A method and system for driving style recognition based on multi-source sensor data

Abstract

This invention discloses a driving style recognition method and system based on multi-source sensor data. The method includes the following steps: S1: acquiring multi-source sensor data and preprocessing it; S2: constructing a symbol requirement set and a symbol regulatory compliance matrix to form a complete symbol library; S3: constructing a hierarchical rule tree; S4: constructing a variational Bayesian analytical inference model; S5: updating the parameters of the variational Bayesian analytical inference model and the complete symbol library; S6: online monitoring and real-time updating, outputting the updated driving style classification and risk assessment results. This invention solves the problem of accurate driving style recognition under multiple constraints by constructing a complete symbol library and a variational Bayesian analytical inference model and fully combining the two.

Description

Technical Field

[0001] This invention relates to the field of driving style recognition technology, and in particular to a driving style recognition method and system based on multi-source sensor data. Background Technology

[0002] Existing driving style recognition technologies are mostly based on the comprehensive processing of vehicle driving status and environmental data. They collect vehicle speed, steering angle, accelerator and brake pedal signals, as well as road conditions and real-time traffic information through onboard sensing devices. The system typically first aggregates the collected multi-channel data into a unified management module, and then uses filtering or interpolation methods to correct missing items and noise, so that the subsequent recognition model can analyze the data on relatively clean input data.

[0003] Existing driving style recognition solutions typically rely on separate processes for data processing and model inference, often lacking more granular, multi-layered representations. Since most systems only map operational behaviors to onboard data once, they struggle to promptly identify different levels of driving intent or energy consumption demands when dealing with complex conditions. Furthermore, some methods remain at a static correction stage when responding to environmental changes and sensor anomalies, failing to flexibly adjust to unexpected emergencies and high power demands. Simultaneously, existing technologies often limit symbolic representation to a limited number of operational labels, lacking necessary adaptive constraints on energy management and traffic regulation compliance in real-world scenarios, making it difficult for the system to balance fine-grained energy consumption scheduling with multi-dimensional behavioral constraints.

[0004] Therefore, there is an urgent need to design a driving style recognition method and system based on multi-source sensor data to solve the problems existing in the above-mentioned technologies. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a driving style recognition method based on multi-source sensor data. The aim is to solve the problem of accurate driving style recognition under multiple constraints by constructing a complete symbol library and a variational Bayesian analytic inference model and fully combining the two.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The first aspect of this invention provides a driving style recognition method based on multi-source sensor data, the method comprising the following steps:

[0008] S1: Acquire multi-source sensor data and preprocess it to obtain high-quality sequences;

[0009] S2: Based on the high-quality sequence, extract key features from multi-source sensor data, construct a symbol requirement set and a symbol regulatory compliance matrix, and obtain a complete symbol library;

[0010] S3: Establish a mapping relationship between the complete symbol library and high-quality sequences, divide the symbol hierarchy according to the three aspects of safety, comfort and economy, define the energy consumption coupling function, and construct a hierarchical rule tree;

[0011] S4: Based on the hierarchical rule tree, combined with high-quality sequences, establish a multi-layer feature fusion function and probability prior distribution, introduce dynamic latent variables to characterize braking recovery and motor response, construct a block-level approximate inference method to accelerate convergence, obtain a variational Bayesian analytical inference model, and output a joint posterior distribution.

[0012] S5: Establish a dynamic mapping between the hierarchical rule tree and the joint posterior distribution to form a multi-level rule loading function, establish a hard threshold filtering boundary to filter high-risk operations, compare the symbol prior and posterior, and update the variational Bayesian analytic inference model parameters and the complete symbol library.

[0013] S6: Monitors vehicle operating status online and updates the variational Bayesian analytical inference model parameters and complete symbol library in real time, outputting updated driving style classification and risk assessment results.

[0014] As an embodiment of this application, step S1 specifically includes:

[0015] S11: Synchronously collect the vehicle's speed during operation via sensor array. acceleration Battery status Road conditions and driving operation The collection period is based on a unified time series. Characterization; the data management unit receives and aggregates the outputs of each sensor, forming a dimension of... initial observation matrix The formula is as follows:

[0016]

[0017] in, Indicates the first Each sensor at time The original measurement value;

[0018] S12: For the initial observation matrix Perform time series alignment to obtain the alignment matrix The formula is as follows:

[0019]

[0020] in, This represents an alignment function, which uses interpolation and interpolation techniques to... Mapping to consistent time , so that the initial observation matrix The corresponding rows are the same Each column corresponds to a base time. ;

[0021] S13: Suppress high-frequency random noise using a sliding window filtering method to form a smooth matrix. The formula is as follows:

[0022]

[0023]

[0024] in, Indicates the first Each sensor at time Smoothed measurement values Indicates the offset;

[0025] S14: Yes and The difference between the measurements is used to determine a threshold. If the threshold is met, the measurement value is marked as a mutation point and stored in the mutation point identification matrix. The formula is as follows:

[0026]

[0027] in, Indicates the threshold for mutation detection;

[0028] By using contextualized labeling, we can improve the identification and retention of transient anomalies and enhance the reliability of data in typical road conditions for electric vehicles.

[0029] Transformation point identification matrix and smoothness matrix Make a comprehensive judgment based on the vehicle operation. With speed acceleration Road conditions Correlation test at each time step To ensure the validity of the data, invalid data was removed and readings with reasonable correction margins were updated, ultimately yielding a high-quality sequence. , .

[0030] As an embodiment of this application, step S2 specifically includes:

[0031] S21: Based on the high-quality sequence Extract vehicle speed acceleration Battery status Road conditions and driving operation The key features are combined with the design goals to construct a set of symbolic requirements, and then a symbolic priority function is used. The key features are quantified, and the specific formula is expressed as follows:

[0032]

[0033] in, Indicates an index for different symbol candidates; Indicates high-quality sequence Indexes for different data channels; Indicates at time Data Channel The output value; Indicates the total number of data channels. Indicates the sampling length. Indicates a data channel With symbols The fusion weight, Indicates to Perform nonlinear transformations to capture vehicle operating characteristics;

[0034] S22: Based on the symbol requirement set constructed in step S21, select common driving behaviors and events such as turning, acceleration, deceleration, intersection waiting, and speed-limited areas to construct an initial symbol set and establish a symbol baseline table. This enables a fast mapping between symbol indexes and semantic parsing, expressed by the following formula:

[0035]

[0036] in, This represents the internal encoding of the basic symbol elements. Indicates the number of symbol types. This indicates the scalable granularity of each symbol; each symbol entry is assigned a system number and corresponds one-to-one with vehicle operation data, providing a semantic benchmark for energy management and driving style analysis.

[0037] S23: Introduces labels representing micro-driving styles, adds descriptions of motor power response and regenerative braking rate, and constructs an energy consumption mapping function. The energy consumption mapping function describes the relationship between energy consumption and symbolic labels. The formula is expressed as follows:

[0038]

[0039] in, Indicates the symbol label index, Symbols and These represent the dimensions of contribution to energy consumption and energy recovery, respectively. and For relevant weights, and Energy management characteristic values ​​in the data channel. and This indicates the total number of channels involved in energy input and output. and These represent nonlinear functions for energy consumption and energy recovery, respectively.

[0040] S24: Introduce regulatory elements, establish a symbol compliance matrix, and assign each symbol based on speed limits, traffic restrictions, and road priority information. Assigning regulatory compliance coefficients The symbol compliance is recorded as :

[0041]

[0042] in, Symbols The corresponding regulatory compliance coefficient, Indicates the time of combination The adjustment amount for the intensity of road supervision at that time. To determine based on vehicle speed With road conditions A filtering function for determining violations or potential hazards;

[0043] S25: To achieve dynamic adaptation of the symbol library under different seasons or energy consumption constraints, construct an adaptive symbol update equation. :

[0044]

[0045] in, Indicates the initial symbol The base attribute values ​​when unaffected by the environment. and For adaptive parameter tuning, and These represent the time variables of battery state and road conditions, respectively. and These represent mapping functions for the battery and the road environment, respectively, used to dynamically adjust symbolic attributes;

[0046] S26: Building upon the progressive steps from S21 to S25, a complete symbol library covering driving behavior, energy management, and traffic regulations is formed. :

[0047]

[0048] The complete symbol library The output includes a basic symbol index, energy conversion coefficient, compliance evaluation, and adaptive management parameters.

