A vehicle health state evaluation method and system based on internet of vehicles big data

Abstract

The present application belongs to the field of big data, and provides a vehicle health state evaluation method and system based on vehicle networking big data, which comprises collecting multi-source vehicle networking data of the vehicle in the running process, constructing a health state vector sequence of the vehicle within a preset time window; based on the continuous subsequence, extracting dynamic change characteristics reflecting the state evolution trend, forming a time sequence health representation of the vehicle individual based on the dynamic change characteristics; constructing a vehicle group similarity graph, using a graph neural network to extract the feature representation of each vehicle in the group structure to form a group structure representation; fusing the time sequence health representation and the group structure representation to obtain a joint health representation vector of the vehicle; based on the joint health representation vector, calculating the difference of each index between the target vehicle and its adjacent vehicles in the vehicle group similarity graph, identifying the significant deviation index of the health state, and evaluating the current health state of the target vehicle according to the significant deviation index, and outputting the health grade and the corresponding explanation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of big data, and in particular relates to a vehicle health state evaluation method and system based on vehicle networking big data. BACKGROUND

[0002] With the rapid development of intelligent transportation and vehicle networking technology, the number of multi-source data that can be collected during vehicle operation is increasing, including vehicle operating condition parameters, environmental perception data, driving behavior data, positioning trajectory data, etc. These data are uploaded to the vehicle networking platform through the vehicle terminal, forming continuous and rich large-scale time series data, which provides a realistic basis for realizing data-driven vehicle health management.

[0003] In the prior art, the evaluation of vehicle health state mainly adopts the following methods: 1) based on sensor measurement values and preset threshold values for static judgment, used to identify whether the key parameters exceed the normal range; 2) based on historical maintenance data and expert knowledge to establish a rule model or a case library for conditional matching or experience reasoning; 3) using time series modeling methods such as ARIMA, LSTM, etc. to predict the evolution trend of vehicle performance; in addition, some methods use clustering analysis technology to group vehicles according to feature similarity, and identify potential anomalies or deviation indicators by comparing the difference between the vehicle and the center of the cluster it belongs to.

[0004] The above methods have certain limitations. The static threshold method is difficult to adapt to the differences in use environment, vehicle condition and driving behavior of different vehicle individuals, and is prone to false positives or false negatives. The rule-based method relies on manual construction and has high maintenance cost, weak generalization ability, and is difficult to adapt to complex and variable operating conditions. Although the clustering method can introduce a group reference mechanism to a certain extent, it usually relies on manual setting of the number of clusters, and the grouping granularity is fixed and lacks flexibility, making it difficult to depict the continuous similarity structure between nodes, and it is also difficult to realize the effective propagation of information within the group, limiting the accuracy and dynamic response capability of the deviation indicator identification. SUMMARY

[0005] In order to solve the problems in the prior art, the present application provides a vehicle health state evaluation method based on vehicle networking big data, comprising the following steps:

[0006] Step S10, collecting multi-source vehicle networking data of the vehicle during operation, and constructing a health state vector sequence of the vehicle within a preset time window according to the multi-source vehicle networking data;

[0007] Step S20, based on a sliding time window, dividing the health state vector sequence into a plurality of continuous subsequences according to time, extracting dynamic change features reflecting state evolution trend based on the continuous subsequences, and forming a time series health representation of the vehicle individual based on the dynamic change features;

[0008] Step S30, a vehicle group similarity graph is constructed, each vehicle is taken as a node in the graph structure, the edge weight between the nodes is calculated based on the similarity of the vehicle type, vehicle age, use environment and driving behavior, and a graph neural network is used to extract the feature representation of each vehicle in the group structure to form a group structure representation;

[0009] Step S40, the time sequence health representation and the group structure representation are fused to obtain a joint health representation vector of the vehicle;

[0010] Step S50, based on the joint health representation vector, the difference between the target vehicle and its adjacent vehicles in the vehicle group similarity graph in each index is calculated, the significant deviation index of the health state is identified, and the current health state of the target vehicle is evaluated according to the significant deviation index, and the health grade and the corresponding explanation are output.

[0011] Further, the step S20 includes the following sub-steps:

[0012] The health state vector sequence is divided into a plurality of time-continuous subsequences in a preset length sliding time window;

[0013] The dynamic change characteristics including fluctuation amplitude, change rate, cycle intensity, outlier density, abnormal jump times and change direction number are extracted for each subsequence;

[0014] The extracted change characteristics are normalized, and a time sequence health representation vector is formed through vector splicing or weighting mechanism.

[0015] Further, the step S30 includes the following sub-steps:

[0016] The vehicles are multi-dimensionally encoded based on the vehicle type configuration, vehicle age, use environment and driving behavior indicators;

[0017] The similarity score is calculated based on the feature distance between each vehicle, and the edge connection and its weight in the graph are constructed according to the set threshold;

[0018] The constructed vehicle similarity graph is input into a graph neural network model, and the structural embedding representation of each vehicle in the graph structure is output.

[0019] Further, the step S40 includes the following sub-steps:

[0020] The time sequence health representation and the group structure representation are dimensionally aligned;

[0021] The joint health representation vector is generated by splicing operation or using neural network model for feature fusion;

[0022] The joint health representation is used for subsequent index deviation analysis and health state evaluation.

[0023] Further, the step S50 comprises the following sub-steps:

[0024] extracting a joint health representation vector of the target vehicle and its adjacent vehicles in the group graph;

[0025] calculating the standard deviation normalized deviation of the value of the target vehicle in each index dimension and the neighbor mean value;

[0026] identifying a significant deviation index based on the deviation result, and determining a health grade and outputting an explanation according to the number of deviation dimensions and the deviation degree.

[0027] Another aspect of the present application also provides a vehicle health state evaluation system based on vehicle networking big data, characterized in that the system comprises the following modules:

[0028] The acquisition module is used for acquiring multi-source vehicle networking data of the vehicle in the running process, and constructing a health state vector sequence of the vehicle within a preset time window according to the multi-source vehicle networking data;

[0029] The time series modeling module is used for dividing the health state vector sequence into a plurality of continuous subsequences according to time based on a sliding time window, extracting dynamic change characteristics reflecting state evolution trend, and forming a time series health representation of the vehicle individual;

[0030] The graph construction module is used for constructing a vehicle group similarity graph, taking each vehicle as a node in the graph structure, calculating the edge weight between the nodes based on the similarity of vehicle type, vehicle age, use environment and driving behavior, extracting the feature representation of each vehicle in the group structure using a graph neural network, and forming a group structure representation;

[0031] The fusion module is used for fusing the time series health representation and the group structure representation to obtain a joint health representation vector of the vehicle;

[0032] The evaluation module is used for calculating the difference of the target vehicle and its adjacent vehicles in the vehicle group similarity graph in each index based on the joint health representation vector, identifying a significant deviation index of the health state, and evaluating the current health state of the target vehicle according to the significant deviation index, and outputting a health grade and a corresponding explanation.

[0033] Further, the time series modeling module comprises:

[0034] The division sub-module is used for dividing the health state vector sequence into a plurality of time continuous subsequences within a preset length of sliding time window;

[0035] The feature extraction submodule is configured to extract dynamic change characteristics including fluctuation amplitude, change rate, cycle intensity, outlier density, abnormal jump times and change direction quantity for each subsequence;

[0036] The time sequence generation submodule is configured to normalize the extracted change characteristics and form a time sequence health representation vector through vector splicing or a weighting mechanism.

