An automobile quality estimation method fusing wasserstein geometric consensus algorithm and deep network

Abstract

A vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks is proposed, involving the fields of artificial intelligence, data processing and algorithms, and vehicle engineering. The method includes the following steps: S1. CAN bus data acquisition; S2. Preprocessing the acquired data; S3. Inputting the preprocessed signal into a pre-trained SARE-Net and outputting a preliminary estimate; S4. Synchronously running the EKF algorithm to output the posterior probability distribution of the vehicle quality. P k EKF S5. Based on the preprocessed signal, a collaborative calculation method for confidence factors based on multi-criteria dynamic game optimization is adopted to generate the optimal confidence factor in real time. λ t * S6. Optimal confidence factor λ t * The invention incorporates the Wasserstein probabilistic consensus fusion framework, generating an optimal consensus probability distribution by solving a weighted centroid problem, and using its mean as the final quality estimate. This invention offers high quality estimation accuracy, outstanding anti-interference capabilities and stability, while also exhibiting strong adaptability to various operating conditions, low model dependency, good economic efficiency, and high real-time performance.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, data processing and algorithms, and vehicle engineering technology, and in particular to a vehicle quality estimation method that integrates the Wasserstein geometric consensus algorithm and deep networks. Background Technology

[0002] Vehicle mass is one of the most critical state parameters in vehicle dynamics, and its accurate estimation is essential for optimizing vehicle performance, economy, and safety. For example, accurate vehicle mass can be used to optimize shifting strategies, develop reasonable regenerative braking strategies, and improve cruise control precision. Especially for large vehicles, accurate and real-time vehicle mass estimation is crucial for braking safety, energy consumption, and powertrain protection.

[0003] Existing vehicle mass estimation methods mainly suffer from the following drawbacks: (1) Physical model-based methods (such as recursive least squares method and traditional Kalman filtering) often treat road slope as zero or estimate it simultaneously with mass. Inaccurate slope estimation can easily lead to poor algorithm convergence or even estimation failure. Furthermore, they are sensitive to filter parameter settings, and their performance heavily depends on an accurate vehicle longitudinal dynamics model and pre-calibrated noise statistics. In actual complex time-varying conditions (such as frequent starts and stops, gear shifts, and road surface disturbances), model mismatch and noise uncertainty are prone to occur, leading to decreased estimation accuracy, slow convergence speed, or even divergence. (2) Pure data-driven methods (such as single deep networks) rely on a large amount of labeled data and have insufficient model generalization ability; pure model-driven methods have weak fitting ability for nonlinear relationships and limited estimation accuracy.

[0004] (3) Most of the existing fusion schemes adopt simple cascading, parallel or fixed weight result fusion, which fails to achieve deep interaction and collaboration between the mechanism model and the data-driven model at the estimation process level. The adaptive ability is limited and cannot intelligently cope with the dynamic changes in the reliability of each estimation source under different working conditions. Therefore, developing a vehicle mass estimation method that can adapt to complex loads, balance real-time performance and accuracy, and has strong adaptability is a core technical challenge that urgently needs to be solved in the field of vehicle mass estimation. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a vehicle quality estimation method that integrates the Wasserstein geometric consensus algorithm and deep networks, so as to overcome the shortcomings of existing technologies, such as lack of economy and real-time performance, weak adaptability, and inaccurate quality estimation.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by this invention is as follows: A vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks, comprising the following steps: S1. CAN bus data acquisition: Vehicle operating information is acquired via the CAN bus, including longitudinal speed, motor output torque, and motor speed; S2. Signal preprocessing: Preprocess the collected data, including anomaly detection and repair, and establish a feedback linkage mechanism with the subsequent master estimator; S3. Deep Network Estimation: The preprocessed signal is input into a pre-trained two-layer structure—the dual adaptive dynamic recursive estimation network (SARE-Net)—which outputs a preliminary estimate m. k NN ; S4. Wasserstein geometry-based probabilistic consensus fusion algorithm: synchronously running extended Kalman filter (EKF) output quality posterior probability distribution. P k EKF Construct a Wasserstein probabilistic consensus fusion framework; S5. Confidence Factor Calculation: Based on the preprocessed signal, a collaborative calculation method for confidence factors based on multi-criteria dynamic game optimization is used to generate the optimal confidence factor in real time. l t * ; S6. Final Estimation and Closed-Loop Optimization: Optimal confidence factor l t * The Wasserstein probabilistic consensus fusion framework is injected, and the optimal consensus probability distribution is generated by solving the weighted centroid problem. The mean of the distribution is used as the final quality estimate.

[0007] A further technical solution of the present invention is that, in step S1, before starting the vehicle, the following preparatory work is performed: the vehicle's OBD interface is connected to the CAN bus, and the CAN bus is connected to the computer to realize the acquisition of vehicle operation data; in step S6, the optimal consensus probability distribution is fed back as a supervision signal to the preprocessing network, SARE-Net, EKF and game optimizer to form a closed-loop self-optimizing system.

[0008] A further technical solution of the present invention is: the specific steps of step S2 include: S21. Constructing a probabilistic graphical model of signal relationships: Define the CAN acquisition signal vector y at time t. t =[v t ,T e,t , n a,t,n t,t ] T v t ,T e,t , n a,t , n t,t Let x be the vehicle speed, engine torque, engine speed, and engine output shaft speed at time T, respectively. After removing abnormal signals, the usable signal is x. t Establish a joint probability distribution as a priori:

[0009] In equation (1), φij is a spatial constraint term based on simplified physical relations, ε is the set of physical associated edges, and ψ i For time smoothing; S22. Spatiotemporal Joint Anomaly Detection and Signal Repair: including spatiotemporal graph encoder construction, hierarchical anomaly identification and repair; S23. Task-oriented adaptive fine-tuning: Establish feedback loops, define task-aware loss, freeze the main network, and fine-tune the preprocessing network.

