An automobile part transportation abnormal behavior identification method and system

Abstract

The application discloses a kind of transport abnormal behavior identification method and system of automobile parts, it is related to logistics transportation monitoring technical field, it solves the technical problem that present technique often uses fixed, general threshold and model to carry out abnormal determination, cannot adapt to the normal behavior baseline difference brought by different road sections, different driving habits, lead to the false alarm rate of transport abnormal behavior identification is higher;By transport multi-source heterogeneous data stream, standardization multi-source time series data set and joint feature vector are constructed, and normal behavior mode model and prediction value are generated accordingly;Subsequently, abnormal behavior identification and risk scoring are carried out to obtain abnormal event results, and differential evaluation and feedback optimization are carried out to obtain comprehensive results, and abnormal determination is upgraded from one-size-fits-all static rule to intelligent dynamic learning process according to circumstances and time, so as to reduce false alarm rate and improve identification accuracy.

Description

Technical Field

[0001] This application belongs to the field of logistics and transportation monitoring technology, specifically a method and system for identifying abnormal transportation behavior of automotive parts. Background Technology

[0002] In the highly sophisticated modern automotive manufacturing system, parts transportation is not only a logistics link but also a lifeline for maintaining the continuous operation of the production line. Anomaly detection in automotive parts transportation refers to the process of real-time monitoring and intelligent judgment of non-standard conditions throughout the entire transportation process using IoT sensing and big data analytics. Its core meaning lies in going beyond traditional trajectory tracking, deeply perceiving the micro-dynamics of the transport carrier and the goods themselves, and accurately identifying potential risks such as route deviations, illegal parking, severe vibrations caused by sudden acceleration or deceleration, excessive temperature and humidity, unauthorized opening of containers, and non-standard loading and unloading operations. Automotive parts are generally characterized by high precision, high value, and fragility; even minor transportation anomalies can lead to hidden damage, resulting in production line shutdowns or vehicle quality defects. By building an anomaly detection mechanism, companies can not only significantly reduce damage rates and claims costs but also effectively avoid huge production stoppage losses due to missing parts, improving the transparency and resilience of the supply chain.

[0003] Existing technologies for transportation monitoring generally use fixed and universal thresholds and models for anomaly detection, which cannot adapt to the differences in normal behavior baselines caused by different road sections and driving habits, resulting in a high false alarm rate for identifying abnormal transportation behavior. Therefore, methods for identifying abnormal transportation behavior of automotive parts still need further improvement. Summary of the Invention

[0004] This application aims to solve at least one of the technical problems existing in the prior art; to this end, this application proposes a method and system for identifying abnormal behavior in the transportation of automotive parts, which solves the technical problem that the prior art often uses fixed and universal thresholds and models for anomaly judgment, which cannot adapt to the differences in normal behavior baselines caused by different road sections and different driving habits, resulting in a high false alarm rate in the identification of abnormal transportation behavior.

[0005] To achieve the above objectives, the first aspect of this application provides a method for identifying abnormal transportation behavior of automotive parts, comprising: Acquire multi-source heterogeneous transportation data streams; the multi-source heterogeneous transportation data streams include positioning data, motion data, environmental data, and status data; the positioning data includes latitude and longitude, speed, and direction angle; the motion data includes triaxial acceleration and triaxial angular velocity; the environmental data includes temperature and humidity; the status data includes the opening and closing status signals of the vehicle door magnetic and electronic locks; Data processing and feature extraction are performed on multi-source heterogeneous data streams in transportation to obtain a standardized multi-source time-series dataset and joint feature vectors; Based on standardized multi-source time-series datasets and joint feature vectors, normal behavior models are constructed and predicted in real time to obtain normal behavior pattern models and predicted values. Abnormal event results are obtained by identifying abnormal behavior and scoring risks based on joint feature vectors, normal behavior pattern models and predicted values. A comprehensive result is obtained by performing differentiated evaluation and feedback optimization based on the results of abnormal events; the comprehensive result includes abnormal results and optimization results.

[0006] This application, through the aforementioned steps, dynamically constructs a personalized normal behavior pattern model and parameter prediction model for the current transportation task based on real-time multi-source data. This generates a distance score for normal behavior and predicted values ​​for each parameter, thereby establishing an adaptive and accurate baseline for normal behavior in the spatiotemporal dimension. Subsequently, anomaly identification is performed by comparing the deviation between real-time data and the predicted values. A risk scoring mechanism is introduced, and feedback data from manual handling is fed back to drive incremental learning of the normal behavior pattern model and parameter prediction model. This allows them to continuously adapt to the evolution of different transportation and interaction characteristics, upgrading anomaly judgment from a one-size-fits-all static rule to an intelligent dynamic learning process that varies depending on the situation and time, thereby reducing the false alarm rate and improving the identification accuracy.

[0007] Furthermore, the process of processing and extracting features from the multi-source heterogeneous data streams of transportation to obtain a standardized multi-source time-series dataset and a joint feature vector includes: Extracting heterogeneous data streams from multiple sources during transportation; A standardized multi-source time-series dataset is obtained by performing spatiotemporal alignment and data cleaning operations on multi-source heterogeneous transportation data streams. The spatiotemporal alignment and data cleaning operations include time alignment, spatial alignment, and data cleaning operations; Define a sliding time window; the size of the sliding time window is set based on experience. A standardized multi-source time-series dataset is segmented using sliding time windows, and several basic features are extracted within each sliding time window. These basic features include trajectory features, kinematic features, and environmental and event features. The trajectory features include mean path deviation, maximum path deviation, average speed, number of emergency braking events, and dwell time. The kinematic features include peak resultant acceleration, high-frequency vibration energy, and the proportion of roll angle exceeding limits. The environmental and event features include duration of temperature exceeding limits and illegal door opening events. Several interactive features are constructed based on several fundamental features; these interactive features include spatiotemporal correlation features, motion environment correlation features, and event sequence features; the spatiotemporal correlation features include high vibration path deviation ratio and abnormal dwell vibration level; the motion environment correlation features include stability during temperature changes; and the event sequence features include emergency braking rear door events. A joint feature vector is obtained by concatenating several basic features and several interactive features.

