A cloud-based monitoring method and system for toll stations

Abstract

This invention relates to the field of road traffic monitoring and maintenance technology, specifically a cloud-based monitoring method and system for toll stations. The method includes: collecting multimodal data at the toll station; performing time-series alignment, feature extraction, and semantic mapping at the edge to generate event semantic fingerprints, which are then compressed and stored; simultaneously performing anomaly measurement and priority determination; calculating event priorities based on a multidimensional scoring mechanism combined with anomaly degree, impact range, and frequency; conducting preliminary anomaly classification through semantic similarity matching and formulating an intelligent transmission strategy; receiving high-priority events at the cloud and integrating the event set; decomposing the semantic fingerprints to construct a cross-modal causal correlation matrix; identifying causal chains and extracting high-causal-strength event combinations to generate synthetic event vectors for anomaly verification; feeding back the cloud causal analysis results to the edge; adaptively adjusting the filtering weight threshold at the edge; and iteratively optimizing the strategy through cloud-edge consistency checks. This invention achieves real-time anomaly response and resource optimization.

Description

Technical Field

[0001] This invention relates to the field of road traffic monitoring and maintenance technology, specifically to a cloud-based monitoring method and system for toll stations. Background Technology

[0002] The intelligent construction of highways is continuously advancing, and the application of cloud computing and edge computing collaborative architecture in traffic management is becoming increasingly in-depth. The maturity of multimodal perception technology provides a technical foundation for the full-element status monitoring of toll stations. With the continuous increase in traffic flow density and the promotion of unmanned operation mode, higher requirements are placed on the real-time response capability of abnormal events and the accuracy of complex scenario discrimination. Traditional localized monitoring systems have room for optimization in terms of data fusion depth and cross-device collaboration, and the centralized processing mode in the cloud is difficult to meet the real-time requirements of data. The industry urgently needs to reduce network transmission load through edge-side intelligent screening, while strengthening the cloud's ability to perform correlation analysis of multi-source events to achieve accurate risk assessment.

[0003] Chinese invention patent application CN120388427A discloses an emergency response system for toll stations integrating a localized AI model. This system comprises five subsystems: an edge computing device, a multi-source sensing module, a localized AI model engine, a dynamic rule engine, and an execution control interface. The edge computing device is an embedded hardware with AI acceleration computing power deployed at the toll station. The multi-source sensing module connects to cameras, millimeter-wave radar, and ETC-RSU equipment. By deploying a lightweight large model on an edge server, the system enables millisecond-level identification and automated handling of emergencies such as vehicles running through toll gates and hazardous material leaks.

[0004] Meanwhile, the standardized construction of smart highways requires technical support for rapid deployment and adaptive optimization across stations, driving the cloud-edge collaboration mechanism towards dynamic feedback and knowledge accumulation. Against this backdrop, cloud-based monitoring technology that integrates semantic compression, causal reasoning, and policy transfer has become a key research direction for improving the operational efficiency of toll stations. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a cloud-based monitoring method and system for toll stations.

[0006] The technical solution of this invention: a cloud-based monitoring method for toll stations, comprising the following specific implementation steps:

[0007] S1. Collect multi-source data on vehicle, personnel, environment and equipment status through multimodal sensing nodes at the toll station side;

[0008] S2. At the edge, perform time-series alignment, feature extraction, and semantic mapping on multi-source data to generate event semantic fingerprints and compress and store them; and at the edge, perform anomaly measurement and priority discrimination on the event semantic fingerprints, and formulate intelligent transmission strategies based on the discrimination results.

[0009] S3: The cloud receives high-priority events uploaded from the edge, integrates the event set and decomposes the event semantic fingerprint to construct a cross-modal causal association matrix, identifies causal chains and extracts high-causal-strength event combinations to generate synthetic event vectors for anomaly verification.

[0010] S4. Feed the cloud-based causal analysis results back to the edge, where the edge adaptively adjusts the filtering weights and thresholds, and iteratively optimizes the monitoring strategy through cloud-edge consistency checks.

