training and testing are completed, the statistical results of each group can be filled into Table 4.

[0088] Table 3 Supplementary Statistics of Original Data Collection from Five Channels of Actual Vehicles Control Group 1 CAN single-mode vehicle dynamic characteristics Verify the baseline effect before the introduction of road environmental information Statistics based on the test set Statistics based on the test set Statistics based on the test set Statistics based on the test set Control Group 2 CAN+GPS / DEM dual-modal features, fixed splicing and fusion Verify the improvement effect after introducing multi-source information Statistics based on the test set Statistics based on the test set Statistics based on the test set Statistics based on the test set experimental group Gated adaptive fusion + dynamic feature enhancement + LSTM-1D-CNN Verify the comprehensive effect of the complete technical chain of this invention. Statistics based on the test set Statistics based on the test set Statistics based on the test set Statistics based on the test set

[0089] After comparing the training and testing results of each model group, the experimental group achieved a significant reduction in the overall MAE index compared to control group 1, while the coefficient of determination R² was significantly increased. This conclusively proves at the underlying logic that cross-modal spatiotemporal fusion of GPS / DEM road spatial environment information and vehicle dynamic temporal information can fundamentally compensate for the lack of physical information in single onboard CAN bus data, achieving a leapfrog improvement in the accuracy of vehicle mass estimation. Furthermore, compared to control group 2, which also has dual-modal input but lacks dynamic adjustment capabilities, the experimental group not only further narrowed the global estimation error but also showed a highly significant reduction in the individual MAPE index under complex terrain samples. This mechanism was implemented step-by-step in actual testing. Context awareness enables the truck to adaptively and dynamically adjust feature weights when facing different transient and steady-state operating conditions, such as flat road cruising, steep slope climbing, and violent acceleration and deceleration. This simultaneously resists the severe disturbances of complex slopes in both physical mechanism and data feature dimensions, effectively ensuring the stability of the estimation results. Based on the above-mentioned tens of millions of real vehicle samples, multi-channel feature statistics, and undisputed evaluation index results, a clear load condition coverage bar chart and a typical operating condition speed / torque time series line chart are formed. From the three dimensions of the breadth of sample coverage, the sensitivity of dynamic physical response, and the superiority of quantitative index evaluation, the invention is mutually corroborated and comprehensively demonstrated to have a highly reliable technical verification foundation, a clear causal derivation path, and excellent industrial application value.

[0090] Example 7

[0091] Reference Figure 6 The seventh embodiment of the present invention provides a new energy truck quality estimation system based on multi-source data fusion, including a multi-source data spatiotemporal matching module 100, a non-stationary working condition sample cleaning module 200, a gated modal adaptive fusion module 300, a dynamic physical mechanism enhancement module 400, and a multi-scale feature regression estimation module 500.

[0092] The multi-source data spatiotemporal matching module 100 is used to perform time alignment and downsampling processing on the road spatial environment feature F2 with the timestamp of the dynamic temporal feature F1 as the unique sample benchmark, so that the road spatial environment feature F2 and the dynamic temporal feature F1 are matched and mapped in a one-to-one spatiotemporal dimension standardization on continuous time steps.

[0093] The non-stationary operating condition sample cleaning module 200 is used to acquire the vehicle's gear shifting status, braking signal, real-time vehicle speed and longitudinal acceleration in real time. When the real-time status meets the set non-stationary operating condition boundary conditions, the corresponding time series data samples are stripped and removed online to retain effective samples of stable driving that satisfy the causal relationship of the vehicle's longitudinal dynamics.

[0094] The gated modal adaptive fusion module 300 is used to split the effective samples into vehicle dynamic feature branches and road environment feature branches. It uses nonlinear gated logic units to dynamically predict the adaptive weight coefficients of the two branches in a time-step manner, and performs time-step context-aware feature weighted fusion to generate an adaptive fusion multivariate temporal feature vector V.

[0095] The dynamic physics mechanism enhancement module 400 is used to derive driving force characteristics, torque-acceleration ratio characteristics, and time-series dynamic difference characteristics that have a strong physical causal relationship with the vehicle mass based on the vehicle's longitudinal dynamic equilibrium equation and by utilizing the vehicle's inherent physical parameters and effective samples. It also eliminates redundant features through linear correlation analysis and constructs an enhanced multivariate time-series feature matrix M.