[0049] As an embodiment of this application, step S3 specifically includes:

[0050] S31: In the complete symbol library With high-quality sequences The mapping relationship between them is established, expressed by the following formula:

[0051]

[0052] in, Represents a globally consistent function, the complete symbol library Include One symbolic entry, high-quality sequence Include One channel, Indicates the first Symbols and channels The fusion weight, Represents a mapping function. Indicates high-quality sequence In the passage ,time The measured value, Represents the complete symbol library The Middle The internal encoding of each symbol;

[0053] S32: Based on the vehicle's operational needs, a hierarchical symbolic division is established for safety, comfort, and economy. The priority order of each level is represented by... , , This means that for each moment... Hierarchical priority The measurement is expressed by the following formula:

[0054]

[0055] in, Indicates a hierarchical index. Indicates hierarchy The normalization factor; ; Indicates channel This level The inhibition coefficient;

[0056] S33: Define the energy consumption coupling function based on the vehicle's motor power and battery recovery characteristics. Indicates hierarchy At any moment The degree of energy consumption impact will Embedding high-level rules allows the system to strike a balance between security and economy, as expressed in the following formula:

[0057]

[0058] in, and These represent the channel indices for energy output and regenerative braking, respectively. Indicates the output power weight. Indicates the weight of the recovery power. and These represent the time of the vehicle. The values ​​of the power output channel and the regenerative force channel, Indicates hierarchy The baseline energy consumption compensation item;

[0059] S34: Defines a set of operations for specific actions such as rapid acceleration and quick steering. Symbolization conditions Make a judgment:

[0060]

[0061] in, Indicates a certain driving operation, Indicates hierarchy The corresponding set of operations, if Then the operation is considered to be at the level. Effective symbolic representation;

[0062] S35: For hierarchy To mitigate the operational risks associated with high-power or sudden braking, a rule-based matching function is constructed. The formula is expressed as follows:

[0063]

[0064] in, Indicates time At the level The control range below, This represents the coefficient that prioritizes amplification of the operation. This represents a coefficient for energy consumption penalty; through the... The value is determined to be greater than or less than zero, and excessive power or sudden braking behavior is identified in real time and energy consumption is limited.

[0065] S36: Prioritize the layering Energy consumption coupling function Operation symbol conditions Rule matching function This process integrates rules to form a hierarchical rule tree that includes high-level rules and operational breakdowns. The formula is expressed as follows:

[0066]

[0067] in, Indicates hierarchy With a certain driving operation The correspondence.

[0068] As an embodiment of this application, step S4 specifically includes:

[0069] S41: Based on the hierarchical rule tree, and combined with high-quality sequences, select energy consumption-related features and driving behavior classification factors, map them to a set of modelable indicators, and establish a multi-layer feature fusion function. The rule constraints and data channels are comprehensively quantified, and the formula is expressed as follows:

[0070]

[0071] in, An index representing a modelable metric. Represents the hierarchical rule tree. The activation strength of the rule, Indicates energy consumption-related characteristics, Indicates the classification factors of driving behavior. This represents the fusion weight constant;

[0072] S42: Order A set of random latent parameters representing driving style, based on a multi-layer feature fusion function. Construct a probability prior distribution from the observations The prior distribution is constrained in the high-dimensional feature space through multiple integrals, as expressed by the following formula:

[0073]

[0074] in, Represents the space of random parameters. This represents the total number of integration indicators. Representation and multi-layer feature fusion function The corresponding prior penalty coefficient;

[0075] S43: Define a set of dynamic hidden variables to address the braking regenerative braking and motor response delay during vehicle operation. ,make To represent the energy-consumption coupled state, a function is introduced. The effect of dynamic latent variables on observations is described by the following formula:

[0076]

[0077] in, Indicates time The observed vector below, This represents the basic observation components that do not include energy consumption dynamics. Indicates the weights of latent variable coupling. Used to simulate the timing characteristics of braking regenerative rate and delayed power response;

[0078] S44: Construct a block-based, hierarchical approximate inference method to accelerate convergence, processing the general variable block and the energy consumption coupled variable block separately, and defining the overall posterior distribution. The factorization model is as follows:

[0079]

[0080] in, Indicates the first The parameter set of a general block, Indicates the first The set of hidden variables for each energy-consumption coupling block. , Indicates the number of blocks;

[0081] Using the variational expectation-maximization method and Continuous iteration and updates;

[0082] S45: Defines the scenario focus function for extreme road conditions or high power requirements. The estimation accuracy of the corresponding latent variables can be refined by increasing the local sampling frequency, as expressed by the following formula:

[0083]

[0084] in, Indicates time In local sampling areas under extreme road conditions Indicates a high-weight amplification factor. Reflects the degree of scene triggering. Indicates the key inhibition coefficient. This indicates a measurement of energy limits or high-load regions.

[0085] S46: Combining the above steps, we obtain a variational Bayesian analytical inference model, which uses a dynamic set of latent variables. With random hidden parameter sets Joint random variables formed by combination With integration as the core Input, using the merged joint posterior distribution The final inference result is expressed by the following formula:

[0086]

[0087] in, This represents the likelihood probability distribution of the observed data. Kullback-Leibler divergence is used to measure the posterior distribution of the population. With the true distribution The differences between them.

[0088] As an embodiment of this application, step S5 specifically includes:

[0089] S51: Convert the hierarchical rule tree With joint posterior distribution Establish dynamic mapping for symbol labels With latent variables Correlation function between A comprehensive evaluation is conducted to generate a multi-layered rule loading function. :

[0090]

[0091] in, This indicates the index of the symbol rule in the hierarchical set. Load weights for symbol rules. For the joint posterior distribution, the correlation function Rules for measuring symbols For latent variables The constraint strength;

[0092] S52: Establish hard threshold filtering boundaries for vehicle operation behavior It is used to determine whether excessive or high-risk behavior violates layered safety requirements. If the value exceeds a predetermined threshold, rule validation is performed immediately. The specific formula is expressed as follows:

[0093]

[0094] in, Indicates hierarchy, Indicates vehicle operation commands. To activate weights, Characterization operations At the level The legitimacy, For latent variables At the level Compensation factor, This represents the network latent variable domain that satisfies the constraints of the current level.

[0095] S53: By comparing time and time The posterior distribution is used to determine the confidence level of the transition, and the distribution variability is defined. ,like If the value exceeds a specified threshold, a significant change in driving style is determined. The specific formula is as follows:

[0096]

[0097] in, , Representing time respectively and time The posterior distribution, Indicates the Kullback-Leibler divergence;

[0098] S54: Distribute prior symbols The sign probability obtained from actual inference Comparison:

[0099]

[0100] Among them, if If the threshold is exceeded, collaborative correction is triggered. By analyzing the causes of the differences, symbol constraints are adjusted or driving risks are indicated, so as to achieve bidirectional updating and iteration of the complete symbol library and the variational Bayesian analytical inference model.

[0101] S55: Define the multinomial loss function This represents the error metric for the variational Bayesian analytical inference model across different scenario dimensions, and the parameters of the variational Bayesian analytical inference model. Dynamic correction is performed, and the formula is expressed as follows:

[0102]

[0103] in, and These represent the error component indices of energy consumption coupling and driving style, respectively. To learn the step size coefficient, Indicates to The gradient;

[0104] S56: Define a warning decision function to address the non-convergence of the variational Bayesian analytical inference model due to extreme driving operations. If the warning determination function If the threshold is exceeded, an alert will be issued and the final revised results of the variational Bayesian analytic inference model and the complete symbol library will be output, as expressed in the following formula:

[0105]

[0106] in, This indicates a focus on high-risk areas involving extreme driving power or speeding. This represents the detection function corresponding to key elements such as braking and acceleration. This is a weighting factor.