[0011] Preferably, generating and compressing the event semantic fingerprint specifically includes:

[0012] Target detection and behavior analysis are performed on the acquired video frame sequences to extract visual feature vectors;

[0013] Perform time-frequency analysis on radar signals to extract radar feature vectors;

[0014] Mel frequency cepstral coefficients and energy abrupt points are extracted from acoustic data to generate acoustic feature vectors;

[0015] The equipment status parameters are standardized and discretized to form equipment status feature vectors;

[0016] The feature vectors of each modality are fused to form a multimodal feature vector;

[0017] The multimodal feature vectors are transformed into event semantic labels using a semantic mapping function;

[0018] Event semantic tags are encoded into dynamically weighted semantic fingerprint vectors to achieve compressed storage.

[0019] Preferably, the anomaly measurement and priority determination of event semantic fingerprints at the edge specifically includes:

[0020] Calculate the deviation between the current event semantic fingerprint and the mean of historical semantic fingerprints to obtain an anomaly metric.

[0021] If the abnormal measurement value exceeds the preset abnormal judgment threshold, it is judged as a potential abnormal event and a local warning is triggered.

[0022] Calculate the impact scope and frequency indicators of an event;

[0023] Construct a multidimensional priority scoring vector based on anomaly metrics, impact range indicators, and frequency indicators;

[0024] The multidimensional priority scoring vector is weighted and aggregated to generate a comprehensive priority score for the event;

[0025] Events are sorted and filtered based on their overall priority scores.

[0026] Preferably, the formulation of intelligent transmission strategies specifically includes:

[0027] The events are categorized into three types based on priority scores and preliminary anomaly assessment results;

[0028] For high-priority emergency anomalies that are initially determined to be abnormal, immediately upload them to the cloud and trigger local alarms;

[0029] For ordinary abnormal events that are initially abnormal but of medium priority, perform periodic batch uploads or delayed uploads.

[0030] Low-priority, low-risk events that are similar to historical events are stored in the edge cache for statistical analysis.

[0031] Preferably, the process of constructing the cross-modal causal correlation matrix includes:

[0032] The received semantic fingerprint is decomposed into visual features, radar features, acoustic features, and device status features;

[0033] Calculate the correlation strength of different events across different modes;

[0034] Based on the weights of each modality, the overall correlation matrix between events is obtained.

[0035] The overall correlation matrix is ​​modeled using Bayesian networks or causal graph models to identify potential causal chains and abnormal triggering sources, and to generate an event causal relationship graph.

[0036] Preferably, generating synthetic event vectors for anomaly verification specifically includes:

[0037] Extract high-causal-strength event combinations from the identified causal chains;

[0038] The feature vectors of the combined events are used to generate a synthetic event vector through a synthesis function.

[0039] Calculate the similarity between the synthesized event vector and the historical event reference vector;

[0040] If the similarity is lower than the set synthetic similarity threshold, it is marked as a potential combined abnormal event and a cloud-based early warning message is generated.

[0041] Preferably, the causal analysis results from the cloud are fed back to the edge, specifically including:

[0042] The cloud remaps the priority parameters at the edge based on the causal strength of different event categories;

[0043] The mapped weight parameters are transmitted to the edge via the feedback interface;

[0044] The edge device adaptively adjusts the weight coefficients and filtering thresholds used when calculating anomalies and priorities locally based on the received weight parameters.

[0045] Preferably, the monitoring strategy is iteratively optimized through cloud-edge consistency checks, specifically including:

[0046] After adjusting the parameters, the edge device re-filters the recent event cache.

[0047] Send some of the re-filtered event results back to the cloud;

[0048] The cloud compares the differences between the edge and global analyses, and if there are systematic biases, it generates compensation parameters and provides secondary feedback.

[0049] After multiple rounds of feedback iterations, the cloud-edge strategy tends to stabilize and converge.