[0037] Further, the atlas construction module specifically comprises:

[0038] The feature encoding submodule is configured to perform multi-dimensional feature encoding on the vehicle based on indicators such as vehicle model configuration, vehicle age, use environment and driving behavior;

[0039] The edge weight calculation submodule is configured to calculate similarity scores based on feature distances between vehicles, and construct edge connections and their weights in the graph according to a set threshold;

[0040] The graph embedding submodule is configured to input the constructed vehicle similarity graph into a graph neural network model, and output a structural embedding representation of each vehicle in the graph structure.

[0041] Further, the fusion module specifically comprises:

[0042] The alignment submodule is configured to perform dimension alignment processing on the time sequence health representation and the group structure representation;

[0043] The fusion submodule is configured to perform feature fusion by splicing operation or using a neural network model to generate a joint health representation vector;

[0044] The representation output submodule is configured to use the joint health representation for subsequent indicator deviation analysis and health status evaluation.

[0045] Further, the evaluation module specifically comprises:

[0046] The adjacency extraction submodule is configured to extract the joint health representation vector of the target vehicle and its adjacent vehicles in the group atlas;

[0047] The difference calculation submodule is configured to calculate the standard deviation normalized deviation of the value of the target vehicle in each indicator dimension and the mean value of the neighbors;

[0048] The grade determination submodule is configured to identify significant deviations in indicators based on the deviation results, and determine the health grade and output an explanation according to the number of deviation dimensions and the degree of deviation.

[0049] The application constructs a vehicle multi-dimensional health state vector sequence in a sliding time window, and extracts dynamic change characteristics in the state evolution process, so that the evaluation system can reflect the continuity and trend of the vehicle running state, effectively avoid misjudgment caused by short-term fluctuations or single-point abnormalities, and enhance the timeliness and individual sensitivity of the health representation.

[0050] The application adopts a graph structure modeling method to construct a vehicle group similarity graph, uses a graph neural network to extract structural embedding features of each vehicle in the graph, and generates a joint health representation vector by combining the time sequence health representation of the vehicle individual. Compared with the traditional clustering method which can only perform rough grouping based on global static features, the graph structure representation can more finely model the multi-dimensional continuous similarity relationship between vehicles, and realize feature propagation and fusion through the graph neural network, significantly improving the accuracy and dynamic adaptability of the group reference.

[0051] On the basis of the joint health representation, the application realizes automatic identification of significantly deviating indicators by calculating the deviation degree of the target vehicle and the adjacent vehicles in each dimension indicator, and generates health grade classification and evaluation description based on the deviation strength, so as to realize more accurate, interpretable and group comparison based health state evaluation, especially suitable for intelligent operation and maintenance and health management scenarios in large-scale Internet of Vehicles environment. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0053] Figure 1 is a flowchart of the method of the application;

[0054] Figure 2 is a schematic diagram of a similarity graph;

[0055] Figure 3 is a schematic diagram of group structure representation generation. DETAILED DESCRIPTION

[0056] The application is described in the preferred manner in combination with the drawings and specific embodiments.

[0057] Reference is made to Figure 1In an embodiment, the present application provides a vehicle health state evaluation method based on Internet of Vehicles big data, which collects multi-source dynamic operation data of vehicles in actual use, and fuses time series modeling and group structure analysis technology to model and compare the health evolution trend of vehicles in a specific time window, so as to realize multi-dimensional intelligent evaluation of the current health state of the target vehicle, and output the interpretable health level and key deviation index information.

[0058] The method is specifically realized by the following steps:

[0059] Step S10, collecting multi-source Internet of Vehicles data of the vehicle in the running process, and constructing a health state vector sequence of the vehicle in a preset time window according to the multi-source Internet of Vehicles data.

[0060] In order to realize dynamic perception and analysis of the health state of the vehicle in the running process, the present application proposes to collect multi-source real-time operation data in the Internet of Vehicles environment, and organize the data into a structured health state vector sequence in a set time window. The core of this method is that time series data modeling can depict the trend and fluctuation of the health state of the vehicle with time, thereby laying a data foundation for subsequent time series feature extraction and group comparison evaluation. Traditional methods mostly rely on single static detection or manual judgment, and it is difficult to capture nonlinear degradation characteristics and implicit fault evolution paths in the running process of the vehicle. The use of dynamic data collection of Internet of Vehicles can realize high-frequency monitoring covering multi-dimensional state parameters, and provide data support for realizing more objective and comprehensive health state evaluation.

[0061] Among them, the multi-source Internet of Vehicles data refers to the multi-type operation parameter data collected by the T-Box module, CAN bus interface, OBD device and other perception and communication nodes in the running process of the vehicle, including but not limited to: engine speed, throttle opening, coolant temperature, oil pressure, vehicle speed, brake state, tire pressure, battery voltage, fault code information (DTC), position information, acceleration information, steering wheel angle, environmental temperature and humidity, driving behavior indicators, etc.; the health state vector sequence refers to synchronously sampling and structuring the above parameters to form multi-dimensional time series data for each vehicle, which is used for subsequent feature analysis.

[0062] In an alternative specific implementation, the step S10 specifically includes the following sub-steps:

[0063] Step S101, setting the sampling period and time window parameters.

[0064] When the system is initialized, two basic parameters are set: sampling period Ts and time window length Lw.

[0065] The sampling period Ts represents the time interval for collecting vehicle state data once per unit time. The optional set values include 1 second, 5 seconds, 10 seconds or more, which can be flexibly configured according to the data transmission bandwidth and the vehicle communication system capability.

[0066] The time window length Lw is used to define the data duration covered by a single evaluation. Common settings include 300 seconds, 600 seconds, 1800 seconds, etc. The actual setting should take into account the data trend capture requirements and the control of computing resources.

[0067] The sampling period Ts and the time window Lw jointly determine the number of sampling points in each evaluation period, denoted as m, which satisfies m = Lw / Ts. The system internally checks the legality of the settings of Ts and Lw to ensure that they are positive integers and can be divided by an integer.

[0068] Step S102, collecting multi-source original data during vehicle operation.

[0069] The system continuously receives and caches various sources of operating parameter data through the on-board intelligent terminal (such as the T-Box module). The data sources include but are not limited to the following modules:

[0070] (1) Engine Control Unit (ECU), which provides engine speed, coolant temperature, throttle opening, intake temperature and other basic operating parameters;

[0071] (2) Brake and Chassis System Control Unit, which provides brake switch status, wheel speed, yaw rate, longitudinal acceleration and other dynamic safety data;

[0072] (3) OBD interface, which reads the current DTC fault codes and related environmental data of the vehicle;

[0073] (4) Position and Environment Sensing Module, which obtains GPS position information, environmental temperature, altitude and other external operating conditions;

[0074] (5) Driving behavior recognition module, which provides sudden acceleration, sudden deceleration, sudden turn frequency, idling time, average speed and other driving habit data.

[0075] All types of data are indexed by time stamp and transmitted to the edge computing unit or cloud platform through standard communication protocols (such as CAN, UDS, MQTT), and are stored synchronously according to the set sampling period Ts.

[0076] Step S103, formatting, time alignment and cleaning processing of original data.

[0077] Due to the inconsistent sampling frequencies and non-uniform time stamps of data from different ECUs and sensors, the system first performs time alignment operation.

[0078] The low-frequency data is uniformly mapped onto the time axis of the sampling period Ts using linear interpolation, Lagrange interpolation, or sample retention.

[0079] Subsequently, anomaly detection and cleaning are performed, specifically including the following two types of operations:

[0080] (1) Missing values are completed using forward filling, backward filling, interpolation, etc.

[0081] (2) Data that exceeds the reasonable physical range (such as engine speed <0 or >10000 rpm) is removed or marked as missing.

