Edge computing gateway data processing method and gateway

Through edge computing gateway processing of road multi-dimensional monitoring data in intelligent traffic systems, the real-time and security problems in traditional centralized mode are solved, efficient and accurate data processing and analysis are achieved, and monitoring accuracy and privacy are enhanced.

CN118736847BActive Publication Date: 2024-11-05SICHUAN HUATI LIGHTING TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411230889.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-04
Publication Date
2024-11-05
Estimated Expiration
2044-09-04

AI Technical Summary

Technical Problem

The traditional centralized data processing model faces real-time, accuracy and safety challenges when processing massive and heterogeneous road monitoring data, especially in intelligent transportation systems, and existing edge computing technologies are used to process complex multidimensional data. There are shortcomings.

Method used

By using edge computing gateway, by obtaining the multi-dimensional monitoring information set of roads, extracting the information flow of the monitoring mode, mapping it into a perceived implicit representation data flow, performing pan-scale extraction and extraction, obtaining perceived features and combining data flow, and performing implicit representation integration. Generate a comprehensive implicit representation of the road to determine the analysis results.

Benefits of technology

It improves the efficiency and privacy of data processing, enhances monitoring accuracy, reduces data transmission bandwidth and delay, prevents data tampering and leaking during transmission, and provides a more comprehensive road state analysis result.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118736847B_ABST
    Figure CN118736847B_ABST
Patent Text Reader

Abstract

The present application relates to the field of data processing, and specifically provides an edge computing gateway data processing method and gateway, which extracts road monitoring information streams of multiple monitoring modes from a road multi-dimensional monitoring information set, maps the extracted multiple road monitoring information streams into multiple road perception implicit representation data streams and performs translation and scaling extraction respectively to obtain a road perception feature merged data stream; then performs implicit representation integration, combines the integration results to obtain a road implicit representation of the road to be analyzed to determine the analysis result of the road to be analyzed. The present application extracts multiple scaled implicit representations from the road perception implicit representation data stream of each monitoring mode based on translation and scaling extraction, and can obtain a more complete combination of each monitoring mode. This can prevent the monitoring data that focuses on one mode in a short period of time from being hidden in the entire road perception implicit representation data stream and not easily detected, thereby improving the monitoring accuracy of single-modality abnormal monitoring data in a short period of time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of data processing, and specifically, to an edge computing gateway data processing method and gateway. Background Art

[0002] With the rapid development of intelligent transportation systems, higher requirements are placed on the real-time processing and analysis capabilities of road monitoring data. The traditional centralized data processing model faces many challenges in processing massive and heterogeneous road monitoring data, especially in complex scenarios such as urban intelligent transportation, highway monitoring, and industrial Internet of Things, which have extremely high requirements for the real-time, accuracy, and security of data. In the prior art, road monitoring data usually comes from a variety of sensors and devices, including high-definition cameras, radar speedometers, environmental sensors, etc. These data cover multiple dimensions such as vehicle speed, density, and environmental parameters, forming a multi-dimensional road monitoring information set. However, how to efficiently collect, integrate, and analyze these multi-dimensional data to accurately reflect the real-time status of the road is a major problem facing the current intelligent transportation system. Traditional data processing methods often rely on centralized servers or cloud platforms, and all monitoring data is uploaded to the central node for centralized processing. This method not only increases the bandwidth and delay of data transmission, but may also cause data to be tampered with or leaked during transmission, thereby causing privacy and security risks. In recent years, edge computing technology has gradually attracted attention. Edge computing reduces the bandwidth and latency of data transmission and improves data processing efficiency and data privacy by pushing computing resources and data storage to the edge nodes of the network, that is, processing and analyzing them close to the data source. However, existing edge computing technologies still have some shortcomings when processing complex and multi-dimensional road monitoring data. Summary of the invention

[0003] The object of the present invention is to provide an edge computing gateway data processing method and a gateway. The embodiment of the present application is implemented as follows:

[0004] In a first aspect, the present application provides an edge computing gateway data processing method, the method comprising: obtaining a road multidimensional monitoring information set of a road to be analyzed; extracting road monitoring information streams of multiple monitoring modes from the road multidimensional monitoring information set, and mapping the extracted multiple road monitoring information streams into multiple road perception implicit representation data streams; using a first perception representation statistic as a scaling quantity and a second perception representation statistic as a scaling step, translating and scaling each of the road perception implicit representation data streams to obtain a first road perception feature merged data stream corresponding to each of the road perception implicit representation data streams; implicitly integrating each of the first road perception feature merged data streams, and combining the implicit representations of the multiple integrated data streams obtained by integration to obtain an implicit road representation of the road to be analyzed; and determining the analysis result of the road to be analyzed based on the road implicit representation.

[0005] In a second aspect, the present application provides an edge computing gateway, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the one or more processors, the method as described above is implemented.

[0006] The beneficial effects of the present application include at least: the edge computing gateway data processing method provided by the present application is based on obtaining a road multi-dimensional monitoring information set of the road to be analyzed; extracting road monitoring information streams of multiple monitoring modes from the road multi-dimensional monitoring information set, and mapping the extracted multiple road monitoring information streams into multiple road perception implicit representation data streams; using the first perception representation statistic as the scaling quantity, and the second perception representation statistic as the scaling step, performing translation scaling extraction on each road perception implicit representation data stream to obtain the first road perception feature merged data stream corresponding to each road perception implicit representation data stream; performing implicit representation integration on each first road perception feature merged data stream, and combining the implicit representations of the multiple integrated data streams obtained by integration to obtain the road implicit representation of the road to be analyzed; determining the analysis result of the road to be analyzed based on the road implicit representation. The present application extracts multiple scaled implicit representations from the road perception implicit representation data stream of each monitoring modality based on translation scaling extraction, and can obtain a more complete combination of each monitoring modality. This can prevent the monitoring data that focuses on one modality in a short period of time from being hidden in the entire road perception implicit representation data stream and not easily detected, thereby improving the monitoring accuracy of single-modal abnormal monitoring data in a short period of time. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a flow chart of an edge computing gateway data processing method provided in an embodiment of the present application.

[0008] Figure 2 It is a schematic diagram of the composition of an edge computing gateway provided in an embodiment of the present application. DETAILED DESCRIPTION

[0009] The execution subject of the edge computing gateway data processing method in the embodiment of the present application is the edge computing gateway, and the network where the edge computing gateway is located includes but is not limited to the Internet, wide area network, metropolitan area network, local area network, VPN network, etc. Figure 1 As shown, the method includes:

[0010] Step S100: Obtain a multi-dimensional road monitoring information set of the road to be analyzed.

[0011] In the intelligent transportation system, the edge computing gateway is a key node for data processing and analysis. Its primary task is to collect and analyze multi-dimensional monitoring information of roads. The edge computing gateway is deployed near major traffic arteries and is responsible for real-time monitoring and collecting data from various sensors, which together constitute a multi-dimensional monitoring information set of roads. This information not only covers traditional basic parameters such as traffic flow and vehicle speed, but also includes richer environmental perception data, such as air quality, noise level, meteorological conditions (such as temperature, humidity, wind direction and speed, etc.) and road conditions (such as slippery, potholes, etc.).

[0012] For example, assume that the edge computing gateway is deployed on the main road of a city to monitor and analyze the real-time traffic and environmental conditions of the road. When executing step S100, the gateway can collect data from at least the following data sources: Traffic monitoring camera: Provides real-time video stream, and traffic parameters such as vehicle quantity, type, and speed can be extracted through video analysis algorithms (such as target detection and tracking models based on deep learning). These parameters are quantified into numerical features, such as vehicle density (the number of vehicles passing a certain section of road per unit time), average speed, etc., and are used as part of the traffic flow feature vector. Radar speed meter: Installed on one side of the road, it measures the speed of passing vehicles by emitting and receiving electromagnetic waves. Radar data directly provides vehicle speed information, and these speed values ​​are arranged in chronological order to form speed time series data. Environmental sensor network: includes air quality monitoring stations, noise monitors, etc., which measure the concentration of pollutants in the air (such as PM2.5, CO, etc.) and noise decibel levels respectively. These data are collected in the form of environmental parameter vectors, such as e=[PM2.5, CO, Noise_dB]. Weather station: Provides real-time weather data, such as temperature, humidity, wind direction and speed, which is crucial for evaluating road conditions. Meteorological data is also organized into feature vectors, such as w=[Temperature, Humidity, Wind_Speed, Wind_Direction]. Road surface sensors: Buried inside the road, used to detect the slipperiness and potholes of the road. These data are usually expressed in binary or numerical form, such as "slippery road: yes / no" or "pothole depth: X cm". The edge computing gateway establishes connections with the above-mentioned sensors and monitoring devices through its built-in communication module (such as Ethernet interface, Wi-Fi, LoRa, etc.), and collects the data generated by these devices in a scheduled or real-time manner. After preliminary verification and cleaning, the collected data is integrated into a multi-dimensional monitoring information set. This information set not only contains rich feature vectors and arrays, but also retains the timestamp information of the data for subsequent time series analysis and pattern recognition.

[0013] Step S200: extracting road monitoring information streams of multiple monitoring modes from a road multi-dimensional monitoring information set, and mapping the extracted multiple road monitoring information streams into multiple road perception implicit representation data streams.

[0014] For example, suppose the edge computing gateway is processing multi-dimensional monitoring data of a busy urban trunk road. A variety of sensors and monitoring equipment are deployed on this road, including high-definition cameras, radar speed guns, environmental sensors (such as air quality monitoring stations, noise monitors), weather stations, and road surface sensors. These devices continuously generate a large amount of raw data, which constitutes a multi-dimensional monitoring information set of the road. First, the edge computing gateway identifies and extracts the data streams of each monitoring mode from the information set according to the preset rules or algorithms. For example: Video stream processing: The video stream captured by the high-definition camera is processed by the built-in video analysis algorithm (such as the target detection model based on the convolutional neural network CNN) to extract the trajectory information, quantity, type, etc. of targets such as vehicles and pedestrians to form a video monitoring information stream. Radar data extraction: The data of the radar speed gun is directly read to extract the real-time speed information of each vehicle to form a speed monitoring information stream. Environmental parameter collection: The data reported regularly by environmental sensors (such as air quality monitoring stations and noise monitors) is integrated to form an environmental monitoring information stream containing parameters such as PM2.5 concentration and noise decibels. Meteorological data extraction: Real-time meteorological data such as temperature, humidity, wind direction and speed provided by the weather station are extracted to form a meteorological monitoring information flow. Road condition perception: Information such as the slipperiness of the road surface and the depth of potholes detected by the road surface sensor is quantified and processed to form a road surface monitoring information flow.

[0015] The extracted road monitoring information streams of each monitoring modality need to be further mapped into higher-level road perception implicit representation data streams. This process usually involves feature engineering, data encoding, and possible machine learning transformations. For example:

[0016] Mapping of video monitoring information stream: Vehicle trajectories, quantity and other information in the video monitoring information stream can be converted into fixed-length feature vectors through encoding techniques (such as One-Hot Encoding or Embedding). At the same time, using models such as recurrent neural networks (RNN) or long short-term memory networks (LSTM), hidden states in time series can be further extracted as implicit representations to form video perception implicit representation data streams.

[0017] Radar data mapping: The speed value in the speed monitoring information stream can be directly used as a feature, but in order to better capture the speed change pattern, the data can be preprocessed using methods such as sliding window averaging and differentiation, and then a simple linear transformation or nonlinear activation function (such as ReLU) can be used to obtain the speed-aware implicit representation data stream.

[0018] Mapping of environmental and meteorological monitoring information flows: Environmental parameters (such as PM2.5, noise decibels) and meteorological parameters (such as temperature, humidity) can be directly used as part of the feature vector, but considering the differences in the dimensions and ranges of different parameters, they can be normalized first. Afterwards, these feature vectors can be further compressed and abstracted through a fully connected layer (Dense Layer) or an autoencoder (Autoencoder) to form an implicit representation data flow of environmental and meteorological perception.

[0019] Mapping of road monitoring information flow: Discrete or quantitative information such as road slipperiness and pothole depth can be converted into continuous and comparable implicit representations through One-Hot Encoding or Embedding Vector to form a road perception implicit representation data flow.

[0020] Based on this, the edge computing gateway transforms the complex multi-dimensional road monitoring information set into a series of structured road perception implicit representation data streams, providing strong support for subsequent data analysis, pattern recognition and decision support. These implicit representations not only reduce the dimension and complexity of the data, but also retain key information.

