A Method and Equipment for Tunnel Anomaly Detection and Response Based on Track Robots

Abstract

This application provides a method and device for tunnel anomaly detection and response based on a track robot, comprising: a pre-set model optimization module that obtains scoring and grading data from a quantitative scoring model, combines historical response records, applies a deep Q-learning network to dynamically adjust the parameters of the scoring model, optimizes the grading threshold to match the actual threat level, and transmits the adjusted grading data to a response scheduling system; based on the adjusted grading data, it automatically matches a pre-set combination of response strategies, and if there is at least one optional strategy combination, it uses a weighted optimization algorithm to comprehensively consider response speed, resource utilization, and disposal cost to determine the optimal strategy combination, and issues execution instructions to the relevant execution units.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and device for tunnel anomaly detection and response based on a track robot. Background Technology

[0002] In tunnel inspection systems, the massive amount of 360° video stream data collected by tracked robots presents challenges for real-time processing. Edge computing nodes have limited computing power, making it difficult to support complex AI models, while lightweight models may affect detection accuracy. The quantitative metrics of anomaly scoring models struggle to comprehensively cover all situations, and fixed thresholds cannot adapt to dynamic environments. Training deep Q-learning networks requires a large amount of sample data, but anomalies are relatively rare, potentially leading to model overfitting. The solution selection problem under multi-objective optimization is NP-hard (Non-deterministic polynomial time), making real-time optimization of the optimal solution difficult. Summary of the Invention

[0003] To achieve the objectives of this invention, in a first aspect, this invention provides a tunnel anomaly detection and response method based on a track robot, mainly comprising:

[0004] A track-based robot performs timed inspections along a pre-defined track, collecting panoramic video stream data from a camera with a pre-defined resolution. This video stream is transmitted in real-time to an edge computing node, where a lightweight AI vision model performs anomaly detection and analysis. If anomalies are detected, abnormal feature data is extracted, a structured message is generated, and this message is input into a pre-defined quantization scoring model to calculate a score. The score determines the anomaly level, which has four levels with priority from highest to lowest. A deep Q-learning network dynamically adjusts the scoring threshold of the pre-defined quantization scoring model, re-classifying the anomalies based on the adjusted thresholds. Finally, a pre-defined response strategy combination is matched to the adjusted classification data. If at least one alternative strategy combination exists, a weighted optimization algorithm is activated. The system comprehensively calculates the optimal strategy combination based on three dimensions: response time, resource utilization, and handling cost, and issues execution instructions to relevant execution units. The tracked robot receives the instructions, moves to the nearest station to the anomaly point, collects anomaly image data from the site, and uploads it to the cloud server. A deep learning image feature matching algorithm is used to compare the data with a pre-stored image database to confirm the anomaly type. Simultaneously, a navigation path is generated based on a simultaneous localization and mapping (SLAM) algorithm, driving the robot to move a second time to the anomaly point. Upon reaching the anomaly point, the tracked robot activates the nearest broadcast node, plays pre-recorded voice instructions, and connects to surrounding surveillance cameras to acquire real-time video stream data. This data is then transmitted to the command center via a real-time communication protocol, establishing a multi-party video communication channel. During anomaly handling, a message queue telemetry transmission protocol is used for instruction data transmission, and all message transmission records are stored in a pre-set data analysis module.

[0005] We extract metrics data, including response time and success rate, from message transmission records, dynamically adjust the level of abnormal conditions, and use log analysis stacks to statistically evaluate the effectiveness of alarm pushes, optimize push strategies, and regularly update the parameters of the deep Q-learning network based on data accumulation, continuously improving response strategies through offline training.

[0006] In a second aspect, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0007] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0008] This invention discloses a method and device for tunnel anomaly detection and response based on a track-mounted robot. The method utilizes a track-mounted robot to perform inspection tasks along a preset track, collecting panoramic video streams within the tunnel, and analyzing and detecting anomalies in real time through edge computing nodes. When an anomaly is detected, feature data is extracted and classified using a quantitative scoring model, with the classification threshold dynamically optimized by a deep Q-learning network. A response scheduling system matches the optimal strategy combination based on the classification data, directing the track-mounted robot to quickly return to the anomaly point, activating broadcast nodes, and establishing multi-party video communication. The entire process employs a message queue telemetry transmission protocol to ensure reliable instruction delivery, and a data analysis module continuously optimizes the anomaly classification and response strategy. This invention achieves intelligent detection, rapid response, and continuous optimization of tunnel anomalies, improving the efficiency and accuracy of tunnel safety management. Attached Figure Description

[0009] Figure 1 This is a flowchart of the tunnel anomaly detection and response method based on a track robot according to the present invention.

[0010] Figure 2 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0012] like Figure 1 The tunnel anomaly detection and response method based on a track robot in this embodiment may specifically include:

[0013] S101, the track robot performs timed inspection tasks along a preset track, and is equipped with a camera of preset resolution to collect panoramic video stream data in the tunnel. The acquired video stream is transmitted to the edge computing node in real time, and the video stream is analyzed for anomaly detection through a lightweight artificial intelligence vision model to obtain the anomaly detection results.

[0014] A track-based robot performs inspections along a pre-set track. During the inspection, a camera with a preset resolution onboard the robot collects panoramic video data of the tunnel environment, obtaining raw video stream information. This raw video stream is then transmitted to an edge computing node for initial storage and processing to obtain basic video data. Based on this basic video data, a lightweight AI vision model is used to decompose the video stream into frames, extracting image features from each frame. Key regions within the images are then identified based on these features. An anomaly detection process is then executed for these key regions to obtain the anomaly detection results.