[0082] After data cleaning, organize into a multi-dimensional synchronous state vector according to the fixed field order (e.g. [engine speed, throttle opening, coolant temperature, vehicle speed, idle time, …]) for subsequent calculation.

[0083] Step S104, construct a health state vector sequence.

[0084] Integrate the data corresponding to each sampling point into a state vector, denoted as Xt=[x1(t),x2(t),...,xn(t)], where:

[0085] xi(t) represents the value of the i-th parameter at time t;

[0086] n is the total number of collected parameters, usually between 10 and 50;

[0087] t is the current sampling time, satisfying t=t0+k*Ts, k∈[0,m-1].

[0088] Within the entire time window Lw, a total of m state vectors are constructed, organized into an m×n-dimensional matrix sequence, defined as:

[0089] H={Xt1,Xt2,...,Xtm}, where each Xti is an n-dimensional vector.

[0090] This sequence is the health state vector sequence of the vehicle within the current evaluation period, which is used as the input of the time series modeling module to describe the running health trajectory of the vehicle within the time range Lw.

[0091] By setting a uniform sampling period and time window, a high-frequency, structured, multi-dimensional vehicle state data collection method is achieved. Compared with traditional manual sampling or fixed-point detection methods, this step has the advantages of high automation, complete information coverage, adjustable sampling density, and unified structure. By constructing the health state vector sequence, the dynamic evolution trajectory of the vehicle during operation can be recorded completely, providing high-quality data support for subsequent feature extraction, anomaly recognition, and health assessment.

[0092] In a specific example, a 2021 A-class sedan uploads 12 indicators including engine speed, vehicle speed, water temperature, battery voltage, fault codes, etc. to the platform at a 1-second cycle during the running process from 10:00 to 10:05 on May 1, 2025. The platform sets a time window Lw of 300 seconds, and a total of 300 complete records are collected. After each record is aligned, cleaned and structured, a 12-dimensional state vector is formed. The system groups the 300 vectors in chronological order to form a health state vector sequence H, which will serve as input data for subsequent vehicle time series health modeling.

[0093] Step S20, based on the sliding time window, the health state vector sequence is divided into multiple continuous subsequences according to time, the dynamic change characteristics reflecting the state evolution trend are extracted based on the continuous subsequences, and the time series health representation of the vehicle individual is formed based on the dynamic change characteristics.

[0094] In order to fully reveal the state change trend of the vehicle during the running process and effectively capture the evolution characteristics of the health parameters in the time dimension, it is necessary to further divide the health state vector sequence into multiple continuous subsequences and extract the dynamic change characteristics of each subsequence. The health state value at a single time point can only reflect the instantaneous characteristics, and it is difficult to reflect the degradation path or mutation trend of the vehicle during the running process. By re-dividing the original health state sequence through the sliding time window mechanism and extracting statistical or embedded features such as change rate, fluctuation degree, trend direction, etc. from it, potential fault signs and running instability factors can be effectively identified, thereby forming a more time series memory individual health representation.

[0095] Among them, the dynamic change characteristics refer to statistical description quantities or time series embedded representations derived from the time evolution trend of the health state parameters in each subsequence. Common features include mean, standard deviation, first derivative, second derivative, maximum rising slope, maximum falling slope, Fourier spectrum features, or hidden space embedded vectors extracted by deep time series models; the time series health representation of the vehicle individual refers to the vectorized expression of the overall running trend of the vehicle within a certain evaluation period, which is the basis for subsequent structural fusion analysis and evaluation judgment.

[0096] In an optional specific implementation, the step S20 specifically includes the following sub-steps:

[0097] Step S201, set the sliding time window length and the sliding step.

[0098] After the system is initialized, the sliding time window length Lw2 and the sliding step Sw are loaded from the configuration file or the strategy parameter, which are used to define the coverage duration of a single subsequence and the interval between adjacent subsequences, respectively.

[0099] The sliding time window length Lw2 represents the time range covered by each consecutive subsequence, and optional implementations include 60 seconds, 90 seconds, or 120 seconds. The sliding step size Sw represents the interval of window movement, used to control the degree of overlap between subsequences, and optional implementations include 10 seconds, 30 seconds, or Lw2.

[0100] Based on the total length Lseq (i.e., the number of sampling points within the total time range) of the health state vector sequence obtained in the previous step S104, the system calculates the number of constructable continuous subsequences Ns using the following formula:

[0101] Ns=floor((Lseq×Ts−Lw2)÷Sw)+1

[0102] Where Ts is the sampling period and floor represents the floor operation.

[0103] Step S202: Divide the health status vector sequence into multiple continuous subsequences.

[0104] The system extracts a vector subsequence of duration Lw2 at the first sliding window position tstart=0, i.e.:

[0105] H1={X0,X1,...,Xm1}, where m1=Lw2÷Ts

[0106] Subsequently, the system advances the window starting point with a sliding step size Sw, generating the following sequentially:

[0107] H2={Xk,Xk+1,...,Xk+m1}

[0108] H3={X{2k},X{2k+1},...,X{2k+m1}},…

[0109] Where k = Sw ÷ Ts

[0110] The process continues until no more complete subsequences can be formed, ultimately yielding a set of subsequences {H1,H2,...,HNs}.

[0111] Each subsequence Hi is a two-dimensional data block with an m1×n structure, where m1 is the number of sampling points in each subsequence and n is the dimension of each vector, i.e., the total number of parameter terms. Each subsequence maintains the original parameter arrangement order and unit consistency.

[0112] Step S203: Extract dynamic change features for each subsequence.

[0113] The system performs one or any combination of the following feature extraction operations on each subsequence Hi:

[0114] (1) Calculate statistical features: For each column in Hi (i.e., each dimension of health parameter), calculate its sample mean (mean), standard deviation (std), maximum value (max), minimum value (min), median, and other basic statistical indicators.

[0115] For example, for the coolant temperature parameter xt, the statistical features are:

[0116] mean = (1 / m1) x ∑xt

[0117] std = sqrt((1 / m1) x ∑(xt-mean)²)

[0118] (2) Calculate trend features: For each column parameter, calculate the first-order difference sequence Δxt=xt−x{t−1}, from which extract the slope, maximum rising speed, maximum falling speed, etc. The slope can be defined as:

[0119] slope = (xlast−xfirst) ÷ Lw2

[0120] where xfirst and xlast are the values at the start and end of the subsequence, respectively.

[0121] (3) Calculate high-order derivative or frequency domain features: For vehicle dynamic response class parameters (such as speed, acceleration, etc.), calculate the second-order difference features or perform a Fast Fourier Transform (FFT) on them, extracting features such as main frequency amplitude, spectral center of gravity, etc.

[0122] (4) Deep embedding feature extraction: input the subsequence Hi into a pre-trained sequence modeling network, such as a Long Short-Term Memory network (LSTM) or an Autoencoder structure (Autoencoder), to extract its hidden layer activation vector as a compressed representation. The vector dimension can be defined by the network architecture, usually between 16 and 128 dimensions.

[0123] Each subsequence feature is finally merged into a fixed-length feature vector Fi, the length of which is determined by the number of feature types and dimensions. For example, each subsequence has 12 parameters, and each parameter extracts 5 features, so Fi is a 60-dimensional vector.

[0124] Step S204, fuse all subsequence features to form a time series health representation of the vehicle individual.

[0125] The system inputs the feature vector set {F1, F2,..., FNs} of all subsequences into the fusion module in chronological order.