[0021] Step S300: Using the first perception representation statistic as the scaling quantity and the second perception representation statistic as the scaling step, performing translation scaling extraction on each road perception implicit representation data stream to obtain a first road perception feature merged data stream corresponding to each road perception implicit representation data stream.

[0022] In step S300, the edge computing gateway first determines two key parameters: the first perception representation statistic (i.e., the number of scales) and the second perception representation statistic (i.e., the scale step). Taking the vehicle density feature as an example, assuming that the implicit representation of each time step contains a scalar value of vehicle density, the edge computing gateway can determine a suitable scale number (e.g., 10) based on prior knowledge or experiments, which means that the vehicle density values ​​of the last 10 time steps are considered in each extraction; at the same time, a scale step (e.g., 2) is determined, which means that the window moves forward 2 time steps after each extraction. Next, the edge computing gateway applies the translation scale extraction technology to each road perception implicit representation data stream. Taking the vehicle density implicit representation data stream as an example, each time point in the data stream corresponds to a vehicle density value, forming a continuous time series. The edge computing gateway first selects the initial window (the vehicle density values ​​of the first 10 time steps). These 10 values ​​constitute a feature vector, which represents the vehicle density change pattern in the current window. Then, according to the stride (2), the window moves forward 2 time steps, and the vehicle density values ​​of the new 10 time steps are extracted again to form a new feature vector. This process is repeated until the entire data stream is covered.

[0023] The feature vectors extracted from each window are concatenated to form the first road perception feature merged data stream of the corresponding road perception implicit representation data stream. This merged data stream not only retains the time order information in the original data stream, but also emphasizes the feature change pattern within a specific length through window translation and extraction.

[0024] For example, suppose that part of the data stream implicitly representing vehicle density is as follows (simplified to scalar value representation): [0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.1, 1.0, 0.9, ...].

[0025] Taking the number of scales = 10 and the scale step = 2 as an example, the feature vector v1 extracted for the first time is: v1 = [0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2]; after the window moves forward 2 steps, the feature vector v2 extracted for the second time is: v2 = [0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.1, 1.0].

[0026] This process is repeated until the end of the data stream, and finally the first road perception feature merged data stream corresponding to the vehicle density implicit representation data stream is obtained.

[0027] Step S400: performing implicit representation integration on each first road perception feature merged data stream, and combining multiple integrated data stream implicit representations obtained by integration to obtain a road implicit representation of the road to be analyzed.

[0028] Step S400 aims to integrate the feature sequences extracted under each monitoring mode (i.e., the first road perception feature merged data stream) through implicit representation through a specific method, and combine these integrated implicit representations to form a comprehensive and integrated road implicit representation of the road to be analyzed.

[0029] After completing step S300, the edge computing gateway has obtained multiple first road perception feature merged data streams, each of which represents the feature changes of a certain monitoring mode (such as vehicle density, speed, environmental parameters, etc.) within a specific time window. In step S400, the gateway further processes these feature merged data streams. For each first road perception feature merged data stream, the edge computing gateway applies a pre-trained machine learning model (such as a convolutional neural network CNN, a recurrent neural network RNN ​​or its variant LSTM / GRU) to perform implicit representation integration. These models are trained to extract higher-level abstract representations from the original features, which often better reflect the essential characteristics of the data.

[0030] Taking the vehicle density feature merged data stream as an example, assume that the data stream is a sequence containing vehicle density values ​​at multiple time steps. The edge computing gateway will input this sequence into a trained RNN model, which processes the input of each time step by loop and uses its internal state to capture the time dependency in the sequence. Finally, the model outputs a fixed-length vector as an implicit representation of the vehicle density feature merged data stream.

[0031] After obtaining the implicit representation of all the first road perception feature merged data streams, the edge computing gateway combines them to form an overall implicit representation of the road to be analyzed. The combination method is usually simple concatenation, but other more complex fusion strategies can also be adopted according to the specific application scenario. Assume that in addition to vehicle density, there are implicit representations of other monitoring modalities such as speed and environmental parameters (such as PM2.5 concentration, noise decibels). The edge computing gateway will concatenate these implicit representation vectors in a certain order to form a longer vector. For example, if the implicit representation of vehicle density is v=[v1,v2,...,v n ], the implicit representation of velocity is s=[s1,s2,...,s m ], the implicit representation of the environment parameters is e=[e1,e2,...,e p ], then the combined road implicit representation is: Road implicit representation = [v, s, e] = [v1, v2, ..., v n ,s1,s2,...,s m ,e1,e2,...,ep]. This combined implicit road representation not only contains the key information of each monitoring modality, but also retains the intrinsic connection and mutual influence between them through implicit representation integration, providing strong data support for subsequent traffic status analysis, accident prediction, environmental assessment and other tasks.

[0032] Step S500: determining the analysis result of the road to be analyzed according to the implicit representation of the road.

[0033] Based on the above steps, the edge computing gateway has completed the collection of multi-dimensional road monitoring information sets, the extraction and mapping of monitoring information flows, the extraction of feature merging data flows, and the integration and combination of implicit representations, and finally obtained a comprehensive road implicit representation. This implicit representation may be a high-dimensional vector that contains information on vehicle density, speed distribution, environmental parameters (such as air quality, noise level), meteorological conditions, and possible traffic events.

[0034] In step S500, the edge computing gateway passes this implicit road representation as input to one or more specially designed machine learning models. These models can be classifiers, regressors, or any other type of prediction models depending on the specific application requirements. The following are some possible application scenarios and corresponding model examples:

[0035] Scenario 1: Traffic congestion prediction.

[0036] In this scenario, the edge computing gateway can use a classification model based on support vector machine (SVM) or random forest (RandomForest) to predict whether the road will be congested. The model receives the implicit representation of the road as input, analyzes key features such as traffic flow and vehicle speed through internal algorithms, and outputs a classification result, such as "congested" or "non-congested". Based on the output of the model, the edge computing gateway can determine whether the current road is congested and feed this information back to the traffic management center or the driver in real time.

[0037] Scenario 2: Accident risk assessment.

[0038] In this scenario, for accident risk assessment, the edge computing gateway may use a deep neural network (such as a convolutional neural network (CNN) combined with a recurrent neural network (RNN)) to better capture complex patterns in space and time. The model first processes the road image data through the CNN layer (if the implicit representation contains image features), then processes the time series data (such as vehicle speed, acceleration, etc.) through the RNN layer, and finally outputs a score representing the accident risk level. Based on the output score, the edge computing gateway can assess the accident risk of the current road and take corresponding preventive measures or issue warnings to relevant parties.

[0039] Scenario 3: Environmental quality monitoring.

[0040] In this scenario, for environmental quality monitoring, the edge computing gateway may use a regression model based on gradient boosting trees to predict the air quality index (AQI) or other environmental parameters. The model predicts the environmental parameter characteristics in the implicit representation of the road (such as PM2.5 concentration, NO2 concentration, etc.) through the set of decision trees obtained through iterative training, and outputs a continuous environmental quality score. Based on the output score, the edge computing gateway can monitor the environmental quality around the road in real time, provide data support for the environmental protection department, or issue environmental warning information to the public.

[0041] Through step S500, the edge computing gateway uses advanced machine learning models to deeply analyze the implicit representation of the road, thereby obtaining accurate and comprehensive analysis results. These results not only help to improve the level of intelligent traffic management, but also provide strong support for environmental protection, public safety and other fields. In practical applications, the edge computing gateway can select appropriate models and algorithms according to specific needs, and improve the accuracy and reliability of analysis results by continuously optimizing model parameters.

[0042] As an implementation manner, step S400, performing implicit representation integration on each first road perception feature merged data stream, and combining multiple integrated data stream implicit representations obtained by integration to obtain a road implicit representation of the road to be analyzed, may specifically include:

[0043] Step S410: determining a first machine learning model corresponding to each first road perception feature combined data stream;

[0044] Step S420: loading the first road perception feature merged data stream corresponding to each road perception implicit representation data stream into the corresponding first machine learning model for implicit representation integration, to obtain an integrated data stream implicit representation corresponding to each road perception implicit representation data stream;

[0045] Step S430: combining multiple implicit representations of integrated data streams to obtain an implicit representation of the road to be analyzed.

[0046] In step S410, the edge computing gateway matches the most suitable machine learning model for each first road perception feature merged data stream according to its characteristics. These feature merged data streams may come from different monitoring modalities, such as video surveillance, radar speed measurement, environmental sensors, etc. The data of each modality has different structures and meanings, and therefore requires different processing strategies. Assume that the edge computing gateway is processing the traffic and environmental monitoring data of a main road in a certain city. These data include: Video surveillance data: The flow of vehicles and pedestrians captured by high-definition cameras is processed by target detection and tracking algorithms to form a feature merged data stream containing information such as vehicle type, quantity, and speed. Radar speed measurement data: The radar device measures the speed of passing vehicles in real time to form a continuous speed feature merged data stream. Environmental sensor data: Including data collected by equipment such as air quality monitoring stations and noise monitors, such as PM2.5 concentration, noise decibel value, etc., which constitute an environmental parameter feature merged data stream. For these different feature merged data streams, the edge computing gateway will adopt the following strategies to determine the corresponding machine learning model:

[0047] Video surveillance data stream: Since video surveillance data usually contains rich image and video information, the edge computing gateway may choose a convolutional neural network (CNN)-based model to process it. CNN is good at extracting high-level feature representations from raw images and is suitable for feature integration of video surveillance data. Radar speed measurement data stream: Radar speed measurement data is continuous time series data, and the edge computing gateway may use recurrent neural network (RNN) or its variants (such as LSTM, GRU) to process it. These models can effectively capture the time dependency in time series data and are suitable for analyzing speed change patterns. Environmental sensor data stream: Environmental sensor data is usually multi-dimensional numerical data, and the edge computing gateway may choose models such as fully connected neural network (DNN) or support vector machine (SVM) to process it. These models can process multi-dimensional inputs and output corresponding predictions or classification results.

[0048] In step S420, the edge computing gateway loads each feature merged data stream into the corresponding machine learning model, and obtains the integrated implicit representation through the calculation of the model. These implicit representations are high-level abstractions of the original data and can better reflect the essential characteristics of the data.

[0049] For video surveillance data streams, the edge computing gateway inputs the processed video frames or feature maps into a CNN-based model. CNN extracts key features such as vehicle type, quantity, and speed from the original image through a combination of convolutional layers, pooling layers, and fully connected layers, and integrates these features into a fixed-length vector as an implicit representation. For example, assume that the model outputs a 128-dimensional feature vector as an implicit representation of the video surveillance data in each time window. For radar speed measurement data streams, the edge computing gateway inputs it into the RNN model. RNN processes the input of each time step (i.e., the speed value at each time point) by looping, and uses its internal state to capture the time dependency of speed changes. Finally, RNN outputs a fixed-length vector as an implicit representation of radar speed measurement data. Assume that the vector length is 64 dimensions and contains key information about speed changes. For environmental sensor data streams, the edge computing gateway may choose a DNN model for processing. DNN uses a multi-layer perceptron (MLP) structure to perform nonlinear transformations on multi-dimensional inputs (such as PM2.5 concentration, noise decibel value, etc.), and finally outputs an implicit representation vector that integrates all environmental parameters. Assume that the vector length is 32 dimensions, covering key environmental characteristics such as air quality and noise level.

[0050] In step S430, the edge computing gateway combines multiple integrated implicit representation vectors to form a comprehensive road implicit representation. This representation vector contains key information from different monitoring modalities and can fully reflect the status and changes of the road.

[0051] After completing the implicit representation integration of video surveillance, radar speed measurement, and environmental sensor data streams, the edge computing gateway concatenates these three implicit representation vectors to form a longer vector. Assume that the implicit representation vector of the video surveillance data is v=[v1,v2,…,v 128 ], the implicit representation vector of radar speed measurement data is s=[s1,s2,…,s 64 ], the implicit representation vector of environmental sensor data is e=[e1,e2,…,e 32 ]. Then the concatenated road implicit representation vector is: Road implicit representation = [v, s, e] = [v1, v2, …, v 128 ,s1,s2,…,s 64 ,e1,e2,…,e 32 ]. The length of this concatenated vector is 128+64+32=224 dimensions. It integrates information from multiple monitoring modes such as video surveillance, radar speed measurement, and environmental sensors, providing comprehensive data support for subsequent traffic status analysis, accident prediction, environmental assessment and other tasks.