[0015] For example, in a scenario where a tracked robot inspects a tunnel, the robot, as the core equipment, moves at a constant speed of 5 kilometers per hour along a pre-set track, carrying a high-definition camera to collect real-time panoramic video data inside the tunnel. This camera has a 1080p resolution and can cover multiple angles inside the tunnel, including the ceiling, side walls, and ground, ensuring no blind spots in the video data. The collected raw video stream information is transmitted to edge computing nodes at a rate of 30 frames per second. This process effectively reduces the latency of data transmission to the cloud and improves real-time performance. Specifically, when the edge computing node performs initial storage and processing of the raw video stream, it compresses the video data and converts it into a basic video data format for subsequent analysis. For example, the node may store the video stream as a temporary file of 500MB per minute, while simultaneously performing preliminary noise reduction processing to filter out blurry parts caused by insufficient lighting inside the tunnel or camera shake. This improves the accuracy of subsequent image analysis and reduces the possibility of misjudgment. In one possible implementation, based on the basic video data, the system uses a lightweight artificial intelligence visual model to perform frame decomposition of the video stream. This lightweight AI vision model is based on the MobileNetV3 or YOLOv8-nano architecture and has been adapted for tunnel scene features (such as low light and fixed background). It retains core feature extraction layers (such as the first 5 layers of the CSPDarknet backbone network) to reduce computational cost. The lightweight AI vision model generation process is as follows: First, a base model is pre-trained based on a publicly available tunnel anomaly dataset (such as TunnelDefect-10K). Then, a structured pruning algorithm (such as L1 regularized channel pruning) is used to remove redundant convolutional kernels—by calculating the sum of the absolute values ​​of the channel weights, pruning 30% of the low-weight channels. Simultaneously, knowledge distillation technology is used, with the original complex model (such as YOLOv8-large) as the teacher model to guide the lightweight model in learning feature mappings, ensuring that the accuracy loss after pruning is ≤5%. After optimization, edge computing node computing power testing (such as the NVIDIA Jetson Nano platform) verifies that the processing time per frame is consistently within 50 milliseconds (including image preprocessing time). This model is typically based on deep learning technology, but to adapt to the limited computing power of edge computing nodes, it undergoes pruning optimization to ensure that the processing time for each image frame is controlled within 50 milliseconds. After decomposition, the system extracts features from each image frame, such as color distribution and edge contours, and uses these features to determine key regions in the image, such as cracks in the tunnel wall or foreign objects on the ground. This approach significantly improves data processing efficiency while reducing computational resource consumption. For example, in analyzing key region information, if the system detects spilled objects in the image, such as scattered plastic bottles, pedestrians entering restricted areas of the tunnel, or even illegally parked vehicles, it will trigger an anomaly detection process.The system compares these abnormal features with preset anomaly templates, such as the shape of spilled objects and the dynamic trajectory of pedestrians, to generate anomaly detection results. These results may include information such as the anomaly type, location coordinates, and severity. For example, if a spilled object is located 200 meters from the tunnel entrance and covers an area of ​​approximately 0.5 square meters, the system would mark it as a medium-risk event. This detection process can promptly identify potential safety hazards, providing accurate information for subsequent handling. In one possible implementation, after the anomaly detection results are generated, the system automatically sends an alarm message to the management center, while simultaneously capturing relevant video clips and images. Figure 1 The data is then uploaded for staff to quickly verify. For example, for illegally parked vehicles, the system provides the specific location and duration of parking; if a vehicle remains in the tunnel for more than 10 minutes, it is escalated to a high-level alarm. This multi-dimensional data support helps managers make quick decisions and reduces the probability of safety accidents in the tunnel. Through these technologies, the track robot inspection system automates the entire process from data collection to anomaly detection, not only improving inspection efficiency but also significantly reducing the risks and costs of manual inspections. For example, a traditional manual inspection might take several hours to complete a tunnel inspection, while the robot system can complete the same task in one hour with higher accuracy. This technological effect is of great significance for ensuring tunnel operation safety and improving management efficiency.

[0016] S102. If there is an abnormal situation in the anomaly detection result, extract the abnormal feature data, generate a structured message, input the structured message into the preset quantitative scoring model to calculate the score value, and determine the level of the abnormal situation based on the score value. There are four levels of abnormal situations, with the first level being the highest and the fourth level being the lowest in terms of processing priority.

[0017] If an anomaly is detected in the anomaly detection results, anomaly feature data including the anomaly image, station number, and time stamp are extracted to generate a structured message. The structured message is then input into a preset quantization scoring model to calculate a score ranging from 0 to 100. Based on preset scoring thresholds and the score values, the level of the anomaly is determined, resulting in an anomaly classification system. A scoring threshold is set for every 25 points.