[0107] As an embodiment of this application, step S6 specifically includes:

[0108] S61: Filter out the vehicle's current main operating characteristics and merge them with the energy consumption safety range to form a basic constraint set, and define the initial composite state. The formula is expressed as follows:

[0109]

[0110] in, This indicates the initial driving style label score identified by the vehicle. This indicates the percentage of remaining energy consumption. and These are adjustable weighting coefficients;

[0111] S62: Real-time sampling of vehicle speed, current output, and pedal operation over short periods, based on a joint posterior distribution. Determine if the vehicle exhibits a style shift; if driving parameters are detected to be inconsistent with... As the differences between them gradually increase, a style revision process is triggered, updating the local posterior state to adapt to the latest observations;

[0112] S63: When high impact power or sudden abnormal operation during emergency braking is detected, a rapid matching relationship is established with the complete symbol library to find the corresponding symbol entries and record the relevant cumulative number of times; if the abnormal entries accumulate rapidly within a short period of time, they are recorded as a high-risk operation trend and bound to the current operating characteristics of the vehicle.

[0113] S64: If the cumulative number of high-risk operation trends exceeds a preset threshold, the situation will be marked as a potential safety or energy consumption risk and an internal alarm signal will be issued.

[0114] S65: When external road conditions and driving style continue to change, the variational Bayesian analytical inference model and the complete symbol library are synchronously corrected online.

[0115] S66: After completing multiple online corrections to the variational Bayesian analytic inference model and the complete symbol library, the vehicle's final energy consumption information and abnormal operation registration are integrated to generate the latest driving style classification and risk assessment results.

[0116] As an embodiment of this application, step S66 further includes determining the final discrimination score. A simplified metric is used to directly determine the degree of deviation between the vehicle's current state and the target safety boundary. The formula is expressed as follows:

[0117]

[0118] in, Indicates the number of abnormal registrations. This is the anomaly count penalty coefficient. If the value is below the safety threshold, it indicates a significant deviation in vehicle behavior, requiring a more robust intervention strategy; otherwise, a normal driving style label and risk assessment result should be output.

[0119] The present invention also provides a driving style recognition system based on multi-source sensor data, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described driving style recognition method based on multi-source sensor data.

[0120] The beneficial effects of this invention are as follows:

[0121] 1. This invention collects data from multiple sensor sources and ensures the integrity and timeliness of the observed information through unified time series management and abrupt change smoothing. During the preprocessing process, data such as speed, acceleration, battery status, and road conditions are aligned to the time axis and filtered by a sliding window. Threshold judgment is used to mark instantaneous anomalies. The resulting high-quality sequence satisfies physical rationality at every moment and provides a stable data foundation for subsequent symbol library creation and hierarchical reasoning.

[0122] 2. This invention extracts features from vehicle speed, acceleration, battery status, road conditions, and driving operations, and combines these with regulatory compliance and dynamic parameter tuning to tightly bind actual driving behaviors with symbol labels. As a result, it not only includes common symbols such as steering, acceleration, and waiting at intersections, but also further subdivides them based on motor power response and regenerative braking rate, so that the symbols reflect the driver's micro-style characteristics at the semantic level. This invention monitors changes in road conditions, seasons, and energy usage constraints, and uses adaptive parameter tuning equations to continuously correct symbol attributes, so that it can still accurately reflect the vehicle status in high-load or energy-limited scenarios.

[0123] 3. This invention establishes a holistic framework that combines hierarchical analysis and variational Bayesian analytical inference models at the rule and reasoning levels. It connects high-level abstract rules with detailed conditions for specific operations such as rapid acceleration or high-speed braking. By constructing a hierarchical rule tree, it represents the needs and constraints at each level of safety, economy, and comfort. Then, it uses a variational Bayesian analytical inference model to model the latent variables of electric vehicle energy consumption coupling state and driving style, and accelerates convergence through block approximation inference. Through collaborative reasoning, this invention can more flexibly capture high-power output or emergency braking behavior when sudden conditions occur, and immediately combine existing symbolic rules to provide short-term energy limits or risk warnings. The entire process is continuously iterated through online updates and local corrections to ensure that the precision of driving style recognition and the consistency of rule execution are maintained under frequent and complex vehicle driving environments. Attached Figure Description

[0124] Figure 1 This is a flowchart illustrating the technical solution of a driving style recognition method based on multi-source sensor data provided in an embodiment of the present invention.

[0125] Figure 2 This is a schematic diagram illustrating the construction of a complete symbol library for a driving style recognition method based on multi-source sensor data provided in an embodiment of the present invention;

[0126] Figure 3 This is a schematic diagram illustrating the construction of a variational Bayesian analytical inference model for a driving style recognition method based on multi-source sensor data, as provided in an embodiment of the present invention. Detailed Implementation

[0127] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0128] Reference Figures 1-3The first aspect of the present invention provides a driving style recognition method based on multi-source sensor data, the method comprising the following steps:

[0129] S1: Acquire multi-source sensor data and preprocess it to obtain high-quality sequences;

[0130] S2: Based on the high-quality sequence, extract key features from multi-source sensor data, construct a symbol requirement set and a symbol regulatory compliance matrix, and obtain a complete symbol library;

[0131] S3: Establish a mapping relationship between the complete symbol library and high-quality sequences, divide the symbol hierarchy according to the three aspects of safety, comfort and economy, define the energy consumption coupling function, and construct a hierarchical rule tree;

[0132] S4: Based on the hierarchical rule tree, combined with high-quality sequences, establish a multi-layer feature fusion function and probability prior distribution, introduce dynamic latent variables to characterize braking recovery and motor response, construct a block-level approximate inference method to accelerate convergence, obtain a variational Bayesian analytical inference model, and output a joint posterior distribution.

[0133] S5: Establish a dynamic mapping between the hierarchical rule tree and the joint posterior distribution to form a multi-level rule loading function, establish a hard threshold filtering boundary to filter high-risk operations, compare the symbol prior and posterior, and update the variational Bayesian analytic inference model parameters and the complete symbol library.

[0134] S6: Monitors vehicle operating status online and updates the variational Bayesian analytical inference model parameters and complete symbol library in real time, outputting updated driving style classification and risk assessment results.

[0135] Specifically, this invention is applicable to all types of motor vehicles, including but not limited to electric vehicles, fuel vehicles, hybrid vehicles, and autonomous vehicles. Through multi-source sensor data fusion and symbolic rule reasoning, it achieves dynamic recognition and optimization of driving styles.

[0136] As an embodiment of this application, step S1 specifically includes:

[0137] S11: Synchronously collect the vehicle's speed during operation via sensor array. acceleration Battery status Road conditions and driving operation The collection period is based on a unified time series. Characterization; the data management unit receives and aggregates the outputs of each sensor, forming a dimension of... initial observation matrix The formula is as follows:

[0138]

[0139] in, Indicates the first Each sensor at time The original measurement value.

[0140] Specifically, the sensor group includes a vehicle speed sensor, an acceleration sensor, a voltage sensor, a current sensor, a road surface friction coefficient or pothole detection sensor, and a brake pressure sensor.

[0141] S12: For the initial observation matrix Perform time series alignment to obtain the alignment matrix The formula is as follows:

[0142]

[0143] in, This represents the alignment function, used to define the mapping relationship. Specifically, it uses interpolation and interpolation techniques to... Mapping to consistent time , so that the initial observation matrix The corresponding rows are the same Each column corresponds to a base time. .