[0096] The multi-scale feature regression estimation module 500 is used to input the enhanced multivariate temporal feature matrix M in parallel into the long temporal hidden layer feature extraction channel and the multi-scale local window feature extraction channel, respectively capturing the temporal long-dependency feature M1 during the continuous driving process of the vehicle and the multi-scale local detail feature M2 within the local time window. The two features are then subjected to high-level deep aggregation and regression mapping to solve, and the continuous estimate of the total mass of the vehicle is output to dynamically adjust the energy management control strategy of the vehicle controller online.

[0097] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the new energy truck quality estimation method based on multi-source data fusion as proposed in the above embodiment.

[0098] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the new energy truck quality estimation method based on multi-source data fusion as proposed in the above embodiment.

[0099] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0100] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0101] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0102] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for estimating the quality of new energy trucks based on multi-source data fusion, characterized in that, include: The dynamic temporal features (F1) representing the vehicle's driving state and the road spatial environment features (F2) representing the vehicle's current geographical location are aligned by mapping and aligning the time and space dimensions to construct a multivariate temporal feature vector (V). By utilizing vehicle driving status information and inherent vehicle parameters, dynamic enhancement features (D) with strong physical causal relationship with vehicle commands are calculated, serving as mechanistic constraints to guide neural network learning. The multivariate temporal feature vector (V) and the dynamic enhancement feature (D) are combined and standardized to generate an enhanced multivariate temporal feature matrix (M) with strong physical mechanism constraints. Using the enhanced multivariate temporal feature matrix (M) as input, a regression algorithm with a multi-scale feature extraction mechanism is used to extract the temporal long-dependency features (M1) of the enhanced multivariate temporal feature matrix (M) in the continuous time dimension and the multi-scale local detail features (M2) in the local time window. The temporal long-dependency feature (M1) and the multi-scale local detail feature (M2) are fused, and a continuous estimate of the total vehicle mass is output by solving the regression mapping.

2. The method for estimating the quality of new energy trucks based on multi-source data fusion as described in claim 1, characterized in that: The process of aligning time and space dimensions includes, Using the timestamp of the dynamic temporal feature (F1) as the unique sample reference, the road spatial environment feature (F2) is subjected to linear interpolation and downsampling processing to perform one-to-one spatiotemporal dimension standardized matching mapping on continuous time steps with the same sampling frequency.

3. The method for estimating the quality of new energy trucks based on multi-source data fusion as described in claim 1 or 2, characterized in that: Before constructing the multivariate time-series feature vector (V), dynamic cleaning of non-stationary operating condition samples is also included. Real-time acquisition of vehicle's current gear status, target gear status, brake opening status, current vehicle speed, and longitudinal acceleration; If the current gear position is not equal to the target gear position, or the brake opening is greater than zero, or the current vehicle speed is less than or equal to the set speed threshold, or the longitudinal acceleration exceeds any one or more of the preset normal driving physical range, the current time step is determined to be a non-stationary noise condition, and the corresponding time series data samples are stripped and removed online.

4. The method for estimating the quality of new energy trucks based on multi-source data fusion as described in claim 3, characterized in that: The road spatial environment features include the kinematically corrected road slope and terrain feature vectors; The kinematically corrected road slope is calculated by forward differential calculation based on the road elevation data of adjacent sampling points to obtain the basic differential slope value. Real-time longitudinal acceleration data of the vehicle is introduced to perform dynamic inertial compensation correction on the basic differential slope value to eliminate the slope perception error caused by the non-uniform speed of the vehicle. The terrain feature vector construction sets several mutually exclusive absolute slope value intervals to define plain terrain, hilly terrain, and hilly terrain respectively. Based on the interval into which the corrected road slope falls, the terrain category of the current vehicle is determined, and the terrain category is converted into a geometric vector without category numerical weight bias as the terrain feature vector.

5. The method for estimating the quality of new energy trucks based on multi-source data fusion as described in any one of claims 1, 2, and 4, characterized in that: The construction of the multivariate time-series feature vector (V) includes, The dynamic temporal features (F1) are used as the vehicle dynamic features branch, and the road spatial environment features (F2) are used as the road environment features branch. Using a nonlinear gated logic unit, the first adaptive weight of the vehicle dynamic feature branch and the second adaptive weight of the road environment feature branch are predicted in real time step by step according to the current driving conditions, wherein the algebraic sum of the first adaptive weight and the second adaptive weight is 1. The features of the two branches are weighted and summed at each time step using the corresponding weights to generate an adaptively fused temporal feature vector.