[0018] For example, in the scenario of a tracked robot inspecting a tunnel, the processing flow for anomaly detection results can be analyzed and illustrated in detail from multiple perspectives to ensure that the logic is rigorous and aligns with actual business needs. For instance, extracting feature information from anomaly images, station locations, and timestamps can be understood as structuring key anomaly segments in the raw video data. Suppose that during tunnel inspection, the system detects a ground spill, the image shows scattered gravel, the station location is recorded as 300 meters from the tunnel entrance, and the timestamp is 10:15 AM. This information, once extracted, forms a data packet containing an image screenshot, the specific location, and the time point. The advantage of this extraction method is that it integrates scattered anomaly information into a traceable record, facilitating subsequent analysis and archiving, while also providing a basis for quickly locating problems. Specifically, the system might compress the image data into a small file, while embedding the station location and timestamp as text in a message. This message data acts like a concise event report, containing a basic description of the anomaly, facilitating subsequent system calls and transmission. This processing method improves data transmission efficiency and lays the foundation for the next stage of evaluation. For example, the process of assessing anomalies using a quantitative scoring model can be analyzed from two dimensions: threat level and impact range. The threat level of the anomaly is determined by a predefined rule base, categorized into three levels: "high," "medium," and "low," with corresponding base scores: High threat: 70 points; Medium threat: 50 points; Low threat: 30 points. The impact range is derived from anomaly image analysis, with a predefined mapping relationship between impact range and score: Impact range ≤ 1 square meter: 20 points; 1 square meter < Impact range ≤ 5 square meters: 6 points per square meter; 5 square meters < Impact range ≤ 10 square meters: 50 points; > 10 square meters: 70 points. Suppose a tracked robot detects scattered debris 300 meters from the tunnel entrance, shown in the image as loose gravel located in the center of the lane. Rule base matching: Anomaly type: Scattered debris; Location parameter: Center of lane. According to the rule base, the anomaly threat level corresponding to "Scattered debris - Center of lane" is medium, with a base score of 50 points. The impact range, based on image analysis, is estimated to cover an area of ​​2 square meters, with a score of 2 × 6 = 12 points. The weighting ratio of threat level to impact range for anomalies is 7:3 (threat level is dominant). The system combines these two parameters with structured messages to calculate a score using the following formula: Threat Level Score × 70% + Impact Range Score × 30% = 50 × 0.7 + 12 × 0.3 = 35 + 3.6 = 38.6 points. The advantage of this scoring mechanism is that it transforms subjective anomalies into objective numerical values, facilitating standardized and hierarchical management. For example, when classifying anomalies based on their scores, 76-100 points correspond to Level 1 events, 51-75 points to Level 2 events, 26-50 points to Level 3 events, and 0-25 points to Level 4 events. Therefore, a score of 38.6 corresponds to a Level 3 event, indicating that the anomaly requires timely handling but is not considered an emergency.Generating hierarchical data helps the management center allocate resources rationally, such as prioritizing Level 1 events while assigning routine cleanup to Level 3 events. This hierarchical approach effectively improves the targeting and efficiency of anomaly handling.

[0019] S103. Apply a deep Q-learning network to dynamically adjust the scoring threshold of the preset quantitative scoring model, and reclassify the abnormal situation according to the adjusted scoring threshold to obtain the adjusted classification data.

[0020] A pre-defined quantitative scoring model is analyzed using a deep Q-learning network to obtain initial configuration data for the scoring thresholds and determine the baseline range of the scoring thresholds. Based on the initial configuration data, a dynamic adjustment mechanism is used to update the scoring thresholds in real time, resulting in adjusted scoring thresholds. Abnormal situations are then re-scored based on the adjusted scoring thresholds. Adjusted classification data for abnormal situations is obtained based on the re-scoring results.

[0021] For example, in the business scenario of tunnel inspection, the process of analyzing a quantitative scoring model using a deep Q-learning network can be discussed from two levels: principle and specific implementation. Deep Q-learning networks are a reinforcement learning method whose core lies in optimizing the threshold configuration of the scoring model through continuous trial and error and feedback. In tunnel inspection, the system may analyze the score distribution of abnormal conditions based on historical inspection data to initially determine the baseline range of the scoring threshold. For example, the score from 0 to 100 might be divided into four level intervals: 0-25, 26-50, 51-75, and 76-100, corresponding to different degrees of anomaly severity. This initial configuration lays the foundation for subsequent dynamic adjustments. For instance, the process of dynamically updating the scoring threshold in real time based on the initial configuration data can be understood as the system continuously revising the threshold range based on new data from actual inspections. In one possible implementation, suppose during a tunnel inspection, the system detects a certain type of anomaly, such as minor cracks in the tunnel wall, with scores frequently falling between 26 and 30 points. However, in actual handling, it is found to pose a low threat. The system might adjust the lower threshold of the corresponding level from 26 to 30 points. This dynamic adjustment mechanism allows the scoring thresholds to better align with actual business needs, improving the accuracy of classification. For example, the reassessment of anomalies based on adjusted scoring thresholds can be explained in detail from an implementation perspective. Suppose that during an inspection, an anomaly of water accumulation is detected in the tunnel, with an initial score of 55 points, falling within the 51-75 point range. However, after adjustment, the upper threshold becomes 70 points. The system will reassess the anomaly, considering factors such as the area and location of the water accumulation to determine whether it still falls within the original level range. This reassessment process helps avoid classification deviations caused by threshold adjustments, ensuring a more reasonable determination of the anomaly's level. For example, the process of obtaining adjusted classification data based on the reassessment results can be illustrated from a business logic perspective. During tunnel inspections, if the aforementioned water accumulation anomaly, after reassessment, scores 65 points, still falling within the 51-70 point range, the system will classify it as a Level 3 anomaly, requiring routine maintenance. This process of adjusting and confirming graded data provides a more accurate basis for subsequent resource scheduling, helps optimize the prioritization of inspection work, and improves overall management efficiency. For example, close coordination between each step is crucial in the implementation of the entire process. From initial threshold configuration to dynamic adjustment, and then to reassessment and updating of graded data, each step revolves around the actual needs of tunnel inspections, ensuring that the judgment and handling of anomalies are more scientific and reasonable. This method can reduce misjudgments and resource waste while ensuring inspection quality, providing strong support for tunnel safety management.