[0008] This application first performs spatiotemporal alignment and cleaning on multi-source heterogeneous data to form a standardized time-series dataset. Then, it uses a sliding time window to segment the data and extracts three basic features—trajectory, kinematics, and environmental events—within each window. Based on these basic features, it further constructs interactive features that reveal cross-dimensional relationships and concatenates the basic and interactive features into a joint feature vector. The joint feature vector not only contains the state variables of each independent dimension but also deeply encodes the spatiotemporal coupling relationship of different sensor data. This enables it to accurately characterize complex abnormal behavior patterns such as severe vibrations occurring while deviating from the path, providing high-information-density input for subsequent normal behavior pattern models and fundamentally improving the ability to perceive and distinguish complex abnormal scenarios.

[0009] Furthermore, the step of constructing and predicting a normal behavior pattern model and predicted values ​​in real time based on a standardized multi-source time-series dataset and joint feature vectors includes: Several historical joint feature vectors of normal transportation labels are obtained and used as training data for the normal behavior model; the normal transportation label refers to the transportation behavior during the transportation process being in a normal state. A normal behavior pattern model is obtained by training a Gaussian mixture model using normal behavior model training data; the normal behavior pattern model is used to characterize the data distribution of the joint feature vectors corresponding to normal transportation behavior. Obtain standardized multi-source time series datasets corresponding to several historical sliding time windows, and integrate them into a multi-source time series composite dataset according to the chronological order of the sliding time windows. The multi-source time series composite dataset is input into the parameter prediction model to obtain the predicted values ​​of several parameters in the standardized multi-source time series dataset; the parameter prediction model is constructed using an LSTM model to predict the values ​​of several parameters in the future time period.

[0010] This application utilizes labeled historical data to model the probability density of high-dimensional joint feature vectors using a Gaussian mixture model, characterizing the static multi-feature joint distribution of normal transportation behavior. Simultaneously, it learns from standardized multi-source time-series data using an LSTM time-series prediction model to predict sensor parameter values ​​in the near future. The Gaussian mixture model defines the normal pattern at the feature space level, while the LSTM model predicts the expected state at the temporal evolution level. Together, they constitute a multi-dimensional, dynamic reference system for normal behavior. This allows anomaly detection to focus not only on whether the real-time state deviates from the normal cluster but also on whether it violates evolutionary patterns. This enables more sensitive and accurate detection of abnormal behavior with deviations in any dimension of static features or dynamic sequences, improving the robustness and accuracy of identification.

[0011] Furthermore, the process of identifying abnormal behavior and scoring risks based on joint feature vectors, normal behavior pattern models, and predicted values ​​to obtain abnormal event results includes: Extract the normal behavior pattern model and predicted values, as well as the joint feature vector of the previous sliding time window; The joint feature vector is input into the normal behavior pattern model to obtain the distance score between the joint feature vector and the data distribution; Real-time acquisition of multi-source heterogeneous data streams in transportation; The root mean square error method is used to calculate the basic residual values ​​between several real-time parameter values ​​and the corresponding predicted values ​​of several parameters in a multi-source heterogeneous data stream. The basic residual values ​​are normalized to obtain standardized basic residual values; the residual normalization operation refers to dividing the basic residual value by the residual standard deviation corresponding to the parameter, and then performing Z-score normalization calculation. The prediction bias score is obtained by weighted fusion of the standardized basic residual values ​​corresponding to several parameters. The comprehensive anomaly score corresponding to the current sliding time window is obtained by weighted summation of the prediction deviation score and the distance score. When the overall anomaly score is greater than the anomaly threshold, an anomaly event label is generated, and the occurrence time and geographic coordinates of the anomaly event label are obtained. Preliminary anomaly types are obtained through rule matching based on joint feature vectors; The results of anomalies are determined based on the event label, preliminary anomaly type, occurrence time, geographic coordinates, and comprehensive anomaly score.

[0012] Furthermore, the method for obtaining the abnormal threshold includes: Obtain the global statistical measure of the comprehensive anomaly score corresponding to several historical sliding time windows under normal transport labels; the global statistical measure includes the mean of the comprehensive anomaly score. and the standard deviation of the composite abnormal score ; Calculate the basic anomaly threshold The formula for calculating the basic anomaly threshold satisfies: Where n represents the global adjustment coefficient, n>0; Obtain the road type corresponding to the current sliding time window and extract its corresponding road segment coefficient. The road types mentioned include expressways, national highways, urban roads, and mountain roads. Get the average score of the comprehensive anomaly scores within the most recent M sliding time windows. and standard deviation of scores Where M is an integer, M>1; Calculate the anomaly threshold YY; the formula for calculating the anomaly threshold satisfies: ; and Represented as weighting coefficients, and ∈(0,1); m represents the local adjustment coefficient, m>0.

[0013] Furthermore, the preliminary anomaly type obtained by rule matching based on the joint feature vector includes: Extract several basic features and several interaction features from the joint feature vector; Based on several basic features and several interactive features, a preliminary anomaly type is obtained by performing rule matching from the feature type mapping rule base; The feature type mapping rule base consists of several conditions and their corresponding preliminary anomaly types; the conditions are constructed from several basic features and several interactive features.

[0014] First, a comprehensive anomaly score is constructed by integrating distance scores based on feature distribution and prediction deviation scores based on prediction bias to fully capture potential risks. Then, by combining global statistical patterns, real-time road type coefficients, and recent fluctuation levels, an adaptive threshold is dynamically calculated, allowing the judgment criteria to be flexibly adjusted according to road conditions and driving behavior. Once an anomaly is triggered, the joint feature vector is quickly matched using a predefined rule base to immediately output the preliminary anomaly type. By dynamically fusing multi-dimensional signals and adopting a context-aware threshold strategy, false alarms caused by differences in road conditions or normal driving fluctuations are effectively suppressed. At the same time, real-time rule matching completes the initial type screening immediately when an anomaly is determined, providing a direct basis for subsequent rapid and differentiated risk assessment and handling.