[0144] S13: Suppress high-frequency random noise using a sliding window filtering method to form a smooth matrix. The formula is as follows:

[0145]

[0146]

[0147] in, Indicates the first Each sensor at time Smoothed measurement values This indicates the offset.

[0148] S14: Yes and The difference between the measurements is used to determine a threshold. If the threshold is met, the measurement value is marked as a mutation point and stored in the mutation point identification matrix. The formula is as follows:

[0149]

[0150] in, Indicates the threshold for mutation detection;

[0151] Specifically, when a sudden change is detected, it is not always discarded as a fault. Instead, the driving environment at that moment is annotated according to typical vehicle operating conditions, such as regenerative braking, uphill braking, rapid acceleration and start-up, and stopping / starting at intersections. In this way, the instantaneous fluctuations that truly belong to the vehicle's normal rapid response or energy recovery process can be identified and retained, rather than simply deleting the contextualized markers. Therefore, by using the above-mentioned contextualized markers, the identification and retention of instantaneous anomalies of vehicles, especially electric vehicles, is improved, and the reliability of the data in typical road conditions of electric vehicles is enhanced.

[0152] Transformation point identification matrix and smoothness matrix A comprehensive assessment is conducted, invalid data is deleted, and readings with reasonable room for correction are updated, based on vehicle operation. With speed acceleration Road conditions Correlation test at each time step The effectiveness of this study ultimately yields high-quality sequences. , .

[0153] The comprehensive judgment involves comparing the mutation point markers generated in the previous step with the smoothed values ​​obtained through the sliding window: if a mutation point is marked as a typical scenario and its numerical deviation is within an acceptable correction range, the original reading is replaced with the smoothed value; if the mutation point neither conforms to any scenario label nor deviates too much from the smoothed values ​​of surrounding times, it is judged as invalid or noise, and the data is directly removed.

[0154] In addition, according to vehicle operation With speed acceleration Road conditions Correlation test at each time step The validity of the rule is as follows: For each moment, the matching degree between the driving operation (accelerator, brake, steering) and the speed / acceleration / road conditions at that moment is checked. For example, when there is braking input, there should be deceleration or negative acceleration; the change in accelerator pedal depth should correspond to positive acceleration; on flat road conditions, the vehicle speed fluctuation should not be too large; if the consistency between the command and the physical feedback is higher than the preset threshold, the data at that moment is considered valid; if the consistency is insufficient, it is considered that there is an anomaly at that point, and the upper-level rule is removed again.

[0155] Specifically, this invention focuses on processing multi-source sensor data during vehicle operation. It ensures the integrity and timeliness of observed information through unified time series management and abrupt change smoothing. During processing, data such as speed, acceleration, battery status, and road conditions are time-axis aligned and filtered using a sliding window, and threshold judgments are used to mark transient anomalies. The resulting high-quality sequences satisfy physical plausibility at every moment, providing a stable data foundation for subsequent symbol creation and hierarchical inference. Compared to conventional single or simplified preprocessing methods, this technical solution also introduces a contextualized labeling mechanism to further assess the credibility of detected potential abrupt changes, ensuring that temporary anomalies are not simply filtered out under typical road conditions, thereby improving the retention of data on extreme vehicle operations.

[0156] As an embodiment of this application, step S2 specifically includes:

[0157] S21: Based on the high-quality sequence Extract vehicle speed acceleration Battery status Road conditions and driving operation The key features are combined with the design goals to construct a set of symbolic requirements, and then a symbolic priority function is used. The key features are quantified, and the specific formula is expressed as follows:

[0158]

[0159] in, Indicates an index for different symbol candidates; Indicates high-quality sequence Indexes for different data channels; Indicates at time Data Channel The output value; Indicates the total number of data channels. Indicates the sampling length. Indicates a data channel With symbols The fusion weight, Indicates to Nonlinear transformations are performed to capture vehicle operating characteristics.

[0160] The design goal refers to ensuring that the extraction of key features and the construction of the symbol requirement set closely align with the application requirements that the system will subsequently address. This means that the selected features should reflect the vehicle's operating status and instantaneous anomalies under typical conditions to the greatest extent possible, while meeting the performance requirements of real-time computation, low false alarms, and stable and reliable system performance. Specifically, this manifests in the following ways: Anomaly recognition accuracy: ensuring that the constructed symbols can prominently reflect instantaneous anomalies or fault symptoms of the vehicle; Scene credibility: symbols should be able to stably extract effective information under typical road conditions and operating modes; Real-time performance and computational overhead: selected features should facilitate rapid computation and updates to meet online operation requirements; System reliability: balancing false alarm and false negative rates to ensure the reliable execution of subsequent diagnostic or control strategies.

[0161] S22: Based on the symbol requirement set constructed in step S21, select common driving behaviors and events such as turning, acceleration, deceleration, intersection waiting, and speed-limited areas to construct an initial symbol set and establish a symbol baseline table. This enables a fast mapping between symbol indexes and semantic parsing, expressed by the following formula:

[0162]

[0163] in, This represents the internal encoding of the basic symbol elements. Indicates the number of symbol types. This indicates the scalable subdivision of each symbol, used for precise differentiation of operations such as steering, acceleration, and deceleration in the system.

[0164] S23: Next, labels representing micro-driving styles are introduced, such as short-term high-output conditions, and descriptions of motor power response and regenerative braking rate are added to construct an energy consumption mapping function. The energy consumption mapping function describes the relationship between energy consumption and symbolic labels. The formula is expressed as follows:

[0165]

[0166] in, Indicates the symbol label index, Symbols; and These represent the dimensions of contribution to energy consumption and energy recovery, respectively. and For relevant weights, and Energy management characteristic values ​​in the data channel. and This indicates the total number of channels involved in energy input and output. and These represent nonlinear functions for energy consumption and energy recovery, respectively.

[0167] Through energy consumption mapping function Energy consumption changes are reflected at the symbolic level, providing a measurable basis for hierarchical reasoning.

[0168] S24: Next, regulatory elements are introduced to establish a symbol compliance matrix, which helps the system identify the matching between driving behavior and road regulations, and assigns each symbol based on information such as speed limits, traffic restrictions, and road priorities. Assigning regulatory compliance coefficients The symbol compliance is recorded as :

[0169]

[0170] in, Symbols The corresponding regulatory compliance coefficient, Indicates the time of combination The adjustment amount for the intensity of road supervision at that time. To determine based on vehicle speed With road conditions A filtering function for determining violations or potential hazards;

[0171] The system uses a symbol compliance matrix to determine whether driving behavior complies with traffic management requirements. Based on the compliance level, violations are flagged or blocked. According to the rules, if a symbol does not comply with the rules, the system rolls back and corrects it: the physical range, logical consistency, and cross-variable correlation of the quantized symbol are checked; if it fails, the symbol at that moment is temporarily restored to a sliding window filter value, and its fusion weight or threshold is adjusted; if it is still not compliant, the symbol at that moment is removed.

[0172] S25: Different seasons or special energy consumption constraints can lead to frequent changes in vehicle status and road demand. Therefore, in order to achieve dynamic adaptation of the symbol library under different seasons or special energy consumption constraints, an adaptive symbol update equation is constructed. :

[0173]

[0174] in, Indicates the initial symbol The base attribute values ​​when unaffected by the environment. and For adaptive parameter tuning, and These represent the time variables of battery state and road conditions, respectively. and These represent mapping functions for the battery and the road environment, respectively, used to dynamically adjust symbolic attributes.

[0175] S26: Building upon the progressive steps from S21 to S25, a complete symbol library covering driving behavior, energy management, and traffic regulations is formed. :

[0176]

[0177] The complete symbol library The output includes a basic symbol index, energy conversion coefficient, compliance evaluation, and adaptive management parameters, providing semantic priors for control strategies and inference algorithms.