[0052] Through the above steps, the edge computing gateway integrates multiple first road perception feature merge data streams into implicit representations, and combines these implicit representations into the road implicit representation of the road to be analyzed. This process fully utilizes the powerful capabilities of machine learning to extract valuable information from complex raw data, providing a solid foundation for decision support for intelligent transportation systems.

[0053] For example, for the processing of video surveillance data streams, in addition to CNN, you can also consider using target detection and tracking algorithms (such as YOLO, SSD, etc.) combined with optical flow methods to improve the accuracy and efficiency of feature extraction.

[0054] In addition, when integrating implicit representations of multiple integrated data streams, in addition to simple splicing, more complex fusion strategies can be considered to improve the accuracy and robustness of implicit road representations. For example, the attention mechanism can be used to assign different weights to data from different monitoring modalities to highlight key information and suppress noise interference. Alternatively, multimodal fusion techniques (such as multimodal deep learning, graph neural networks, etc.) can be used to model the intrinsic connections and interactions between different monitoring modalities, thereby generating a more comprehensive and accurate implicit road representation.

[0055] As an implementation manner, step S430, combining multiple implicit representations of integrated data streams to obtain an implicit representation of a road to be analyzed, may include:

[0056] Step S431: determining the first importance coefficient of the implicit representation of each integrated data stream;

[0057] Step S432: adjusting each integrated data stream implicit representation according to the first importance coefficient to obtain a plurality of adjusted data stream implicit representations;

[0058] Step S433: combining multiple implicit representations of regulated data streams to obtain an implicit representation of the road to be analyzed.

[0059] In step S431, when determining the importance coefficient (i.e., weight) of each integrated data stream implicitly represented, the edge computing gateway comprehensively considers multiple factors to ensure that the weight distribution is both reasonable and meets the needs of the actual application scenario. For example, the importance coefficient can be determined from the following perspectives, and an adaptive selection can be made according to the actual situation:

[0060] Data quality: The data quality of different monitoring modalities may vary. For example, video surveillance data is greatly affected by weather, lighting and other conditions, while radar speed measurement data is relatively stable. Therefore, the edge computing gateway can assign a higher weight to high-quality data based on the quality evaluation results of historical data.

[0061] Data timeliness: Data of certain monitoring modes may have higher timeliness requirements, such as traffic flow data, which is particularly important in real-time traffic management. Therefore, for such data with strong timeliness, the edge computing gateway can give higher weight.

[0062] Task relevance: Different analysis tasks may be more sensitive to data from a specific monitoring modality. For example, when predicting an accident, vehicle speed and location information may be more important than air quality data; while when assessing environmental quality, the opposite is true. Therefore, the edge computing gateway adjusts the weight of each monitoring modality data according to the needs of the specific task.

[0063] Data complementarity: There may be a complementary relationship between the data of certain monitoring modalities, that is, they together provide a comprehensive description of the road state. In this case, the edge computing gateway recognizes and exploits this complementarity to enhance the accuracy and robustness of the implicit representation of the road by reasonably assigning weights.

[0064] Expert knowledge: In some cases, expert experience or domain knowledge can also serve as an important basis for determining weights. By combining experts’ judgment on the importance of the monitoring modality, the edge computing gateway can assign weights more accurately.

[0065] For example, suppose the edge computing gateway is processing traffic and environmental monitoring data on a city's main roads, including data from three monitoring modalities: video surveillance, radar speed measurement, and environmental sensors. When determining the weights, the edge computing gateway can consider the following factors:

[0066] Video surveillance data: Although video surveillance data provides rich visual information, it is greatly affected by factors such as weather and lighting. Therefore, a higher weight (such as 0.4) can be given in clear weather, and a lower weight (such as 0.2) can be appropriately given in bad weather.

[0067] Radar speed data: Radar speed data is stable and reliable, and is crucial for real-time traffic management. Therefore, a high fixed weight (such as 0.3) can be given to ensure its importance in the implicit representation of the road.

[0068] Environmental sensor data: Environmental sensor data is important for evaluating road environment quality, but may not be the most critical in some traffic management tasks. Therefore, a moderate weight (such as 0.3) can be given and adjusted according to specific task requirements.

[0069] After determining the importance coefficients of the implicit representations of each integrated data stream, the edge computing gateway adjusts each implicit representation according to these coefficients. The purpose of the adjustment is to make the data of different monitoring modalities reasonably reflected and weighted in the combination process, so as to generate a more accurate and comprehensive implicit representation of the road.

[0070] One adjustment method is to use a weighted summation method. For each implicit representation hi of the integrated data stream (where i represents a different monitoring mode), the edge computing gateway first obtains its corresponding importance coefficient wi, and then calculates the adjusted implicit representation using the following formula:

[0071] ;

[0072] in, Represents the adjusted implicit representation. In this way, the monitoring modality data with a larger importance coefficient will occupy a larger proportion in the combination process. Let's continue with the previously mentioned urban trunk road traffic and environmental monitoring data as an example. Assume that the importance coefficients of video surveillance data, radar speed measurement data, and environmental sensor data have been determined to be 0.4, 0.3, and 0.3, respectively. Now, the edge computing gateway will use these coefficients to adjust the implicit representation of each integrated data stream:

[0073] Implicit representation of video surveillance data: Assume that its implicit representation after integration is , then the adjusted implicit expression is .

[0074] Implicit representation of radar speed measurement data: Assume that its implicit representation after integration is , then the adjusted implicit expression is .

[0075] Implicit representation of environmental sensor data: Assume that its implicit representation after integration is , then the adjusted implicit expression is .

[0076] In step S433, after completing the adjustment of the implicit representations of each integrated data stream, the edge computing gateway combines these adjusted implicit representations to form a comprehensive road implicit representation of the road to be analyzed. The combination method is usually concatenation, but other more complex fusion strategies can also be considered, such as weighted fusion, attention mechanism, etc. In the concatenation combination, the edge computing gateway simply concatenates the adjusted implicit representations into a longer vector in a certain order. Continuing with the previous example: Assume that the adjusted video surveillance data is implicitly represented as v', the radar speed measurement data is implicitly represented as s', and the environmental sensor data is implicitly represented as e'. Then the concatenated road implicit representation is: The length of this concatenated vector is 80 (rounded off here, and the vector length may need to be adjusted to an integer in actual implementation). It integrates information from multiple monitoring modes such as video surveillance, radar speed measurement, and environmental sensors, and performs reasonable weighting processing through importance coefficients.

[0077] In addition to simple concatenation, edge computing gateways can also consider using more complex fusion strategies to combine multiple regulated data streams implicitly. For example:

[0078] Weighted fusion: Based on the splicing, a global weight vector can be further introduced to perform weighted processing on the entire spliced ​​vector. This global weight vector can be determined based on the performance evaluation results of the entire road monitoring system.

[0079] Attention mechanism: The attention mechanism is a mechanism that can dynamically adjust the importance of different parts. When combining multiple implicit representations, an attention network can be introduced to calculate the importance score of each implicit representation (i.e., attention weight), and perform weighted fusion of the implicit representations based on these weights. This approach can more flexibly cope with data changes in different scenarios.

[0080] Through the detailed implementation of step S430 and the expanded explanation of its application scenarios, it can be seen that it is very necessary to optimize the combination process of implicit representation of multiple integrated data streams by introducing importance coefficients (weights) in intelligent transportation systems. This not only helps to improve the accuracy and comprehensiveness of implicit representation of roads, but also can better adapt to changes in requirements in different application scenarios.

[0081] As an implementation manner, step S200, after extracting a plurality of road monitoring information streams of monitoring modes from the road multi-dimensional monitoring information set and mapping the extracted plurality of road monitoring information streams into a plurality of road perception implicit representation data streams, may further include:

[0082] Step S201: using the third perception representation statistic as the scaling quantity and the fourth perception representation statistic as the scaling step, performing translation scaling extraction on each road perception implicit representation data stream to obtain a second road perception feature merged data stream corresponding to each road perception implicit representation data stream;

[0083] Based on this, step S400, performing implicit representation integration on each first road perception feature merged data stream, and combining multiple integrated data stream implicit representations obtained by integration to obtain the road implicit representation of the road to be analyzed, may include:

[0084] Step S400a: performing implicit representation integration on each first road perception feature merged data stream, and combining the implicit representations of multiple first integrated sub-data streams obtained by integration to obtain a first sub-road implicit representation, and performing implicit representation integration on each second road perception feature merged data stream, and combining the implicit representations of multiple second integrated sub-data streams obtained by integration to obtain a second sub-road implicit representation;

[0085] Step S400b: combining the first sub-road implicit representation with the second sub-road implicit representation to obtain the road implicit representation of the road to be analyzed.

[0086] In the embodiment of the present application, step S200 is a key step for the edge computing gateway to extract and map road monitoring information from a complex multi-dimensional road monitoring information set. This process not only involves data extraction from different monitoring modalities, but also involves converting raw data into a high-level implicit representation for subsequent data analysis and processing. In addition, in order to capture the dynamic change pattern in the data more carefully, step S200 can be further expanded to include step S201, that is, to translate and extract the road perception implicit representation data stream to generate a second road perception feature merged data stream.

[0087] The edge computing gateway establishes connections with various sensors and monitoring devices through preset protocols and interfaces to collect data in real time or regularly. For example, it extracts video stream data from high-definition cameras, extracts trajectory information, quantity, type and other features of targets such as vehicles and pedestrians through built-in video analysis algorithms (such as target detection and tracking models based on deep learning); directly reads vehicle speed data from radar speed guns; collects environmental parameters such as air quality (such as PM2.5 concentration) and noise decibel value from environmental sensors. The extracted raw data needs to be further mapped into road perception implicit representation data streams. This process usually involves feature engineering and data encoding. For example, vehicle trajectory information in video stream data can be encoded into a series of vectors containing features such as position, speed, and direction; radar speed measurement data can be directly used as one-dimensional time series data; and environmental sensor data may be normalized to form multi-dimensional feature vectors. These implicit representation data streams not only reduce the dimension and complexity of the data, but also retain key information, which facilitates subsequent data analysis.

[0088] In step S201, after the mapping is completed, the edge computing gateway can also perform translation and scaling extraction on each road perception implicit representation data stream to generate a second road perception feature merged data stream. This process is similar to the sliding window technology in signal processing, but here it focuses more on extracting feature sequences of a specific length from the implicit representation.

[0089] Taking the vehicle trajectory implicit representation data stream as an example, assume that each time point corresponds to a vector containing features such as vehicle position and speed. The edge computing gateway sets the third perception representation statistic (such as window size W) and the fourth perception representation statistic (such as step size S) as the scaling quantity and scaling step. Then, the gateway uses a fixed-size window (length W) to slide on the data stream, with a sliding step size of S each time. At each window position, the gateway extracts all feature vectors in the window and merges them into a new feature merged data stream (that is, the second road perception feature merged data stream). For example, assume that part of the data of the vehicle trajectory implicit representation data stream is as follows (simplified representation):

[0090] Original data stream: [f1,f2,f3,f4,f5,f6,…]

[0091] Among them, f iRepresents the feature vector at the i-th time point. Set the window size W=3 and the step size S=2, then the translation and scaling extraction process is as follows: the initial window contains [f1,f2,f3], and the first element of the extracted second road perception feature merged data stream is the merge of these three vectors (the specific merging method may involve splicing, averaging or other aggregation operations). The window slides forward two steps to [f3,f4,f5], and the second element after extraction is the merge of these three vectors. Repeat the above process until the end of the data stream. In this way, the edge computing gateway can extract multiple second road perception feature merged data streams from each road perception implicit representation data stream. These data streams further refine the feature sequence information in the original data and provide more options for subsequent implicit representation integration.

[0092] Step S400 is a key step for the edge computing gateway to implicitly represent and integrate multiple first road perception feature merged data streams and combine them into a road implicit representation. In the expanded implementation, step S400 is subdivided into two sub-steps, S400a and S400b, to process feature merged data streams from different scaling strategies (i.e., original scaling and extended scaling).

[0093] After completing step S200 and its extended step S201, the edge computing gateway obtains multiple first road perception feature merged data streams (from the original scale) and multiple second road perception feature merged data streams (from the extended scale). These data streams represent road state information at different time scales and feature combinations.