[0010] A further technical solution of the present invention is: the specific steps of step S22 include: S221. Spatiotemporal Graph Encoder Construction: The spatiotemporal graph encoder treats each time step signal as a graph node, with the node feature being the signal value; it defines spatial edges based on the probabilistic graphical model and temporal edges based on temporal adjacency; and it learns the hidden state h, which incorporates spatiotemporal neighborhood information, through a spatiotemporal graph convolutional network. i,t ; S222. Hierarchical anomaly identification and repair: S2221. Node-level Anomaly Scoring: Light Quantum Networks Based on h i,t Calculate the anomaly score a i,t ∈[0,1] ; S2222. Graph context repair: For a i,t For suspected abnormal nodes > τ1, initiate repair; repair value The attention-based read head reads from all normal nodes in its spatiotemporal neighborhood (a i,t Generated from the hidden state <τ2), using the core computation of the attention mechanism, a new representation containing global dependencies is generated:

[0011] In equation (2), Attention(h) i,t , h j,t’ Let be the attention weight function, representing the hidden state h of the normal node j in its spatiotemporal neighborhood at time t' when repairing the abnormal node i. j,t’The importance of the information provided; h i,t It is learned by the spatiotemporal graph encoder, and the hidden state features of node i at time t contain the association information of its spatiotemporal neighborhood; thus, "h" is calculated under the weight function. i,t h j,t’ The similarity between y and y is calculated, and a weight value between 0 and 1 is output. j,t’ For with h j,t’ The corresponding "value" vector; S2223. Physical constraint projection: Projecting the repaired complete signal vector Project the signal onto the constrained manifold defined by the probabilistic graphical model in step S21, and solve the optimization problem to maximize the satisfaction of the physical prior.

[0012] A further technical solution of the present invention is that step S3 includes the following specific steps: S31. Input Feature Selection: Select the vehicle longitudinal speed, longitudinal acceleration, motor output torque, transmission ratio, main reducer ratio, transmission efficiency, wheel effective radius, air density coefficient, frontal area, air resistance, rolling resistance coefficient, and road slope as input feature vectors; S32. Network Design: A pre-trained two-layer structure-network dual adaptive dynamic recursive estimation network (SARE-Net) is adopted as a deep network. The pre-processed input feature vector is input, and forward inference is used to obtain the prediction process noise covariance Q at the current time step. k and preliminary estimate m k NN ; The aforementioned two-layer structure-network dual adaptive dynamic recursive estimation network has the following structure:

[0013] In equation (3), x m For input; f m For the Gate of Oblivion; i m For input gate; c m Let t be the cell unit state at time t; o m h represents the cell state output at time t. m The network output at time t; σ is the sigmoid function; tanh is the hyperbolic tangent function; b i b f b c b o W is the bias vector; xi W hi W xf W hf W xc W hc Wxo W ho These are the weighting coefficients.

[0014] A further technical solution of the present invention is: the specific steps of step S4 include: S41. Extended Kalman Filter (EKF): EKF selects driving speed, road gradient, and vehicle mass as state variables, constructs and updates the state-space expression, and outputs a Gaussian posterior distribution P. k EKF =N(m k EKF , σ k 2,EKF );m k EKF Let σ be the mean of the normal distribution under EKF. k 2,EKF Let V be the variance of the normal distribution under EKF; S42. SARE-Net obtains the data-side uncertainty σ through a small network mapping. k NN Output Gaussian prediction distribution P k NN =N(m k NN , σ k 2,NN ), where the mean m k NN The original network output has a variance σ. k 2,NN The network’s “confidence” in this estimate is characterized by an online prediction made by an additional small branch network connected to the hidden layer; S43. Construct the Wasserstein consensus optimization problem to find the optimal consensus distribution P. k * This minimizes the sum of the weighted Wasserstein distances. S44. Design a consensus network ψ ϑ Online solution for P k * The mean and variance of.

[0015] A further technical solution of the present invention is as follows: Step S43 specifically includes: for two one-dimensional Gaussian distributions, which have closed-form solutions based on distribution information from two different sources, namely, a neural network (NN) and an extended Kalman filter (EKF), finding a consensus Gaussian distribution P. k * =N(m k * ,σ k 2,*The goal is to minimize the sum of the weighted distances to the two source distributions; the smaller the distance, the closer the distributions are. (4) In equation (4): P k * Represents the optimal distribution; argmin p W2 represents the parameter that minimizes the expression within the parentheses when solving for distribution P. 2 The Wasserstein-2 distance is used to measure the similarity between two distributions. The specific content of step S44 includes designing a learnable consensus network ψ to approximate the mapping and introducing more complex nonlinear relationships to adapt to quality estimation under complex working conditions. (5) In equation (5): m k * ,σ k 2,* Let m represent the optimal distribution parameter pair for the k-th object, where m k * It is the optimal mean, σ k 2,* It is the optimal variance, corresponding to the optimal distribution P. k * Core statistical characteristics; ψ θ It is a parameter fusion mapping function, where the subscript θ represents a specific type of weighted average or nonlinear combination of functions. Its function is to integrate multiple input parameters into the optimal output parameters; m k NN σ k 2,NN The first reference distribution is the mean and variance of the neural network (NN) distribution; m k EKF σ k 2,EKF The second reference distribution is obtained by using the Extended Kalman Filter (EKF) to determine the mean and variance of the distribution; λ k It is a weighting coefficient, with a value between 0 and 1, used to balance the relative importance of the two distribution bias terms; The training objective (loss function) of the consensus network ψ is the same as the optimization objective in step S43: (6) In equation (6), L ψ This represents the weighted loss value based on the fusion function ψ, which is the specific value of the objective function corresponding to the "minimize" operation in formula (4); the found P k * Substituting the values ​​into the objective function yields the final weighted loss value, which intuitively reflects the overall matching effect between the optimal distribution and the two reference distributions.

[0016] A further technical solution of the present invention is that step S5 includes the following specific steps: S51. Construct a real-time multi-dimensional performance index spectrum, including instantaneous consistency index, short-term smoothness index, and long-term local credit index; S52. Establish a dynamic game optimization model, and define the solution of the confidence factor as a constrained convex optimization problem; S53. Solving for the optimal confidence factor λ online using the Lagrange multiplier method combined with interval search. t * .