[0050] Preferably, the monitoring strategy, which iteratively optimizes through cloud-edge consistency checks, also includes:

[0051] The results of the strategy after stable convergence are solidified into a transferable strategy template;

[0052] The strategy template is stored in a cloud knowledge base for rapid deployment and adaptation at this toll station and other toll stations.

[0053] The technical solution of the present invention: a cloud-based monitoring system for toll stations, used to execute the above-mentioned cloud-based monitoring method for toll stations, comprising:

[0054] The edge-aware and semantic micro-agent module is used to deploy multimodal sensors and perform data acquisition, temporal alignment, feature extraction, and semantic fingerprint generation.

[0055] The edge priority and initial screening decision module is used to calculate the event anomaly degree and multi-dimensional priority score, perform preliminary anomaly judgment and formulate hierarchical transmission strategy;

[0056] The cloud-based causal association and synthesis verification module is used to build a global event library and perform cross-modal causal chain inference and synthesis event verification.

[0057] The cloud-edge linkage feedback and adaptive optimization module is used to transform cloud analysis conclusions into edge policy instructions, perform consistency checks, and accumulate transferable policy templates.

[0058] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects:

[0059] This invention designs a cloud-based monitoring method and system for toll stations. It significantly reduces network transmission burden through semantic fingerprint compression and real-time filtering of multimodal data at the edge, while utilizing local anomaly measurement and priority grading mechanisms to achieve rapid response to emergency events. In the cloud, cross-modal causal chain analysis reveals hidden risk correlations, and combined with synthetic event verification, enhances the ability to identify complex anomalies. A cloud-edge collaborative feedback mechanism dynamically optimizes edge strategy parameters, achieving system adaptive convergence through consistency checks and difference compensation. The final, transferable strategy templates support rapid deployment across sites, ensuring real-time monitoring response while reducing manual reliance and maintenance costs, providing reliable technical support for intelligent operation and maintenance of highways. Attached Figure Description

[0060] Figure 1 This is a flowchart of a cloud-based monitoring method for toll stations proposed in this invention.

[0061] Figure 2 This is a system architecture diagram of a cloud-based monitoring system for toll stations proposed in this invention. Detailed Implementation

[0062] Example 1, as Figure 1 As shown, the cloud-based monitoring method for toll stations proposed in this invention includes the following specific implementation steps:

[0063] S1. By deploying multimodal sensing nodes at toll stations, multi-source data such as vehicle, personnel, environment, and equipment status are collected. After unified temporal alignment, feature extraction, and semantic mapping at the edge, event-level semantic fingerprints are formed and compressed for storage. At the same time, anomaly measurement and priority judgment are performed locally, thereby achieving efficient screening of uploaded data and rapid response to emergency events. The specific implementation process is as follows:

[0064] S11. In the toll lane and toll booth areas, deploy multimodal sensing nodes such as video cameras, millimeter-wave radar, microphone arrays, and equipment status sensors to simultaneously acquire vehicle motion characteristics, traffic sound characteristics, equipment operating parameters, and surrounding environmental status information. These data types include, but are not limited to:

[0065] Video frame sequence V t This mainly describes the vehicle's appearance, license plate location, and the actions of the people involved.

[0066] Radar signal R t This is used to supplement vehicle speed and distance characteristics;

[0067] Acoustic eigenvector A t It can detect signals such as horns, collisions, and abnormal equipment noise.

[0068] Equipment status parameter E t It covers the opening and closing status of the barrier gate, power load, and the working status of the toll collection terminal;

[0069] Based on this, a complete multimodal observation set X is formed. t ={V t ,R t A t E t};

[0070] S12. Asynchronous data of different modalities are uniformly aligned through a weighted time window to ensure temporal consistency. Specifically, the weighted time window function maps multimodal data within the same time period to a unified observation time t.