[0178] Specifically, this application constructs a complete symbol library for driving behavior and energy management. By extracting features from vehicle speed, acceleration, and battery parameters, and combining regulatory compliance with dynamic parameter tuning, it tightly binds actual driving behaviors with symbol labels. As a result, it not only includes common symbols such as steering, acceleration, and waiting at intersections, but also further subdivides them based on motor power response and regenerative braking rate. This allows the symbols to reflect the driver's micro-style characteristics at the semantic level. Compared with the traditional approach that relies solely on static symbols for judgment, this solution monitors changes in road conditions, seasons, and energy usage constraints, and uses adaptive parameter tuning equations to continuously correct symbol attributes, ensuring that it can accurately reflect the vehicle's state even in high-load or energy-limited scenarios.

[0179] As an embodiment of this application, step S3 specifically includes:

[0180] S31: In the complete symbol library With high-quality sequences A mapping relationship is established between them, and the consistency between symbolic features and data is measured globally to form a comprehensive input suitable for subsequent hierarchical parsing. The formula is as follows:

[0181]

[0182] in, Represents the global consistency function, used to measure... Each symbol in the high-quality sequence The degree of coupling between them; the complete symbol library Include One symbol entry; high-quality sequence Include One channel; Indicates the first Symbols and channels The fusion weight; This represents a mapping function used to calculate the degree of association between symbols and data; Indicates high-quality sequence In the passage ,time The measured value; Represents the complete symbol library The Middle The internal encoding of each symbol;

[0183] Through calculation The degree of coupling between symbols and data is evaluated, and symbol entries with high consistency are selected.

[0184] S32: Based on the vehicle's operational needs, a hierarchical symbolic division is established for safety, comfort, and economy. The priority order of each level is represented by... , , This means that for each moment... Hierarchical priority To measure, The hierarchical index is expressed by the following formula:

[0185]

[0186] in, Indicates hierarchy The normalization factor; ; Indicates channel This level The suppression coefficient is used to control the fluctuation of priority at each level under specific operating conditions. A higher value indicates a higher priority for that level.

[0187] S33: Define an energy consumption coupling function for the unique characteristics of electric vehicles, specifically considering the motor power and battery recycling features. Indicates hierarchy At any moment The degree of energy consumption impact will Embedding high-level rules allows the system to strike a balance between security and economy, as expressed in the following formula:

[0188]

[0189] in, and These represent the channel indices for energy output and regenerative braking, respectively. Indicates the output power weight. Indicates the weight of the recovery power. and These represent the time of the vehicle. The values ​​of the power output channel and the regenerative force channel, Indicates hierarchy The baseline energy consumption compensation item is used to maintain overall energy consumption stability;

[0190] Through energy consumption coupling function With hierarchical priority They will jointly guide the priority of high-level rules, taking into account both security and economic needs.

[0191] S34: Based on the high-level rules and energy consumption coupling results from the previous step, define an operation set for specific operations such as rapid acceleration and quick steering. Symbolization conditions Make a judgment:

[0192]

[0193] in, Indicates a certain driving operation, Indicates hierarchy The corresponding set of operations, if Then the operation is considered to be at the level. Effective symbolic representation;

[0194] The downward refinement approach enables the hierarchical rules to be applied to actual operational scenarios, combining high-level abstract requirements with vehicle dynamics processes, thus solving the problem of insufficient precision in rule subdivision in existing technologies.

[0195] S35: For hierarchy To mitigate the operational risks associated with high-power or sudden braking, a rule-based matching function is constructed. In the event of a large current surge or a sudden braking increase, a short-term energy limit will be triggered, as expressed by the following formula:

[0196]

[0197] in, Indicates time At the level The control range below, This represents the coefficient that prioritizes amplification of the operation. This represents a coefficient for energy consumption penalty; through the... The system determines whether the value is greater than or less than zero, identifies excessive power or sudden braking behavior in real time and implements energy consumption limits, overcomes the blind spot problem under external environmental fluctuations, and achieves refined energy management.

[0198] S36: Prioritize the hierarchical structure described in the above steps. Energy consumption coupling function Operation symbol conditions Rule matching function This process integrates rules to form a hierarchical rule tree that includes high-level rules and operational breakdowns. The formula is expressed as follows:

[0199]

[0200] in, Indicates hierarchy With a certain driving operation The correspondence, hierarchical rule tree Output the prior constraints in a graph structure for use in the inference process of the subsequent variational Bayesian inferential analytical model.

[0201] As an embodiment of this application, step S4 specifically includes:

[0202] S41: Based on the hierarchical rule tree, and combined with high-quality sequences, select energy consumption-related features and driving behavior classification factors, map them to a set of modelable indicators, and establish a multi-layer feature fusion function. This is used to comprehensively quantify rule constraints and data channels, and the formula is expressed as follows:

[0203]

[0204] in, An index representing a modelable metric. Represents the hierarchical rule tree. The activation strength of the rule, Indicates energy consumption-related characteristics, Indicates the classification factors of driving behavior. This represents the fusion weight constant.

[0205] S42: Based on the historical driving records of step S1 and the hierarchical rule tree of step S3, let A set of random latent parameters representing driving style, based on a multi-layer feature fusion function. Construct a probability prior distribution from the observations The prior distribution is constrained in the high-dimensional feature space through multiple integrals, as expressed by the following formula:

[0206]

[0207] in, Represents the space of random parameters. This represents the total number of integration indicators. Representation and multi-layer feature fusion function The corresponding prior penalty coefficient is applied during network initialization, while ensuring the diversity of vehicle behavior. Normalization is performed so that the initial distribution can reflect driving style characteristics and provide sufficient discrimination to support variational inference.

[0208] S43: Define a set of dynamic hidden variables to address behaviors such as regenerative braking and motor response delay during vehicle operation. ,make Represents the energy-consumption coupled state, a set of dynamic hidden variables. By using random implicit parameter sets Combined to form joint random variables To further refine the characterization of energy consumption, a function is introduced. The effect of dynamic latent variables on observations is described by the following formula:

[0209]

[0210] in, Indicates time The observed vector below, This represents the basic observation components that do not include energy consumption dynamics. Indicates the weights of latent variable coupling. Used to simulate the timing characteristics of braking regeneration rate and follow-through power response.

[0211] S44: For joint random variables To address the limitations of traditional inference speed due to increased dimensionality, a block-based, hierarchical approximate inference method is constructed to accelerate convergence. This method processes general variable blocks and energy consumption coupled variable blocks separately, defining the overall posterior distribution. The factorization model is as follows:

[0212]

[0213] in, Indicates the first The parameter set of a general block, Indicates the first A set of latent variables for an energy consumption coupled variable block. , Indicates the number of blocks. Indicates the first A general block The variational posterior distribution approximation factor, Indicates the first Energy consumption coupled variable block The variational posterior distribution approximation factor;

[0214] Using the variational expectation-maximization method and Continuously iterate and update.

[0215] S45: Define a scenario focus function for extreme road conditions or high power demands detected during the inference process. The estimation accuracy of the corresponding latent variables can be refined by increasing the local sampling frequency, as expressed by the following formula:

[0216]

[0217] in, Indicates time In local sampling areas under extreme road conditions Indicates a high-weight amplification factor. Reflects the degree of scene triggering. Indicates the key inhibition coefficient. This indicates a measurement of energy limits or high-load regions.

[0218] Whenever scene focus function When the value exceeds a set threshold, a specific scene focusing mode is entered, increasing the sampling density in the corresponding latent variable dimension and performing further refined inference. The set threshold is an empirical hyperparameter, which can be initially set... and Normalization, and then in history The sample is defined using quantiles (such as the 95th percentile).