[0017] A further technical solution of the present invention is that step S51 includes: The instantaneous consistency index of S511.EKF is: , I A ins The term represents the target metric, with the superscript "ins" indicating "Insensitivity" and the subscript "A" indicating the model identifier. Together, they measure the model's sensitivity to observations; ε t The innovation at time t, which is the residual between the observed value and the model prediction, is the core input of the Kalman filter. H t Represents the observation matrix; P t│t-1 Let be the prior state covariance matrix at time t, which is a prediction of the state covariance at time t based on information from time t-1; δ is a regularization constant used to avoid singularity or zero denominators in the matrix; the instantaneous consistency index of the LSTM network is: I B ins =(σ 2 NN,t +δ) -1 When the observed value is abnormal, Wasserstein-2 increases, resulting in a larger value within the parentheses, which leads to... I A ins A smaller Wasserstein-2 value indicates that the model is highly sensitive to the anomalous observation; conversely, when the observation is normal, a smaller Wasserstein-2 value indicates a lower sensitivity. I A ins A large value indicates that the model has low sensitivity to this normal observation; S512. Short-term smoothness index: Calculates the estimator over a past time window W sFirst-order difference norm of the internal output quality estimate: The superscript smo stands for "Smoothness"; the larger the index value, the smoother the series fluctuations and the higher the stability. within the sliding window Ws Summing is performed on each sample, and two adjacent time steps are taken ( t−i+1 and t−i The squared Euclidean distance between adjacent time points measures the amplitude of fluctuations. When the series fluctuations are gentle: a smaller squared distance between adjacent time points indicates a smaller average fluctuation, i.e., a smaller value in parentheses. I smo A larger value indicates higher stability; when the sequence fluctuates wildly: a larger squared distance between adjacent elements indicates a larger average fluctuation, i.e., a larger value in parentheses. I smo Small indicates low stability; S513. Long-term local credit indicators: I cred =C cred (M current ), I cred The credit metric representing the final output. C cred It is a credit mapping function. M current The current market state C represents the input. cred (M j ) and operating mode M j The estimation error statistic is inversely proportional to the current market state data (Mcurrent), which is a multi-dimensional input describing the market environment. Then, a pre-defined credit mapping function is used. C cred Calculations are performed to convert market conditions into credit metrics that can be directly used for risk assessment. I cred The function can be linear or nonlinear, and its purpose is to quantify the level of credit risk in the current market environment.

[0018] A further technical solution of the present invention is that, in step S6: Forward fusion via λ t * Modulate the distribution shape, construct a weighted Wasserstein barycenter optimization objective, and obtain the consensus distribution. P t * =N ( m t * , (σ t * ) 2 ); In the formula: P t * The optimal probability distribution at time t represents the final result obtained by fusing two reference distributions; N() represents the normal distribution. m t * The optimal mean of this normal distribution is given by the fusion function ψ in formula (5). θ Calculated; (σ) t * ) 2 The optimal variance of this normal distribution, also output by the fusion function of formula (5), is the optimal standard deviation σ. t * The square of.

[0019] By adopting the above technical solution, the vehicle quality estimation method of the present invention, which integrates the Wasserstein geometric consensus algorithm and deep networks, has the following advantages compared with the prior art: 1. Significantly Improved Quality Estimation Accuracy: This invention performs Wasserstein consensus optimization in the probability distribution space, fundamentally seeking the optimal estimation point from the outputs of two heterogeneous estimators (SARE-Net and EKF). This method is mathematically more rigorous, simultaneously optimizing the position (mean) and uncertainty (variance) of the estimated value. Its statistical properties outperform any numerical fusion method based on fixed rules, thus fundamentally improving estimation accuracy. This invention provides reliable and accurate quality estimates, comprehensively enhancing the robustness of vehicle quality estimation.

[0020] 2. Outstanding anti-interference and stability: This invention utilizes a noise covariance adaptive algorithm with dual-source uncertainty perception and stability constraints to estimate and correct the measured noise covariance matrix in real time. It also combines this with a structure-network dual adaptive dynamic recursive estimation network for forward inference to obtain the prediction process noise covariance Q at the current moment. k This provides a more accurate posterior distribution P for probabilistic consensus fusion algorithms. k This effectively improves the system's anti-interference capability and enables the system to have better estimation performance.

[0021] Furthermore, this invention relies on LSTM neural networks to achieve real-time process noise estimation and dynamically adjusts the process noise covariance to adapt to time-varying noise environments. Simultaneously, when an estimator (such as SARE-Net in untrained mode) consistently performs poorly, the game-theoretic optimizer automatically reduces its confidence through a penalty term. This change is immediately mitigated through the Wasserstein consensus framework by adjusting the distribution and optimizing weights, thus gently reducing the influence of anomalous estimators in the final decision, preventing sudden changes or collapses in the overall system performance, and achieving a smooth process of "fault mitigation" rather than "fault switching."

[0022] 3. Significantly enhanced adaptability to different working conditions: The collaborative calculation method of confidence factor based on multi-criteria dynamic game optimization introduced in the method framework of this invention enables the estimated value to better fit the situation under different working conditions.

[0023] 4. Significantly reduced model dependence: This invention breaks through the limitations of traditional methods by deeply integrating data-driven and model-driven approaches. It reduces the dependence on accurate models through deep networks, and embeds quality constraints into the loss function part of the network design, enabling the system to fine-tune network parameters based on new data to adapt to changes in vehicle characteristics.

[0024] 5. Good economic efficiency: This invention does not require additional sensors and collects existing vehicle operation data through the CAN bus, reducing system cost and complexity.

[0025] 6. High Real-Time Performance: This invention avoids the data accumulation and waiting time of batch processing algorithms. The core algorithm is based on optimization theory and probabilistic calculations, avoiding the black-box uncertainty and potential numerical instability of deep neural networks. Each module has a clearly defined function, and can operate as an overall intelligent system. Its sub-modules (such as high-performance preprocessing and game optimizers) can also independently serve other vehicle estimation tasks. Each step is processed in real time, meeting the real-time requirements of vehicle control.

[0026] The technical features of a vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks, as described below, are further explained with reference to the accompanying drawings and embodiments. Attached Figure Description

[0027] Figure 1 This is a flowchart of the overall process of a vehicle quality estimation method that integrates the Wasserstein geometric consensus algorithm and deep networks according to the present invention. Figure 2 This is a flowchart of the feedback linkage mechanism provided in step S2 of embodiment one of this application; Figure 3 This is a flowchart of the collaborative calculation of confidence factors based on multi-criteria dynamic game optimization provided in step S5 of embodiment one of this application; Figure 4 This is the Wasserstein geometry mechanism probability consensus fusion algorithm provided in step S6 of embodiment one of this application, which generates the flowchart of the optimal consensus probability distribution. Detailed Implementation Example 1

[0028] A vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks is described in the following general process: Figure 1 As shown, it mainly includes a four-layer progressive structure and multiple feedback optimization links: 1. Data Layer: Intelligent Signal Preprocessing Guided by Probabilistic Graphical Models Input: Raw CAN bus data, including multi-source heterogeneous signals during vehicle operation, such as vehicle speed, acceleration, engine speed, and braking pressure.

[0029] Module: Intelligent signal preprocessing module guided by probabilistic graphical model.