[0094] For each first road perception feature merged data stream and second road perception feature merged data stream, the edge computing gateway applies pre-trained machine learning models to perform implicit representation integration. These models may include deep neural networks (such as CNN, RNN, etc.), autoencoders, or other algorithms that can extract high-level feature representations. The purpose of the integration is to convert the low-level features in the original feature merged data stream into more abstract, higher-level implicit representations.

[0095] After completing the implicit representation integration, the edge computing gateway combines (e.g., splices) multiple implicit representations of the first integrated sub-data streams to obtain the first sub-road implicit representation; and simultaneously combines multiple implicit representations of the second integrated sub-data streams to obtain the second sub-road implicit representation. These two sub-representations reflect the road state information at the original scale and the extended scale, respectively.

[0096] After obtaining the first sub-road implicit representation and the second sub-road implicit representation, the edge computing gateway further combines them to generate a comprehensive road implicit representation. This process needs to comprehensively consider the importance and complementarity of the two sub-representations.

[0097] The edge computing gateway can combine the implicit representation of the first sub-road and the implicit representation of the second sub-road in a simple splicing manner. However, it is more common to introduce a weight mechanism to adjust the contribution of different sub-representations in the final combination. This can be achieved by setting different weight coefficients, which can be set according to the needs of the actual application scenario and the importance of the data.

[0098] For example, suppose the first sub-road is implicitly represented as r1, the second sub-road is implicitly represented as r2, and they are assigned weights w1 and w2 respectively (satisfying w1+w2=1). Then the final road implicit representation R can be calculated by the following formula:

[0099] ;

[0100] Or, use a weighted splicing approach:

[0101] R=[A,B];

[0102] in, ;

[0103] ;

[0104] Among them, r 1,i and r 2,j denote the elements (may be vectors or scalars) in the implicit representation of the first subpath and the implicit representation of the second subpath, respectively, and n1 and n2 denote their lengths, respectively. Specifically, for r 1,i , 1 indicates that it belongs to the first sub-path implicit representation, i indicates the order, for example, when i=1, it is r 1,1 , represents the first element in the implicit representation of the first subpath. Similarly, for r 2,j , 2 indicates that it belongs to the second sub-path implicit representation, j indicates the order, for example, when j=1, it is r 2,1 , represents the first element in the implicit representation of the second sub-path. w1 and w2 are the weights of the implicit representation of the first sub-path and the implicit representation of the second sub-path, respectively.

[0105] In this way, the edge computing gateway integrates and combines feature information from different monitoring modalities and different scaling strategies into a comprehensive implicit road representation. This representation not only contains rich road state information, but also improves the accuracy and reliability of the information through implicit representation integration and weighted combination.

[0106] As an implementation manner, step S500, determining the analysis result of the road to be analyzed according to the implicit representation of the road, may include:

[0107] Step S510: loading the implicit representation of the road into the second machine learning model for full connection mapping processing to obtain a first full connection mapping result;

[0108] Step S520: determining a road analysis result of the road to be analyzed according to the first fully connected mapping result.

[0109] In step S510, the edge computing gateway uses the implicit road representation generated in step S400 (or its extended version) as input and loads it into a pre-trained second machine learning model. This model is usually a fully connected neural network (FCNN), also known as a densely connected neural network (Dense Neural Network), which can perform complex nonlinear transformations on input data to achieve multiple tasks such as classification and regression.

[0110] Assume that an edge computing gateway is processing traffic flow monitoring data on a highway. Through the previous steps, the gateway has generated an implicit representation of the road containing multi-dimensional information such as vehicle density, speed distribution, lane occupancy, etc. Now, the gateway inputs this implicit representation into the second machine learning model to determine the current road congestion.

[0111] For the task of classifying congestion conditions, the edge computing gateway may choose a fully connected neural network with multiple hidden layers. This network can extract features that are highly relevant to congestion conditions from the implicit representation of the road through layer-by-layer nonlinear transformations. In each layer of the network, nonlinear activation functions (such as ReLU, Sigmoid, or Tanh) are usually used to increase the expressive power of the model. These activation functions can map the input of neurons to a nonlinear space, allowing the model to learn more complex patterns. In order to train this fully connected neural network, a loss function needs to be defined to measure the difference between the model prediction and the true label. For classification tasks, commonly used loss functions include Cross-Entropy Loss and Negative Log-Likelihood Loss.

[0112] First, the implicit road representation is organized into the format required for model input (such as batch processing, normalization, etc.). The implicit road representation is input into a fully connected neural network, and the final predicted probability distribution is obtained by layer-by-layer calculation. The cross entropy loss is calculated based on the predicted probability distribution and the true label. The network parameters are updated using optimization algorithms such as gradient descent to minimize the loss function.

[0113] In step S520, after the full connection mapping process is completed, the edge computing gateway determines the specific analysis result of the road to be analyzed based on the first full connection mapping result (i.e., the predicted probability distribution of the model). This process usually involves parsing the predicted probability and formulating the decision rules.

[0114] Continuing with the classification of highway congestion conditions as an example, after obtaining the predicted probability distribution of the model, the edge computing gateway will select a category with the highest probability as the final analysis result. For example, if the model predicts that the probability of severe congestion is 0.8, and the probabilities of other categories are all low, the gateway determines that the current road is in a state of severe congestion. Some possible analysis results are: no congestion, light congestion, moderate congestion, severe congestion, and so on. In addition to the classification of congestion conditions, depending on the specific application scenario, the road analysis results may also include accident warning, environmental quality assessment, abnormal behavior detection and other aspects. For example, in environmental quality assessment, the analysis results may include air quality levels (excellent, good, light pollution, moderate pollution, heavy pollution, etc.); in abnormal behavior detection, the type of abnormal behavior and its location information can be directly output.

[0115] In order to accurately extract the analysis results from the predicted probability distribution, the edge computing gateway formulates a set of clear decision rules. These rules may be based on simple threshold judgments (such as selecting the category with the highest probability as the result), or may involve more complex logical processing (such as considering the joint probability distribution of multiple categories). In practical applications, the edge computing gateway can also evaluate the reliability of the prediction results by calculating confidence intervals; or integrate the prediction results of multiple models through ensemble learning methods (such as voting, averaging, etc.) to improve the overall performance. In summary, the process of performing full-connection mapping processing on the implicit representation of the road and determining the road analysis results through the second machine learning model in step S500 is one of the core links of data processing and analysis in the intelligent transportation system. By reasonably selecting machine learning models, formulating clear decision rules, and considering the interpretability and credibility of the results, the edge computing gateway can realize real-time and accurate monitoring and analysis of road conditions to provide strong support for traffic management and decision-making.

[0116] As an implementation mode, the plurality of first machine learning models and the second machine learning models are debugged together according to the prior knowledge template, and the debugging process of the plurality of first machine learning models and the second machine learning models includes:

[0117] Step S10: Acquire prior knowledge template data, where the prior knowledge template data includes example road monitoring information of a plurality of example monitoring roads and a priori marks of road analysis results of each example monitoring road;

[0118] Step S20: extracting sample road monitoring information streams of multiple monitoring modalities from the sample road monitoring information, and mapping the extracted multiple sample road monitoring information streams into multiple sample road perception implicit representation data streams;

[0119] Step S30: using the first perception representation statistic as the scaling quantity and the second perception representation statistic as the scaling step, performing translation scaling extraction on each example road perception implicit representation data stream to obtain a first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream;

[0120] Step S40: loading the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream into the corresponding first machine learning model for implicit representation integration, and combining the example integrated data stream implicit representations corresponding to each example road perception implicit representation data stream obtained by integration to obtain the example road implicit representation;

[0121] Step S50: loading the implicit representation of the example road into the second machine learning model for full-connection mapping processing to obtain a second full-connection mapping result, and determining the model debugging error according to the second full-connection mapping result and the corresponding road analysis result prior mark;

[0122] Step S60: Optimize the internal model parameters of the multiple first machine learning models and the second machine learning model according to the model debugging error.

[0123] In step S10, the edge computing gateway obtains prior knowledge template data from a reliable data source. These data usually come from historical monitoring records, field tests, or simulations, and contain detailed monitoring information of multiple example monitoring roads and their corresponding road analysis result prior markings. The form of these data can refer to the aforementioned step S100, which will not be repeated here.

[0124] Based on historical traffic flow, accident records, environmental monitoring reports and other information, detailed road analysis results are marked for each example monitored road, such as congestion level (no congestion, light congestion, moderate congestion, heavy congestion), accident risk level (low, medium, high), environmental quality level (excellent, good, medium, poor), etc.

[0125] In step S20, the edge computing gateway extracts sample road monitoring information streams of multiple monitoring modalities from the example road monitoring information, and maps these original data streams into higher-level example road perception implicit representation data streams.

[0126] The edge computing gateway communicates with various monitoring devices through preset interfaces and protocols to collect sample road monitoring information in real time or periodically. For video stream data, target detection and tracking algorithms are used to extract the trajectory and attribute information of targets such as vehicles and pedestrians; for radar speed measurement data, the speed value is directly read and parsed; for environmental sensor data, parameter values ​​such as air quality and noise level are collected. The extracted raw data stream is mapped into an implicit representation data stream through feature engineering and encoding techniques. For example, vehicle trajectory information in video stream data can be encoded as a vector sequence containing features such as position, speed, and direction; radar speed measurement data is directly used as one-dimensional time series data; and environmental sensor data is normalized to form a multi-dimensional feature vector.

[0127] In step S30, the edge computing gateway uses translation and scaling technology to extract a feature sequence of a specific length from each example road perception implicit representation data stream to form a first example road perception feature merged data stream. Using a sliding window method, translation and scaling extraction is performed on each example road perception implicit representation data stream according to the set scaling number and stride. Each time the window slides, the feature vectors in the window are merged into a new feature sequence as part of the first example road perception feature merged data stream. The process of step S30 can refer to the aforementioned step S300 and will not be repeated here.

[0128] In step S40, the edge computing gateway loads each first example road perception feature merged data stream into the corresponding first machine learning model for implicit representation integration, and combines the integrated implicit representation into the example road implicit representation. For each first example road perception feature merged data stream, the edge computing gateway inputs it into the corresponding first machine learning model. These models may be built based on algorithms such as convolutional neural networks (CNN), recurrent neural networks (RNN) or their variants (such as LSTM, GRU), which can extract and integrate key features in the data stream. The model outputs the integrated implicit representation vector or matrix.

[0129] The integrated implicit representations output by all the first machine learning models are combined according to certain rules (such as splicing) to form a complete implicit representation of the example road. This representation vector or matrix integrates the feature information of multiple monitoring modalities and reflects the overall state of the example road. The process of step S40 can refer to the aforementioned step S400 and will not be repeated here.

[0130] In step S50, the edge computing gateway loads the implicit representation of the example road into the second machine learning model for full connection mapping processing (i.e., classification or regression task), and calculates the model debugging error based on the processing result and the prior mark. The second machine learning model is usually a neural network containing multiple fully connected layers. It takes the implicit representation of the example road as input and outputs the prediction result (such as congestion level, accident risk level, etc.) by layer-by-layer calculation. The prediction result may be in the form of a probability distribution vector or a specific numerical value. The prediction result of the second machine learning model is compared with the prior mark of the road analysis result in the prior knowledge template to calculate the model debugging error. Common error calculation methods include mean square error (MSE), cross-entropy loss, etc.

[0131] In step S60, the edge computing gateway optimizes internal parameters of multiple first machine learning models and second machine learning models according to the model debugging error to improve the prediction accuracy and generalization ability of the model.

[0132] Commonly used optimization algorithms include Gradient Descent, Stochastic Gradient Descent (SGD), Adam, etc. These algorithms minimize the debugging error by iteratively adjusting the model parameters. During the training process, the back propagation algorithm is used to calculate the gradient of the debugging error with respect to the model parameters. The gradient information is used to guide the update direction of the parameters. The model parameters are updated according to the calculated gradient information. The update rules may include simple additive updates (such as SGD) or more complex adaptive learning rate adjustment strategies (such as Adam). Through the above-mentioned derivative implementation methods, multiple machine learning models in the edge computing gateway can be fully and deeply debugged and optimized based on rich prior knowledge template data. This process not only improves the model's ability to monitor and analyze road conditions, but also provides solid technical support for the stable operation of intelligent transportation systems.