[0015] Furthermore, the comprehensive result obtained by differential evaluation and feedback optimization based on the results of abnormal events includes: Obtain the results of abnormal events and waybill information; Determine the abnormal outcome based on the abnormal event results and waybill information; Obtain feedback data for abnormal results; the feedback data refers to the data obtained from the final verification of the abnormal results, including feedback tags and abnormal results; the feedback tags include confirmed abnormality and confirmed false alarm. Several abnormal results labeled as confirmed false alarms are stored in the optimization data pool to obtain the optimization dataset; The optimization results were obtained by optimizing the normal behavior pattern model and the parameter prediction model based on the optimized dataset; The overall result is determined based on the abnormal results and the optimization results.

[0016] Furthermore, the determination of the abnormal result based on the abnormal event result and the waybill information includes: Extract the preliminary anomaly type, occurrence time, geographic coordinates, and overall anomaly score from the results of abnormal events; An anomaly risk level mapping operation is performed on the preliminary anomaly type, occurrence time, and geographic coordinates to obtain the anomaly type level, time risk level, and location risk level. The anomaly risk level mapping maps numerical data and categorical data to the same numerical range, which facilitates the subsequent determination of anomaly results. Extract the value tags of the transported goods from the waybill information; the value tags of the transported goods include high value, medium value, and low value. The value rating of transported goods is obtained by mapping the value tags of the transported goods to different levels. The comprehensive risk score is obtained by weighted summation of the anomaly type level, time risk level, location risk level, cargo value level, and normalized comprehensive anomaly score. When the comprehensive risk score is greater than or equal to the high-risk threshold, the final anomaly label will be set to high-risk anomaly. When the overall risk score is less than the high-risk threshold and the overall risk score is greater than or equal to the medium-risk threshold, the final anomaly label will be set to medium-risk anomaly. Otherwise, set the final anomaly label to low-risk anomaly; The anomaly results are determined based on the final anomaly label, the anomaly event results, the corresponding joint feature vector, distance score, several predicted values, and the standardized multi-source time series dataset.

[0017] Furthermore, the optimization results obtained by optimizing the normal behavior pattern model and parameter prediction model based on the optimized dataset include: Obtain the optimized dataset and count the number of samples of outlier results in the optimized dataset; When the number of samples reaches the optimized number threshold Extract the joint feature vectors and their corresponding distance scores from several outlier results in the optimization dataset, and integrate them into a pattern optimization dataset. Extract the standardized multi-source time series datasets corresponding to the sliding time windows of several outlier results in the optimized dataset and their corresponding predicted values; Using each sliding time window in the optimized dataset as the center of the time window, the standardized multi-source time series datasets of several activity time windows before and after the sliding time window and their corresponding predicted values ​​are obtained, and combined with the standardized multi-source time series datasets and their corresponding predicted values ​​of each sliding time window in the optimized dataset to form multi-source time series composite data and time series predicted values. A prediction optimization dataset is composed of several multi-source time-series integrated data and time-series predicted values; The normal behavior pattern model and the parameter prediction model were updated using incremental learning methods on the pattern optimization dataset and the prediction optimization dataset, respectively, to obtain the optimized normal behavior pattern model and parameter prediction model. The optimization results are determined based on the optimized normal behavior pattern model and parameter prediction model.

[0018] This application first integrates multiple dimensions such as anomaly type, spatiotemporal data, and cargo value to calculate a comprehensive risk score, enabling accurate classification and differentiated handling of high-risk events. Then, it collects feedback data after manual verification, especially storing samples confirmed as false alarms and their corresponding feature vectors along with the original time-series data into an optimized dataset. When the data accumulates to a certain scale, the normal behavior pattern model and parameter prediction model are optimized and updated using incremental learning techniques based on the feature vectors of these samples and the original time-series data within the extended time window. This not only upgrades anomaly alarms to risk warnings that can guide response priorities but also utilizes feedback data generated in actual operations, especially false alarm samples, to drive the model to continuously correct its cognitive boundaries regarding normal behavior, thereby continuously reducing the false alarm rate.

[0019] Another aspect of this application provides a system for identifying abnormal transportation behavior of automotive parts, comprising: a data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected together; The data acquisition module transports multi-source heterogeneous data streams through data acquisition devices; the transported multi-source heterogeneous data streams include positioning data, motion data, environmental data, and status data; The data analysis module includes a feature construction unit, an anomaly initial judgment unit, and a result generation unit; The feature construction unit performs data processing and feature extraction on multi-source heterogeneous transportation data streams to obtain a standardized multi-source time-series dataset and a joint feature vector. The anomaly initial judgment unit: constructs and predicts normal behavior patterns in real time based on standardized multi-source time-series datasets and joint feature vectors to obtain normal behavior pattern models and predicted values; and identifies and scores abnormal behavior based on joint feature vectors, normal behavior pattern models, and predicted values ​​to obtain abnormal event results. The result generation unit performs differentiated evaluation and feedback optimization based on the results of abnormal events to obtain a comprehensive result.

[0020] Compared with the prior art, the beneficial effects of this application are: 1. This application obtains a standardized multi-source time-series dataset and joint feature vector by processing and extracting features from multi-source heterogeneous transportation data streams; it constructs and predicts normal behavior patterns based on the standardized multi-source time-series dataset and joint feature vector, obtaining normal behavior pattern models and predicted values; it identifies abnormal behavior and scores risks based on the joint feature vector, normal behavior pattern model, and predicted values, obtaining abnormal event results; it performs differentiated evaluation and feedback optimization based on abnormal event results to obtain comprehensive results; it dynamically constructs personalized normal behavior pattern models and parameter prediction models for the current transportation task based on real-time multi-source data, and generates distance scores for normal behavior and predicted values ​​for each parameter, thereby establishing an adaptive and accurate normal behavior baseline in the spatiotemporal dimension; it then identifies anomalies by comparing the deviation between real-time data and the predicted values, and introduces a risk scoring mechanism to feed back feedback data from manual handling, driving incremental learning of the normal behavior pattern model and parameter prediction model, enabling them to continuously adapt to the evolution of different transportation and interaction characteristics, upgrading anomaly judgment from a one-size-fits-all static rule to an intelligent dynamic learning process that varies depending on the event and time, thereby reducing the false alarm rate and improving the identification accuracy.