[0022] S104. Match the preset response strategy combination according to the adjusted hierarchical data. If there is at least one optional strategy combination, start the weighted optimization algorithm to perform comprehensive calculation from three dimensions: response time, resource utilization and disposal cost, determine the optimal strategy combination, and issue the execution instruction to the relevant execution unit.

[0023] The abnormal situation level is obtained from the adjusted hierarchical data, and matched with preset response strategy combinations to generate a candidate strategy set. Based on the candidate strategy set, all strategy combinations within the set are scored using a weighted optimization algorithm, comprehensively calculating the optimal strategy combination based on three dimensions: response time, resource utilization, and handling cost. If the response time of the optimal strategy combination exceeds a preset response time threshold, a supplementary strategy is obtained from a preset backup strategy library to update the candidate strategy set. Based on the updated candidate strategy set, the weighted optimization score is recalculated to determine the final strategy combination. Execution instructions are extracted from the final strategy combination to generate an instruction sequence. The instruction sequence is then sent to the corresponding execution unit through a preset communication interface.

[0024] For example, in the business scenario of tunnel inspection, the adjusted classification data provides an important basis for subsequent response. After obtaining the event level, the system matches it with preset response strategy combinations to form a candidate strategy set. Suppose that an inspection finds a structural deformation in the tunnel, scoring 88 points, which is classified as a high-level anomaly. The system matches three strategy combinations based on the level, focusing on rapid response, resource conservation, and cost control, respectively. This matching process ensures the targeted nature of the strategies. For example, for the candidate strategy set, the system uses a weighted optimization algorithm to evaluate the performance of each combination. Suppose that the first strategy has a shorter response time but lower resource utilization; the second strategy has higher resource utilization but higher cost; and the third strategy has lower cost but less than ideal response speed. Through weighted calculation, the system comprehensively considers the weights of response speed (40%), resource utilization (30%), and handling cost (30%), and finally selects the first strategy as the optimal combination. This process reflects the balance of strategy selection. For example, if the response speed of the optimal strategy combination is lower than the preset response speed threshold, such as requiring a response time of 30 minutes but actually expecting 40 minutes, the system will extract a supplementary strategy from the backup strategy library. Assuming there is a strategy in the backup library involving calling external emergency resources, the system adds it to the candidate set, updates it, recalculates the weighted score, and finally determines a new combination that includes external resource calls. This mechanism ensures response efficiency. For example, extracting execution instructions from the final strategy combination and generating an instruction sequence is crucial for implementing the strategy. In the aforementioned structural deformation case, the instruction sequence might include dispatching two maintenance devices to a designated location within one hour and notifying monitoring personnel to track it in real time. This sequence is sent to the execution unit through a preset communication interface, ensuring accurate instruction transmission. For example, after the instruction is issued, triggering the response process is the final manifestation of the entire process. Assuming the instruction is issued to the field equipment, the maintenance equipment arrives at the designated point in the tunnel on time and begins repair work, while monitoring personnel confirm the progress through system feedback. This process ensures seamless connection from classification to execution, improving the timeliness of anomaly handling. It should be noted that the implementation of each of the above stages is closely designed around the tunnel inspection requirements. For example, the parameters of the weighted optimization algorithm can be dynamically adjusted based on historical data. If it is found that the response speed weight needs to be increased, the system will automatically adjust it to 50% to adapt to emergency situations. This flexibility provides a more reliable guarantee for tunnel safety. For example, in one possible implementation, updating the backup strategy library is also crucial. The system can periodically analyze historical response cases, adding effective strategies to the library and removing outdated solutions to ensure that the strategies in the library always meet actual needs. This continuous optimization provides more possibilities for responses in complex scenarios.

[0025] S105. The track robot receives the execution command, moves to the nearest station number position of the anomaly point, collects the abnormal image data on site and uploads it to the cloud server. It compares the anomaly type with the pre-stored image database through a deep learning image feature matching algorithm. At the same time, it generates a navigation path based on the simultaneous localization and map building algorithm, and drives the robot to move to the anomaly point a second time.

[0026] The track-mounted robot receives execution commands, extracts relevant information about the abnormal location and station number from these commands, generates task initialization data, and transmits it to the internal control module of the track-mounted robot to receive a task start signal. Based on the task start signal, the track-mounted robot moves to the station number closest to the abnormal situation, uses built-in sensors to collect on-site environmental data, generates preliminary environmental perception data, and transmits it to the cloud server. The image acquisition module takes multi-angle photos of the abnormal location, generates an on-site image dataset, uploads it to the cloud storage unit, and determines whether the image data clarity meets a preset clarity threshold. If the image data clarity meets the preset clarity threshold, feature matching is performed between the image data and a pre-stored image database using image comparison technology to extract key features of the abnormal type and generate an abnormality classification result. Based on the abnormality classification result, combined with simultaneous localization and mapping (SLAM) technology, a navigation path is generated for the abnormal location, and the path data is transmitted to the track-mounted robot drive unit, driving the robot to perform a secondary movement to the abnormal point.