[0133] As an implementation manner, step S20, after extracting sample road monitoring information streams of multiple monitoring modalities from the sample road monitoring information and mapping the extracted multiple sample road monitoring information streams into multiple sample road perception implicit representation data streams, may also include:

[0134] Step S20A: Using the third perception representation statistic as the scaling quantity and the fourth perception representation statistic as the scaling step, perform translation scaling extraction on each example road perception implicit representation data stream to obtain a second example road perception feature merged data stream corresponding to each example road perception implicit representation data stream.

[0135] Step S20A introduces the translation and scaling extraction technology to capture the dynamic change characteristics in the data stream in a more refined way, providing richer information for subsequent implicit representation integration and decision support.

[0136] The third perceptual representation statistic (scaling number) defines the size of the window in each translation and scaling operation, that is, the number of features extracted from the data stream each time. For example, if it is set to 10, it means that each operation will extract feature data for 10 consecutive time steps. The fourth perceptual representation statistic (scaling step) determines the distance the window slides on the data stream. For example, if it is set to 2, after each extraction of 10 features, the window will move forward 2 time steps and then extract again.

[0137] The edge computing gateway performs a sliding window operation on each example road perception implicit representation data stream according to the set number of scales and scale steps. Specifically:

[0138] First, place a window of size equal to the number of scale steps at the beginning of the data stream. Extract feature data of all time steps within the window to form a feature merged data stream segment (i.e., part of the second example road perception feature merged data stream). Move the window forward according to the scale step and repeat the feature extraction process until the entire data stream is covered.

[0139] Assume that the edge computing gateway is processing traffic flow monitoring data on a highway. Through step S20, the gateway has generated an example road perception implicit representation data stream containing multi-dimensional information such as vehicle speed, density, lane occupancy, etc. Now, the gateway analyzes the dynamic change characteristics in these data streams in more detail. Assume that each time point in the data stream corresponds to a feature vector, such as f t =[v t ,ρ t ,o t ], where v t represents the average vehicle speed, ρ t represents the vehicle density, o t Indicates lane occupancy. The edge computing gateway sets the number of scales to 10 (i.e., extracts features of 10 time steps each time) and the scale step to 2 (i.e., the window moves 2 time steps each time). The initial window contains feature vectors from time step 1 to time step 10: [f1,f2,…,f 10 ]. Extract these 10 feature vectors as the first second example road perception feature merged data stream segment. Move the window forward 2 time steps to time step 3 to time step 12, and extract a new feature merged data stream segment. Repeat the above process until the entire data stream is covered, generating a complete second example road perception feature merged data stream.

[0140] Through the translation and scaling extraction in step S20A, the edge computing gateway generates multiple feature merged data stream segments for each example road perception implicit representation data stream. These data stream segments not only contain the key information in the original data stream, but also reveal the local dynamic change characteristics in the data stream through fine scaling operations. These features are of great significance for subsequent implicit representation integration, pattern recognition and decision support.

[0141] Based on this, step S40, loading the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream into the corresponding first machine learning model for implicit representation integration, and combining the example integrated data stream implicit representations corresponding to each example road perception implicit representation data stream obtained by integration to obtain the example road implicit representation, may include:

[0142] Step S41: Load the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream into the corresponding first machine learning model for implicit representation integration, and combine the first example integrated sub-data stream implicit representations corresponding to each example road perception implicit representation data stream obtained by the integration to obtain the first example sub-road implicit representation.

[0143] In the preceding step of step S40, the edge computing gateway has extracted the first example road perception feature merged data stream from each example road perception implicit representation data stream. These data streams contain continuous feature sequences within a specific time window, reflecting the dynamic changes of road conditions within a certain time period. Assume that the first example road perception feature merged data stream corresponding to an example road perception implicit representation data stream is composed of multiple feature vectors, each feature vector contains multi-dimensional information such as vehicle speed, density, lane occupancy, etc. For example, the first feature vector may be f1=[v1,ρ1,o1], where v1 represents the average vehicle speed at the start of the time window, ρ1 represents the vehicle density, and o1 represents the lane occupancy.

[0144] Each first example road perception feature merged data stream has its corresponding first machine learning model. These models may be feature extractors or encoder-decoder structures based on deep learning, which can learn and extract higher-level abstract features based on the input feature sequence.

[0145] Taking the vehicle speed feature as an example, it is assumed that the corresponding first machine learning model is a recurrent neural network (RNN) or its variants (such as LSTM, GRU). This model can capture the time dependency in time series data and model the continuous changes in vehicle speed.

[0146] After loading the first example road perception feature merge data stream into the corresponding first machine learning model, the model will process these feature sequences to extract and integrate higher-level implicit representations. This integration process may involve calculations of multiple network layers, including convolutional layers, pooling layers, fully connected layers, etc., depending on the architecture of the model. Assume that after processing the feature sequence, the RNN model outputs a fixed-length vector as the implicit representation of the sequence. This vector contains key information in the sequence, such as the trend of speed changes, periodicity, etc.

[0147] For each example road perception implicit representation data stream, the edge computing gateway will obtain a corresponding first example integrated sub-data stream implicit representation (i.e., the output vector of the RNN model). Next, the gateway combines these implicit representations to form the first example sub-road implicit representation.

[0148] The combination method is usually simple concatenation, that is, all the implicit representations of the first-instance integrated sub-data streams are concatenated in order into a longer vector. For example, if there are three implicit representations of the first-instance integrated sub-data streams, h1, h2, and h3, the implicit representation of the concatenated first-instance sub-road is [h1, h2, h3].

[0149] Assume that the edge computing gateway is processing traffic flow monitoring data of a city's main road. Through the previous steps, the gateway has extracted feature sequences such as vehicle speed, density, and lane occupancy in multiple time windows, and formed a first-example road perception feature merge data stream. For the vehicle speed feature, the gateway selected an LSTM-based RNN model for processing. By analyzing the continuous speed change sequence, the model learns the trend and periodicity of speed changes and outputs an implicit representation vector of fixed length. For other features (such as density and lane occupancy), the gateway may use different first machine learning models for processing, such as convolutional neural networks (CNN) for image features (if the speed data is obtained through a camera), fully connected neural networks (DNN) for non-time series features, etc. Finally, the gateway splices all the implicit representations of the first-example integrated sub-data streams to form the implicit representation of the first-example sub-road. This representation integrates the traffic flow feature information in multiple time windows, providing strong data support for subsequent tasks such as traffic status analysis and accident prediction.

[0150] Step S42: Load the second example road perception feature merged data stream corresponding to each example road perception implicit representation data stream into the corresponding first machine learning model for implicit representation integration, and combine the second example integrated sub-data stream implicit representations corresponding to each example road perception implicit representation data stream obtained by integration to obtain the second example sub-road implicit representation.

[0151] In step S20A, the edge computing gateway extracts a second example road perception feature merged data stream from each example road perception implicit representation data stream using a translation scale extraction technique with a smaller scale number and stride. These data streams contain a more fine-grained feature sequence than the first example road perception feature merged data stream, which can reveal the rapid changes in road conditions in a short period of time. Assume that the second example road perception feature merged data stream corresponding to an example road perception implicit representation data stream consists of a series of shorter feature vectors, each feature vector still contains multi-dimensional information such as vehicle speed, density, lane occupancy, etc., but with higher temporal resolution.

[0152] Similar to step S41, each second example road perception feature merged data stream is loaded into the corresponding first machine learning model for implicit representation integration. These models are trained to identify and extract key information in the feature sequence and convert it into a higher level of abstract representation. Since the second example road perception feature merged data stream and the first example road perception feature merged data stream may be similar in data structure and feature dimension, they can usually be processed using the same machine learning model. However, in some cases, depending on the characteristics of the data and the analysis requirements, the edge computing gateway may select or adjust model parameters for the second example road perception feature merged data stream. After the second example road perception feature merged data stream is input into the corresponding first machine learning model, the model will process the feature sequence to extract and integrate the implicit representation. This process may involve the calculation of multiple layers of neural networks, including operations such as convolution, pooling, activation, and full connection, depending on the architecture and training objectives of the model. Taking the vehicle speed feature as an example, if a recurrent neural network (RNN) or its variants (such as LSTM, GRU) is used, the model will capture the details of the speed change by analyzing the time dependency in the feature sequence. The output of the model will be an implicit representation vector containing the integrated information.

[0153] For each example road perception implicit representation data stream, the edge computing gateway will obtain a corresponding second example integrated sub-data stream implicit representation. Similar to step S41, the gateway combines these implicit representations to form a second example sub-road implicit representation.

[0154] The combination method can be concatenation, that is, concatenating all the implicit representations of the second example integrated sub-data streams into a longer vector in sequence. For example, if there are N implicit representations of the second example integrated sub-data streams, respectively h1′, h2′, …, h N ′, then the second example sub-road after splicing is implicitly represented as [h1′,h2′,…,h N ′].

[0155] Assume that the edge computing gateway is processing traffic flow monitoring data of a highway. Through the extended translation and scaling extraction technology of step S20A, the gateway obtains multiple second example road perception feature merged data streams, which record the changes of vehicle speed, density and other features with higher time resolution.

[0156] In order to deeply analyze the hidden information in these fine-grained feature sequences, the gateway selected an LSTM-based RNN model for processing. By analyzing the time dependency in the feature sequence, the model learned detailed features such as small fluctuations in speed changes and uneven density distribution, and output the corresponding implicit representation vector. Finally, the gateway spliced ​​all the implicit representations of the second example integrated sub-data streams to form an implicit representation of the second example sub-road. This representation not only contains the key information in the implicit representation of the first example sub-road, but also adds more detailed features about the rapid changes in road conditions, providing more comprehensive and accurate data support for subsequent tasks such as traffic status analysis and accident prediction.

[0157] Step S43: combining the first example sub-road implicit representation with the second example sub-road implicit representation to obtain an example road implicit representation of the example monitored road.

[0158] After completing steps S41 and S42, the edge computing gateway has obtained the implicit representation of the first example sub-road and the implicit representation of the second example sub-road, respectively. These two implicit representations capture the state information of the example monitored road from different time scales and feature granularities. In order to make full use of this information, the gateway combines them.

[0159] The implicit representation of the first example sub-road: Assuming it is represented as R1, it is a long vector or matrix concatenated from multiple implicit representations of the first example integrated sub-data stream. These implicit representations are extracted based on a larger number of scales and strides, reflecting the changing trend of the road status over a longer period of time.

[0160] The second example sub-road implicit representation: Assuming it is represented by R2, it is also a long vector or matrix, which is concatenated by multiple second example integrated sub-data stream implicit representations. These implicit representations are based on a small number of scales and strides, and can capture subtle changes in road conditions in a short period of time.

[0161] The edge computing gateway combines the implicit representation R1 of the first example sub-road with the implicit representation R2 of the second example sub-road. The combination can be concatenation, but other fusion strategies such as weighted fusion and attention mechanism can also be considered, depending on the needs of the application scenario and the characteristics of the data. Assume that the edge computing gateway is processing traffic flow monitoring data of a busy urban main road. Through steps S41 and S42, the gateway generates the implicit representation R1 of the first example sub-road and the implicit representation R2 of the second example sub-road respectively.

[0162] The implicit representation R1 of the first example sub-road may contain the average value or statistics of features such as traffic flow, average vehicle speed, lane occupancy rate, etc. over a longer period of time (such as every hour), reflecting the macro-changing trend of road conditions. The implicit representation R2 of the second example sub-road may contain detailed information such as traffic flow changes and vehicle speed fluctuations over a shorter time interval (such as every minute), revealing the micro-dynamics of road conditions.

[0163] Through the splicing combination in step S43, the gateway obtains a complete implicit representation R of the monitored road that contains both macro trends and micro dynamics. This representation not only provides comprehensive data support for subsequent tasks such as traffic status analysis and accident prediction, but also helps to discover some patterns and anomalies that are difficult to detect under a single scale strategy.

[0164] Step S43 is an important part of the edge computing gateway in the data processing flow of the intelligent transportation system. By effectively combining the implicit representations generated under different scaling strategies, the gateway can generate a more comprehensive and detailed road status representation, providing strong support for subsequent decision support.

[0165] As an implementation mode, each first machine learning model includes an implicit representation mining component and an implicit representation integration component. Based on this, step S40, loading the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream into the corresponding first machine learning model for implicit representation integration, and combining the example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream obtained by integration to obtain the example road implicit representation, may include:

[0166] Step S40a: performing implicit representation pre-mining on the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream according to the implicit representation mining component to obtain a first example pre-mined implicit representation.