[0030] Function: It uses a probabilistic graphical model to model the spatiotemporal dependencies between CAN data, and performs preprocessing operations such as intelligent denoising, outlier detection and correction, and key feature extraction to output high-quality signals and provide reliable input for subsequent estimation layers.

[0031] 2. Estimation Layer: Dual-path parallel quality estimation architecture Module 1: Two-layer adaptive recursive estimation network (SARE-Net) Input: High-quality signal output from the data layer.

[0032] Function: Construct a two-layer adaptive recurrent neural network structure to automatically learn the nonlinear mapping relationship between vehicle mass and operating signals, realize data-driven mass estimation, and output the first estimation result.

[0033] Module 2: Extended Kalman Filter (EKF) Input: High-quality signal and vehicle dynamics prior model.

[0034] Function: Construct a state-space model based on vehicle dynamics equations, recursively estimate mass parameters using the extended Kalman filter algorithm, and outputs the second estimation result.

[0035] 3. Decision-making level: Multi-criteria dynamic game optimization to generate confidence factors Input: The two estimation results output by the estimation layer and the corresponding operating condition information.

[0036] Module: Multi-criteria dynamic game optimizer.

[0037] Function: Based on multiple criteria such as estimation accuracy, robustness, and computational efficiency, a dynamic game model is constructed. Through game optimization algorithms, reliability factors are assigned to the estimation results of SARE-Net and EKF, quantifying the reliability of different estimation methods under the current working conditions.

[0038] 4. Fusion and Optimization Layer: Wasserstein probabilistic consensus is used to fuse and generate the final estimate. Inputs: Confidence factors of the decision layer output, and two estimation distributions of the estimation layer.

[0039] Module: Wasserstein probabilistic consensus fusion module.

[0040] Function: Models the multi-path estimation results as a probability distribution, uses the Wasserstein distance to measure the differences between the distributions, and fuses the different distributions through a geometric consensus algorithm to generate a globally consistent final quality estimate m; at the same time, it outputs the consensus distribution as a supervision signal for feedback optimization.

[0041] 5. Feedback Optimization Link: Iterative Optimization of the Entire Link Link 1: The consensus distribution output by the fusion layer is used as a supervision signal and sent back to the data layer to guide the probabilistic graphical model to optimize the preprocessing strategy and improve signal quality.

[0042] Link 2: Through the meta-learning optimization framework, the final estimate and consensus distribution are fed back to the estimation layer and decision layer to update the network parameters of SARE-Net and the criterion weights of the dynamic game model, thereby achieving adaptive optimization across the entire link and improving the robustness and accuracy of the method under complex conditions.

[0043] The specific steps of this invention are as follows:

[0044] S1. CAN Bus Data Acquisition: Vehicle operating information is acquired via the CAN bus, including longitudinal speed, motor output torque, and motor speed. Before starting the vehicle, the vehicle's OBD interface must be connected to the CAN bus, and the CAN bus must be connected to the computer to acquire vehicle operating data. S2. Signal preprocessing: Preprocess the collected data, including anomaly detection and repair, and establish a feedback linkage mechanism with the subsequent master estimator; S3. Deep Network Estimation: The preprocessed signal is input into a pre-trained two-layer structure—the dual adaptive dynamic recursive estimation network (SARE-Net)—to output a preliminary estimate. m k NN ; S4. Wasserstein geometry-based probabilistic consensus fusion algorithm: synchronously running extended Kalman filter (EKF) output quality posterior probability distribution. P k EKF Construct a Wasserstein probabilistic consensus fusion framework; S5. Confidence Factor Calculation: Based on the preprocessed signal, a collaborative calculation method for confidence factors based on multi-criteria dynamic game optimization is used to generate the optimal confidence factor in real time. l t * ; S6. Final Estimation and Closed-Loop Optimization: Optimal confidence factor l t * The Wasserstein probabilistic consensus fusion framework is injected, and an optimal consensus probability distribution is generated by solving the weighted centroid problem. The mean of this distribution is used as the final quality estimate. This optimal consensus probability distribution serves as a supervisory signal, feeding back to the preprocessing network, SARE-Net, EKF, and game optimizer to form a closed-loop self-optimizing system.

[0045] The specific steps of step S2 include: S21. Constructing a probabilistic graphical model of signal relationships: Define the CAN acquisition signal vector y at time t. t =[v t ,T e,t , n a,t ,n t,t ] T v t ,T e,t , n a,t , n t,t Let x be the vehicle speed, engine torque, engine speed, and engine output shaft speed at time T, respectively. After removing abnormal signals, the usable signal is x. t Establish a joint probability distribution as a priori:

[0046] In equation (1), φ ij ε is a spatial constraint term based on simplified physical relationships, measuring whether the values ​​of signals i and j conform to their physical relationships (such as linear scaling, equation constraints); ε is a set of physically associated edges, defining which signal pairs have known physical constraint relationships (such as torque and speed), which is manually defined based on prior knowledge of the vehicle's longitudinal dynamics equations; ψ i This is a time smoothing term to prevent drastic changes in the same signal at adjacent times, forcing time smoothness based on the reasonable assumption that "the signal changes continuously within a short time window"; S22. Spatiotemporal Joint Anomaly Detection and Signal Repair: This includes spatiotemporal graph encoder construction, hierarchical anomaly identification and repair; the specific steps of this step S22 include: S221. Spatiotemporal Graph Encoder Construction: The spatiotemporal graph encoder treats each time step signal as a graph node, with the node feature being the signal value; spatial edges are defined based on the probabilistic graphical model, and temporal edges are defined based on temporal adjacency; through learning via a spatiotemporal graph convolutional network, the hidden state h of each node, i.e., each signal, at each time step is obtained. i,t It integrates spatiotemporal neighborhood information; S222. Hierarchical anomaly identification and repair: S2221. Node-level Anomaly Scoring: Light Quantum Networks Based on h i,t Calculate the anomaly score a i,t ∈[0,1] ; S2222. Graph context repair: For a i,t For suspected abnormal nodes > τ1, initiate repair; repair value The attention-based read head reads from all normal nodes in its spatiotemporal neighborhood (a i,t Generated from the hidden state <τ2), using the core computation of the attention mechanism, a new representation containing global dependencies is generated:

[0047] In equation (2), Attention(h) i,t h j,t’ Let be the attention weight function, representing the hidden state h of the normal node j in its spatiotemporal neighborhood at time t' when repairing the abnormal node i. j,t’ The importance of the information provided; h i,t It is learned by the spatiotemporal graph encoder, and the hidden state features of node i at time t contain the association information of its spatiotemporal neighborhood; thus, "h" is calculated under the weight function. i,t h j,t’ The similarity between y and y is calculated, and a weight value between 0 and 1 is output. j,t’ For with h j,t’ The corresponding "value" vector; the essence of the attention mechanism is "to match all values ​​with the query, obtain weights, and then sum the values ​​with weights", where y i,t’ It is the object that is weighted.