[0021] 2. This application utilizes labeled historical data. On the one hand, it uses a Gaussian mixture model to model the probability density of high-dimensional joint feature vectors, characterizing the static multi-feature joint distribution of normal transportation behavior. On the other hand, it uses an LSTM time-series prediction model to learn standardized multi-source time-series data and predict sensor parameter values ​​in the short term. The Gaussian mixture model defines what constitutes a normal pattern at the feature space level, while the LSTM model predicts what constitutes an expected state at the time evolution level. Together, they constitute a multi-dimensional, dynamic reference system for normal behavior. This allows anomaly detection to focus not only on whether the real-time state deviates from the normal cluster, but also on whether it violates the evolutionary rules. This enables more sensitive and accurate detection of abnormal behavior that deviates in any dimension of static features or dynamic sequences, improving the robustness and accuracy of identification.

[0022] 3. This application first integrates distance scores based on feature distribution and prediction deviation scores based on prediction deviation to form a comprehensive anomaly score to fully capture anomalies; then, it combines global statistics, real-time road type coefficients, and recent fluctuation levels to dynamically calculate adaptive thresholds, enabling the judgment criteria to be flexibly adjusted according to road conditions and recent behavior; once an anomaly is triggered, the joint feature vector is quickly matched through a predefined feature rule library to immediately output the preliminary anomaly type. By dynamically fusing multi-dimensional anomaly signals and using context-aware thresholds, false alarms caused by differences in road conditions or normal driving fluctuations are reduced; at the same time, the real-time rule matching mechanism completes the initial type screening immediately when an anomaly is determined, providing a direct basis for subsequent rapid and differentiated risk assessment and handling. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a method for identifying abnormal transportation behavior of automotive parts according to this application; Figure 2 This is a schematic diagram illustrating the principle of an abnormal transportation behavior recognition system for automotive parts according to this application. Detailed Implementation

[0025] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] Please see Figure 1 The first aspect of this application provides a method for identifying abnormal transportation behavior of automotive parts, including: Acquire multi-source heterogeneous data streams for transportation; these data streams include location data, motion data, environmental data, and status data; location data includes latitude and longitude, speed, and orientation angle; motion data includes triaxial acceleration and triaxial angular velocity; environmental data includes temperature and humidity; and status data includes the opening and closing status signals of the vehicle's door magnetic sensors and electronic locks. Data processing and feature extraction are performed on multi-source heterogeneous data streams in transportation to obtain a standardized multi-source time-series dataset and joint feature vectors; Based on standardized multi-source time-series datasets and joint feature vectors, normal behavior models are constructed and predicted in real time to obtain normal behavior pattern models and predicted values. Abnormal event results are obtained by identifying abnormal behavior and scoring risks based on joint feature vectors, normal behavior pattern models and predicted values. A comprehensive result is obtained by differential evaluation and feedback optimization based on the results of abnormal events; the comprehensive result includes abnormal results and optimization results.

[0027] In this embodiment, data processing and feature extraction are performed on multi-source heterogeneous data streams in transportation to obtain a standardized multi-source time-series dataset and a joint feature vector, including: Extracting heterogeneous data streams from multiple sources during transportation; A standardized multi-source time-series dataset is obtained by performing spatiotemporal alignment and data cleaning operations on multi-source heterogeneous transportation data streams. Spatiotemporal alignment and data cleaning operations include time alignment, spatial alignment, and data cleaning. In this embodiment, time alignment refers to correcting and synchronizing the timestamps of all data streams using a unified high-precision clock source; spatial alignment refers to converting vehicle motion data to a geographic coordinate system for joint analysis with positioning data; and data cleaning refers to removing outliers and filling in missing values ​​for lost data. Define a sliding time window; the size of the sliding time window is set based on experience, and in this embodiment, the size of the sliding time window is set to 60 seconds. The standardized multi-source time-series dataset is segmented using a sliding time window, and several basic features are extracted within each sliding time window. These basic features include trajectory features, kinematic features, and environmental and event features. Trajectory features include mean path deviation, maximum path deviation, average speed, number of emergency braking events, and dwell time. Kinematic features include peak resultant acceleration, high-frequency vibration energy, and the proportion of roll angle exceeding limits. Environmental and event features include the duration of temperature exceeding limits and illegal door opening events. In this embodiment, the mean path deviation is calculated using the path deviation degree, which represents the shortest vertical distance between the transport trajectory and the preset planned path within the sliding time window. The maximum path deviation refers to the maximum value in the path deviation degree. The average speed refers to the average speed in the positioning data within the sliding time window. The number of emergency braking events refers to the number of times the acceleration of the speed falls below the emergency braking threshold, which is set empirically; in this embodiment, the emergency braking threshold is set to -2.5 m / s². Dwell time refers to the duration of a dwell point. Dwell points are identified by spatial clustering analysis of transport trajectory points within a sliding time window, identifying dense clusters of points with speeds close to 0, and these clusters are used as dwell points. The peak value of the resultant acceleration is determined by the maximum value of the resultant acceleration, which is determined by the vector magnitude of the triaxial acceleration. High-frequency vibration energy refers to the energy proportion of a specific frequency band after FFT transformation of the vector magnitude of the triaxial acceleration. In this embodiment, the specific frequency band is the 5Hz-15Hz range. The roll angle exceeding the limit ratio refers to the ratio of the duration during which the vehicle's roll angle exceeds the safe angle within the sliding time window, with the safe angle set to 10 degrees. The temperature exceeding the limit duration refers to the ratio of the duration during which the temperature exceeds the temperature threshold within the sliding time window, with the temperature threshold set to 30 degrees Celsius. An illegal door opening event refers to an event where the magnetic and electronic lock signals of the vehicle door are in the open state at unplanned dwell points or during travel. Several interactive features are constructed based on several basic features. These interactive features include spatiotemporal correlation features, motion environment correlation features, and event sequence features. The spatiotemporal correlation features include high vibration path deviation ratio and abnormal stop vibration level. In this embodiment, the high vibration path deviation ratio refers to the average value of the corresponding peak resultant acceleration during the period when the path deviation exceeds a threshold. The average value of the peak resultant acceleration can effectively identify rough handling that occurs on remote sections of the deviated route. The abnormal stop vibration level refers to the average value of the high-frequency vibration energy during the identified stop point. Normal parking results in low vibration, while abnormal opening or transfer results in high vibration. The motion environment correlation features include stability during temperature changes. During the period when the rate of temperature change exceeds its corresponding threshold, the standard deviation corresponding to the average tilt angle is calculated as the stability during temperature changes. The threshold corresponding to the rate of temperature change is set to 15%. The event sequence features include emergency braking back door event. In this embodiment, the emergency braking back door event refers to whether an illegal door opening event is triggered within the time window after the emergency deceleration event. A joint feature vector is obtained by concatenating several basic features and several interactive features.