[0167] Step S40a aims to extract more critical and abstract information from the original feature merge data stream through pre-extraction of features, i.e., implicit representation pre-mining. The implicit representation mining component is one of the machine learning modules integrated in the edge computing gateway, which is specifically used to extract valuable implicit features from complex data streams. These features not only reflect the surface properties of the data, but also reveal the inherent connections and potential laws between the data.

[0168] The edge computing gateway obtains the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream. These data streams are composed of feature vectors in multiple time windows, and each feature vector contains data from multiple monitoring modalities. The implicit representation mining component receives these feature merged data streams as input and applies a specific algorithm or model to pre-extract features. This process can be regarded as a refinement of the data stream, the purpose of which is to remove redundant information and retain the most critical features for subsequent analysis. The implicit representation mining component may use deep learning models such as convolutional neural networks (CNNs), autoencoders, or variational autoencoders (VAEs). Taking CNN as an example, the model can automatically learn and extract spatial hierarchical features in the data stream through a combination of convolutional layers, pooling layers, and activation functions. After being processed by the implicit representation mining component, each feature merged data stream is converted into a more compact and information-dense implicit representation vector or matrix, that is, the first example pre-mined implicit representation. This representation not only reduces the dimensionality of the data, but also enhances the distinction between data, which facilitates subsequent data integration. For example, suppose the edge computing gateway is processing traffic flow data on a city arterial road. Each feature vector in the original feature merge data stream may contain the following information: [vehicle speed, vehicle density, lane occupancy, PM2.5 concentration, noise decibel]. After being processed by the implicit representation mining component (using CNN as an example), each feature vector is converted into a fixed-length implicit representation vector, such as [0.5, -0.2, 0.8]. Each element in this implicit representation vector corresponds to the abstract feature value of the original feature vector after nonlinear transformation, which together describe the road status in the current time window.

[0169] Step S40a pre-extracts features from the example road perception implicit representation data stream through the implicit representation mining component, realizing the conversion from the original feature merging data stream to the first example pre-mined implicit representation. This process not only simplifies the complexity of the data, but also improves the data representation capability and analysis value, laying a solid foundation for subsequent data integration and decision support.

[0170] Step S40b: perform implicit representation integration on the first example pre-mined implicit representation corresponding to each example road perception implicit representation data stream according to the implicit representation integration component to obtain an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream.

[0171] The implicit representation integration component is another key module in the edge computing gateway, which is specifically used to further integrate the implicit representation after the pre-mining process. This step aims to fuse the implicit representations under multiple time windows or different monitoring modalities into a unified and comprehensive implicit representation to better reflect the overall picture of the road status.

[0172] In step S40a, the first example pre-mined implicit representation corresponding to each example road perception implicit representation data stream has been extracted. These implicit representations may exist in the form of vectors or matrices, and contain key feature information in the original data stream. The edge computing gateway takes each first example pre-mined implicit representation as input and passes it to the implicit representation integration component for processing. The integration component may contain multiple network layers (such as fully connected layers, convolutional layers, recurrent neural network layers, etc.), depending on the complexity of the integration task and the characteristics of the data. The implicit representation integration component may use models such as multi-layer perceptron (MLP), long short-term memory network (LSTM) or graph neural network (GNN). For example, if the data has obvious time series characteristics, LSTM may be a suitable choice; if the relationship between different monitoring modalities needs to be considered, GNN may be more suitable. After being processed by the implicit representation integration component, the implicit representation corresponding to each example road perception implicit representation data stream is integrated into an example integrated data stream implicit representation. This representation not only contains all the key information in the original data stream, but also eliminates redundancy and enhances the correlation between features through the integration process. Taking traffic flow and environmental monitoring at intersections as an example, assume that the first example pre-mined implicit representation in each time window is a vector containing vehicle speed, density, lane occupancy, and environmental parameters (such as PM2.5 concentration, noise decibel). The implicit representation integration component may use an LSTM model to process these vector sequences. LSTM captures long-term dependencies in the sequence through its internal state (memory unit). For the implicit representation vector in each time window, the LSTM model updates its internal state and calculates the output of the current time step based on the new input and internal state. This process is repeated until the data of all time windows are processed. Finally, the LSTM model outputs an integrated implicit representation vector or matrix, which integrates the traffic flow changes and environmental parameter fluctuations at the intersection in different time windows. This integrated representation can be used for subsequent tasks such as traffic state analysis, anomaly detection, and environmental assessment.

[0173] Step S40b further processes the first example pre-mined implicit representation through the implicit representation integration component to form an example integrated data stream implicit representation. This process not only improves the data representation capability, but also provides a more comprehensive and accurate data foundation for subsequent decision support.

[0174] Step S40c: Combine the implicit representations of the example integrated data streams corresponding to each example road perception implicit representation data stream to obtain an example road implicit representation.

[0175] After completing the implicit representation pre-mining of step S40a and the implicit representation integration of step S40b, the edge computing gateway has obtained the implicit representation of the example integrated data stream corresponding to each example road perception implicit representation data stream. These representations may exist in the form of vectors, matrices or higher-dimensional data structures, containing key information in each data stream.

[0176] The edge computing gateway first collects all the implicit representations of the sample integrated data streams. Assuming that the highway is divided into multiple monitoring segments, each monitoring segment corresponds to an example road perception implicit representation data stream, and then corresponds to an example integrated data stream implicit representation. According to the actual application scenario and data characteristics, the edge computing gateway selects a suitable combination strategy to integrate these example integrated data stream implicit representations. Common combination strategies include concatenation, weighted average, dynamic fusion based on attention mechanism, etc. The simplest and most direct way is to sequentially concatenate all the implicit representations of the sample integrated data streams into a longer vector or matrix. For example, if each implicit representation of the sample integrated data stream is a 100-dimensional vector and there are 5 monitoring segments on the highway, the final concatenated implicit representation of the sample road will be a 500-dimensional vector. The edge computing gateway integrates all the implicit representations of the sample integrated data streams into a unified implicit representation of the sample road according to the selected combination strategy. This representation not only contains the key information of each monitoring segment on the highway, but also realizes the global integration of information through combination operations.

[0177] Taking the splicing strategy as an example, suppose there are three monitoring sections A, B, and C on the highway, and the corresponding example integrated data streams are implicitly represented as h A ,h B ,h C Each implicit representation is a 100-dimensional vector, i.e. h A ,h B ,h C ∈R 100 .

[0178] The edge computing gateway concatenates these three vectors in order to obtain the final implicit representation H of the example road: H = [h A ,hB ,h C ]∈R 300 Among them, [⋅,⋅,⋅] represents the vector concatenation operation, and H is a vector containing 300 elements, which fully reflects the status information of the three monitoring sections A, B, and C of the highway.

[0179] Step S40c integrates multiple implicit representations of example integrated data streams into a unified implicit representation of the example road through a combination strategy, thereby achieving a comprehensive description of the state of the entire example road.

[0180] As an implementation mode, each first machine learning model further includes a distribution embedding component. Step S40a, after performing implicit representation pre-mining on the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream according to the implicit representation mining component and obtaining the first example pre-mined implicit representation, may further include:

[0181] Step S40a1: performing distribution information embedding processing on the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream according to the distribution embedding component to obtain the first example road perception distribution feature merged data stream;

[0182] Step S40b, performing implicit representation integration on the first example pre-mined implicit representation corresponding to each example road perception implicit representation data stream according to the implicit representation integration component to obtain an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream, may include:

[0183] Step S40b': performing implicit representation integration on the first example road perception distribution feature merged data stream corresponding to each example road perception implicit representation data stream according to the implicit representation integration component to obtain an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream.

[0184] In the field of intelligent transportation, road perception data often has significant temporal and spatial dependencies. In order to fully capture these dependencies, the edge computing gateway performs distributed information embedding processing on the first example road perception feature merge data stream through the distributed embedding component after the implicit representation pre-mining, namely positional embedding. This process aims to encode the spatial or temporal location information of the data into the feature representation, thereby improving the model's ability to understand sequence data or spatial data. Assume that the edge computing gateway is processing traffic flow monitoring data on a highway. The data consists of real-time data from multiple monitoring stations, each of which records the number, speed, model, and other information of vehicles passing through the station within a specific time period. These data are first integrated into the first example road perception feature merge data stream, each of which represents the change of traffic status over time in a certain area on the highway.

[0185] For time series data, the position embedding method can include sine and cosine function encoding (such as position encoding in the Transformer model), learnable position embedding matrix, etc. Here, it is assumed that sine and cosine function encoding is used because it can naturally handle sequences of arbitrary lengths and does not require a predefined maximum length of the sequence. For each time step feature vector, the distribution embedding component generates a corresponding position embedding vector. This vector is usually the same dimension as the feature vector, but each element is calculated based on the position of the time step. After the distribution information embedding process, the original first example road perception feature merge data stream is converted into the first example road perception distribution feature merge data stream. Each feature vector in this new data stream contains a comprehensive representation of the original feature and time position information, providing a richer source of information for subsequent data integration. After step S40a1, the edge computing gateway obtains the first example road perception distribution feature merge data stream containing distribution information. Next, the implicit representation integration component will further process these data streams to generate an implicit representation of the example integrated data stream. This step aims to integrate the feature vectors of multiple time steps or spatial positions into a unified implicit representation that can fully reflect the dynamic changes in road conditions.

[0186] The implicit representation integration component may use a variety of algorithms or models, such as recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU), Transformers, etc. These models all have the ability to process sequence data, but each has its own characteristics. For example, LSTM and GRU solve the long-term dependency problem of RNN by introducing a gating mechanism; Transformer achieves parallel computing and more distant dependency capture through a self-attention mechanism. Taking LSTM as an example, the implicit representation integration component takes each first example road perception distribution feature merge data stream as the input sequence. The LSTM network captures the temporal dependency in the sequence through its internal state (memory unit and hidden state). At each time step, the LSTM unit receives the position embedding feature vector of the current time step and the hidden state of the previous time step as input, and calculates the new hidden state and memory unit state. This process is repeated until the entire sequence is processed.

[0187] The last hidden state of the LSTM network or some aggregation of all hidden states of the entire sequence (such as average, maximum pooling, etc.) can be used as an implicit representation of the sample integrated data stream. This representation integrates the information of all time steps in the sequence and can reflect the dynamic changes of road conditions.

[0188] After processing in step S40b', the edge computing gateway obtains the implicit representation of the example integrated data stream corresponding to each example road perception implicit representation data stream. These representations not only contain the original feature information, but also incorporate the time or space location information and the dynamic change relationship between sequences. They can be used in subsequent traffic status analysis, anomaly detection, traffic flow prediction and other tasks, providing a strong basis for decision support of intelligent transportation systems.

[0189] For example, in the task of traffic flow prediction, the edge computing gateway can use the implicit representation of the example integrated data stream as input to train a prediction model to estimate the traffic flow in a certain period of time in the future. This model can capture the periodicity, trend and random changes of traffic flow, thereby providing more accurate and reliable prediction results.

[0190] As an implementation manner, S40a1, performing distribution information embedding processing on the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream according to the distribution embedding component to obtain the first example road perception distribution feature merged data stream may include:

[0191] S40a11: Determine the perception weight corresponding to each first example road perception feature combined data stream;

[0192] S40a12: Perform distribution information embedding processing on the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream according to the perception weight, to obtain the first example road perception distribution feature merged data stream.

[0193] In intelligent transportation systems, data from different monitoring modes and different locations may have different description capabilities for the overall road state. Therefore, before performing distributed information embedding processing, the edge computing gateway first needs to determine the importance of each first example road perception feature merged data stream, that is, the perception weight. These weights reflect the contribution of different data streams to the overall road state assessment.

[0194] For example, suppose the edge computing gateway is processing traffic flow data on a main urban road that contains multiple monitoring stations. These monitoring stations are distributed at different locations on the road, and each station is equipped with high-definition cameras, radar speed guns, and other equipment to capture features such as the number, speed, and type of vehicles. These features are integrated into the first example road perception feature merged data stream, each of which corresponds to the real-time data of a monitoring station.

[0195] The method for determining the perception weight is selected according to actual needs, for example, it can include the following methods: The calculation process of the attention mechanism can be expressed as:

[0196] Statistical analysis based on historical data: The edge computing gateway can analyze historical traffic flow data and calculate the accuracy or relevance of each monitoring station's prediction of overall road congestion in different time periods. Stations with high accuracy will be given higher weights.