[0027] For example, in a tunnel inspection scenario, after receiving an anomaly alarm, the response scheduling system generates an execution instruction containing a return-to-base task. The instruction might explicitly require the tracked robot to proceed to station K12+300 within the tunnel to address a suspected crack anomaly. After parsing the instruction, the system extracts the anomaly location as K12+300 and generates task initialization data, including target coordinates and task priority. This data is transmitted wirelessly to the tracked robot's internal control module, triggering a task start signal. Upon receiving the signal, the robot's drive motors start, and it moves along the track to K12+300. Upon arrival, built-in sensors such as lidar and temperature / humidity sensors begin collecting environmental data, such as a temperature of 28°C and humidity of 65%, generating preliminary environmental perception data and uploading it to a cloud server. The cloud server verifies the data integrity, ensuring no missing or incorrect data. In one possible implementation, after receiving the environmental perception data, the cloud server activates the image acquisition module. The tracked robot is equipped with a 1080p high-definition camera, capturing images of the anomaly location from multiple angles, such as the front, side, and top views of the crack, generating a dataset of on-site images. Assuming 10 images are captured at a resolution of 1920x1080 and uploaded to a cloud storage unit, the system uses an algorithm to determine image clarity, setting a clarity threshold of 80%. If all images meet the clarity standard, the process proceeds to the next step. Image comparison technology matches the on-site images with crack features in a pre-stored database, such as crack width, depth, and extension direction, generating anomaly classification results, determining them as "minor cracks" or "serious cracks." For example, based on the anomaly classification results, the system combines simultaneous localization and mapping (SLAM) technology to generate a navigation path. Assuming the crack is located on the left side of the tunnel wall, the system calculates the path from the current location to the crack point, avoiding obstacles such as pipes or cables. The path data includes a 30° turning angle and a straight-line distance of 5 meters. The path data is transmitted to the robot drive unit, driving the robot to move to the crack point with centimeter-level positioning accuracy. After arriving, the robot can perform further detection or marking tasks. It is understandable that the above process ensures a closed-loop operation from instruction parsing to path navigation. Each step is closely integrated. For example, sensor data provides environmental information for image acquisition, image clarity assessment provides reliable input for anomaly classification, and the navigation path ensures the robot accurately reaches the target point. This design improves inspection efficiency, reduces the need for manual intervention, and provides technical support for tunnel safety management.

[0028] S106. After the track robot arrives at the anomaly point, it activates the nearest broadcast node, plays the pre-recorded voice command, and connects to the monitoring cameras around the anomaly point to obtain real-time video stream data. The data is then transmitted to the command center through a real-time communication protocol to establish a multi-party video communication channel.

[0029] After arriving at the anomaly location, the track-based robot automatically scans the surrounding environment, identifies the nearest broadcast node, and sends an activation signal via its built-in communication module. This activates the broadcast node to play pre-recorded voice messages, providing initial on-site prompts. Based on the broadcast node's activation status, the track-based robot simultaneously connects to monitoring equipment around the anomaly location, acquires real-time video data, generates an initial video stream, and checks its integrity. If the initial video stream is complete, it compresses the video data using a preset communication protocol, generating a compressed video stream, which is then uploaded to the command center. The command center receives the compressed video stream, decompresses it to generate restored video data, and connects it to the video interaction system, establishing a multi-party video interaction channel. Based on the connection status of the video interaction channel, the command center generates command data for on-site coordination needs and sends it to the track-based robot via the communication protocol, updating task execution parameters. After receiving the command data, the track-based robot parses the task execution parameters, combines them with the real-time video data from the anomaly location, adjusts its position or acquisition angle, and generates an updated on-site information stream. This updated on-site information stream is continuously transmitted from the track-based robot to the command center, completing the closed-loop processing of command transmission and on-site coordination.

[0030] For example, after arriving at an anomaly location, the track-based robot automatically scans the surrounding environment and identifies the nearest broadcast node, providing an efficient on-site interaction method. For instance, in a tunnel inspection scenario, after reaching a suspected seepage point, the robot scans for broadcast nodes within a 5-meter radius, locking onto the nearest node with coordinates X:120, Y:50. The robot sends an activation signal via its built-in 4G module, triggering the node to play a pre-recorded 10-second message: "Attention, an anomaly detected, do not approach." This method ensures that on-site personnel receive timely warning information. It should be noted that the activation of the broadcast node relies on the robot's precise positioning capabilities, typically based on an environmental map generated by LiDAR, with an accuracy of up to 0.1 meters, ensuring accurate signal transmission. Simultaneously, the robot connects to monitoring equipment around the anomaly location to acquire real-time video data and generate an initial video stream. For example, a fixed camera installed inside the tunnel, covering the anomaly location, can be connected to by the robot via Wi-Fi, acquiring a 1280x720 resolution video stream at 30fps. Preferably, integrity assessment is based on timestamp verification to confirm that the video stream covers the entire abnormal location, providing reliable data for subsequent analysis. Video data compression is a crucial step to reduce transmission bandwidth requirements. For example, the robot uses the H.264 encoding protocol to compress the initial video stream to one-third of its size, generating a compressed video stream with a file size reduced from 100MB to approximately 30MB. The compressed video stream is uploaded to the command center via a 5G network at a transmission rate of 50Mbps, ensuring real-time data arrival. Upon receiving the data, the command center decompresses and restores the video data, then connects it to the video interaction system. For example, the command center uses dedicated software to decompress the data and generate a clear video stream, which is displayed on a multi-screen interactive interface for remote experts to analyze the seepage point situation in real time. The command center generates instruction data based on the video interaction system and updates the robot's task parameters. For example, if experts discover that the seepage point is located at the top of the tunnel, the robot's camera angle needs to be adjusted to a 45-degree elevation angle to re-acquire video. The instruction data is sent via an encrypted protocol, containing specific parameters such as "adjust angle: 45 degrees, acquisition duration: 15 seconds". After parsing the instructions, the robot adjusts its robotic arm angle and regenerates the on-site information stream, maintaining a resolution of 1280x720 to cover the details of the seepage point. Preferably, the information stream is transmitted via the Real-Time Streaming Protocol (RTSP) to maintain low-latency interaction with the command center, with latency controlled within 200 milliseconds. Through the updated on-site information stream, the robot continuously maintains data interaction with the command center. For example, based on the 15-second video stream collected by the robot according to the instructions, the flow rate and range of the seepage point are clearly displayed, and the command center determines the seepage level accordingly and generates further handling instructions. This closed-loop processing method relies on the efficiency of the video interaction channel to ensure real-time synchronization between instructions and on-site data. It should be noted that stable communication between the robot and the command center depends on the 5G base station deployed in the tunnel, with a signal coverage rate of 95%, ensuring the continuity of data transmission.