[0126] Optional fusion methods include:

[0127] (1) Average pooling: average merge all subsequence features by dimension:

[0128] Z = (1 / Ns) x∑Fi

[0129] (2) Max pooling: take the maximum value of each dimension and merge:

[0130] Zj = max(F1j, F2j,..., FNsj), for all j ∈ [1, length(F)]

[0131] (3) MLP fusion: concatenate the subsequence features and input them into a multi-layer perceptron network, output a compressed fusion vector of uniform length.

[0132] The final vector Z is a time-series health representation of the vehicle individual, which is the unified input feature for subsequent steps of graph fusion and evaluation.

[0133] This step extracts the micro change trend under high time resolution, effectively enhancing the sensitivity and dynamic response of the system to the vehicle operating state. Through multi-scale feature extraction and multi-mode feature fusion, this time-series health representation not only reflects the current value of the parameter, but also contains its evolution direction, change speed and historical fluctuation information, thereby providing a more discriminative and predictive feature basis for subsequent health state analysis.

[0134] In one specific example, a 2021 A-class sedan equipped with T-Box terminal, ECU, CAN bus and other submodules, during the operation from 10:00 to 10:05 on May 1, 2025, 12 operating parameters were continuously collected at a sampling period Ts of 1 second:

[0135] Engine speed (unit: rpm)

[0136] Coolant temperature (unit: °C)

[0137] Vehicle speed (unit: km / h)

[0138] Accelerator pedal opening (unit: %)

[0139] Brake switch state (0 / 1)

[0140] Longitudinal acceleration (unit: m / s²)

[0141] Lateral acceleration (unit: m / s²)

[0142] Battery voltage (unit: V)

[0143] Throttle opening (unit: %)

[0144] Fault code number (unit: pieces)

[0145] DTC severity rating (unit: 0~100)

[0146] Ambient temperature (unit: °C)

[0147] In the 300-second evaluation period, 300 groups of 12-dimensional synchronous vectors are collected, forming a health state vector sequence H with a dimension of 300x12.

[0148] The system sets a sliding time window Lw2 of 60 seconds and a sliding step Sw of 30 seconds, resulting in 9 overlapping sub-sequences, each containing 60 sampling points.

[0149] For each sub-sequence, extract dynamic change features:

[0150] Take sub-sequence H1 (corresponding to 10:00:00~10:01:00) as an example. For each parameter, extract the following 5 dynamic change features:

[0151] Mean (mean): arithmetic mean of 60 sampling points;

[0152] Standard deviation (std): measures the fluctuation amplitude of the parameter within the sub-sequence;

[0153] Maximum upward slope (maxslopeup): the maximum positive rate of change of the parameter between any two consecutive time points;

[0154] Maximum downward slope (maxslopedown): the maximum negative rate of change of the parameter between any two time points;

[0155] First-order derivative mean (meandiff): the mean value of the first-order difference (rate of change) sequence of the parameter, reflecting the trend direction.

[0156] For example, assume that in sub-sequence H1, the engine speed features are as follows:

[0157] mean = 1735.2 rpm

[0158] std = 112.6 rpm

[0159] maxslopeup = +148.0 rpm / s

[0160] maxslopedown = -205.0 rpm / s

[0161] meandiff = -15.3 rpm / s

[0162] Coolant temperature features are as follows:

[0163] mean = 88.5 °C

[0164] std = 1.2 °C

[0165] maxslopeup = +0.5 °C / s

[0166] maxslopedown = -0.4 °C / s

[0167] meandiff = +0.02 °C / s

[0168] …

[0169] Similarly, 5 features are extracted for each of the 12 parameters, forming a 12x5 = 60-dimensional feature vector F1, denoted as:

[0170] F1 = [1735.2, 112.6, 148.0, -205.0, -15.3,

[0171] 88.5, 1.2, 0.5, -0.4, +0.02,

[0172] … (5 features for the remaining 10 parameters are arranged in turn)

[0173] The system continues to extract feature vectors F2 to F9 for sub-sequences H2 to H9, forming a feature set {F1, F2,..., F9}, each of which is 60-dimensional.

[0174] Fusion of all sub-sequence features forms the time-series health representation:

[0175] The system performs an average pooling operation on the above 9 feature vectors, and averages by dimension to generate a unified time-series health representation vector Z for the current evaluation period of the vehicle:

[0176] Zj = (1 / 9) x ∑{i=1 to 9} Fij, j ∈ [1, 60]

[0177] For example, if the mean of the engine speed in each sub-sequence is:

[0178] [1735.2, 1780.3, 1812.0, 1750.8, 1695.4, 1662.1, 1630.2, 1650.0, 1705.6]

[0179] Then the 1st dimension in the final Z is:

[0180] Z1 = (1735.2 +... + 1705.6) / 9 ≈ 1702.4 rpm

[0181] The final time-series health representation Z is:

[0182] Z = [1702.4, 105.7, 142.1, -198.3, -12.8,

[0183] 88.3, 1.1, 0.4, -0.3, 0.01,

[0184] …

[0185] [The remaining 50 dimensions are sequentially arranged]

[0186] This example shows how to extract fine-grained dynamic features from a real-sampled health state vector sequence according to time division sliding window, and construct a complete process of unified time sequence health representation. Each dimension of the feature has physical meaning, traceability and interpretability. Through this processing method, the system can not only describe the current vehicle running state, but also depict the stability, change rate and direction of the state, providing data basis for subsequent health level assessment and anomaly identification.

[0187] Step S30, constructing a vehicle group similarity graph, taking each vehicle as a node in the graph structure, calculating the edge weight between nodes based on the similarity of vehicle type, vehicle age, use environment and driving behavior, and using a graph neural network to extract the feature representation of each vehicle in the group structure to form a group structure representation.

[0188] In order to more accurately evaluate the health state of a single vehicle, it is necessary to establish a context information reference system within the vehicle group. By constructing a vehicle group similarity graph, a large number of vehicles with different characteristics can be organized into a graph structure, so as to use the feature association relationship between similar vehicles to structureally enhance the representation of the health state of individual vehicles. The basic principle of this method is that the running state of a vehicle is not only determined by its own parameters, but also affected by the common behavior patterns and technical attributes of the group it belongs to. By modeling the similarity between vehicles using a graph structure, the system can refer to the running rules of neighboring vehicles when evaluating individual states, thereby improving the discriminability and generalizability of the model.

[0189] Among them, the vehicle group similarity graph refers to a graph structure constructed by multiple vehicles as nodes, each node in the graph represents a vehicle, and the edge represents the similarity connection relationship between vehicles; the graph structure refers to a mathematical structure composed of a node set and an edge set, used to represent the relationship between entities; the node edge weight refers to the weight information carried by the edge connecting two nodes, reflecting the similarity between the two vehicles; the group structure representation refers to the embedding vector of the target vehicle in the group graph structure, used to supplement its individual characteristics and enhance its understanding of the structural environment.

[0190] In an optional specific implementation, the step S30 specifically includes the following sub-steps:

[0191] Step S301, determining the node set of the vehicle graph structure.

[0192] The system first acquires all vehicle data participating in health state modeling in the current evaluation period, denoted as a vehicle set V = {v1, v2,..., vN}, where N is the total number of vehicles, and each vehicle vi is a node in the graph. For each node vi, the system extracts the following types of information to construct its initial attributes:

[0193] Typei: Based on the VIN code, the vehicle brand, series, displacement, fuel type, and drive form parameters are parsed to form a fixed-length one-hot vector. For example, the type code of the BYD Qin PLUS plug-in hybrid version may be [0, 0, 1, 0, 0,...].

[0194] Agei: The age parameter is calculated by subtracting the vehicle's first registration date from the current evaluation time point, and the result is expressed in months, such as 26 months.