[0197] Evaluation based on expert knowledge: Traffic engineers or domain experts can make subjective assessments of the importance of monitoring sites based on factors such as road layout and traffic flow characteristics, and give corresponding weight recommendations.

[0198] Machine learning model prediction: The edge computing gateway can also train a dedicated machine learning model, with the input being the data features of the monitoring sites and the output being the weight of each site. This model can automatically learn which sites are more critical to road condition assessment during the training process.

[0199] After determining the perception weights of each first example road perception feature merged data stream, the edge computing gateway then uses these weights to guide the distribution information embedding process at S40a12. Distribution information embedding aims to encode the spatial or temporal location information of the data into the feature representation to improve the model's ability to understand sequence data or spatial data.

[0200] For intelligent transportation systems, distributed information embedding usually includes two types: time position embedding and spatial position embedding. Time position embedding is used to process time series data and capture the dependency of data over time; spatial position embedding is used to process spatial distribution data and capture the spatial correlation of data.

[0201] For time series data (such as changes in traffic flow data over time), the edge computing gateway can use sine and cosine function encoding as a method of time position embedding. However, after adding perception weights, the embedding process can be slightly adjusted. Specifically, the weight factor can be considered when calculating the position embedding vector, so that the data of the key time step is more emphasized in the embedding process. Assume that the position embedding vector pt of time step t was originally calculated by the following formula:

[0202] ;

[0203] Where d is the dimension of the embedding vector. After adding the perceptual weights, the weights can be incorporated into the embedding process, but directly adjusting the parameters of the sine and cosine functions may not be intuitive. An alternative is to multiply the embedding vector with the weights after embedding, that is:

[0204] ;

[0205] in, is the perception weight corresponding to the time step $t$. As another implementation, the weight factor can be considered before the embedding calculation, which needs to be designed according to the specific embedding method and weight determination mechanism. The perception weight can be combined with the embedded feature vector instead of directly adjusting the embedding process. For example, before adding or concatenating the embedding vector with the original feature vector, the embedding vector can be multiplied by the corresponding weight. For spatially distributed data (such as vehicle density at different locations on the road), the edge computing gateway can use a similar method to handle spatial position embedding. However, unlike temporal position embedding, spatial position embedding needs to consider the relative position relationship of the data in space. Assume that the road is divided into multiple grids, and the vehicle density data in each grid is integrated into a feature vector. In order to capture the spatial dependency between grids, the edge computing gateway can generate a spatial position embedding vector for each grid. This vector can be calculated based on factors such as the coordinates of the grid, the relationship between adjacent grids, etc. Similarly, after adding the perception weight, the weight can be multiplied or otherwise combined with the spatial position embedding vector to emphasize the importance of key grids. However, in intelligent transportation systems, the implementation of spatial location embedding may be more complicated, because the road network is often not a simple grid structure, but a complex network consisting of multiple road sections, intersections, etc. Therefore, spatial location embedding can adopt more advanced methods, such as graph neural networks (GNNs), to capture the complex spatial relationships in road networks.

[0206] After processing in step S40a12, the edge computing gateway obtains the weighted first example road perception distribution feature merged data stream. These data streams not only contain the original feature information but also incorporate the time and space location information and perception weight factors so that the subsequent data integration and analysis process can more accurately reflect the dynamic changes of road conditions. These weighted distribution feature merged data streams can be used in a variety of intelligent transportation application scenarios such as traffic flow prediction, congestion warning, accident detection, etc. In the traffic flow prediction task, the weighted feature representation can provide more accurate historical data support to help the prediction model more accurately estimate the trend of future traffic flow changes; in the congestion warning task, the weighted feature representation can highlight the congestion of key sections and provide timely warning information for the traffic management department so that corresponding measures can be taken to alleviate the congestion; in the accident detection task, the weighted feature representation can enhance the recognition ability of abnormal traffic patterns and improve the accuracy and timeliness of accident detection. Step S40a1 adds a new dimension to the data processing process of the intelligent transportation system by determining the perception weight and embedding the distribution information, so that the system can better understand and analyze the dynamic changes of road conditions and provide more reliable and valuable information for subsequent decision support.

[0207] As an implementation mode, the implicit representation integration component includes a saliency focusing component, a first cross-layer normalization component, a multi-layer perceptron component, and a second cross-layer normalization component. Step S40b, according to the implicit representation integration component, performs implicit representation integration on the first example pre-mined implicit representation corresponding to each example road perception implicit representation data stream to obtain an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream, which may include:

[0208] Step S40b1: performing saliency focusing on the first example pre-mined implicit representation corresponding to each example road perception implicit representation data stream according to the saliency focusing component to obtain a first saliency focused implicit representation.

[0209] In the embodiment of the present application, the saliency focus component uses the attention mechanism to dynamically weight the input data to highlight the features that are most important to the task goal. In traffic data processing, this means that the model can automatically identify and focus on key information such as traffic flow changes, congestion, and abnormal events, while ignoring factors that are irrelevant to the task or have little impact.

[0210] The edge computing gateway receives the first example pre-mined implicit representation from step S40a as input. These representations are usually in the form of vectors or matrices, and contain abstract feature information of the road state.

[0211] The attention mechanism in the saliency focus component starts working by first defining a series of learnable query, key, and value vectors. In practical applications, these vectors may be learned from the input data through a multi-layer perceptron (MLP), or the input data itself may be used as the key and value, and another set of learnable parameters may be used as the query. For each element (or feature) in the input data, the attention mechanism calculates the similarity between the query and the key (usually using dot product, cosine similarity, etc.). Then, the softmax function is used to convert these similarities into probability distributions, namely attention weights. Finally, the values ​​are weighted and summed according to these weights to obtain the weighted output.

[0212] Specifically, assume that the input data is a matrix H∈R n×d (where n is the time step or the number of monitoring sites, and d is the feature dimension), the calculation process of the attention mechanism can be expressed as:

[0213] ;

[0214] in, , represents the query matrix, is the key matrix, is the value matrix, T represents the transpose, dk Indicates the dimensions of the key matrix.

[0215] After weighted processing by the attention mechanism, the edge computing gateway obtains the first significant focused implicit representation, which not only retains the key information in the original data, but also enhances the expressiveness of important features through dynamic weighting.

[0216] For example, suppose the first example pre-mined implicit representation in a time window is a vector h=[0.5,20,0.8] containing features such as vehicle density, average speed, lane occupancy, etc. During the saliency focus process, the attention mechanism may identify that the average speed (20) is a key feature of the current road state (for example, on a highway, a too high average speed may mean heavy traffic or a risk of speeding). Therefore, the attention mechanism assigns a higher weight to the average speed feature, while the weights of other features are relatively low. The weighted first saliency focused implicit representation may become a=[0.3,2.5,0.7], where the feature value of the average speed is significantly amplified.

[0217] Step S40b1 dynamically weights the first example pre-mined implicit representation through the saliency focus component to generate the first saliency focus implicit representation. This process effectively highlights key features and suppresses irrelevant information, providing more accurate and efficient input for subsequent data integration and analysis.

[0218] Step S40b2: performing a standardization operation on the first salient focus implicit representation and the first example front mining implicit representation corresponding to each example road perception implicit representation data stream according to the first cross-layer standardization component to obtain a first standardized implicit representation.

[0219] For example, suppose the edge computing gateway is processing traffic flow and environmental monitoring data on the main roads of a city. These data come from multiple monitoring sites, including high-definition cameras, radar speed guns, and environmental sensors. After preliminary processing and implicit representation mining, the gateway has obtained the first example pre-mined implicit representation of each monitoring site (denoted as h i ) and the first significant focused implicit representation (denoted as a i ), where i represents different monitoring sites or time windows.

[0220] The first cross-layer normalization component is often used in deep learning models to improve training stability and accelerate convergence. It scales and translates the input data so that the mean of the data is close to 0 and the variance is close to 1 (or a fixed value), thereby eliminating the scale differences between different features. In step S40b2, the component considers both the first significant focus implicit representation and the first example pre-mining implicit representation, aiming to integrate them into a unified scale.

[0221] In step S40b2, the edge computing gateway receives the first salient focus implicit representation a from step S40b1. i and the first example pre-mined implicit representation h from step S40a i as input.

[0222] In some implementations, the first cross-layer normalization component may further include a cross-layer connection (such as a residual connection), that is, i With h i Perform some form of combination (e.g., element-wise addition or concatenation). i , the first cross-layer normalization component performs the following operations:

[0223] First, calculate all a in the current batch (or the entire dataset) i Then, use the following formula for each a i To standardize:

[0224] ;

[0225] in, is a small constant used to prevent division by zero.

[0226] In practical applications, especially when using batch normalization, the mean and variance are usually calculated on each mini-batch rather than on the entire dataset. In addition, the normalized data is usually transformed by a learnable scaling factor γ and offset β to restore the expressiveness of the data:

[0227] ;

[0228] The standardized si is used as the output of step S40b2 and is passed to the subsequent implicit representation integration component.

[0229] For example, suppose the first significant focus implicit representation of a monitoring station is a=[2.5,1.2,−0.8], with mean μ=1.0 and variance σ2=1.5. Without considering the scaling factor and offset, the first standardized implicit representation may be s=[(2.5−1.0) / 1.5+10−5,…]≈[1.11,…].

[0230] Step S40b2 performs a standardization operation on the first salient focus implicit representation (and possibly the first example pre-mined implicit representation, depending on implementation details) through the first cross-layer standardization component to improve the consistency and stability of the data. This step provides a more standardized input data format for subsequent data integration and analysis, which helps to improve the efficiency and accuracy of the entire processing flow.

[0231] Step S40b3: perceive the first standardized implicit representation corresponding to each example road perception implicit representation data stream according to the multi-layer perceptron component to obtain a first perception implicit representation.

[0232] Step S40b3 extracts higher-level abstract features from the standardized data through multi-layer nonlinear transformations to provide richer information for subsequent decision support.

[0233] For example, suppose the edge computing gateway is processing traffic flow data of an urban area with multiple intersections. These data come from monitoring equipment such as high-definition cameras and radar speed guns deployed at various intersections. After preliminary processing and implicit representation mining, a standardized implicit representation data stream is formed. In step S40b2, these representations have been standardized through the first cross-layer normalization component to eliminate scale differences and improve data consistency and stability.

[0234] Multilayer Perceptron (MLP) is a feedforward neural network consisting of multiple fully connected layers, each containing multiple neurons. Each neuron receives the output of the previous layer as input and obtains its own output through weighted summation and nonlinear activation function calculation. MLP can learn complex nonlinear mapping relationships and extract useful feature representations from input data.

[0235] In step S40b3, the multi-layer perceptron component receives the first standardized implicit representation as input, and outputs the first perceptual implicit representation through calculation of the multi-layer network. This process is essentially to further abstract and transform the input data in order to better capture the inherent patterns and regularities in the data.

[0236] The edge computing gateway uses the first standardized implicit representation output by step S40b2 as the input of the multilayer perceptron component. These representations may be a series of vectors or matrices containing standardized road perception features. The multilayer perceptron component contains multiple hidden layers, each of which consists of multiple neurons. The input data first enters the first hidden layer, multiplied by the weight matrix and added with the bias term, and then the output of the layer is calculated by a nonlinear activation function (such as ReLU, Sigmoid, or Tanh). This process is repeated until the last hidden layer or output layer is reached. Assuming that the multilayer perceptron component has two hidden layers, each hidden layer contains a number of neurons, the calculation process can be described as:

[0237] ;

[0238] ;

[0239] ;

[0240] Among them, s is the first normalized implicit representation, σ is the nonlinear activation function, and W (i) and b (i) are the weight matrix and bias term of the i-th layer, z (i) is the output of the i-th layer (also called the activation value), and p is the final output of the multilayer perceptron component (i.e., the first perceptual implicit representation). Note that nonlinear activation functions are usually no longer applied after the output layer unless there is a specific need (such as Softmax activation in classification tasks). For example, in the above formula, W (1) represents the weight matrix of the first layer, b (1) represents the bias term of the first layer; W (2) represents the weight matrix of the first layer, b (2) represents the bias term of the first layer; W (3) represents the weight matrix of the first layer, b (3) represents the bias term of the first layer; z (1) is the activation value of layer 1; z (2) is the activation value of layer 2.