[0031] S107. During the anomaly handling process, the message queue telemetry transmission protocol is used to transmit command data, and all message transmission records are stored in the preset data analysis module.

[0032] The service quality level (SQL) is determined based on the severity of the anomaly. A SQL of Level 2 corresponds to an anomaly of Level 1 or 2, while a SQL of Level 1 corresponds to an anomaly of Level 3 or 4. Based on the SQL, command data is transmitted using a pre-defined message queue telemetry transmission protocol. Specifically: When the SQL is Level 2, a dedicated transmission path is established, and command data for the anomaly event is distributed along this path. Distribution status information is obtained to determine if the command data has reached the target node. If the command data has not reached the target node, a new transmission request is generated, and the undelivered command data is distributed through a backup transmission channel to obtain new distribution status information. When the SQL is Level 1, a shared transmission channel is generated, and command data is distributed according to pre-defined message routing rules, with distribution status information obtained. The distribution status information is integrated into a transmission record, which is then stored in a pre-defined data analysis module.

[0033] For example, in the scenario of handling anomalies in a tracked robot, the instruction transmission mechanism for events of different levels can be managed hierarchically through a pre-defined message queue telemetry transmission protocol. The core of the message queue telemetry transmission protocol lies in the structured processing of instruction data and status information to ensure the orderly and reliable transmission of information. In the anomaly handling scenario, abnormal situations are classified into defined levels. A level 1 event might be a sudden structural crack in a tunnel, a level 2 event might be an equipment malfunction, while level 3 and 4 events might be minor deviations or non-emergency alarms during routine inspections. For example, regarding the priority level of instructions, the system automatically allocates resources based on the event level. For level 1 and 2 events, the quality of service level is set to 2, meaning a higher transmission priority and dedicated path are required. Assuming a level 1 event, where a crack is found in the tunnel, and the instruction data is "immediately stop all equipment in the area," the system will generate a dedicated transmission path, prioritizing bandwidth resources to ensure the instruction is delivered in the shortest possible time. If the path is established, the instruction will be distributed to the target node, such as the tracked robot or field control equipment, and the distribution status information will be fed back in real time. If an instruction fails to be delivered, for example due to transmission interruption caused by signal interference, the system will activate a message retention mechanism, regenerate the delivery request, and switch to a backup dedicated channel to ensure that the instruction is ultimately delivered to the target node. For example, for instruction data for Level 3 and Level 4 events, with a service quality level of 1, the system will distribute it through a shared transmission channel. Suppose that in a Level 3 event, a minor dust accumulation is detected during inspection, and the instruction data is "Start cleaning mode." The system will send the instruction to the robot through the shared channel according to preset message routing rules. Although this method has a lower priority, it is sufficient to meet the needs of non-emergency scenarios while saving dedicated resources. Distribution results are recorded in real time to confirm whether the instruction has been delivered. Specifically, the integration and storage of delivery records is a crucial part of the entire process. The system will summarize the distribution status information of each level of event into delivery records; for example, the delivery time for a Level 1 event instruction is 5 seconds, and the delivery time for a Level 3 event instruction is 10 seconds. These records will be stored in the data analysis module for subsequent traceability and optimization of transmission strategies. This layered transmission mechanism not only ensures the priority delivery of emergency instructions but also guarantees the stable transmission of routine instructions, improving overall efficiency in emergency response scenarios. In one possible implementation, the design of dedicated transmission paths and shared transmission channels fully considers the complex communication environment within tunnels. Dedicated paths can prioritize base stations with high signal coverage, while shared channels dynamically allocate remaining bandwidth resources to ensure maximum resource utilization. This approach effectively improves the reliability of instruction transmission, especially in emergency situations, enabling rapid response to on-site needs.

[0034] S108. Obtain indicator data, including response time and handling success rate, from message transmission records, dynamically adjust the abnormal situation level, and use log analysis stack to statistically evaluate the alarm push effect, optimize the push strategy, and regularly update the deep Q-learning network parameters based on data accumulation, and continuously improve the response strategy through offline training.