[0195] Envi: Contains the geographical type (city / mountain / highland), temperature interval (normal / high / low), humidity level (dry / medium / humid), etc. of the vehicle's main operating area, and is represented by a combined coding method.

[0196] Driveri: A floating-point vector generated by the previous module statistics, including the number of sudden accelerations, sudden decelerations, steering angle change rates, idle speed ratios, and average brake intensities per unit time.

[0197] Zi: A 60-dimensional time feature representation of the vehicle's running state formed in the previous step, used to describe the recent running dynamic trend of the vehicle.

[0198] The system concatenates the above 5 types of information into the node initial attribute vector:

[0199] hi⁽ 0 ⁾=[Typei∥Agei∥Envi∥Driveri∥Zi]

[0200] Where ∥ represents the vector concatenation operation.

[0201] Step S302, calculate the similarity between vehicles and generate the edge set of the graph.

[0202] For any two vehicles vi and vj, the system calculates their comprehensive similarity Sij according to the following sub-steps:

[0203] Step S3021, type similarity calculation (S1ij)

[0204] If the type codes of vi and vj are completely identical, then S1ij = 1; if only the brand is consistent but the configuration is different, then S1ij = 0.5; otherwise, set it to 0.

[0205] The implementation is to compare whether the one-hot vectors of Typei and Typej are completely matched or prefix matched.

[0206] Step S3022, age similarity calculation (S2ij)

[0207] The age similarity is defined using the normalized reciprocal function:

[0208] S2ij=1 / (1+|Agei−Agej|)

[0209] where Agei and Agej are the ages (in months) of vi and vj, respectively. The closer the value range is, the higher the similarity is, and the maximum is 1.

[0210] Step S3023, environment similarity calculation (S3ij)

[0211] Take Envi and Envj as three-classification combined vectors, and use the Hamming distance method to calculate the dimension difference denv:

[0212] denv=Hamming distance(Envi,Envj)

[0213] S3ij=1−(denv / maxd)

[0214] where maxd is the maximum possible dimension number, ensuring that S3ij∈[0,1].

[0215] Step S3024, driving behavior similarity calculation (S4ij)

[0216] The cosine similarity method is used:

[0217] S4ij=(Driveri•Driverj) / (||Driveri||×||Driverj||)

[0218] where • is the dot product operation, and ||·|| is the L2 norm. The result is in the interval of -1 to +1, which is then linearly mapped to the interval of [0,1].

[0219] Step S3025, comprehensive similarity calculation (Sij)

[0220] The weighted average method is used to summarize:

[0221] Sij=w1·S1ij+w2·S2ij+w3·S3ij+w4·S4ij

[0222] where w1, w2, w3, and w4 are system-set weight coefficients, and the default values can be [0.3, 0.2, 0.2, 0.3], satisfying ∑w k =1.

[0223] Step S3026: Determine whether to establish an edge connection.

[0224] Set a similarity threshold Tsim, for example, Tsim=0.75. When Sij≥Tsim, establish an edge eij between vi and vj, and record the edge weight Aij=Sij. The edge weight is written into the adjacency matrix A. For node pairs where no edge has been established, set Aij=0.

[0225] Step S303: Construct the graph structure and initialize the graph neural network input.

[0226] The graph structure is defined as G=(V,E,A), where:

[0227] V is the set of vehicle nodes;

[0228] E is the set of connecting edges;

[0229] A is an N×N adjacency matrix that records the weights of all edges.

[0230] Initialize each node as hi⁽ 0 Arranged row by row, they form the input feature matrix H⁽ 0 ⁾∈ℝ^{N×d}, where d is the dimension of the attribute vector of each node.

[0231] like Figure 2 The diagram shown is a schematic representation of a similarity map according to the present invention.

[0232] Step S304: Extract structure-aware features using a graph neural network, the principle of which is as follows. Figure 3 As shown, this is achieved through the following steps.

[0233] The system uses a graph convolutional neural network (GCN) for structure-aware representation learning, and the computation rules are as follows:

[0234] H⁽ l+1 ⁾=σ(·H⁽ l ⁾·W⁽ l ⁾)

[0235] in:

[0236] H⁽ l ⁾∈ℝ^{N×dl} is the matrix representing the nodes of the l-th layer;

[0237] W⁽ l ⁾∈ℝ^{dl×d{l+1}} are the trainable weights of the l-th layer;

[0238] σ is the activation function, such as ReLU;

[0239] The symmetric normalized form of the adjacency matrix is: D^(-1 / 2)·(A+I)·D^(-1 / 2), where D is the degree matrix and I is the identity matrix.

[0240] After performing L-layer propagation, the final structural representation vector of each node is obtained:

[0241] Si=hi⁽ L ⁾

[0242] where Si∈ℝ^{dL} is the embedding representation of the target vehicle vi in the group graph structure, i.e., the group structure representation.

[0243] In the batch evaluation task at 10:00 on May 1, 2025, the vehicle fleet management platform system collected time-series health vectors and driving behavior data from 1000 vehicles. The target vehicle v325 is a BYD Qin PLUS DM-i with a vehicle age of 26 months, running in an urban area, and its driving behavior is mainly characterized by low-frequency sudden acceleration and high-frequency idling.

[0244] The system calculates the comprehensive similarity between v325 and the remaining 999 vehicles, and the results show that Sij for v127, v402, and v678 are 0.82, 0.76, and 0.85, respectively, all higher than Tsim=0.75, so three edges are established.

[0245] Under the two-layer GCN structure, with parameter configuration of input dimension 100 and output dimension 64, the system finally extracts the structural representation S325∈ℝ^{64} of the target vehicle v325. This structural vector will be fused with the time-series health vector Z325 in the subsequent health deviation analysis and grade evaluation process.

[0246] Compared with traditional clustering methods, graph structure representation has higher modeling ability and adaptability in vehicle health state evaluation. Traditional clustering methods usually treat vehicles as independent samples and only group vehicles based on distance metrics (such as Euclidean distance, cosine similarity, etc.) in the global space. The results are highly dependent on the pre-set of the number of clusters, the selection of the initial center, and the assumption of the distance function, making it difficult to capture complex local structure relationships and heterogeneous feature distributions, and lacking the ability to propagate information across groups.

[0247] In contrast, graph structure representation models each vehicle as a node and quantifies its similarity with other vehicles in multiple dimensions (such as vehicle configuration, vehicle age, driving behavior, and usage environment) through flexible edge weight definition, thereby constructing a structured graph that reflects the relationship between vehicle groups. This representation not only preserves individual characteristics but also explicitly depicts the local relationship network between vehicles, allowing nodes to aggregate features and enhance context through graph neural networks.

[0248] The graph neural network allows each node to aggregate the features of the neighbor nodes in multiple propagation layers, realizing the near neighbor learning. This means that the system can utilize the health state trend of the structural neighbors to assist the state judgment of the current vehicle, and has a natural context enhancement mechanism, while the clustering method can only give the classification result and cannot refine the comparison within the group.

[0249] The clustering method generally assumes that the sample distribution within the cluster has a certain simple structure (such as Gaussian distribution), which is not suitable for nonlinear boundaries. The graph structure representation cooperates with the multi-layer graph neural network to capture complex nonlinear correlation, which is especially suitable for describing the non-convex similarity structure formed by vehicles in a high-dimensional and heterogeneous parameter space.

[0250] The graph structure can obtain embedding representation through structural feature self-encoding or contrast learning without relying on labels; at the same time, it can easily introduce partial labeled information to realize semi-supervised learning, while the clustering method is generally difficult to naturally integrate with label information.