[0219] S46: Combining the above steps, a variational Bayesian analytical inference model is obtained that considers both the coupling relationship between driving operation and energy consumption, and retains multi-layered rule constraints. The model uses joint random variables... With integration as the core The input provides a valid prior for subsequent real-time style recognition and collaborative correction, utilizing the merged joint posterior distribution. The final inference result is expressed by the following formula:

[0220]

[0221] in, This represents the likelihood probability distribution of the observed data. Kullback-Leibler divergence is used to measure the posterior distribution of the population. With prior distribution The differences between them.

[0222] As an embodiment of this application, step S5 specifically includes:

[0223] S51: Convert the hierarchical rule tree With joint posterior distribution Establish dynamic mapping for symbol labels With latent variables Correlation function between A comprehensive evaluation is conducted to generate a multi-layered rule loading function. :

[0224]

[0225] in, This indicates the index of the symbol rule in the hierarchical set. Load weights for symbol labels. For the joint posterior distribution, the correlation function Used for measuring symbol labels For latent variables The constraint strength;

[0226] Loading functions via multi-level rules Obtain a coupled representation of symbolic rules and network inference process to ensure that vehicle dynamics and hierarchical semantics are taken into account.

[0227] S52: Establish hard threshold filtering boundaries for vehicle operation behavior It is used to determine whether excessive or high-risk behavior violates layered safety requirements. If the value exceeds the predetermined threshold (95th percentile), rule validation is immediately performed. The specific formula is expressed as follows:

[0228]

[0229] in, Indicates hierarchy, Indicates a certain driving operation, To activate weights, Representation of symbolic conditions, For latent variables At the level Compensation factor, This represents the network latent variable domain that satisfies the constraints of the current level.

[0230] S53: By comparing time and time The posterior distribution is used to determine the confidence level of the transition, and the distribution variability is defined. ,like If the value exceeds a specified threshold (95th percentile), a significant change in driving style is determined, and the monitoring process will change direction, combining magnitude and duration. The specific formula is expressed as follows:

[0231]

[0232] in, , Representing time respectively and time The posterior distribution, Indicates the Kullback-Leibler divergence;

[0233] S54: When there is a significant difference between the posterior inference and the symbolic prior, the identification of high-risk driver actions or rule tolerance bias will affect the prior symbolic probability. The sign probability obtained from actual inference Comparison:

[0234]

[0235] in, This represents the distance between the prior symbol probability vector and the posterior symbol probability vector obtained from the actual inference. If the threshold (95th percentile) is exceeded, collaborative correction is triggered. By analyzing the causes of the difference, the symbol constraints are adjusted or driving risks are indicated, so as to achieve bidirectional update and iteration of the complete symbol library and the variational Bayesian analytical inference model.

[0236] S55: Define the multinomial loss function This represents the error metric for the variational Bayesian analytical inference model across different scenario dimensions, and the parameters of the variational Bayesian analytical inference model. Dynamic correction is performed, and the formula is expressed as follows:

[0237]

[0238] in, and These represent the error component indices of energy consumption coupling and driving style, respectively. To learn the step size coefficient, Indicates to The gradient;

[0239] When the bias is severe or the external environment fluctuates drastically, variational Bayesian analytical inference of model parameters is possible. Perform dynamic corrections and continuous adjustments. It can maintain the accuracy of coupling energy consumption calculation with driving behavior in high-speed scenarios and under changing road conditions.

[0240] Specifically, this application establishes a hierarchical parsing and variational Bayesian synergy framework at the rule and inference levels. It connects high-level abstract rules with detailed conditions for specific operations such as rapid acceleration or high-speed braking. A hierarchical rule tree is generated to represent the needs and constraints at each level of safety, economy, and comfort. Then, a variational Bayesian network is used to model the coupled state of electric vehicle energy consumption and the latent variables of driving style. Block approximation inference accelerates convergence. Compared to using only probabilistic models or only symbolic classification, the synergistic inference approach can more flexibly capture high-power output or emergency braking behaviors in the event of sudden conditions and immediately combine existing symbolic rules to provide short-term energy limiting or risk warnings. The entire process is continuously iterated through online updates and local corrections to ensure that the system maintains the precision of driving style recognition and the consistency of rule execution under frequent and complex vehicle driving environments.

[0241] S56: If extreme driving maneuvers cause the variational Bayesian analytical inference model to fail to converge, a risk warning will be triggered, and the final updated model parameters will be updated accordingly. Write the posterior distribution back to the system for downstream modules to query; define an alert determination function. If the warning determination function If the threshold is exceeded, an alert will be issued and the final revised results of the variational Bayesian analytic inference model and the complete symbol library will be output, as expressed in the following formula:

[0242]

[0243] in, This indicates a focus on high-risk areas involving extreme driving power or speeding. Let represent the detection functions corresponding to braking, acceleration, and actual wheel-end power, respectively. For the weighting factor;

[0244] when If the threshold (95th percentile) is exceeded, an alert will be issued and the final revised results of the variational Bayesian analytic inference model and the complete symbol library will be output, providing a direct basis for subsequent real-time style recognition and risk intervention.

[0245] As an embodiment of this application, real-time online updates are performed while ensuring the collaboration between multi-level rules and variational Bayesian posterior distribution. Step S6 specifically includes:

[0246] S61: After completing the multi-level rule loading and risk correction in the above steps, the main operating characteristics of the vehicle at present are selected, including real-time vehicle speed, acceleration, steering angle and recent abnormal operations. These characteristics are then combined with the energy consumption safety range to form a basic constraint set, which is used to guide real-time judgment and updates in subsequent short cycles.

[0247] To strengthen the uniformity of constraints, an initial composition state is defined. This includes a weighted combination of driving style weights and energy consumption surplus metrics, expressed by the following formula:

[0248]

[0249] in, This indicates the initial driving style label score identified by the vehicle. This indicates the percentage of remaining energy consumption. and These are adjustable weighting coefficients;

[0250] S62: Real-time sampling of vehicle speed, current output, and pedal operation over short periods, based on a joint posterior distribution. Determine if the vehicle exhibits a style shift; if driving parameters are detected to be inconsistent with... As the differences between the elements gradually increase, a style revision process is triggered, updating the local posterior state to adapt to the latest observations; this is achieved by preserving the initial synthetic state. The constraints enable the fusion of short-period observations with historically accumulated information, taking into account both instantaneous anomalies and the continuity of overall behavior.

[0251] S63: When high impact power or sudden abnormal operation during emergency braking is detected, a rapid matching relationship is established with the complete symbol library to find the corresponding symbol entries and register the relevant cumulative number of times; if the abnormal entries accumulate rapidly within a short period of time, they are recorded as a high-risk operation trend and bound to the current operating characteristics of the vehicle.

[0252] S64: If the cumulative number of high-risk operation trends exceeds a preset threshold, the situation will be marked as a potential safety or energy consumption risk and an internal alarm signal will be issued. Based on the actual vehicle acceleration and electrical fault alarm information, it will be assessed whether it is necessary to immediately limit the vehicle speed, reduce torque, or issue a warning, thereby intervening in high energy consumption and safety hazards in the early stage and improving driving reliability.

[0253] The pre-set threshold is an empirical threshold based on the severity of the event:

[0254] Mild high-risk operations: ≥5 times within 10 minutes;

[0255] Moderate to high-risk procedures: ≥3 times within 10 minutes;

[0256] Highly dangerous operations: ≥2 times within 30 minutes.

[0257] S65: When external road conditions and driving style continue to change, the variational Bayesian analytical inference model and the complete symbol library are synchronously corrected online; ensuring that the system can maintain accuracy under high load or extreme environment; setting an online update rate, using the latest sampling of vehicle status and abnormal operation, to fine-tune the key latent variables of the variational Bayesian analytical inference model and the weights of the complete symbol library, so that it iterates with the fluctuation of monitoring data.

[0258] S66: After completing multiple online corrections to the variational Bayesian analytic inference model and the complete symbol library, the vehicle's final energy consumption information and abnormal operation registration are integrated to generate the latest driving style classification and risk assessment results.