6. The method for estimating the quality of new energy trucks based on multi-source data fusion as described in claim 5, characterized in that: The dynamic enhancement feature (D) includes, Based on the vehicle's longitudinal dynamics equilibrium equation, the time-series driving force characteristics of the vehicle are calculated by using real-time motor torque, gear ratio, and final drive ratio parameters. Within the time step when the vehicle's longitudinal acceleration is greater than zero, the ratio of real-time motor torque to longitudinal acceleration is calculated to construct the torque-acceleration ratio feature. Calculate the first-order difference of key dynamic time series parameters between adjacent time steps, and construct the time series dynamic difference feature; By introducing a linear correlation coefficient, the derived features are statistically validated, and weakly correlated redundant features with absolute values ​​below a set threshold are eliminated to form dynamic enhancement features (D) that are strongly correlated with the overall vehicle quality.

7. The method for estimating the quality of new energy trucks based on multi-source data fusion as described in any one of claims 1, 2, 4, and 6, characterized in that: The extraction of the enhanced multivariate temporal feature matrix (M) in terms of temporal long-term dependency features (M1) in the continuous time dimension and multi-scale local detail features (M2) within the local time window includes, The enhanced multivariate temporal feature matrix (M) is connected in parallel to the long temporal hidden layer feature extraction channel and the multi-scale local window feature extraction channel. The long temporal hidden layer feature extraction channel is used to capture the macro-temporal dynamic dependence trend of the vehicle during the entire process of starting, cruising or continuous climbing, and generate a temporal long dependence feature (M1). Simultaneously, by utilizing the multi-scale local window feature extraction channel, multiple parallel time-domain sliding windows of different sizes are used to capture the microscopic local significant working condition features caused by rapid changes in terrain or sudden changes in driving force, generating multi-scale local detail features (M2). After eliminating the time axis size difference by global average pooling of the time-series long-dependency feature (M1) and multi-scale local detail feature (M2), channel aggregation is performed, and the continuous estimate of the total mass of the vehicle is directly output through the regression mapping layer. This estimate is then used as a control parameter and input into the truck's vehicle controller in real time to dynamically adjust the vehicle's energy management control strategy.

8. A new energy truck quality estimation system based on multi-source data fusion, employing the new energy truck quality estimation method based on multi-source data fusion as described in any one of claims 1 to 7, characterized in that: It includes a multi-source data spatiotemporal matching module (100), a non-stationary working condition sample cleaning module (200), a gated mode adaptive fusion module (300), a dynamic physical mechanism enhancement module (400), and a multi-scale feature regression estimation module (500). The multi-source data spatiotemporal matching module (100) is used to perform time alignment and downsampling processing on the road spatial environment feature (F2) with the timestamp of the dynamic temporal feature (F1) as the unique sample benchmark, so that the road spatial environment feature (F2) and the dynamic temporal feature (F1) are matched and mapped in a one-to-one spatiotemporal dimension standardization on continuous time steps. The non-stationary operating condition sample cleaning module (200) is used to acquire the vehicle's gear shifting status, braking signal, real-time vehicle speed and longitudinal acceleration in real time. When the real-time status meets the set non-stationary operating condition boundary conditions, the corresponding time series data samples are stripped and removed online to retain effective samples of stable driving that satisfy the causal relationship of the vehicle's longitudinal dynamics drive. The gated modal adaptive fusion module (300) is used to split the effective samples into vehicle dynamic feature branches and road environment feature branches, and use nonlinear gated logic units to dynamically predict the adaptive weight coefficients of the two branches in a step-by-step manner, and perform step-by-step context-aware feature weighted fusion to generate an adaptive fusion multivariate temporal feature vector (V). The dynamic physical mechanism enhancement module (400) is used to derive driving force characteristics, torque-acceleration ratio characteristics and time-series dynamic difference characteristics with strong physical causal relationship with the vehicle mass based on the vehicle longitudinal dynamic balance equation, using the vehicle's inherent physical parameters and effective samples, and to eliminate redundant features through linear correlation analysis to construct an enhanced multivariate time-series feature matrix (M). The multi-scale feature regression estimation module (500) is used to input the enhanced multivariate temporal feature matrix (M) into the long temporal hidden layer feature extraction channel and the multi-scale local window feature extraction channel in parallel, respectively capturing the temporal long-dependency feature (M1) during the continuous driving process of the vehicle and the multi-scale local detail feature (M2) within the local time window. The two features are then subjected to high-level deep aggregation and regression mapping to solve, and the continuous estimate of the total mass of the vehicle is output to dynamically adjust the energy management control strategy of the vehicle controller online.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the new energy truck quality estimation method based on multi-source data fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the new energy truck quality estimation method based on multi-source data fusion as described in any one of claims 1 to 7.

Images