[0251] Further, the graph structure can be dynamically updated with the addition of vehicles, and the incremental construction of adjacent edges can be calculated locally without the need for global retraining; while the clustering method often needs to be re-divided as a whole, which has high computational cost and high update cost.

[0252] The graph structure representation in this step not only has richer expression ability and stronger context modeling ability, but also can realize more efficient, robust and interpretable health state evaluation in a heterogeneous, dynamic and complex Internet of Vehicles environment, which has obvious practical value superior to traditional clustering algorithms in large-scale vehicle fleet management, predictive maintenance, intelligent travel and other applications.

[0253] Step S40, the time sequence health representation is fused with the group structure representation to obtain a joint health representation vector of the vehicle.

[0254] In the process of vehicle health state evaluation, relying only on the time sequence features of the vehicle itself may not fully reflect its relative health level or abnormality degree, especially when the individual numerical value is at the edge but not easy to be identified. Therefore, this step fuses the time sequence health representation of the vehicle with its graph embedding representation in the group structure to generate a joint health representation vector containing individual dynamic features and group structure features, thereby realizing more comprehensive and accurate modeling of the vehicle health state. The fusion operation makes the model retain the running trend information of the vehicle itself and introduces the relative position and relationship information between similar vehicles, has context awareness ability, and is beneficial to identifying relatively deviated individuals or potential abnormal samples.

[0255] The joint health representation vector in this step refers to the unified feature representation formed by integrating the above two representations in the vector space, which is used for subsequent health state analysis, deviation detection or grade evaluation tasks.

[0256] In an alternative implementation, the step S40 specifically comprises the following sub-steps:

[0257] Step S401, unify the dimension range of the time sequence representation and the structure representation.

[0258] The system first acquires the time sequence health representation vector Zi and the group structure representation vector Si of the target vehicle vi, respectively, Zi ∈ ℝ^d1, Si ∈ ℝ^d2, which are usually inconsistent in dimension due to different sources. To achieve effective fusion, it is necessary to unify them to the same vector dimension d.

[0259] Step S4011, acquire the original vector dimension.

[0260] The system reads the storage module or intermediate cache structure and acquires:

[0261] Zi, the time sequence health representation generated in step S20, for example, dimension d1 = 60;

[0262] Si, the output of the graph neural network in step S30, for example, dimension d2 = 64.

[0263] Step S4012, construct a linear transformation network for dimension alignment.

[0264] The system initializes two sets of linear mapping networks (optionally single-layer or multi-layer perceptrons) for dimension conversion of the vectors. Taking single-layer linear projection as an example, the calculation is as follows:

[0265] Zi'=Wz×Zi+bz

[0266] Si'=Ws×Si+bs

[0267] Wherein:

[0268] Wz ∈ ℝ^{d×d1} is the weight matrix of Zi;

[0269] Ws ∈ ℝ^{d×d2} is the weight matrix of Si;

[0270] bz ∈ ℝ^d, bs ∈ ℝ^d are bias terms;

[0271] d is the target unified dimension, which can be set to 48, 64, or 128, etc. commonly used embedding length.

[0272] The weights Wz, Ws and the bias terms are given by the initialization strategy and are updated during the training process.

[0273] Step S4013, perform transformation and generate standard representation.

[0274] After the above transformation, the two vectors after standardization are obtained:

[0275] Zi'∈ℝ^d, Si'∈ℝ^d

[0276] Both are in the same vector space, have additivity or concatenation consistency, and are convenient for subsequent fusion.

[0277] Step S402, performing vector fusion operation to generate joint health representation.

[0278] The system fuses Zi' and Si' to form a joint health representation vector Hi, which is used to comprehensively represent the dynamic trend and structural position characteristics of the vehicle individual.

[0279] Step S4021, selecting a fusion strategy.

[0280] The system selects one of the following fusion strategies according to the deployment strategy or model configuration file:

[0281] Method one, concatenation fusion:

[0282] Hi=[Zi'∥Si'], obtaining a joint vector with a dimension of 2d; suitable for scenarios where no information is lost, but the subsequent calculation cost is increased.

[0283] Method two, weighted fusion:

[0284] Hi=α×Zi'+(1−α)×Si'

[0285] Where α∈[0,1] is the fusion weight, the default value is 0.5 or is automatically learned from training data; suitable for keeping the vector dimension constant and balancing the weight of the two kinds of information.

[0286] Method three, attention fusion (optional implementation scheme):

[0287] Use attention mechanism to calculate fusion weight and construct the following expression:

[0288] α=Softmax(Watt×[Zi'∥Si'])

[0289] Hi=α1×Zi'+α2×Si'

[0290] Where Watt is the attention weight matrix, and α1+α2=1. This method dynamically determines whether the vehicle should focus more on its own features or the structural context in the current period.

[0291] Step S4022, performing fusion operation and outputting the result.

[0292] The system completes vector fusion operation according to the set strategy to generate a fusion vector Hie R^d or R^2d (depending on the strategy), and retains its identifier in the memory for subsequent steps.

[0293] In step S403, the joint health representation vector is output and normalized.

[0294] In order to improve the numerical stability and consistency of the subsequent evaluation module, the system performs normalization processing on the fused vector H.

[0295] In step S4031, the vector norm is calculated.

[0296] The system calculates the L2 norm of H, i.e.:

[0297] ||H|| = sqrt(∑j(Hij)^2)

[0298] Where Hij represents the jth component in H.

[0299] In step S4032, vector normalization is performed.

[0300] The normalized representation is calculated as:

[0301] Hnorm = H ÷ ||H||

[0302] After normalization, Hinorm e R^d (or 2d), each dimension of which is in a unified scale range, avoiding domination of small dimensions by large dimensions.

[0303] In step S4033, it is output to the subsequent module.

[0304] Hnorm is used as the final joint health representation vector and written into the intermediate representation storage structure for health deviation analysis and level judgment in step S50.

[0305] This step realizes the complementary integration of multi-dimensional information by fusing the running trend characteristics of the vehicle itself with its behavior pattern in the group structure. This joint health representation not only has the ability to depict time evolution, but also has the ability to express relative position under group reference, making the system more sensitive to small deviations and more robust to individual fluctuations when performing subsequent health status evaluation, and supporting accurate and efficient health level inference in high-dimensional data environment.

[0306] In one specific example, taking the aforementioned target vehicle v325 as an example, its time series health vector Z325 is 60-dimensional, and its structure representation vector S325 is 64-dimensional. The system sets the target dimension d = 48, and uses the following linear mapping:

[0307] Z325' = Wz x Z325 + bz

[0308] S325'=Ws×S325+bs

[0309] Two normalized vectors Z325' and S325' are obtained, both of which are 48-dimensional. The fusion strategy is set to weighted summation, and the weight a = 0.6. The following is calculated:

[0310] H325=0.6×Z325'+0.4×S325'

[0311] After the vector H325 is subjected to L2 normalization processing, H325norm is obtained. The joint vector is recorded to the vehicle evaluation buffer area and is used as the health state deviation identification input of the next step S50. The vector contains not only the continuous operation performance of v325 in the current operation cycle, but also the relative structural position of v325 in all similar vehicles

[0312] In step S50, based on the joint health representation vector, the difference between the target vehicle and its adjacent vehicles in the vehicle group similarity graph in each index is calculated, the significant deviation index of the health state is identified, and the current health state of the target vehicle is evaluated according to the significant deviation index, and the health level and the corresponding explanation are output.