[0071] S13. At the edge, edge computing technology is used to extract features from video, radar, acoustic, and device status data, and these features are then transformed into high-level event labels through a semantic mapping function. Specifically:

[0072] For the acquired video frame sequence V t Object detection and behavior analysis are performed by extracting vehicle edge contours, license plate character blocks, and human motion vectors using a lightweight convolutional network to form a feature set F. v ;

[0073] Radar signal R t The velocity amplitude distribution and distance curve characteristics F were obtained through time-frequency analysis. r ;

[0074] Acoustic data A t Extract Mel-frequency cepstral coefficients (MFCCs) and energy abrupt change points to generate acoustic features F. a ;

[0075] Status parameter E collected from the toll booth equipment t Standardization and discretization are performed to form the device status fingerprint F. e ;

[0076] The modal feature sets F v F r F a and F e The fusion is performed at the edge to form a multimodal feature vector F. t =[F v ,F r ,F a ,F e ];

[0077] Through semantic mapping function (This embodiment uses a lightweight classifier) ​​to convert the feature vector into an event semantic label Y. t=S(F t );

[0078] Examples include: "Vehicles forcibly pass through without paying tolls", "Machining malfunction of the barrier gate", and "Abnormal noise".

[0079] S14. Encode the event semantic labels into semantic fingerprint vectors, and achieve efficient data expression and transmission through dynamic weighted compression, that is: encode each event semantic label Y t Encoded as a semantic fingerprint vector Its construction method is as follows:

[0080] ;

[0081] in, f represents a semantic fingerprint vector; v f r f a and f e These represent the feature vectors of visual, radar, acoustic, and device modalities, respectively; α1, α2, α3, and α4 represent modal weights, characterizing the relative importance of each modality within the event window, and are adaptively adjusted based on edge data.

[0082] S15. Perform anomaly measurement on semantic fingerprints. If the anomaly score exceeds the threshold, issue a direct warning at the edge and prioritize uploading the data to ensure real-time response to potential emergencies. Specifically:

[0083] semantic fingerprints Anomaly measurement is performed using the following distance function: ;

[0084] Among them, D t σ represents the anomaly metric, which measures the degree of deviation between the current event window's semantic fingerprint and the historical mean, and is used for anomaly detection; μ represents the historical mean of the semantic fingerprint, which is the mean based on past event fingerprints; σ represents the historical standard deviation of the semantic fingerprint, which measures the fluctuation range of the fingerprint and is used to normalize the anomaly degree.

[0085] Therefore: If D t If the value exceeds the set anomaly detection threshold θ, it is identified as a potential anomaly event, triggering a local alert and prioritizing its upload to the cloud for processing.

[0086] S2. By performing multi-dimensional scoring on semantic fingerprints and anomaly metrics, and combining the event's impact range and duration to calculate priority, events are sorted and filtered. At the edge, semantic similarity matching is used for preliminary anomaly judgment, and events are classified into urgent, ordinary, and low-risk categories. An intelligent transmission strategy is then formulated, the specific implementation process of which is as follows:

[0087] S21. Generate a comprehensive score for each event to quantify its importance and ensure reasonable resource allocation at the edge. Specifically:

[0088] semantic fingerprints The corresponding anomaly degree D t This reflects the degree to which an event deviates from the normal pattern;

[0089] Calculate the event's impact scope index C t Specifically:

[0090] For vehicle incidents, count the number of lanes and vehicles affected;

[0091] For equipment malfunctions, count the number of affected toll collection machines and the traffic capacity.

[0092] For environmental incidents (such as noise or abnormal behavior), count the surrounding people and the area scope;

[0093] Calculate the event frequency index S t By using edge caching to record similar events in the past T w Number of occurrences per second;

[0094] Based on this, a multidimensional priority scoring vector P is constructed. t : ;

[0095] Among them, w d w c and w s The weights of each indicator are represented and adaptively adjusted through initial trial data statistics or online learning to ensure reasonable scoring in different scenarios.