[0028] This embodiment first performs spatiotemporal alignment and cleaning on multi-source heterogeneous data to construct a standardized time-series dataset. Then, it uses a sliding time window for slicing, extracting three basic features—trajectory, kinematics, and environmental events—within each window. Based on this, it further mines cross-dimensional correlations to construct interactive features, and merges the basic features and interactive features into a joint feature vector. This vector not only encompasses the state quantities of each independent dimension but also deeply encodes the spatiotemporal coupling relationship of multi-sensor data, thereby accurately characterizing complex abnormal behavior patterns such as path deviation accompanied by severe vibration. This high-information-density feature representation provides high-quality input for subsequent normal behavior pattern models, fundamentally enhancing the perception sensitivity and discrimination ability of complex abnormal scenarios.

[0029] This embodiment describes the construction and real-time prediction of a normal behavior model based on a standardized multi-source time-series dataset and joint feature vectors, resulting in a normal behavior pattern model and predicted values. Obtain several historical joint feature vectors of normal transportation labels and use them as training data for the normal behavior model; the normal transportation label refers to the transportation behavior during the transportation process being in a normal state. The normal behavior pattern model is obtained by training the Gaussian mixture model with normal behavior model training data; the normal behavior pattern model is used to characterize the data distribution of the joint feature vector corresponding to normal transportation behavior. Obtain standardized multi-source time series datasets corresponding to several historical sliding time windows, and integrate them into a multi-source time series composite dataset according to the chronological order of the sliding time windows. The multi-source time series composite dataset is input into the parameter prediction model to obtain the predicted values ​​of several parameters in the standardized multi-source time series dataset. The parameter prediction model is constructed using an LSTM model to predict the values ​​of several parameters in the future. In this embodiment, the Gaussian mixture model and the LSTM model used are both general models, and their training process is consistent with that of general models.

[0030] In this embodiment, abnormal behavior identification and risk scoring based on joint feature vectors, normal behavior pattern models, and predicted values ​​yield abnormal event results, including: Extract the normal behavior pattern model and predicted values, as well as the joint feature vector of the previous sliding time window; The joint feature vector is input into the normal behavior pattern model to obtain the distance score between the joint feature vector and the data distribution; Real-time acquisition of multi-source heterogeneous data streams in transportation; The root mean square error method is used to calculate the basic residual values ​​between several real-time parameter values ​​and the corresponding predicted values ​​of several parameters in a multi-source heterogeneous data stream. The basic residual values ​​are normalized to obtain standardized basic residual values. The residual normalization operation is to divide the basic residual value by the residual standard deviation corresponding to the parameter, and then perform Z-score normalization calculation. In this embodiment, the numerical range and physical meaning of the parameter are different. Therefore, it is necessary to perform residual normalization operation to facilitate the subsequent calculation of prediction deviation score. The prediction bias score is obtained by weighting and fusing the standardized basic residual values ​​corresponding to several parameters; in this embodiment, the weight coefficients corresponding to several parameters are set based on experience. The comprehensive anomaly score corresponding to the current sliding time window is obtained by weighted summation of the prediction deviation score and the distance score; the weight coefficients corresponding to the prediction deviation score and the distance score are set according to experience, and in this embodiment they are set to 0.5 and 0.5 respectively; When the overall anomaly score is greater than the anomaly threshold, an anomaly event label is generated, and the occurrence time and geographic coordinates of the anomaly event label are obtained. Preliminary anomaly types are obtained through rule matching based on joint feature vectors; The results of anomalies are determined based on the event label, preliminary anomaly type, occurrence time, geographic coordinates, and comprehensive anomaly score.

[0031] The method for obtaining the abnormal threshold in this embodiment includes: Obtain the global statistical measure of the comprehensive anomaly score corresponding to several historical sliding time windows under the normal transport label; the global statistical measure includes the mean of the comprehensive anomaly score. and the standard deviation of the composite abnormal score ; Calculate the basic anomaly threshold The formula for calculating the basic anomaly threshold satisfies: Where n represents the global adjustment coefficient, n>0; the specific value is set according to experience, and in this embodiment it is set to 3; Obtain the road type corresponding to the current sliding time window and extract its corresponding road segment coefficient. Road types include expressways, national highways, urban roads, and mountain roads; the road segment coefficients correspond to the road types. In this embodiment, the road segment coefficients for expressways, national highways, urban roads, and mountain roads are set to 1, 1.3, 1.8, and 2.5, respectively. Get the average score of the comprehensive anomaly scores within the most recent M sliding time windows. and standard deviation of scores Where M is an integer, M>1; the specific value is set according to experience, and in this implementation it is set to 10; Calculate the anomaly threshold YY; the formula for calculating the anomaly threshold satisfies: ; and Represented as weighting coefficients, and ∈(0,1), the specific values ​​are set according to experience. In this embodiment, they are set to 0.7 and 0.3 respectively to adjust the importance of global information and local information; m represents the local adjustment coefficient, m>0; the specific values ​​are set according to experience. In this embodiment, it is set to 2.

[0032] In this embodiment, the preliminary anomaly type is obtained by rule matching based on the joint feature vector, including: Extract several basic features and several interaction features from the joint feature vector; Based on several basic features and several interactive features, a preliminary anomaly type is obtained by performing rule matching from the feature type mapping rule base; The feature type mapping rule base consists of several conditions and their corresponding preliminary anomaly types; the conditions are constructed from several basic features and several interactive features; in this embodiment, some rules of the feature type mapping rule base include: Rule 1: When the mean path deviation is high and the peak resultant acceleration is high, the preliminary anomaly type is: path deviation accompanied by severe vibration; Rule 2: When there is an illegal door opening incident, a long dwell time, and a low average path deviation, the preliminary anomaly type is: unplanned stop and opening of container.