[0048] S2223. Physical constraint projection: Projecting the repaired complete signal vector Projected onto the constrained manifold defined by the probabilistic graphical model in step S21, the repaired signal is optimized to satisfy the physical priors to the maximum extent by solving the optimization problem. S23. Task-oriented adaptive fine-tuning: Establish a feedback loop, define the task-aware loss, freeze the main network, and fine-tune the preprocessing network, as detailed below: S231. Feedback Link Establishment: The output X of the preprocessing network is then... t use The data is fed into the pre-trained main SARE-Net+EKF fusion system to obtain the quality estimate m. t * .

[0049] S232. Define task-aware loss: Calculate the feedback loss from the main task:

[0050] S233. Freeze the main network and fine-tune the preprocessing network: use L task Fine-tuning the last few layers or specific parameters of the preprocessing network can achieve end-to-end adaptation with the main system.

[0051] In step S2, the raw data first enters the "probabilistic graphical model-guided preprocessing network". This network not only uses the spatiotemporal graphical model to repair anomalies, but more importantly, it ensures that the output clean signal strictly follows the basic vehicle dynamics relationship through physical constraint projection, thus guaranteeing the reliability of the input of the subsequent physical model (EKF) from the source.

[0052] Step S3 includes the following specific steps: S31. Input Feature Selection: Select the vehicle longitudinal speed, longitudinal acceleration, motor output torque, transmission ratio, main reducer ratio, transmission efficiency, wheel effective radius, air density coefficient, frontal area, air resistance, rolling resistance coefficient, and road slope as input feature vectors; S32. Network Design: A pre-trained two-layer structure-network dual adaptive dynamic recursive estimation network (SARE-Net) is adopted as a deep network. The pre-processed input feature vector is input, and forward inference is used to obtain the prediction process noise covariance Q at the current time step. k and preliminary estimate m k NN ; The aforementioned two-layer structure-network dual adaptive dynamic recursive estimation network has the following structure:

[0053] In equation (3), x m For input; f m For the Gate of Oblivion; i m For input gate; c m Let t be the cell unit state at time t; o m h represents the cell state output at time t.m The network output at time t; σ is the sigmoid function; tanh is the hyperbolic tangent function; b i b f b c b o W is the bias vector; xi W hi W xf W hf W xc W hc W xo W ho These are the weighting coefficients.

[0054] Step S32 includes the following specific contents: S321. Two-layer network structure adaptive dynamic recursive estimation network 1. Based on the standard LSTM, we define an extended standard LSTM, first selecting a set of candidate gate vectors:

[0055] 2. A differentiable decision function φ based on the EKF state determines which gating structure is activated at the current time step:

[0056]

[0057] In the above formula, Sofa tmax is the normalized exponential function. The output of the MLP is transformed into a probability distribution vector; the MLP is a lightweight multilayer perceptron. For feature vectors φ with different dimensions... t Nonlinear fusion and feature upscaling are performed to extract higher-level pattern information related to structural decision-making.

[0058] In the above formula, trace(P) k│k-1 ) is EKF at time k The prediction error covariance matrix P k│k-1 trace; ε k For EKF at time k The new information vector. φ t The system state feature vector includes: prediction uncertainty trace, innovation energy, fuzzy confidence, and state transition matrix F. k condition number S322. Dynamic Parameter Generation Layer for Neural Networks 1. Real-time tracking of target parameter network input and design

[0059] In the above formula, Encoder is a small neural network; Standardized information weighted by the predicted uncertainty is used to confirm whether there are sudden changes in the system and adjust the stability of the network. Indicates vector concatenation; svd(P k│k-1 ) represents the singular values ​​of the predicted covariance matrix; Let be the value of the Jacobian matrix of the observation equation at the latest state.

[0060] S323. Co-training and Loss Function of Physically Constrained Embedding 1. Loss due to physical consistency constraints

[0061] Where matrix B is the time-varying weight matrix, a i The acceleration F is calculated using the finite difference method. i tot The total driving force is calculated based on the CAN signal.

[0062] 2. Cooperative adaptive loss

[0063] Among them, P k [m,m] Let V be the variance of the quality status.

[0064] 3. Total loss function:

[0065] Where L pre This is the standard prediction error.

[0066] In step S3, the purified signal is simultaneously fed to two core estimators. EKF, based on an accurate longitudinal dynamics model, recursively calculates and outputs its optimal estimate and covariance. SARE-Net then uses the system features ψ extracted from EKF... t and It dynamically adjusts its internal structure and parameters to output a data-driven prediction and its uncertainty σ. 2 NN Thus, both have completed the upgrade from point estimation to probability distribution.

[0067] The specific steps of step S4 include: SS41. Extended Kalman Filter (EKF): EKF selects driving speed, road gradient, and vehicle mass as state variables, constructs and updates the state-space expression, and outputs a Gaussian posterior distribution P. k EKF =N(m k EKF , σ k2,EKF );m k EKF Let σ be the mean of the normal distribution under EKF. k 2,EKF Let V be the variance of the normal distribution under EKF; S42. SARE-Net obtains the data-side uncertainty σ through a small network mapping. k NN Output Gaussian prediction distribution P k NN =N(m k NN , σ k 2,NN ), where the mean m k NN The original network output has a variance σ. k 2,NN The network’s “confidence” in this estimate is characterized by an online prediction made by an additional small branch network connected to the hidden layer; S43. Construct the Wasserstein consensus optimization problem to find the optimal consensus distribution P. k * To minimize the sum of weighted Wasserstein distances; specific details include: Given two one-dimensional Gaussian distributions, one based on a neural network (NN) and the other on an extended Kalman filter (EKF), which have closed-form solutions, we need to find a consensus Gaussian distribution P. k * =N(m k * ,σ k 2,* The goal is to minimize the sum of the weighted distances to the two source distributions; the smaller the distance, the closer the distributions are.