[0313] In the vehicle health state evaluation process, the operation characteristics of an individual vehicle only provide its own state expression, and lack of reference benchmark for whether it is in a healthy deviation state. Therefore, in order to enhance the relativity and context rationality of the evaluation, this step compares the joint health representation between the target vehicle and its adjacent vehicles in the vehicle group similarity graph, identifies the significant difference in some index dimensions, and then evaluates whether its current operation state is abnormal or deviates from the average level of the group, and finally outputs the specific health level and evaluation explanation. This method can avoid misjudgment due to differences in vehicle characteristics or group average deviation, and improve the accuracy, robustness and interpretability of health state judgment.

[0314] In an alternative specific implementation, the step S50 specifically includes the following sub-steps:

[0315] In step S501, the joint health representation vectors of the target vehicle and its adjacent vehicles are extracted.

[0316] In step S5011, the joint health representation of the target vehicle is read.

[0317] The system finds and extracts the vector Hi∈ℝ^d in the joint health representation vector library through the target vehicle identifier vi, where d represents the dimension of the health representation vector, such as 48 or 64.

[0318] In step S5012, the neighbor set of the target vehicle in the graph is obtained.

[0319] The system uses the adjacency matrix A of the vehicle group similarity graph to identify the set of directly connected nodes Ni of the target vehicle in the graph, defined as:

[0320] Ni = {vj|Aij>0 and j≠i}

[0321] Here, Aij represents the edge weight between vehicle vi and vj, which is usually calculated based on indicators such as vehicle type, vehicle age, and driving behavior. If the number of nodes in Ni is less than a set threshold (e.g., 3), the system can optionally expand outward to two-hop adjacent nodes.

[0322] Step S5013: Extract the health vector set of adjacent vehicles.

[0323] The system sequentially extracts the joint health representation Hj of neighboring vehicles vj∈Ni from the representation vector library, forming an adjacency vector set:

[0324] HN={Hj∈ℝ^d|∀vj∈Ni}

[0325] The system caches this collection for subsequent statistical analysis operations.

[0326] Step S502: Calculate the degree of difference in each indicator dimension.

[0327] Step S5021: Calculate the mean and standard deviation of adjacent vehicles dimension by dimension.

[0328] For each dimension k∈{1,…,d} in the set HN, the system calculates the neighbor mean μk and standard deviation σk respectively:

[0329] μk=(1 / |Ni|)×Σ{vj∈Ni}Hj[k]

[0330] σk=sqrt[(1 / |Ni|)×Σ{vj∈Ni}(Hj[k]−μk)^2]

[0331] If σk=0 is encountered (meaning that the neighboring dimensions are completely identical), the system will automatically skip that dimension to avoid division by zero anomalies.

[0332] Step S5022: Calculate the standardized deviation of the target vehicle in each dimension.

[0333] For each dimension k of the target vehicle vector Hi, calculate the deviation from the neighbor mean:

[0334] δk=|Hi[k]−μk| / σk

[0335] The system generates a set of deviation vectors Δ={δ1,δ2,...,δd}, which are used to determine whether the behavior deviates from the group behavior.

[0336] Step S503: Identify significant deviations from health status.

[0337] Step S5031, set the deviation threshold.

[0338] The system sets a deviation threshold θ for determining whether a significant deviation is formed. θ can be a static threshold (e.g. 1.5), or can be automatically adjusted by a training data set.

[0339] Step S5032, determine whether the deviation is dimension by dimension.

[0340] The system checks the deviation δk of each dimension k in turn, and if it satisfies:

[0341] δk≥θ, then the k dimension is added to the significant deviation dimension set Kabn.

[0342] Finally, Kabn={k|δk≥θ and σk≠0} is obtained.

[0343] Step S5033, record the deviation index information.

[0344] The system generates a structured record for each significant deviation dimension, including:

[0345] Index name (mapped from dimension number)

[0346] Deviation value δk

[0347] Hi[k] current value

[0348] μk neighbor mean

[0349] Relative deviation percentage: (|Hi[k]−μk| / μk)×100%

[0350] The record will be used to evaluate the explanation generation module.

[0351] Step S504, output the health level and corresponding explanation.

[0352] Step S5041, set the health level classification criteria.

[0353] The system defines health level criteria, divided into normal, mild abnormality and severe abnormality, according to the following criteria:

[0354] If Kabn=∅, the level is determined to be normal

[0355] If |Kabn|∈[1,m] and max(δk)∈[θ,θhigh], the level is mild abnormality

[0356] If max(δk)≥θhigh or |Kabn|≥m, the level is severe abnormality

[0357] Where m is the upper limit of the number of significantly deviated dimensions (e.g., 3), and θhigh is the threshold for severe anomalies (e.g., 3.0).

[0358] Step S5042: Generate evaluation description text.

[0359] The system generates descriptions for each indicator in Kabn, including the indicator name, direction of deviation, numerical difference, and recommendations. For example:

[0360] The deviation from the group average in the 12th dimension of engine thermal efficiency fluctuation is 38%, and it is recommended to check whether there is periodic instability in the fuel injection system.

[0361] The description can be automatically generated based on the mapping template for system display or report export.

[0362] Step S5043: Structure the output results.

[0363] The system will structure the assessment levels and descriptions as follows:

[0364] {

[0365] vehicleid:v325,

[0366] health level: Mildly abnormal.

[0367] Deviation metrics: [Engine thermal efficiency fluctuation, battery voltage mean drift]

[0368] details:[

[0369] {dimension:12,name:engine thermal efficiency fluctuation,delta:2.5,recommendation:check fuel injection system}

[0370] {dimension:23,name:Battery voltage mean drift,delta:1.9,recommendation:Check charging management unit} ]

[0372] }

[0373] The results can be sent to the vehicle management system, the visualization front-end, or the maintenance dispatch system.

[0374] In a specific example, for the target vehicle v325, its joint health representation H325 is a 48-dimensional vector, and its neighboring vehicles in the graph are v127, v402, and v678. The system obtains the joint representation vectors H127, H402, and H678 of the three neighboring vehicles respectively, and calculates the neighbor mean μk and standard deviation σk for each dimension.

[0375] On the 9th dimension of battery discharge fluctuation characteristics, H325[9]=0.82, while the neighbor mean μ9=0.55, the standard deviation σ9=0.10, so the deviation δ9=|0.82−0.55|÷0.10=2.7, which is greater than the set threshold θ=1.5, and it is determined that the dimension deviates significantly.

[0376] According to the number of deviating dimensions and the maximum δ value, the system determines the health level of v325 as mild abnormality, and outputs the evaluation statement as: there is a large deviation in the battery discharge fluctuation characteristics, and the target vehicle index is 47% higher than the average of similar vehicles, and it is recommended to detect the consistency of the battery pack thermal management system and the controller.

[0377] In another embodiment, the present application also provides a vehicle health state evaluation system based on Internet of Vehicles big data, comprising:

[0378] The acquisition module is configured to acquire multi-source Internet of Vehicles data of the vehicle during operation, and construct a health state vector sequence of the vehicle within a preset time window according to the multi-source Internet of Vehicles data;

[0379] The time series modeling module is configured to divide the health state vector sequence into a plurality of continuous subsequences according to time based on a sliding time window, extract dynamic change characteristics reflecting state evolution trend, and form a time series health representation of the vehicle individual;

[0380] The graph construction module is configured to construct a vehicle group similarity graph, take each vehicle as a node in the graph structure, calculate the edge weight between nodes based on the similarity of vehicle type, vehicle age, use environment and driving behavior, extract the feature representation of each vehicle in the group structure using a graph neural network, and form a group structure representation;

[0381] The fusion module is configured to fuse the time series health representation and the group structure representation to obtain a joint health representation vector of the vehicle;

[0382] The evaluation module is configured to calculate the difference between the target vehicle and its adjacent vehicles in the vehicle group similarity graph in each index based on the joint health representation vector, identify the significant deviation index of the health state, and evaluate the current health state of the target vehicle according to the significant deviation index, and output the health level and the corresponding statement.