[0033] In this embodiment, the comprehensive result obtained through differentiated evaluation and feedback optimization based on the results of abnormal events includes: Obtain the results of abnormal events and waybill information; Determine the abnormal outcome based on the abnormal event results and waybill information; Obtain feedback data for abnormal results; feedback data refers to the data obtained from the final verification of abnormal results, including feedback labels and abnormal results; feedback labels include confirmed abnormalities and confirmed false alarms; Several abnormal results labeled as confirmed false alarms are stored in the optimization data pool to obtain the optimization dataset; The optimization results were obtained by optimizing the normal behavior pattern model and the parameter prediction model based on the optimized dataset; The overall result is determined based on the abnormal results and the optimization results.

[0034] In this embodiment, determining the abnormal result based on the abnormal event result and waybill information includes: Extract the preliminary anomaly type, occurrence time, geographic coordinates, and overall anomaly score from the results of abnormal events; An anomaly risk level mapping operation is performed on the preliminary anomaly type, occurrence time, and geographic coordinates to obtain the anomaly type level, time risk level, and location risk level. Anomaly risk level mapping maps numerical and categorical data to the same numerical range, facilitating subsequent anomaly determination. In this embodiment, the anomaly type levels for path deviation accompanied by severe vibration and unplanned parking / unloading among several preliminary anomaly types are set to 0.7 and 0.9, respectively. Geographic coordinates are used to classify the areas into remote areas, suburbs, and urban areas, and the location risk levels for these areas are set to 1, 0.7, and 0.3, respectively. Events are classified into holiday nights, holiday daytimes, weekday nights, and weekday daytimes, and their corresponding time risk levels are set to 1, 0.7, 0.6, and 0.3, respectively. Extract the cargo value tags from the waybill information; cargo value tags include high value, medium value, and low value. The value tags of transported goods are mapped to different levels to obtain the value levels of the goods. In this embodiment, the value levels of high value, medium value, and low value are set to 1, 0.6, and 0.3, respectively. The comprehensive risk score is obtained by weighted summation of the anomaly type level, time risk level, location risk level, cargo value level, and normalized comprehensive anomaly score. The weight coefficients corresponding to the anomaly type level, time risk level, location risk level, cargo value level, and normalized comprehensive anomaly score are set based on experience. In this embodiment, they are set to 0.2, 0.05, 0.2, 0.15, and 0.4, respectively. When the comprehensive risk score is greater than or equal to the high-risk threshold, the final anomaly label will be set to high-risk anomaly. When the overall risk score is less than the high-risk threshold and the overall risk score is greater than or equal to the medium-risk threshold, the final anomaly label will be set to medium-risk anomaly. Otherwise, the final anomaly label is set to low-risk anomaly; the high-risk threshold and medium-risk threshold are set based on experience, and in this embodiment they are set to 0.8 and 0.5 respectively. The anomaly results are determined based on the final anomaly label, the anomaly event results, the corresponding joint feature vector, distance score, several predicted values, and the standardized multi-source time series dataset.

[0035] In this embodiment, the optimization results obtained by optimizing the normal behavior pattern model and parameter prediction model based on the optimized dataset include: Obtain the optimized dataset and count the number of samples of outlier results in the optimized dataset; When the number of samples reaches the optimization threshold, the optimization threshold is set based on experience, and in this embodiment it is set to 1000. Extract the joint feature vectors and their corresponding distance scores from several outlier results in the optimization dataset, and integrate them into a pattern optimization dataset. Extract the standardized multi-source time series datasets corresponding to the sliding time windows of several outlier results in the optimized dataset and their corresponding predicted values; Using each sliding time window in the optimized dataset as the center of the time window, we obtain the standardized multi-source time series datasets of several activity time windows before and after the sliding time window and their corresponding predicted values, and combine them with the standardized multi-source time series datasets and their corresponding predicted values ​​corresponding to each sliding time window in the optimized dataset to form multi-source time series composite data and time series predicted values. A prediction optimization dataset is composed of several multi-source time-series integrated data and time-series predicted values; The normal behavior pattern model and the parameter prediction model are updated using an incremental learning method with the pattern optimization dataset and the prediction optimization dataset, respectively, to obtain the optimized normal behavior pattern model and the parameter prediction model. In this embodiment, the optimized normal behavior pattern model and the parameter prediction model are used in the generation of predicted values ​​and distance scores in the subsequent sliding time window. The optimization results are determined based on the optimized normal behavior pattern model and parameter prediction model.

[0036] This embodiment first integrates multiple factors such as anomaly type, spatiotemporal context, and cargo value to calculate a comprehensive risk score, enabling accurate classification and differentiated handling of high-risk events. Subsequently, it collects feedback from manual verification, focusing on incorporating samples confirmed as false alarms and their corresponding feature vectors along with the original time-series data into the optimized dataset. When the accumulated data reaches a preset scale, incremental learning technology is used to iteratively update the normal behavior pattern model and parameter prediction model based on the feature vectors of these samples and the original data within the extended time window. This not only upgrades simple anomaly alarms to risk warnings that can guide response priorities, but also continuously corrects the model's cognitive boundaries of normal behavior through feedback data from actual operations, thereby significantly reducing the false alarm rate in dynamic evolution.

[0037] Please see Figure 2 Another embodiment of this application provides a system for identifying abnormal transportation behavior of automotive parts, including: a data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected. Data acquisition module: transports multi-source heterogeneous data streams through data acquisition equipment; the transported multi-source heterogeneous data streams include positioning data, motion data, environmental data, and status data; the data acquisition equipment in this embodiment includes several sensors, etc. The data analysis module includes a feature construction unit, an anomaly initial judgment unit, and a result generation unit; Feature construction unit: Performs data processing and feature extraction on multi-source heterogeneous transportation data streams to obtain a standardized multi-source time-series dataset and joint feature vector; Anomaly Preliminary Judgment Unit: Based on standardized multi-source time-series datasets and joint feature vectors, a normal behavior model is constructed and predicted in real time to obtain a normal behavior pattern model and predicted value; based on joint feature vectors, normal behavior pattern model and predicted value, abnormal behavior is identified and risk scoring is performed to obtain abnormal event results; Result generation unit: Based on the results of abnormal events, differentiated evaluation and feedback optimization are performed to obtain comprehensive results.