[0068] In equation (4): P k * Represents the optimal distribution; argminp represents the parameter that minimizes the expression within the parentheses when solving for distribution P; W22 is the Wasserstein-2 distance, used to measure the similarity between two distributions; S44. Design a consensus network ψ ϑ Online solution for P k * The mean and variance of the mean and variance, the specific content of this step S44 includes designing a learnable consensus network ψ to approximate the mapping, and introducing more complex nonlinear relationships to adapt to quality estimation under complex working conditions;

[0069] m k * ,σ k 2,* Let m represent the optimal distribution parameter pair for the k-th object, where m k * It is the optimal mean, σ k 2,* It is the optimal variance, corresponding to the optimal distribution P. k * Core statistical characteristics; ψ θ It is a parameter fusion mapping function, where the subscript θ represents a specific type of weighted average or nonlinear combination of functions. Its function is to integrate multiple input parameters into the optimal output parameters; m k NN σ k 2,NN The first reference distribution is the mean and variance of the neural network (NN) distribution; m k EKF σ k 2,EKF The second reference distribution is obtained by using the Extended Kalman Filter (EKF) to determine the mean and variance of the distribution; λ k It is a weighting coefficient, with a value between 0 and 1, used to balance the relative importance of the two distribution bias terms; The training objective (loss function) of the consensus network ψ is the same as the optimization objective in step S43:

[0070] In equation (6), Lψ represents the weighted loss value based on the fusion function ψ, which is the specific value of the objective function corresponding to the "minimize" operation in equation (4); the found P k * Substituting the values ​​into the objective function yields the final weighted loss value, which intuitively reflects the overall matching effect between the optimal distribution and the two reference distributions.

[0071] Step S5 includes the following specific steps: S51. Construct a real-time multi-dimensional performance index spectrum, including instantaneous consistency index, short-term smoothness index, and long-term local credit index; S52. Establish a dynamic game optimization model, and define the solution of the confidence factor as a constrained convex optimization problem; S53. Solving for the optimal confidence factor λ online using the Lagrange multiplier method combined with interval search. t * .

[0072] Step S51 includes: The instantaneous consistency index of S511.EKF is: , I A ins The term represents the target metric, with the superscript "ins" indicating "Insensitivity" and the subscript "A" indicating the model identifier. Together, they measure the model's sensitivity to observations; ε t The innovation at time t, which is the residual between the observed value and the model prediction, is the core input of the Kalman filter. H t Represents the observation matrix; P t│t-1 Let be the prior state covariance matrix at time t, which is a prediction of the state covariance at time t based on information from time t-1; δ is a regularization constant used to avoid singularity or zero denominators in the matrix; the instantaneous consistency index of the LSTM network is: I B ins =(σ 2 NN,t +δ) -1 When the observed value is abnormal, Wasserstein-2 increases, resulting in a larger value within the parentheses, which leads to... I A ins A smaller Wasserstein-2 value indicates that the model is highly sensitive to the anomalous observation; conversely, when the observation is normal, a smaller Wasserstein-2 value indicates a lower sensitivity. I A ins A large value indicates that the model has low sensitivity to this normal observation; S512. Short-term smoothness index: Calculates the first-order difference norm of the quality estimate output by the estimator over the past time window Ws. The superscript smo stands for "Smoothness"; the larger the value, the smoother the series fluctuations and the higher the stability. within the sliding window Ws Summing is performed on each sample, and two adjacent time steps are taken ( t−i+1 and t−i The squared Euclidean distance between adjacent time points measures the amplitude of fluctuations. When the series fluctuations are gentle: a smaller squared distance between adjacent time points indicates a smaller average fluctuation, i.e., a smaller value in parentheses. I smo A larger value indicates higher stability; when the sequence fluctuates wildly: a larger squared distance between adjacent elements indicates a larger average fluctuation, i.e., a larger value in parentheses. I smo Small size indicates low stability; S513. Long-term local credit indicators: I cred =C cred (Mcurrent ), I cred The credit metric representing the final output. C cred It is a credit mapping function. M current The current market state C represents the input. cred (M j ) and operating mode M j The estimation error statistic is inversely proportional to the current market state data (Mcurrent), which is a multi-dimensional input describing the market environment. Then, a pre-defined credit mapping function is used. C cred Calculations are performed to convert market conditions into credit metrics that can be directly used for risk assessment. I cred The function can be linear or nonlinear, and its purpose is to quantify the level of credit risk in the current market environment.

[0073] Step S52, establishing the dynamic game optimization model, includes: The solution to the confidence factor λ is defined as the following constrained convex optimization problem:

[0074]

[0075] In the formula: max(0,.) is the truncation penalty term; IAtot and IBtot are the total indicators of categories A and B, respectively; w1, w2, and w3 are the weight coefficients of different sub-items.

[0076] Step S53 specifically includes: S531. Real-time Solution: The objective function of the above optimization problem is a piecewise convex function with linear constraints. At each time t, the Lagrange multiplier method combined with interval search is used to quickly solve for a closed-form solution or an approximate solution. This method has a light computational burden, meets the real-time requirements of vehicle operation, and yields the optimal solution λ. t * .

[0077] In step S5, based on the preprocessed signal, a collaborative calculation method for the confidence factor based on multi-criteria dynamic game optimization is adopted. This method receives multi-dimensional performance index spectra from the two estimators, as well as a signal quality report from the preprocessing stage. By solving a constrained convex optimization problem, the optimal confidence factor λt* is output. This factor is not only a weight but also a fusion strategy instruction; its value directly determines whether subsequent fusion will favor "aggressive learning" or a "conservative trust model," generating the optimal confidence factor in real time.

[0078] In step S6, the optimal confidence factor is injected into the Wasserstein probabilistic consensus fusion framework. The optimal consensus probability distribution is generated by solving the weighted centroid problem, and its mean is used as the final quality estimate, with its variance σ being taken. t 2 This serves as the confidence metric for the estimation. Forward fusion is achieved through λ. t * Modulate the distribution shape, construct a weighted Wasserstein barycenter optimization objective, and obtain the consensus distribution. P t * =N ( m t * , (σ t * ) 2 ); In the formula: P t * The optimal probability distribution at time t represents the final result obtained by fusing two reference distributions; N() represents the normal distribution. m t * The optimal mean of this normal distribution is given by the fusion function ψ in formula (5). θ Calculated; (σ) t * ) 2 The optimal variance of this normal distribution, also output by the fusion function of formula (5), is the optimal standard deviation σ. t * The square of.

[0079] Specifically, on the forward path, the λ generated by the dynamic game optimizer... t * Deeply integrated into the Wasserstein consensus optimization framework, the "style weights" not only function as weights but also determine their magnitude by modulating the input distribution and the optimization objective function. This allows the game conclusions of the middle layers to directly and profoundly influence the mathematical process of top-level decision-making, rather than simply being a post-weighted approach.