[0259] As an embodiment of this application, step S66 further includes determining the final discrimination score. A simplified metric is used to directly determine the degree of deviation between the vehicle's current state and the target safety boundary. The formula is expressed as follows:

[0260]

[0261] in, Indicates the number of abnormal registrations. This is the anomaly count penalty coefficient. If the value is below the safety threshold, it indicates that the vehicle behavior has deviated significantly and requires a more robust intervention strategy. Otherwise, it outputs a normal driving style label and risk assessment results to provide dynamic support for downstream adaptive driving strategies and energy consumption management.

[0262] The present invention also provides a driving style recognition system based on multi-source sensor data, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described driving style recognition method based on multi-source sensor data.

[0263] This invention fully integrates symbolic representations of driving behavior with variational Bayesian inference strategies, enabling the system to retain the differentiated requirements of multi-layered rules for safety, economy, and comfort, while also tracking energy consumption factors in real time under different operating conditions and making rapid judgments on abnormal operations. This solves the technical challenge of accurately recognizing driving styles under multiple constraints.

[0264] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A driving style recognition method based on multi-source sensor data, characterized in that, The method includes the following steps: S1: Acquire multi-source sensor data and preprocess it to obtain high-quality sequences; S2: Based on the high-quality sequence, extract key features from multi-source sensor data, construct a symbol requirement set and a symbol regulatory compliance matrix, and obtain a complete symbol library; S3: Establish a mapping relationship between the complete symbol library and high-quality sequences, divide the symbol hierarchy according to the three aspects of safety, comfort and economy, define the energy consumption coupling function, and construct a hierarchical rule tree; S4: Based on the hierarchical rule tree, combined with high-quality sequences, establish a multi-layer feature fusion function and probability prior distribution, introduce dynamic latent variables to characterize braking recovery and motor response, construct a block-level approximate inference method to accelerate convergence, obtain a variational Bayesian analytical inference model, and output a joint posterior distribution. S5: Establish a dynamic mapping between the hierarchical rule tree and the joint posterior distribution to form a multi-level rule loading function, establish a hard threshold filtering boundary to filter high-risk operations, compare the symbol prior and posterior, and update the variational Bayesian analytic inference model parameters and the complete symbol library. S6: Monitors vehicle operating status online and updates the parameters of the variational Bayesian analytic inference model and the complete symbol library in real time, outputting updated driving style classification and risk assessment results; Step S1 includes: S11: Synchronously collect the vehicle's speed during operation via sensor array. acceleration Battery status Road conditions and driving operation The collection period is based on a unified time series. Characterization; the data management unit receives and aggregates the outputs of each sensor, forming a dimension of... initial observation matrix The formula is as follows: in, Indicates the first Each sensor at time The original measurement value.

2. The driving style recognition method based on multi-source sensor data according to claim 1, characterized in that, Step S1 further includes: S12: For the initial observation matrix Perform time series alignment to obtain the alignment matrix The formula is as follows: in, This represents an alignment function, which uses interpolation and interpolation techniques to... Mapping to consistent time ; S13: Suppress high-frequency random noise using a sliding window filtering method to form a smooth matrix. The formula is as follows: in, Indicates the first Each sensor at time Smoothed measurement values Indicates the offset; S14: Yes and The difference between the measurements is used to determine a threshold. If the threshold is met, the measurement value is marked as a mutation point and stored in the mutation point identification matrix. The formula is as follows: in, Indicates the threshold for mutation detection; Transformation point identification matrix and smoothness matrix Make a comprehensive judgment based on the vehicle operation. With speed acceleration Road conditions Correlation test at each time step To ensure the validity of the data, invalid data was removed and readings with reasonable correction margins were updated, ultimately yielding a high-quality sequence. , .

3. The driving style recognition method based on multi-source sensor data according to claim 2, characterized in that, Step S2 specifically includes: S21: Based on the high-quality sequence Extract vehicle speed acceleration Battery status Road conditions and driving operation Key features, constructing a symbol requirement set, and using a symbol priority function The key features are quantified, and the specific formula is expressed as follows: in, Indicates an index for different symbol candidates; Indicates high-quality sequence Indexes for different data channels; Indicates at time Data Channel The output value; Indicates the total number of data channels. Indicates the sampling length. Indicates a data channel With symbols The fusion weight, Indicates to Perform nonlinear transformations to capture vehicle operating characteristics; S22: Based on the symbol requirement set constructed in step S21, select common driving behaviors and events such as turning, acceleration, deceleration, intersection waiting, and speed-limited areas to construct an initial symbol set and establish a symbol baseline table. This enables a fast mapping between symbol indexes and semantic parsing, expressed by the following formula: in, This represents the internal encoding of the basic symbol elements. Indicates the number of symbol types. This indicates the scalable subdivision of each symbol; S23: Introduces labels representing micro-driving styles, adds descriptions of motor power response and regenerative braking rate, and constructs an energy consumption mapping function. The energy consumption mapping function describes the relationship between energy consumption and symbolic labels. The formula is expressed as follows: in, Indicates the symbol label index, Symbols and These represent the dimensions of contribution to energy consumption and energy recovery, respectively. and For relevant weights, and Energy management characteristic values ​​in the data channel. and This indicates the total number of channels involved in energy input and output. and These represent nonlinear functions for energy consumption and energy recovery, respectively. S24: Introduce regulatory elements, establish a symbol compliance matrix, and assign each symbol based on speed limits, traffic restrictions, and road priority information. Assigning regulatory compliance coefficients The symbol compliance is recorded as : in, Symbols The corresponding regulatory compliance coefficient, Indicates the time of combination The adjustment amount for the intensity of road supervision at that time. To determine based on vehicle speed With road conditions A filtering function for determining violations or potential hazards; S25: Constructing Symbolic Adaptive Update Equations This enables dynamic adaptation of the symbol library to different seasons or energy consumption constraints, expressed by the following formula: in, Indicates the initial symbol The base attribute values ​​when unaffected by the environment. and For adaptive parameter tuning, and These represent the time variables of battery state and road conditions, respectively. and These represent mapping functions for the battery and the road environment, respectively, used to dynamically adjust symbolic attributes; S26: Building upon the progressive steps from S21 to S25, a complete symbol library covering driving behavior, energy management, and traffic regulations is formed. : The complete symbol library The output includes a basic symbol index, energy conversion coefficient, compliance evaluation, and adaptive management parameters.

4. The driving style recognition method based on multi-source sensor data according to claim 3, characterized in that, Step S3 specifically includes: S31: In the complete symbol library With high-quality sequences The mapping relationship between them is established, expressed by the following formula: in, Represents a globally consistent function, the complete symbol library Include One symbolic entry, high-quality sequence Include One channel, Indicates the first Symbols and channels The fusion weight, Represents a mapping function. Indicates high-quality sequence In the passage ,time The measured value, Represents the complete symbol library The Middle The internal encoding of each symbol; S32: Based on the vehicle's operational needs, a hierarchical symbolic division is established for safety, comfort, and economy. The priority order of each level is represented by... , , This means that for each moment... Hierarchical priority The measurement is expressed by the following formula: in, Indicates a hierarchical index. Indicates hierarchy The normalization factor; ; Indicates channel This level The inhibition coefficient; S33: Define the energy consumption coupling function based on the vehicle's motor power and battery recovery characteristics. Indicates hierarchy At any moment The degree of energy consumption impact will Embedding high-level rules allows the system to strike a balance between security and economy, as expressed in the following formula: in, and These represent the channel indices for energy output and regenerative braking, respectively. Indicates the output power weight. Indicates the weight of the recovery power. and These represent the time of the vehicle. The values ​​of the power output channel and the regenerative force channel, Indicates hierarchy The baseline energy consumption compensation item; S34: Defines a set of operations for specific actions such as rapid acceleration and quick steering. Symbolization conditions Make a judgment: in, Indicates a certain driving operation, Indicates hierarchy The corresponding set of operations, if Then the operation is considered to be at the level. Effective symbolic representation; S35: For hierarchy To mitigate the operational risks associated with high-power or sudden braking, a rule-based matching function is constructed. The formula is expressed as follows: in, Indicates time At the level The control range below, This represents the coefficient that prioritizes amplification of the operation. This represents a coefficient for energy consumption penalty; through the... The value is determined to be greater than or less than zero, and excessive power or sudden braking behavior is identified in real time and energy consumption is limited. S36: Prioritize the layering Energy consumption coupling function Operation symbol conditions Rule matching function This process integrates rules to form a hierarchical rule tree that includes high-level rules and operational breakdowns. The formula is expressed as follows: in, Indicates hierarchy With a certain driving operation The correspondence.