[0096] S22. Sort the events detected by the edge nodes and filter out high-priority events to prepare data for cloud or local alarms, specifically:

[0097] Multidimensional score P t Perform weighted aggregation to generate a single priority score U t : ;

[0098] Where n represents the rating vector P t In this embodiment, n=3, corresponding to the anomaly degree D. t Event impact scope index C t and event frequency index S t ;U t This represents the overall priority score of the events, used for sorting and filtering; w i This represents the weight of the corresponding indicator, which can be dynamically adjusted through online learning at the edge nodes or based on management experience; P t,iThis represents the value of the i-th indicator in the rating vector;

[0099] Press U for all events within the same time window t Sort the values ​​in descending order;

[0100] Set the threshold U based on the computing and transmission capabilities of the edge nodes. min Only retain U t ≥U min The event will be entered into the subsequent anomaly detection or cloud upload queue;

[0101] S23. Quickly identify potential anomalies in the filtered events to provide a basis for edge early warning and cloud-based in-depth analysis, specifically:

[0102] Selected semantic fingerprints Semantic fingerprint database of historical normal events Similarity Sim t calculate: ;

[0103] Set a similarity threshold λ, if the similarity Sim t If the value is less than λ, it is considered a preliminary anomaly.

[0104] For events initially determined to be abnormal, local alarm information is generated or uploaded to the cloud first, depending on the type of abnormality (such as vehicle violation, equipment failure, or environmental abnormality).

[0105] S24. Based on priority and anomaly detection results, formulate an intelligent transmission strategy from the edge to the cloud to optimize resources and maximize response efficiency, specifically as follows:

[0106] The events are divided into three categories:

[0107] Category 1, Emergency Abnormal Events: High priority and initially determined to be abnormal, immediately uploaded to the cloud and trigger local alarms;

[0108] Category 2, Common Abnormal Events: Initial abnormality but medium priority, can be uploaded in batches periodically or with a delay;

[0109] Category 3, Low-risk events: These events have low priority and are similar to historical events. They are stored only in the edge cache for statistical analysis.

[0110] For emergency events, the edge module records the complete semantic fingerprint of the event and the original multimodal data fragments to ensure sufficient information for further analysis in the cloud; low-risk events can be uploaded by periodically compressing and summarizing to optimize the use of data resources.

[0111] S3: The cloud receives high-priority abnormal events and their semantic fingerprints uploaded from the edge. By constructing an event set and comparing it with historical events, it performs cross-modal causal correlation analysis to identify potential causal chains. It also performs synthetic verification on event combinations to confirm anomalies and assess potential impacts. The analysis results are intelligently fed back to the edge nodes to achieve edge-cloud collaborative optimization, improve the accuracy of anomaly detection and the overall system response capability. The specific implementation process is as follows:

[0112] S31. To integrate uploaded events into a unified event set and perform semantic fingerprint comparison with the historical event database to generate historical reference vectors, providing prior information for causal analysis, specifically:

[0113] The cloud receives event data uploaded from the edge device, including but not limited to semantic fingerprints. Anomaly degree D t Event timestamps, lane and vehicle information are integrated to form a unified event set. :

[0114] ;

[0115] Among them, e i Represents the i-th event; n represents the number of events;

[0116] For each event e i Retrieve similar events H from the historical event database H i Through semantic fingerprint similarity Sim(e i ,h j )calculate: ;

[0117] Retain the set of historical events with a similarity higher than a set threshold. ;

[0118] Generate event history reference vector : ;

[0119] Among them, h j Represents historical event samples; H i Indicates the relationship with event e i A set of potentially similar historical events; Represents a set of highly similar historical events;

[0120] S32. Decompose the event semantic fingerprint into visual, radar, acoustic, and device state modal features, construct a modal correlation matrix, and generate an overall causal correlation matrix through weighted mutual information or conditional probability. Then, use a Bayesian network or causal graph model to identify event chains and potential triggering sources. Specifically:

[0121] Decompose the semantic fingerprint of an event into feature vectors of different modalities:

[0122] Visual features (license plate recognition, vehicle type, abnormal behavior) F v ;

[0123] Radar or lidar characteristics (vehicle speed, trajectory, abnormal stopping) F r ;

[0124] Acoustic characteristics (abnormal sounds, whistles) F s ;

[0125] Device status characteristics (gate status, card reader status) F e ;

[0126] Construct an event modality association matrix C, whose elements Indicates event e i and e j The degree of correlation on mode m is calculated using weighted mutual information: ;

[0127] in, Indicates event e i With e j The correlation strength in mode m; The weights represent the modal m.

[0128] The overall correlation matrix is ​​obtained by combining all modes: ;

[0129] in, Represents the elements of the overall event correlation matrix;

[0130] Model C using Bayesian networks or causal graph models, identify potential causal chains and abnormal triggering sources, and generate an event causal graph G=(V,E), where nodes V represent events and edges E represent causal effects.

[0131] S33. Extract high-causal-strength event combinations from the causal chain, generate synthetic event vectors, and compare them with historical event vectors to confirm abnormal combinations. Generate cloud-based early warning information and assess potential risks. Specifically:

[0132] Extracting high-causal-strength event combinations (such as e) from causal networks i →e j →e k Generate synthetic event vectors. : ;

[0133] For the synthetic event vector With historical event vectors Perform similarity matching:

[0134] ;

[0135] in, Sim represents the composition function; syn Indicates the similarity between the synthesized event and historical references;

[0136] If Sim syn <λ syn If a potential combined abnormal event is identified (based on a set similarity threshold), it is marked as such.

[0137] S4. Feedback strategies are generated based on the causal analysis results in the cloud and mapped to the edge. The edge adaptively adjusts the screening weights and thresholds based on the feedback, and the cloud-edge strategies gradually converge through consistency checks and difference compensation. Finally, the optimization results are precipitated into transferable strategy templates, realizing dynamic optimization and long-term knowledge accumulation under cloud-edge linkage. The specific implementation process is as follows:

[0138] S41. Based on the causal network G and synthetic event similarity results output in step S3, the cloud generates a set of feedback strategies for the edge side. Specifically, the cloud will adjust the causal strength C for different event categories. ij The risk level of the event is re-labeled by weighting and sorting the events and combining the results of historical verification.

[0139] This causal strength is compared with the original priority parameter (w) at the edge. d w c w s Mapping is performed so that the edge can automatically adjust its computational weights in future monitoring;

[0140] For example, when a certain modality exhibits a significant triggering effect in causal analysis, the cloud will amplify its weight and transmit it to the edge via a feedback interface, thereby changing the real-time filtering strategy at the edge.

[0141] S42. After receiving feedback from the cloud, the edge device will adaptively reconstruct the monitoring parameters locally.

[0142] Specifically, the edge device adjusts the filtering threshold and weight coefficient, and re-filters recent events through a local event caching mechanism to test the applicability of the cloud correction strategy.

[0143] S43. Construct a cloud-edge consistency verification mechanism: In the initial stage after policy reconstruction, the edge end will send back some of the re-selected event results to the cloud again. The cloud will compare the analysis differences between the local and global sides. If a systematic bias is found at the edge end (e.g., a certain type of event is continuously underestimated by the edge), the cloud will dynamically generate compensation parameters and provide secondary feedback to reduce the policy bias between the cloud and the edge.

[0144] S44. When the cloud edge reaches a stable convergence after multiple rounds of feedback iteration, the strategy results of this stage are solidified into a template and stored in the knowledge base in the cloud.

[0145] These strategy templates can be applied to this toll station scenario for a long time and can be migrated to toll stations in other regions or extended to similar intelligent duty scenarios, achieving cross-domain reuse and rapid adaptation.