[0080] In the feedback path, the probability consensus distribution generated at the top level serves as high-level knowledge, providing bidirectional nourishment downwards: on the one hand, it acts as a correction signal, performing "soft training" and "soft guidance" on the underlying SARE-Net and EKF with adaptive strength to enhance their individual capabilities; on the other hand, it serves as an evaluation benchmark, optimizing the internal evaluation parameters of the mid-level game optimizer through meta-learning loops, enabling its decision logic to continuously evolve in pursuit of long-term optimal system performance.

[0081] 1. Forward fusion: The optimization objective of traditional probabilistic consensus fusion is:

[0082] Weighted Wasserstein centroid optimization: It has an approximately closed-form solution for a Gaussian distribution. The final consensus distribution P t * =N(m t * ,(σ t * ) 2 The mean and variance of λ are both explicitly dependent on λ. t * Take m t * For the final quality estimate, its uncertainty (σ) t * ) 2 The feedback will be sent to the system.

[0083] 2. Feedback optimization: Consistency loss: .

Claims

1. A vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks, characterized in that, Includes the following steps: S1. CAN bus data acquisition: Vehicle operating information is acquired via the CAN bus, including longitudinal speed, motor output torque, and motor speed; S2. Signal preprocessing: Preprocess the collected data, including anomaly detection and repair, and establish a feedback linkage mechanism with the subsequent master estimator; S3. Deep Network Estimation: The preprocessed signal is input into a pre-trained two-layer structure—the dual adaptive dynamic recursive estimation network (SARE-Net)—to output a preliminary estimate. m k NN ; S4. Wasserstein geometry-based probabilistic consensus fusion algorithm: synchronously running extended Kalman filter (EKF) output quality posterior probability distribution. P k EKF Construct a Wasserstein probabilistic consensus fusion framework; S5. Confidence Factor Calculation: Based on the preprocessed signal, a collaborative calculation method for confidence factors based on multi-criteria dynamic game optimization is used to generate the optimal confidence factor in real time. λ t * ; S6. Final Estimation and Closed-Loop Optimization: Optimal confidence factor λ t * The Wasserstein probabilistic consensus fusion framework is injected, and the optimal consensus probability distribution is generated by solving the weighted centroid problem. The mean of the distribution is used as the final quality estimate.

2. The vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks as described in claim 1, characterized in that, In step S1, before starting the vehicle, the following preparatory work is also performed: the vehicle's OBD interface is connected to the CAN bus, and the CAN bus is connected to the computer to realize the acquisition of vehicle operation data; in step S6, the optimal consensus probability distribution is fed back as a supervision signal to the preprocessing network, SARE-Net, EKF and game optimizer to form a closed-loop self-optimizing system.

3. The vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks according to claim 1, characterized in that: The specific steps of step S2 include: S21. Constructing a probabilistic graphical model of signal relationships: Define the CAN acquisition signal vector y at time t. t =[v t ,T e,t , n a,t ,n t,t ] T v t ,T e,t , n a,t , n t,t Let x be the vehicle speed, engine torque, engine speed, and engine output shaft speed at time T, respectively. After removing abnormal signals, the usable signal is x. t Establish a joint probability distribution as a priori: ， In equation (1), φ ij Let ε be the spatial constraint term based on simplified physical relationships, and let ψ be the set of physically associated edges. i For time smoothing; S22. Spatiotemporal Joint Anomaly Detection and Signal Repair: including spatiotemporal graph encoder construction, hierarchical anomaly identification and repair; S23. Task-oriented adaptive fine-tuning: Establish feedback loops, define task-aware loss, freeze the main network, and fine-tune the preprocessing network.

4. The vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks according to claim 3, characterized in that: The specific steps of step S22 include: S221. Spatiotemporal Graph Encoder Construction: The spatiotemporal graph encoder treats each time step signal as a graph node, with the node feature being the signal value; it defines spatial edges based on the probabilistic graphical model and temporal edges based on temporal adjacency; and it learns the hidden state h, which incorporates spatiotemporal neighborhood information, through a spatiotemporal graph convolutional network. i,t ; S222. Hierarchical anomaly identification and repair: S2221. Node-level Anomaly Scoring: Light Quantum Networks Based on h i,t Calculate the anomaly score a i,t ∈[0,1] ; S2222. Graph context repair: For a i,t For suspected abnormal nodes > τ1, initiate repair; repair value The attention-based read head reads from all normal nodes in its spatiotemporal neighborhood (a i,t Generated from the hidden state <τ2), using the core computation of the attention mechanism, a new representation containing global dependencies is generated: ， In equation (2), Attention(h) i,t , h j,t’ Let be the attention weight function, representing the hidden state h of the normal node j in its spatiotemporal neighborhood at time t' when repairing the abnormal node i. j,t’ The importance of the information provided; h i,t It is learned by the spatiotemporal graph encoder, and is the hidden state feature of node i at time t, which contains the association information of its spatiotemporal neighborhood; y j,t’ For with h j,t’ The corresponding "value" vector; S2223. Physical constraint projection: Projecting the repaired complete signal vector Project the signal onto the constrained manifold defined by the probabilistic graphical model in step S21, and solve the optimization problem to maximize the satisfaction of the physical prior.

5. The vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks according to claim 1, characterized in that: Step S3 includes the following specific steps: S31. Input Feature Selection: Select the vehicle longitudinal speed, longitudinal acceleration, motor output torque, transmission ratio, main reducer ratio, transmission efficiency, wheel effective radius, air density coefficient, frontal area, air resistance, rolling resistance coefficient, and road slope as input feature vectors; S32. Network Design: A pre-trained two-layer structure-network dual adaptive dynamic recursive estimation network (SARE-Net) is adopted as a deep network. The pre-processed input feature vector is input, and forward inference is used to obtain the prediction process noise covariance Q at the current time step. k and preliminary estimate m k NN ; The aforementioned two-layer structure-network dual adaptive dynamic recursive estimation network has the following structure: ， In equation (3), x m For input; f m For the Gate of Oblivion; i m For input gate; c m Let t be the cell unit state at time t; o m h represents the cell state output at time t. m The network output at time t; σ is the sigmoid function; tanh is the hyperbolic tangent function; b i b f b c b o W is the bias vector; xi W hi W xf W hf W xc W hc W xo W ho These are the weighting coefficients.