5. The driving style recognition method based on multi-source sensor data according to claim 4, characterized in that, Step S4 specifically includes: S41: Based on the hierarchical rule tree, and combined with high-quality sequences, select energy consumption-related features and driving behavior classification factors, map them to a set of modelable indicators, and establish a multi-layer feature fusion function. The rule constraints and data channels are comprehensively quantified, and the formula is expressed as follows: in, An index representing a modelable metric. Represents the hierarchical rule tree. The activation strength of the rule, Indicates energy consumption-related characteristics, Indicates the classification factors of driving behavior. This represents the fusion weight constant; S42: Order A set of random latent parameters representing driving style, based on a multi-layer feature fusion function. Construct a probability prior distribution from the observations The prior distribution is constrained in the high-dimensional feature space through multiple integrals, as expressed by the following formula: in, Represents the space of random parameters. This represents the total number of integration indicators. Representation and multi-layer feature fusion function The corresponding prior penalty coefficient; S43: Define a set of dynamic hidden variables to address the braking regenerative braking and motor response delay during vehicle operation. ,make To represent the energy-consumption coupled state, a function is introduced. The effect of dynamic latent variables on observations is described by the following formula: in, Indicates time The observed vector below, This represents the basic observation components that do not include energy consumption dynamics. Indicates the weights of latent variable coupling. Used to simulate the timing characteristics of braking regenerative rate and delayed power response; S44: Construct a block-based, hierarchical approximate inference method to accelerate convergence, processing the general variable block and the energy consumption coupled variable block separately, and defining the overall posterior distribution. The factorization model is as follows: in, Indicates the first The parameter set of a general block, Indicates the first The set of hidden variables for each energy-consumption coupling block. , Indicates the number of blocks; Using the variational expectation-maximization method and Continuous iteration and updates; S45: Defines the scenario focus function for extreme road conditions or high power requirements. The estimation accuracy of the corresponding latent variables can be refined by increasing the local sampling frequency, as expressed by the following formula: in, Indicates time In local sampling areas under extreme road conditions Indicates a high-weight amplification factor. Reflects the degree of scene triggering. Indicates the key inhibition coefficient. This indicates a measurement of energy limits or high-load regions. S46: Combining the above steps, we obtain a variational Bayesian analytical inference model, which uses a dynamic set of latent variables. With random hidden parameter sets Joint random variables formed by combination With integration as the core Input, using the merged joint posterior distribution The final inference result is expressed by the following formula: in, This represents the likelihood probability distribution of the observed data. Kullback-Leibler divergence is used to measure the posterior distribution of the population. With prior distribution The differences between them.

6. The driving style recognition method based on multi-source sensor data according to claim 5, characterized in that, Step S5 specifically includes: S51: Convert the hierarchical rule tree With joint posterior distribution Establish dynamic mapping for symbol labels With latent variables Correlation function between A comprehensive evaluation is conducted to generate a multi-layered rule loading function. : in, This indicates the index of the symbol rule in the hierarchical set. Load weights for symbol rules. For the joint posterior distribution, the correlation function Rules for measuring symbols For latent variables The constraint strength; S52: Establish hard threshold filtering boundaries for vehicle operation behavior It is used to determine whether excessive or high-risk behavior violates layered safety requirements. If the value exceeds a predetermined threshold, rule validation is performed immediately. The specific formula is expressed as follows: in, This represents a hierarchical semantic index. Indicates vehicle operation commands. To activate weights, Representation of symbolic conditions, For latent variables At the level Compensation factor, This represents the network latent variable domain that satisfies the constraints of the current level. S53: By comparing time and time The posterior distribution is used to determine the confidence level of the transition, and the distribution variability is defined. ,like If the value exceeds a specified threshold, a significant change in driving style is determined. The specific formula is as follows: in, , Representing time respectively and time The posterior distribution, Indicates the Kullback-Leibler divergence; S54: Distribute prior symbols The sign probability obtained from actual inference Comparison: Among them, if If the threshold is exceeded, collaborative correction is triggered. By analyzing the causes of the differences, symbol constraints are adjusted or driving risks are indicated, so as to achieve bidirectional updating and iteration of the complete symbol library and the variational Bayesian analytical inference model. S55: Define the multinomial loss function This represents the error metric for the variational Bayesian analytical inference model across different scenario dimensions, and the parameters of the variational Bayesian analytical inference model. Dynamic correction is performed, and the formula is expressed as follows: in, and These represent the error component indices of energy consumption coupling and driving style, respectively. To learn the step size coefficient, Indicates to The gradient; S56: Define a warning decision function to address the non-convergence of the variational Bayesian analytical inference model due to extreme driving operations. If the warning determination function If the threshold is exceeded, an alert will be issued and the final revised results of the variational Bayesian analytic inference model and the complete symbol library will be output, as expressed in the following formula: in, This indicates a focus on high-risk areas involving extreme driving power or speeding. This represents the detection function corresponding to braking, acceleration, and actual wheel-end power. This is a weighting factor.

7. The driving style recognition method based on multi-source sensor data according to claim 6, characterized in that, Step S6 specifically includes: S61: Filter out the vehicle's current main operating characteristics and merge them with the energy consumption safety range to form a basic constraint set, and define the initial composite state. The formula is expressed as follows: in, This indicates the initial driving style label score identified by the vehicle. This indicates the percentage of remaining energy consumption. and These are adjustable weighting coefficients; S62: Real-time sampling of vehicle speed, current output, and pedal operation over short periods, based on a joint posterior distribution. Determine if the vehicle exhibits a style shift; if driving parameters are detected to be inconsistent with... As the differences between them gradually increase, a style revision process is triggered, updating the local posterior state to adapt to the latest observations; S63: When high impact power or sudden abnormal operation during emergency braking is detected, a rapid matching relationship is established with the complete symbol library to find the corresponding symbol entries and record the relevant cumulative number of times; if the abnormal entries accumulate rapidly within a short period of time, they are recorded as a high-risk operation trend and bound to the current operating characteristics of the vehicle. S64: If the cumulative number of high-risk operation trends exceeds a preset threshold, the situation will be marked as a potential safety or energy consumption risk and an internal alarm signal will be issued. S65: When external road conditions and driving style continue to change, the variational Bayesian analytical inference model and the complete symbol library are synchronously corrected online. S66: After completing multiple online corrections to the variational Bayesian analytic inference model and the complete symbol library, the vehicle's final energy consumption information and abnormal operation registration are integrated to generate the latest driving style classification and risk assessment results.

8. The driving style recognition method based on multi-source sensor data according to claim 7, characterized in that, Step S66 also includes calculating the final discrimination score. A simplified metric is used to directly determine the degree of deviation between the vehicle's current state and the target safety boundary. The formula is expressed as follows: in, Indicates the number of abnormal registrations. This is the anomaly count penalty coefficient. If the value is below the safety threshold, it indicates a significant deviation in vehicle behavior, requiring enhanced intervention strategies; otherwise, a normal driving style label and risk assessment result will be output.

9. A driving style recognition system based on multi-source sensor data, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the driving style recognition method based on multi-source sensor data as described in any one of claims 1 to 8.

Images