[0146] Example 2, as Figure 2 As shown, the present invention proposes a toll station cloud monitoring system, which is used to execute a toll station cloud monitoring method proposed in Embodiment 1, including: an edge perception and semantic micro-agent module, an edge priority and initial screening decision module, a cloud causal association and synthesis verification module, and a cloud-edge linkage feedback and adaptive optimization module.

[0147] The edge-aware and semantic micro-agent module is used to deploy multimodal sensors at toll lanes and toll booths and perform local time synchronization, noise suppression and lightweight feature extraction. The module has a built-in semantic-first compressor and a minimum forensic kernel selector to convert video, radar, acoustic and device status into structured event summaries and irreversible semantic fingerprints. At the same time, it saves short-term forensic fragments in a local circular cache and generates retrieval pointers to meet the requirements of on-demand retrieval and privacy protection.

[0148] The edge priority and initial screening decision module calculates multi-dimensional priority scores based on semantic fingerprints and historical fingerprints. It integrates anomaly intensity, impact range and persistence, and performs dynamic threshold screening and semantic similarity matching to complete the initial anomaly identification. It is responsible for generating hierarchical transmission strategies and adaptively adjusting compression rate and upload scheduling according to available bandwidth to ensure that high-risk events are transmitted first while low-risk events are stored locally for statistical summary.

[0149] The cloud-based causal association and synthesis verification module is responsible for receiving high-priority events uploaded from the edge, building a global event library, and conducting cross-modal causal chain inference based on an interpretable causal attention mechanism. It uses a synthetic event engine to generate near-realistic combined samples under physical constraints to enhance the ability to distinguish rare anomalies. At the same time, the cloud maintains a model container marketplace for canary deployment and performance orchestration, and distributes the evidence package and risk level output from causal analysis as disposal suggestions.

[0150] The cloud-edge linkage feedback and adaptive optimization module is responsible for transforming cloud-based causal conclusions into executable strategy instructions at the edge and implementing them at the edge through a progressive strategy reconstruction mechanism. It includes a consistency verification subprocess to compare the differences between edge trial operation and cloud-based global analysis and generate compensation factors. Finally, the converged strategy is precipitated as a transferable knowledge template for other sites to quickly adapt and use.

[0151] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A cloud-based monitoring method for toll stations, characterized in that, The specific implementation steps include the following: S1. Collect multi-source data on vehicle, personnel, environment and equipment status through multimodal sensing nodes at the toll station side; S2. Perform time-series alignment, feature extraction, and semantic mapping on multi-source data at the edge, generate event semantic fingerprints, and compress and store them. Furthermore, anomaly measurement and priority discrimination are performed on the semantic fingerprint of the event at the edge, and an intelligent transmission strategy is formulated based on the discrimination results. Specifically, the formulation of intelligent transmission strategies includes: The events are categorized into three types based on priority scores and preliminary anomaly assessment results; For high-priority emergency anomalies that are initially determined to be abnormal, immediately upload them to the cloud and trigger local alarms; For ordinary abnormal events that are initially abnormal but of medium priority, perform periodic batch uploads or delayed uploads. Low-priority, low-risk events that are similar to historical events are stored in the edge cache for statistical analysis. S3: The cloud receives high-priority events uploaded from the edge, integrates the event set and decomposes the event semantic fingerprint to construct a cross-modal causal association matrix, identifies causal chains and extracts high-causal-strength event combinations to generate synthetic event vectors for anomaly verification. The process of constructing the cross-modal causal association matrix includes: The received semantic fingerprint is decomposed into visual features, radar features, acoustic features, and device status features; Calculate the correlation strength of different events across different modes; Based on the weights of each modality, the overall correlation matrix between events is obtained. The overall correlation matrix is ​​modeled using Bayesian networks or causal graph models to identify potential causal chains and abnormal triggering sources, and to generate an event causal relationship graph. S4. Feed the cloud-based causal analysis results back to the edge, where the edge adaptively adjusts the filtering weights and thresholds, and iteratively optimizes the monitoring strategy through cloud-edge consistency checks.