[0241] After multi-layer network calculations, the edge computing gateway obtains the first perception implicit representation p. This representation contains higher-level abstract features of the input data, which is of great significance for subsequent tasks such as traffic status analysis, flow prediction, and congestion warning.

[0242] Step S40b3 further abstracts and transforms the first standardized implicit representation through a multi-layer perceptron component to generate a first perceptual implicit representation.

[0243] Step S40b4: performing a standardization operation on the first standardized implicit representation and the first perceptual implicit representation corresponding to each example road perception implicit representation data stream according to the second cross-layer standardization component to obtain an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream.

[0244] In step S40b4, the gateway further standardizes the first standardized implicit representation and the first perceived implicit representation through the second cross-layer normalization component, aiming to improve the stability and generalization ability of the data, and finally generates an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream.

[0245] The second cross-layer normalization component is designed to re-normalize the data after being processed by the multi-layer perceptron to ensure a more even distribution of the data and reduce the problem of internal covariate shift, thereby helping to improve the stability and efficiency of subsequent processing steps. In addition, through cross-layer connections (such as residual connections), this component can also retain some of the original information to avoid losing important features in the deep network.

[0246] Specifically, the edge computing gateway receives the first perceived implicit representation (denoted as pi) from step S40b3 and the first normalized implicit representation (denoted as si) from step S40b2 (or after appropriate processing). Here, i represents different example road perceived implicit representation data streams or time windows.

[0247] As some embodiments, a second cross-layer normalization component may combine pi and si via a cross-layer connection. This combination may be a simple concatenation or an element-wise addition (if both have the same dimensions or have the same dimensions after appropriate transformation). For pi and (optionally) si corresponding to each example road perception implicit representation data stream, the second cross-layer normalization component performs normalization. This process is similar to the first cross-layer normalization, but may be performed for different representations or combined representations. The normalization operation typically involves calculating the mean and variance (possibly based on the entire dataset or the current mini-batch), then subtracting the mean from each element and dividing by the standard deviation (plus a small constant to prevent division by zero).

[0248] Assume that directly i After normalization (ignoring the cross-layer connection), the normalized representation can be recorded as f i ′, its calculation process can be expressed as:

[0249] ;

[0250] ;

[0251] ;

[0252] Where N is p i The dimension, p ij Yes i The jth element of , μ is the mean, σ2 is the variance, ϵ is a small constant, and 1 is the same as p i An all-ones vector of the same dimension.

[0253] The representation after the second cross-layer normalization (may be f i ′ or a more complex representation containing cross-layer connection information) is output as an example integrated data stream implicit representation. This representation combines the information of the first standardized implicit representation and the first perceptual implicit representation, and improves the stability and consistency of the data through standardization.

[0254] Step S40b4 further standardizes the first standardized implicit representation and the first perceived implicit representation through the second cross-layer standardization component to generate an example integrated data stream implicit representation. This process not only improves the stability and consistency of the data, but also helps to improve the efficiency and accuracy of subsequent processing steps. In intelligent transportation systems, such a data processing flow is of great significance for achieving efficient and accurate traffic monitoring and management.

[0255] The present application also provides an edge computing gateway, such as Figure 2 As shown, the edge computing gateway 100 includes: a processor 101 and a memory 103. Among them, the processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the edge computing gateway 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the edge computing gateway 100 does not constitute a limitation on the embodiments of the present application. The embodiment of the present application provides an edge computing gateway, and the edge computing gateway in the embodiment of the present application includes: one or more processors; a memory; one or more computer programs, wherein one or more computer programs are stored in the memory and configured to be executed by one or more processors, and when one or more programs are executed by one or more processors, the above method is implemented.

Claims

1. A data processing method for an edge computing gateway, characterized in that: The method comprises: Obtain a multi-dimensional road monitoring information set of the road to be analyzed; Extracting a plurality of road monitoring information streams of monitoring modes from the multi-dimensional road monitoring information set, and mapping the extracted plurality of road monitoring information streams into a plurality of road perception implicit representation data streams; Using the first perception representation statistic as the scaling quantity and the second perception representation statistic as the scaling step, performing translation scaling extraction on each of the road perception implicit representation data streams to obtain a first road perception feature merged data stream corresponding to each of the road perception implicit representation data streams; Performing implicit representation integration on each of the first road perception feature merged data streams, and combining multiple implicit representations of integrated data streams obtained by integration to obtain an implicit road representation of the road to be analyzed; Determining an analysis result of the road to be analyzed according to the road implicit representation; The step of performing implicit representation integration on each of the first road perception feature merged data streams, and combining the implicit representations of multiple integrated data streams obtained by integration to obtain the implicit representation of the road to be analyzed, includes: Determine a first machine learning model corresponding to each of the first road perception feature combined data streams; Loading the first road perception feature merged data stream corresponding to each of the road perception implicit representation data streams into the corresponding first machine learning model for implicit representation integration, to obtain an integrated data stream implicit representation corresponding to each of the road perception implicit representation data streams; Combining a plurality of implicit representations of the integrated data streams to obtain an implicit representation of the road to be analyzed; Wherein, determining the analysis result of the road to be analyzed based on the implicit representation of the road includes: Loading the road implicit representation into a second machine learning model for full connection mapping processing to obtain a first full connection mapping result; Determining a road analysis result of the road to be analyzed according to the first fully connected mapping result; The plurality of first machine learning models and the second machine learning models are debugged together according to a priori knowledge template, and the debugging process of the plurality of first machine learning models and the second machine learning models includes: Acquire prior knowledge template data, wherein the prior knowledge template data includes example road monitoring information of a plurality of example monitoring roads and a priori mark of a road analysis result of each example monitoring road; Extracting sample road monitoring information streams of multiple monitoring modalities from the example road monitoring information, and mapping the extracted multiple sample road monitoring information streams into multiple example road perception implicit representation data streams; Using the first perception representation statistic as the scaling quantity and the second perception representation statistic as the scaling stride, performing translation scaling extraction on each of the example road perception implicit representation data streams to obtain a first example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams; Loading the first example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams into the corresponding first machine learning model for implicit representation integration, and combining the example integrated data stream implicit representations corresponding to each of the example road perception implicit representation data streams obtained by integration to obtain an example road implicit representation; Loading the implicit representation of the example road into a second machine learning model for full-connection mapping processing to obtain a second full-connection mapping result, and determining a model debugging error based on the second full-connection mapping result and a priori marking of a corresponding road analysis result; The internal model parameters of the multiple first machine learning models and the second machine learning model are optimized based on the model debugging error.

2. The method according to claim 1, characterized in that The step of combining the plurality of implicit representations of the integrated data streams to obtain the implicit representation of the road to be analyzed includes: Determining a first importance coefficient implicitly represented by each of the integrated data streams; Adjusting each of the integrated data stream implicit representations according to the first importance coefficient to obtain a plurality of adjusted data stream implicit representations; The plurality of implicit representations of the regulated data streams are combined to obtain an implicit representation of the road to be analyzed.

3. The method according to claim 1, characterized in that After extracting a plurality of road monitoring information streams of monitoring modes from the road multi-dimensional monitoring information set and mapping the extracted plurality of road monitoring information streams into a plurality of road perception implicit representation data streams, the method further includes: Using the third perception representation statistic as the scaling quantity and the fourth perception representation statistic as the scaling step, performing translation scaling extraction on each of the road perception implicit representation data streams to obtain a second road perception feature merged data stream corresponding to each of the road perception implicit representation data streams; The step of performing implicit representation integration on each of the first road perception feature merged data streams, and combining the implicit representations of multiple integrated data streams obtained by integration to obtain the implicit representation of the road to be analyzed, includes: Performing implicit representation integration on each of the first road perception feature merged data streams, and combining the implicit representations of multiple first integrated sub-data streams obtained by integration to obtain a first sub-road implicit representation, and performing implicit representation integration on each of the second road perception feature merged data streams, and combining the implicit representations of multiple second integrated sub-data streams obtained by integration to obtain a second sub-road implicit representation; The first sub-road implicit representation and the second sub-road implicit representation are combined to obtain the road implicit representation of the road to be analyzed.

4. The method according to claim 1, characterized in that: After extracting a plurality of sample road monitoring information streams of monitoring modalities from the example road monitoring information and mapping the extracted plurality of sample road monitoring information streams into a plurality of example road perception implicit representation data streams, the method further includes: Using the third perception representation statistic as the scaling quantity and the fourth perception representation statistic as the scaling stride, performing translation scaling extraction on each of the example road perception implicit representation data streams to obtain a second example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams; The step of loading the first example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams into the corresponding first machine learning model for implicit representation integration, and combining the example integrated data stream implicit representations corresponding to each of the example road perception implicit representation data streams obtained by integration to obtain an example road implicit representation, includes: Loading the first example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams into the corresponding first machine learning model for implicit representation integration, and combining the first example integrated sub-data stream implicit representations corresponding to each of the example road perception implicit representation data streams obtained by integration to obtain a first example sub-road implicit representation; Loading the second example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams into the corresponding first machine learning model for implicit representation integration, and combining the second example integrated sub-data stream implicit representations corresponding to each of the example road perception implicit representation data streams obtained by integration to obtain a second example sub-road implicit representation; Combining the first example sub-road implicit representation with the second example sub-road implicit representation to obtain an example road implicit representation of the example monitored road; Each of the first machine learning models includes an implicit representation mining component and an implicit representation integration component, wherein the first example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams is loaded into the corresponding first machine learning model for implicit representation integration, and the example integrated data stream implicit representation corresponding to each of the example road perception implicit representation data streams obtained by integration is combined to obtain an example road implicit representation, including: Performing implicit representation pre-mining on a first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream according to the implicit representation mining component to obtain a first example pre-mined implicit representation; Performing implicit representation integration on the first example pre-mined implicit representation corresponding to each example road perception implicit representation data stream according to the implicit representation integration component to obtain an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream; The example integrated data stream implicit representations corresponding to each of the example road perception implicit representation data streams are combined to obtain an example road implicit representation.

5. The method according to claim 4, characterized in that Each of the first machine learning models further includes a distribution embedding component, and after performing implicit representation pre-mining on the first example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams according to the implicit representation mining component to obtain the first example pre-mined implicit representation, it also includes: Performing distribution information embedding processing on the first example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams according to the distribution embedding component to obtain the first example road perception distribution feature merged data stream; The implicit representation integration component is used to perform implicit representation integration on the first example pre-mined implicit representation corresponding to each example road perception implicit representation data stream to obtain an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream, including: The implicit representation integration component performs implicit representation integration on the first example road perception distribution feature merged data stream corresponding to each example road perception implicit representation data stream to obtain an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream.

6. The method according to claim 5, characterized in that The step of performing distribution information embedding processing on the first example road perception feature merged data stream corresponding to each example road perception implicit representation data stream according to the distribution embedding component to obtain the first example road perception distribution feature merged data stream comprises: Determine the perception weight corresponding to each of the first example road perception feature combined data streams; A distribution information embedding process is performed on the first example road perception feature merged data stream corresponding to each of the example road perception implicit representation data streams according to the perception weight to obtain a first example road perception distribution feature merged data stream.

7. The method according to claim 4, characterized in that The implicit representation integration component includes a saliency focusing component, a first cross-layer normalization component, a multi-layer perceptron component, and a second cross-layer normalization component. The implicit representation integration component is used to perform implicit representation integration on the first example pre-mined implicit representation corresponding to each example road perception implicit representation data stream to obtain an example integrated data stream implicit representation corresponding to each example road perception implicit representation data stream, including: performing saliency focusing on the first example front-mined implicit representation corresponding to each of the example road perception implicit representation data streams according to the saliency focusing component to obtain a first saliency focused implicit representation; performing a normalization operation on the first salient focus implicit representation and the first example front mining implicit representation corresponding to each of the example road perception implicit representation data streams according to the first cross-layer normalization component to obtain a first normalized implicit representation; Perceiving the first standardized implicit representation corresponding to each of the example road perception implicit representation data streams according to the multi-layer perceptron component to obtain a first perception implicit representation; The first standardized implicit representation and the first perceptual implicit representation corresponding to each of the example road perception implicit representation data streams are standardized according to the second cross-layer normalization component to obtain an example integrated data stream implicit representation corresponding to each of the example road perception implicit representation data streams.

8. An edge computing gateway, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the one or more processors, the method according to any one of claims 1 to 7 is implemented.