[0038] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0039] The working principle of this application is as follows: First, acquire multi-source heterogeneous transportation data streams. Second, process and extract features from these data streams to obtain a standardized multi-source time-series dataset and joint feature vectors. Third, construct and predict normal behavior patterns in real time based on the standardized multi-source time-series dataset and joint feature vectors to obtain normal behavior pattern models and predicted values. Fourth, identify and score abnormal behavior based on the joint feature vectors, normal behavior pattern models, and predicted values ​​to obtain abnormal event results. Fifth, perform differentiated evaluation and feedback optimization based on the abnormal event results to obtain a comprehensive result. Finally, dynamically construct personalized normal behavior pattern models and parameter prediction models for the current transportation task based on real-time multi-source data, and generate distance scores for normal behavior and predicted values ​​for each parameter. This approach establishes an adaptive and accurate baseline for normal behavior in the spatiotemporal dimensions. Subsequently, anomalies are identified by comparing the deviation between real-time data and the predicted value. A risk scoring mechanism is introduced, and feedback data from manual handling is fed back to drive incremental learning of the normal behavior pattern model and the parameter prediction model. This allows the model to continuously adapt to the evolution of different transportation and interaction characteristics, upgrading anomaly judgment from a one-size-fits-all static rule to an intelligent dynamic learning process that varies from event to event and time. This reduces the false alarm rate and improves the recognition accuracy. It avoids the problem that existing technologies often use fixed and universal thresholds and models for anomaly judgment, which cannot adapt to the differences in normal behavior baselines caused by different road sections and driving habits, resulting in a high false alarm rate for transportation anomaly behavior recognition.

[0040] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A method for identifying abnormal transportation behavior of automotive parts, characterized in that, include: Acquire multi-source heterogeneous data streams from transportation sources; The multi-source heterogeneous data stream for transportation includes location data, motion data, environmental data, and status data; Data processing and feature extraction are performed on multi-source heterogeneous data streams in transportation to obtain a standardized multi-source time-series dataset and joint feature vector; Normal behavior pattern model and predicted value are obtained by constructing a normal behavior model and making real-time prediction based on standardized multi-source time series datasets and joint feature vectors. Abnormal event results are obtained by identifying abnormal behavior and scoring risks based on joint feature vectors, normal behavior pattern models and predicted values. A comprehensive result is obtained by differential evaluation and feedback optimization based on the results of abnormal events.

2. The method for identifying abnormal transportation behavior of automotive parts according to claim 1, characterized in that, The process of processing and extracting features from multi-source heterogeneous data streams in transportation to obtain a standardized multi-source time-series dataset and joint feature vectors includes: Extracting heterogeneous data streams from multiple sources during transportation; A standardized multi-source time-series dataset is obtained by performing spatiotemporal alignment and data cleaning operations on multi-source heterogeneous transportation data streams. The spatiotemporal alignment and data cleaning operations include time alignment, spatial alignment, and data cleaning operations; Define a sliding time window; the size of the sliding time window is set based on experience. A standardized multi-source time-series dataset is segmented using sliding time windows, and several basic features are extracted within each sliding time window. These basic features include trajectory features, kinematic features, and environmental and event features. The trajectory features include mean path deviation, maximum path deviation, average speed, number of emergency braking events, and dwell time. The kinematic features include peak resultant acceleration, high-frequency vibration energy, and the proportion of roll angle exceeding limits. The environmental and event features include duration of temperature exceeding limits and illegal door opening events. Several interactive features are constructed based on several fundamental features; these interactive features include spatiotemporal correlation features, motion environment correlation features, and event sequence features; the spatiotemporal correlation features include high vibration path deviation ratio and abnormal dwell vibration level; the motion environment correlation features include stability during temperature changes; and the event sequence features include emergency braking rear door events. A joint feature vector is obtained by concatenating several basic features and several interactive features.

3. The method for identifying abnormal transportation behavior of automotive parts according to claim 1, characterized in that, The process of constructing and predicting normal behavior patterns and forecasts based on standardized multi-source time-series datasets and joint feature vectors includes: Several historical joint feature vectors of normal transportation labels are obtained and used as training data for the normal behavior model; the normal transportation label refers to the transportation behavior during the transportation process being in a normal state. A normal behavior pattern model is obtained by training a Gaussian mixture model using normal behavior model training data; the normal behavior pattern model is used to characterize the data distribution of the joint feature vectors corresponding to normal transportation behavior. Obtain standardized multi-source time series datasets corresponding to several historical sliding time windows, and integrate them into a multi-source time series composite dataset according to the chronological order of the sliding time windows. The multi-source time series composite dataset is input into the parameter prediction model to obtain the predicted values ​​of several parameters in the standardized multi-source time series dataset; the parameter prediction model is constructed using an LSTM model to predict the values ​​of several parameters in the future time period.

4. The method for identifying abnormal transportation behavior of automotive parts according to claim 1, characterized in that, The process of identifying abnormal behavior and scoring risks based on joint feature vectors, normal behavior pattern models, and predicted values ​​to obtain abnormal event results includes: Extract the normal behavior pattern model and predicted values, as well as the joint feature vector of the previous sliding time window; The joint feature vector is input into the normal behavior pattern model to obtain the distance score between the joint feature vector and the data distribution; Real-time acquisition of multi-source heterogeneous data streams in transportation; The root mean square error method is used to calculate the basic residual values ​​between several real-time parameter values ​​and the corresponding predicted values ​​of several parameters in a multi-source heterogeneous data stream. The basic residual values ​​are obtained by performing residual normalization. The prediction bias score is obtained by weighted fusion of the standardized basic residual values ​​corresponding to several parameters. The comprehensive anomaly score corresponding to the current sliding time window is obtained by weighted summation of the prediction deviation score and the distance score. When the overall anomaly score is greater than the anomaly threshold, an anomaly event label is generated, and the occurrence time and geographic coordinates of the anomaly event label are obtained. Preliminary anomaly types are obtained through rule matching based on joint feature vectors; The results of anomalies are determined based on the event label, preliminary anomaly type, occurrence time, geographic coordinates, and comprehensive anomaly score.