[0383] It should be noted that the above explanation of the vehicle health state evaluation method based on Internet of Vehicles big data is also applicable to the device of the present application, which will not be described here.

[0384] Those skilled in the art can clearly understand that the units and algorithm steps described in the embodiments disclosed herein can be realized by electronic hardware, computer software and a combination of the two. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0385] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0386] In several embodiments provided in the present application, any function realized in the form of a software function unit and sold or used as an independent product can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory; hereinafter referred to as: ROM), a random access memory (Random Access Memory; hereinafter referred to as: RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0387] The above is only a specific implementation of the present application. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. The part of the module structure not specifically mentioned in the present application is subject to the content recorded in the prior art. The prior art mentioned in the foregoing background section and the specific embodiment section of the present application can be used as a part of the present application to understand the meaning of some technical features or parameters.

Claims

1. A method for assessing vehicle health status based on Internet of Vehicles big data, characterized in that, The method includes the following steps: Step S10: Collect multi-source vehicle network data during vehicle operation, and construct a health status vector sequence of the vehicle within a preset time window based on the multi-source vehicle network data; The multi-source vehicle network data includes engine speed, throttle opening, coolant temperature, oil pressure, vehicle speed, braking status, tire pressure, battery voltage, fault code information, location information, acceleration information, steering wheel angle, ambient temperature and humidity, and driving behavior indicators. Step S20: Based on the sliding time window, the health state vector sequence is divided into multiple continuous subsequences according to time. Dynamic change features reflecting the state evolution trend are extracted based on the continuous subsequences. A time-series health representation of an individual vehicle is formed based on the dynamic change features. Step S20 includes the following sub-steps: Within a sliding time window of a preset length, the health status vector sequence is divided into multiple time-continuous subsequences; For each subsequence, dynamic change features including fluctuation amplitude, rate of change, periodic intensity, outlier density, number of anomalous jumps, and number of change directions are extracted; The extracted change features are normalized and then formed into a time-series health representation vector through vector concatenation or weighting mechanisms. Step S30: Construct a vehicle group similarity graph, taking each vehicle as a node in the graph structure. Calculate the edge weights between nodes based on the similarity of vehicle type, vehicle age, usage environment, and driving behavior. Use a graph neural network to extract the feature representation of each vehicle in the group structure to form a group structure representation. Step S40: The temporal health representation and the group structure representation are fused to obtain the joint health representation vector of the vehicle; Step S50: Based on the joint health representation vector, calculate the differences between the target vehicle and its neighboring vehicles in the vehicle group similarity map in each indicator, identify significant deviation indicators of health status, evaluate the current health status of the target vehicle according to the significant deviation indicators, and output the health level and corresponding description.

2. The vehicle health status assessment method based on vehicle network big data according to claim 1, characterized in that, Step S30 includes the following sub-steps: Multidimensional feature encoding of vehicles is performed using indicators based on vehicle configuration, vehicle age, usage environment, and driving behavior; Similarity scores are calculated based on the feature distances between vehicles, and edge connections and their weights are constructed in the graph according to a set threshold. The constructed vehicle similarity map is input into the graph neural network model, which outputs the structural embedding representation of each vehicle in the graph structure.

3. The vehicle health status assessment method based on vehicle network big data according to claim 1, characterized in that, Step S40 includes the following sub-steps: Dimensional alignment is performed on the temporal health representation and the population structure representation; A joint health representation vector is generated by employing concatenation operations or using a neural network model for feature fusion. The combined health representation is used for subsequent indicator bias analysis and health status assessment.

4. The vehicle health status assessment method based on vehicle network big data according to claim 1, characterized in that, Step S50 includes the following sub-steps: Extract the joint health representation vector of the target vehicle and its neighboring vehicles in the group graph; Calculate the normalized deviation of the standard deviation of the target vehicle's values ​​for each indicator dimension from the neighbor mean; Based on the deviation results, significant deviation indicators are identified, and the health level and output description are determined according to the number of deviation dimensions and the degree of deviation.

5. A vehicle health status assessment system based on Internet of Vehicles big data, characterized in that, The system includes the following modules: The acquisition module is used to collect multi-source vehicle network data during vehicle operation and construct a health status vector sequence of the vehicle within a preset time window based on the multi-source vehicle network data. The multi-source vehicle network data includes engine speed, throttle opening, coolant temperature, oil pressure, vehicle speed, braking status, tire pressure, battery voltage, fault code information, location information, acceleration information, steering wheel angle, ambient temperature and humidity, and driving behavior indicators. The temporal modeling module is used to divide the health state vector sequence into multiple continuous subsequences according to time based on a sliding time window, extract dynamic change features that reflect the state evolution trend, and form a temporal health representation of an individual vehicle. The time series modeling module specifically includes: The sub-module is used to divide the health status vector sequence into multiple time-continuous sub-sequences within a sliding time window of a preset length. The feature extraction submodule is used to extract dynamic change features for each subsequence, including fluctuation amplitude, rate of change, periodic intensity, outlier density, number of anomalous jumps, and number of change directions. The time series generation submodule is used to normalize the extracted change features and form a time series health representation vector through vector concatenation or weighting mechanism; The graph construction module is used to construct a vehicle group similarity graph. Each vehicle is treated as a node in the graph structure. The edge weights between nodes are calculated based on the similarity of vehicle type, vehicle age, usage environment and driving behavior. A graph neural network is used to extract the feature representation of each vehicle in the group structure to form a group structure representation. The fusion module is used to fuse the temporal health representation with the group structure representation to obtain the joint health representation vector of the vehicle; The evaluation module is used to calculate the differences between the target vehicle and its neighboring vehicles in the vehicle group similarity map on various indicators based on the joint health representation vector, identify significant deviation indicators of health status, evaluate the current health status of the target vehicle based on the significant deviation indicators, and output the health level and corresponding description.

6. The vehicle health status assessment system based on Internet of Vehicles big data according to claim 5, characterized in that, The map construction module specifically includes: The feature encoding submodule is used to perform multi-dimensional feature encoding on vehicles using indicators based on vehicle configuration, vehicle age, usage environment, and driving behavior. The edge weight calculation submodule is used to calculate similarity scores based on the feature distances between vehicles and to construct edge connections and their weights in the graph according to a set threshold. The graph embedding submodule is used to input the constructed vehicle similarity graph into the graph neural network model and output the structural embedding representation of each vehicle in the graph structure.

7. The vehicle health status assessment system based on Internet of Vehicles big data according to claim 5, characterized in that, The fusion module specifically includes: The alignment submodule is used to perform dimensional alignment between the temporal health representation and the population structure representation. The fusion submodule is used to perform feature fusion by concatenation operations or using a neural network model to generate a joint health representation vector. The output submodule is used to apply the combined health representation to subsequent indicator deviation analysis and health status assessment.

8. The vehicle health status assessment system based on Internet of Vehicles big data according to claim 5, characterized in that, The evaluation module specifically includes: The adjacency extraction submodule is used to extract the joint health representation vector of the target vehicle and its neighboring vehicles in the group graph; The difference calculation submodule is used to calculate the standard deviation normalization deviation of the target vehicle's values ​​in each indicator dimension from the neighbor mean; The rating determination submodule is used to identify significant deviation indicators based on the deviation results, and to determine the health level and output description according to the number of deviation dimensions and the degree of deviation.

Images