[0035] Obtain the response time-consuming and disposal success rate data from the message passing record, calculate the average response time-consuming and the success rate distribution, and obtain the real-time performance indicators of abnormal events. If the response time-consuming in the real-time performance indicators is higher than the preset time-consuming threshold, then adjust the abnormal condition level according to the correlation between the time-consuming and the success rate, and generate new classification data. Among them, the preset time-consuming threshold is determined based on historical data statistics and safety standards: for level 1 events, take the 80th percentile value of the response time-consuming of similar events (such as structural cracks) in the past 6 months (such as 0.3 seconds), and calibrate it according to the 'emergency event response time limit' requirement in the 'Safety Regulations for Highway Tunnel Operations'; the thresholds for level 2 / 3 / 4 events are 0.5 seconds, 1.0 seconds, and 2.0 seconds respectively, which are also set by combining historical percentile values with industry standards and recalculated and updated quarterly according to new data. The correlation between the time-consuming and the success rate is a negative correlation. For example, through Pearson coefficient calculation, the correlation coefficient r = -0.7, that is, for every 0.1 second increase in the time-consuming, the success rate decreases by an average of 5%, etc. The rule for adjusting the abnormal condition level is: if the real-time response time-consuming t > the preset threshold t0, and the corresponding disposal success rate s < the preset success rate threshold s0 (s0 for level 1 = 95%, s0 for level 2 = 90%), then trigger a level adjustment; the specific adjustment method: when t > t0 and s < s0, if the current event is a level 2 event, then upgrade it to level 1; if it is a level 3 event and t > 1.2×t0, then upgrade it to level 2; after adjustment, generate new classification data, and record the levels, time-consuming, success rate, and trigger conditions before and after the adjustment. Extract the alarm push effect indicators from the push log data through log analysis stack, analyze the coverage rate and response rate of the push strategy, and obtain the optimized push strategy configuration. According to the optimized push strategy configuration, update the execution parameters of the alarm push and generate new push log data. Obtain the push log data and classification data within the past three months, extract the abnormal event characteristics, and obtain the offline training dataset. Through the deep Q-learning algorithm, update the network parameters based on the offline training dataset and generate an optimized response strategy model. According to the optimized response strategy model, adjust the push strategy configuration to obtain new business execution parameters. Specifically, the optimized response strategy model is a deep Q network (DQN) model, whose input is abnormal feature data (such as type, location, current level), real-time performance indicators (response time-consuming, success rate), and the output is the adjustment threshold of the quantization scoring model (such as the correction value of the scoring threshold for level 1 / level 2 events); the network structure includes 3 fully connected layers (the hidden layer dimensions are 128, 64, 32 respectively), the activation function uses ReLU, and the output layer uses linear activation to output continuous threshold adjustment amounts.The process of generating a model using the deep Q-learning algorithm is as follows: The state space includes the current anomaly level, historical average response time, and the success rate of handling the last three similar events; the action space includes the adjustment step size of the scoring threshold (e.g., ±5 points, with each adjustment not exceeding 20% ​​of the original threshold); the reward function is R = α × (1 - response time / threshold) + β × success rate (α = 0.6, β = 0.4, reward range [-1, 1]); the training process involves using a 3-month offline dataset (containing 5000+ anomaly event records), employing an ε-greedy strategy (ε linearly decays from 0.9 to 0.1) to select actions, updating the target network every 1000 steps, minimizing the Q-value loss using the Adam optimizer (learning rate 0.001), iteratively training until the loss function (MSE) stabilizes below 0.01, ultimately generating a converged response policy model.

[0036] For example, in the scenario of handling abnormal events during tunnel inspection, obtaining response time and success rate data from message transmission records is a crucial step in optimizing system performance. The average response time can be determined by analyzing the duration from sending to receiving each instruction; assuming the records show an average time of 0.5 seconds over a certain period, the success rate distribution reflects the handling effectiveness of different levels of abnormal events, such as a 98% success rate for level 1 events and 90% for level 4 events. These real-time performance metrics provide a basis for subsequent adjustments. If the time exceeds a preset threshold by 0.3 seconds, the system will analyze the correlation with success rate data and dynamically adjust the abnormal event classification threshold, such as elevating some level 2 events to level 1 to ensure faster response and generating new classification rule data. For example, by extracting alarm push effectiveness metrics from push logs using the ELK stack, the coverage and response rate of the push strategy can be analyzed. Coverage reflects whether alarms reach all relevant nodes, while the response rate measures the actual response rate of nodes to alarms. Suppose the analysis finds a coverage rate of 95% but a response rate of only 80%, the system will optimize the push strategy configuration, such as increasing the push frequency or adjusting the push channel, generating new execution parameters, and updating the log data. This approach effectively improves the reach and feedback efficiency of alerts. For example, acquiring push logs and hierarchical rule data from the past three months, and extracting abnormal event features to form an offline training dataset, forms the basis for further optimization. Features may include information such as event type, time period, and node location. Using a deep Q-learning algorithm, the system updates network parameters based on the dataset, generating an optimized response strategy model. If the model training finds that response times for nighttime events are longer, the system will adjust its strategy, prioritizing more resources for nighttime tasks. This method allows the system to better adapt to real-world scenario requirements. For example, adjusting the abnormal hierarchical rules and push strategy configuration based on the optimized response strategy model can generate new business execution parameters. Specifically, if the model shows a low response rate for level 3 events, the system may increase the push priority in some scenarios and optimize push path selection to ensure rapid instruction delivery. This adjustment significantly improves overall processing efficiency, especially under resource constraints. In one possible implementation, the combined application of the above methods can form a complete optimization loop. From monitoring real-time performance metrics to constructing the offline training dataset, and then to updating the strategy model and adjusting parameters, each step is closely aligned with the tunnel inspection scenario. Suppose an analysis reveals that the response time to a Level 1 seepage event is extended to 0.8 seconds due to signal issues. The system will then review historical data and, based on model recommendations, prioritize the use of base stations with more stable signals for command transmission. This closed-loop mechanism continuously improves the system's adaptability and reliability, providing strong support for rapid response to abnormal events.

[0037] Reference Figure 2This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 2 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor in this computer is designed to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device is used to store data such as on-site anomaly image data and message transmission records. The network interface of the computer device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the personalized recommendation method based on real-time user behavior of any of the above embodiments.