6. The vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks according to claim 1, characterized in that: The specific steps of step S4 include: S41. Extended Kalman Filter (EKF): EKF selects driving speed, road gradient, and vehicle mass as state variables, constructs and updates the state-space expression, and outputs a Gaussian posterior distribution P. k EKF =N(m k EKF , σ k 2,EKF );m k EKF Let σ be the mean of the normal distribution under EKF. k 2,EKF Let V be the variance of the normal distribution under EKF; S42. SARE-Net obtains the data-side uncertainty σ through a small network mapping. k NN Output Gaussian prediction distribution P k NN =N(m k NN , σ k 2,NN ), where the mean m k NN The original network output has a variance σ. k 2,NN The network's "confidence" in this estimate is characterized by an online prediction made by an additional small branch network connected to the hidden layer; S43. Construct the Wasserstein consensus optimization problem to find the optimal consensus distribution P. k * This minimizes the sum of the weighted Wasserstein distances. S44. Design a consensus network ψ ϑ Online solution for P k * The mean and variance of.

7. The vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks according to claim 6, characterized in that: Step S43 specifically includes: for two one-dimensional Gaussian distributions, which have closed-form solutions based on distribution information from two different sources—a neural network (NN) and an extended Kalman filter (EKF)—find a consensus Gaussian distribution P. k * =N(m k * ,σ k 2,* The goal is to minimize the sum of the weighted distances to the two source distributions; the smaller the distance, the closer the distributions are. (4） In equation (4): P k * Represents the optimal distribution; argmin p W2 represents the parameter that minimizes the expression within the parentheses when solving for distribution P. 2 The Wasserstein-2 distance is used to measure the similarity between two distributions. The specific content of step S44 includes designing a learnable consensus network ψ to approximate the mapping and introducing more complex nonlinear relationships to adapt to quality estimation under complex working conditions. (5) In equation (5): m k * ,σ k 2,* Let m represent the optimal distribution parameter pair for the k-th object, where m k * It is the optimal mean, σ k 2,* It is the optimal variance, corresponding to the optimal distribution P. k * Core statistical characteristics; ψ θ It is a parameter fusion mapping function, where the subscript θ represents a specific type of function weighted average or nonlinear combination. Its function is to integrate multiple input parameters into the optimal output parameters; m k NN σ k 2,NN The first reference distribution is the mean and variance of the neural network (NN) distribution; m k EKF σ k 2,EKF The second reference distribution is obtained by using the Extended Kalman Filter (EKF) to determine the mean and variance of the distribution; λ k It is a weighting coefficient, with a value between 0 and 1, used to balance the relative importance of the two distribution bias terms; The training objective (loss function) of the consensus network ψ is the same as the optimization objective in step S43: (6) In equation (6), L ψ This represents the weighted loss value based on the fusion function ψ, which is the specific value of the objective function corresponding to the "minimize" operation in formula (4); the found P k * Substituting the values ​​into the objective function yields the final weighted loss value, which intuitively reflects the overall matching effect between the optimal distribution and the two reference distributions.

8. The vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks according to claim 1, characterized in that: Step S5 includes the following specific steps: S51. Construct a real-time multi-dimensional performance index spectrum, including instantaneous consistency index, short-term smoothness index, and long-term local credit index; S52. Establish a dynamic game optimization model, and define the solution of the confidence factor as a constrained convex optimization problem; S53. Solving for the optimal confidence factor λ online using the Lagrange multiplier method combined with interval search. t * .

9. The vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks according to claim 8, characterized in that: Step S51 includes: The instantaneous consistency index of S511.EKF is: , I A ins The term represents the target metric, with the superscript "ins" indicating "Insensitivity" and the subscript "A" indicating the model identifier. Together, they measure the model's sensitivity to observations; ε t The innovation at time t, which is the residual between the observed value and the model prediction, is the core input of the Kalman filter. H t Represents the observation matrix; P t│t-1 Let be the prior state covariance matrix at time t, which is a prediction of the state covariance at time t based on information from time t-1; δ is a regularization constant used to avoid singularity or zero denominators in the matrix; the instantaneous consistency index of the LSTM network is: I B ins =(σ 2 NN,t +δ) -1 When the observed value is abnormal, Wasserstein-2 increases, resulting in a larger value within the parentheses, leading to... I A ins A smaller Wasserstein-2 value indicates that the model is highly sensitive to the anomalous observation; conversely, when the observation is normal, a smaller Wasserstein-2 value indicates a lower sensitivity. I A ins A large value indicates that the model has low sensitivity to this normal observation; S512. Short-term smoothness index: Calculates the estimator over a past time window W s The first-order difference norm of the internal output quality estimate: ; superscript smo It represents "Smoothness"; the larger the value, the smoother the fluctuation of the sequence and the higher the stability. within the sliding window Ws Summing is performed on each sample, and two adjacent time steps are taken ( t−i+1 and t−i The squared Euclidean distance between adjacent time points measures the amplitude of fluctuations. When the series fluctuations are gentle: a smaller squared distance between adjacent time points indicates a smaller average fluctuation, i.e., a smaller value in parentheses. I smo A larger value indicates higher stability; when the sequence fluctuates wildly: a larger squared distance between adjacent elements indicates a larger average fluctuation, i.e., a larger value in parentheses. I smo Small size indicates low stability; S513. Long-term local credit indicators: I cred =C cred (M current ), I cred The credit metric representing the final output. C cred It is a credit mapping function. M current The current market state C represents the input. cred (M j ) and operating mode M j The estimation error statistic is inversely proportional to the current market state data (Mcurrent), which is a multi-dimensional input describing the market environment. Then, a pre-defined credit mapping function is used. C cred Calculations are performed to convert market conditions into credit metrics that can be directly used for risk assessment. I cred The function can be linear or nonlinear, and its purpose is to quantify the level of credit risk in the current market environment.

10. The vehicle quality estimation method integrating the Wasserstein geometric consensus algorithm and deep networks according to claim 2, characterized in that, In step S6: Forward fusion via λ t * Modulate the distribution shape, construct a weighted Wasserstein barycenter optimization objective, and obtain the consensus distribution. P t * =N ( m t * , (σ t * ) 2 ); In the formula: P t * The optimal probability distribution at time t represents the final result obtained by fusing two reference distributions; N() represents the normal distribution. m t * The optimal mean of this normal distribution is given by the fusion function ψ in formula (5). θ Calculated; (σ) t * ) 2 The optimal variance of this normal distribution, also output by the fusion function of formula (5), is the optimal standard deviation σ. t * The square of.

Images