5. The method for identifying abnormal transportation behavior of automotive parts according to claim 4, characterized in that, The method for obtaining the abnormal threshold includes: Obtain the global statistical measure of the comprehensive anomaly score corresponding to several historical sliding time windows under normal transport labels; the global statistical measure includes the mean of the comprehensive anomaly score. and the standard deviation of the composite abnormal score ; Calculate the basic anomaly threshold The formula for calculating the basic anomaly threshold satisfies: Where n represents the global adjustment coefficient, n>0; Obtain the road type corresponding to the current sliding time window and extract its corresponding road segment coefficient. The road types mentioned include expressways, national highways, urban roads, and mountain roads. Get the average score of the comprehensive anomaly scores within the most recent M sliding time windows. and standard deviation of scores Where M is an integer, M>1; Calculate the anomaly threshold YY; the formula for calculating the anomaly threshold satisfies: ; and Represented as weighting coefficients, and ∈(0,1); m represents the local adjustment coefficient, m>

0.

6. The method for identifying abnormal transportation behavior of automotive parts according to claim 4, characterized in that, The preliminary anomaly type obtained by rule matching based on joint feature vectors includes: Extract several basic features and several interaction features from the joint feature vector; Based on several basic features and several interactive features, a preliminary anomaly type is obtained by performing rule matching from the feature type mapping rule base; The feature type mapping rule base consists of several conditions and their corresponding preliminary anomaly types; the conditions are constructed from several basic features and several interactive features.

7. The method for identifying abnormal transportation behavior of automotive parts according to claim 1, characterized in that, The comprehensive result obtained by differential evaluation and feedback optimization based on the results of abnormal events includes: Obtain the results of abnormal events and waybill information; Determine the abnormal outcome based on the abnormal event results and waybill information; Obtain feedback data for abnormal results; the feedback data refers to the data obtained from the final verification of the abnormal results, including feedback tags and abnormal results; the feedback tags include confirmed abnormality and confirmed false alarm. Several abnormal results labeled as confirmed false alarms are stored in the optimization data pool to obtain the optimization dataset; The optimization results were obtained by optimizing the normal behavior pattern model and the parameter prediction model based on the optimized dataset; The overall result is determined based on the abnormal results and the optimization results.

8. The method for identifying abnormal transportation behavior of automotive parts according to claim 7, characterized in that, The determination of abnormal results based on abnormal event results and waybill information includes: Extract the preliminary anomaly type, occurrence time, geographic coordinates, and overall anomaly score from the results of abnormal events; An anomaly risk level mapping operation is performed on the preliminary anomaly type, occurrence time, and geographic coordinates to obtain the anomaly type level, time risk level, and location risk level; Extract the value tags of the transported goods from the waybill information; the value tags of the transported goods include high value, medium value, and low value. The value rating of transported goods is obtained by mapping the value tags of the transported goods to different levels. The comprehensive risk score is obtained by weighted summation of the anomaly type level, time risk level, location risk level, cargo value level, and normalized comprehensive anomaly score. When the comprehensive risk score is greater than or equal to the high-risk threshold, the final anomaly label will be set to high-risk anomaly. When the overall risk score is less than the high-risk threshold and the overall risk score is greater than or equal to the medium-risk threshold, the final anomaly label will be set to medium-risk anomaly. Otherwise, set the final anomaly label to low-risk anomaly; The anomaly results are determined based on the final anomaly label, the anomaly event results, the corresponding joint feature vector, distance score, several predicted values, and the standardized multi-source time series dataset.

9. The method for identifying abnormal transportation behavior of automotive parts according to claim 7, characterized in that, The optimization results obtained by optimizing the normal behavior pattern model and parameter prediction model based on the optimized dataset include: Obtain the optimized dataset and count the number of samples of outlier results in the optimized dataset; When the number of samples reaches the optimized number threshold Extract the joint feature vectors and their corresponding distance scores from several outlier results in the optimization dataset, and integrate them into a pattern optimization dataset. Extract the standardized multi-source time series datasets corresponding to the sliding time windows of several outlier results in the optimized dataset and their corresponding predicted values; Using each sliding time window in the optimized dataset as the center of the time window, the standardized multi-source time series datasets of several activity time windows before and after the sliding time window and their corresponding predicted values ​​are obtained, and combined with the standardized multi-source time series datasets and their corresponding predicted values ​​of each sliding time window in the optimized dataset to form multi-source time series composite data and time series predicted values. A prediction optimization dataset is composed of several multi-source time-series integrated data and time-series predicted values; The normal behavior pattern model and the parameter prediction model were updated using incremental learning methods on the pattern optimization dataset and the prediction optimization dataset, respectively, to obtain the optimized normal behavior pattern model and parameter prediction model. The optimization results are determined based on the optimized normal behavior pattern model and parameter prediction model.

10. A system for identifying abnormal transportation behavior of automotive parts, characterized in that, include: A data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected to each other; The data acquisition module: transports multi-source heterogeneous data streams through data acquisition equipment; The multi-source heterogeneous data stream for transportation includes location data, motion data, environmental data, and status data; The data analysis module includes a feature construction unit, an anomaly initial judgment unit, and a result generation unit; The feature construction unit performs data processing and feature extraction on multi-source heterogeneous transportation data streams to obtain a standardized multi-source time-series dataset and a joint feature vector. The anomaly initial judgment unit: constructs and predicts normal behavior patterns in real time based on standardized multi-source time-series datasets and joint feature vectors to obtain normal behavior pattern models and predicted values; and identifies and scores abnormal behavior based on joint feature vectors, normal behavior pattern models, and predicted values ​​to obtain abnormal event results. The result generation unit performs differentiated evaluation and feedback optimization based on the results of abnormal events to obtain a comprehensive result.

Images