2. The cloud-based monitoring method for toll stations according to claim 1, characterized in that, Generating and compressing event semantic fingerprints specifically includes: Target detection and behavior analysis are performed on the acquired video frame sequences to extract visual feature vectors; Perform time-frequency analysis on radar signals to extract radar feature vectors; Mel frequency cepstral coefficients and energy abrupt points are extracted from acoustic data to generate acoustic feature vectors; The equipment status parameters are standardized and discretized to form equipment status feature vectors; The feature vectors of each modality are fused to form a multimodal feature vector; The multimodal feature vectors are transformed into event semantic labels using a semantic mapping function; Event semantic tags are encoded into dynamically weighted semantic fingerprint vectors to achieve compressed storage.

3. The cloud-based monitoring method for toll stations according to claim 2, characterized in that, The specific steps for anomaly measurement and priority determination of event semantic fingerprints at the edge include: Calculate the deviation between the current event semantic fingerprint and the mean of historical semantic fingerprints to obtain an anomaly metric. If the abnormal measurement value exceeds the preset abnormal judgment threshold, it is judged as a potential abnormal event and a local warning is triggered. Calculate the impact scope and frequency indicators of an event; Construct a multidimensional priority scoring vector based on anomaly metrics, impact range indicators, and frequency indicators; The multidimensional priority scoring vector is weighted and aggregated to generate a comprehensive priority score for the event; Events are sorted and filtered based on their overall priority scores.

4. The cloud-based monitoring method for toll stations according to claim 3, characterized in that, Generate synthetic event vectors for anomaly verification, specifically including: Extract high-causal-strength event combinations from the identified causal chains; The feature vectors of the combined events are used to generate a synthetic event vector through a synthesis function. Calculate the similarity between the synthesized event vector and the historical event reference vector; If the similarity is lower than the set synthetic similarity threshold, it is marked as a potential combined abnormal event and a cloud-based early warning message is generated.

5. The cloud-based monitoring method for toll stations according to claim 4, characterized in that, Feeding the cloud-based causal analysis results back to the edge, specifically including: The cloud remaps the priority parameters at the edge based on the causal strength of different event categories; The mapped weight parameters are transmitted to the edge via the feedback interface; The edge device adaptively adjusts the weight coefficients and filtering thresholds used when calculating anomalies and priorities locally based on the received weight parameters.

6. The cloud-based monitoring method for toll stations according to claim 5, characterized in that, The monitoring strategy is iteratively optimized through cloud-edge consistency checks, specifically including: After adjusting the parameters, the edge device re-filters the recent event cache. Send some of the re-filtered event results back to the cloud; The cloud compares the differences between the edge and global analyses, and if there are systematic biases, it generates compensation parameters and provides secondary feedback. After multiple rounds of feedback iterations, the cloud-edge strategy tends to stabilize and converge.

7. The cloud-based monitoring method for toll stations according to claim 6, characterized in that, The monitoring strategy is iteratively optimized through cloud-edge consistency checks, which also includes: The results of the strategy after stable convergence are solidified into a transferable strategy template; The strategy template is stored in a cloud knowledge base for rapid deployment and adaptation at this toll station and other toll stations.

8. A cloud-based monitoring system for toll stations, used to execute the cloud-based monitoring method for toll stations as described in any one of claims 1 to 7, characterized in that, include: The edge-aware and semantic micro-agent module is used to deploy multimodal sensors and perform data acquisition, temporal alignment, feature extraction, and semantic fingerprint generation. The edge priority and initial screening decision module is used to calculate the event anomaly degree and multi-dimensional priority score, perform preliminary anomaly judgment and formulate hierarchical transmission strategy; The cloud-based causal association and synthesis verification module is used to build a global event library and perform cross-modal causal chain inference and synthesis event verification. The cloud-edge linkage feedback and adaptive optimization module is used to transform cloud analysis conclusions into edge policy instructions, perform consistency checks, and accumulate transferable policy templates.

Images