[0038] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0039] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A tunnel anomaly detection and response method based on a track robot, characterized in that, The method includes: A track-based robot performs timed inspections along a pre-defined track, collecting panoramic video stream data from a camera with a pre-defined resolution. This video stream is transmitted in real-time to an edge computing node, where a lightweight AI vision model performs anomaly detection and analysis. If anomalies are detected, abnormal feature data is extracted, a structured message is generated, and this message is input into a pre-defined quantization scoring model to calculate a score. The score determines the anomaly level, which has four levels with priority from highest to lowest. A deep Q-learning network dynamically adjusts the scoring threshold of the pre-defined quantization scoring model, re-classifying the anomalies based on the adjusted thresholds. Finally, a pre-defined response strategy combination is matched to the adjusted classification data. If at least one alternative strategy combination exists, a weighted optimization algorithm is activated. The system comprehensively calculates the optimal strategy combination based on three dimensions: response time, resource utilization, and handling cost, and issues execution instructions to relevant execution units. The tracked robot receives the instructions, moves to the nearest station to the anomaly point, collects anomaly image data from the site, and uploads it to the cloud server. A deep learning image feature matching algorithm is used to compare the data with a pre-stored image database to confirm the anomaly type. Simultaneously, a navigation path is generated based on a simultaneous localization and mapping (SLAM) algorithm, driving the robot to move a second time to the anomaly point. Upon reaching the anomaly point, the tracked robot activates the nearest broadcast node, plays pre-recorded voice instructions, and connects to surrounding surveillance cameras to acquire real-time video stream data. This data is then transmitted to the command center via a real-time communication protocol, establishing a multi-party video communication channel. During anomaly handling, a message queue telemetry transmission protocol is used for instruction data transmission, and all message transmission records are stored in a pre-set data analysis module. We extract metrics data, including response time and success rate, from message transmission records, dynamically adjust the level of abnormal conditions, and use log analysis stacks to statistically evaluate the effectiveness of alarm pushes, optimize push strategies, and regularly update the parameters of the deep Q-learning network based on data accumulation, continuously improving response strategies through offline training.

2. The tunnel anomaly detection and response method based on a track robot according to claim 1, characterized in that, The track-mounted robot performs timed inspection tasks along a preset track, and is equipped with a camera of preset resolution to collect panoramic video stream data inside the tunnel. The acquired video stream is transmitted in real time to an edge computing node, where a lightweight artificial intelligence vision model is used to perform anomaly detection and analysis on the video stream, obtaining anomaly detection results, including: The track robot performs inspections along a preset track. During the inspection, it collects panoramic video data of the tunnel environment through a camera with a preset resolution mounted on the track robot, obtains raw video stream information, and transmits it to an edge computing node for preliminary storage and processing to obtain basic video data. Based on the basic video data, a lightweight artificial intelligence visual model is used to decompose the video stream into frames, extract the image features of each frame, and determine the key region information in the image based on the image features. For information in key areas, an anomaly detection process is executed to obtain anomaly detection results.

3. The tunnel anomaly detection and response method based on a track robot according to claim 1, characterized in that, If an anomaly is detected in the anomaly detection results, the anomaly feature data is extracted, a structured message is generated, and the structured message is input into a preset quantitative scoring model to calculate a score. The level of the anomaly is determined based on the score, where there are four levels of anomalies, with level one being the highest and level four being the lowest processing priority, including: If an anomaly is detected in the anomaly detection results, the anomaly feature data containing the anomaly image, station number, and time stamp is extracted and a structured message is generated. The structured message is input into the preset quantitative scoring model to calculate a score ranging from 0 to 100. The level of the abnormal situation is determined based on the preset scoring threshold and score value, and the classification data of the abnormal situation is obtained. A scoring threshold is set for every 25 points.

4. The tunnel anomaly detection and response method based on a track robot according to claim 1, characterized in that, The application of a deep Q-learning network dynamically adjusts the scoring threshold of a preset quantization scoring model, and reclassifies abnormal situations based on the adjusted scoring threshold to obtain adjusted classification data, including: The preset quantitative scoring model is analyzed by using a deep Q-learning network to obtain the initial configuration data of the scoring threshold and determine the baseline range of the scoring threshold. Based on the initial configuration data, a dynamic adjustment mechanism is used to update the scoring threshold in real time to obtain the adjusted scoring threshold; Based on the adjusted scoring thresholds, abnormal situations are re-scored; Based on the results of the re-scoring, obtain the adjusted classification data for the abnormal situations.

5. The tunnel anomaly detection and response method based on a track robot according to claim 1, characterized in that, The step involves matching a preset response strategy combination with the adjusted hierarchical data. If at least one alternative strategy combination exists, a weighted optimization algorithm is activated to comprehensively calculate the optimal strategy combination based on three dimensions: response time, resource utilization, and processing cost. The execution instructions are then sent to the relevant execution units, including: Obtain the abnormal situation level from the adjusted hierarchical data, match it with the preset response strategy combination, and generate a candidate strategy set; Based on the candidate strategy set, all strategy combinations within the set are scored. The scoring is obtained by using a weighted optimization algorithm to comprehensively calculate the optimal strategy combination from three dimensions: response time, resource utilization, and disposal cost. If the response time of the optimal strategy combination is greater than the preset response time threshold, then a supplementary strategy is obtained from the preset backup strategy library to update the candidate strategy set. Based on the updated set of candidate strategies, the weighted optimization score is recalculated to determine the final strategy combination; Extract execution instructions from the final strategy combination to generate an instruction sequence; The instruction sequence is sent to the corresponding execution unit through a preset communication interface.

6. The tunnel anomaly detection and response method based on a track robot according to claim 1, characterized in that, The tracked robot receives execution commands, moves to the nearest station position to the anomaly point, collects on-site anomaly image data and uploads it to the cloud server, compares it with a pre-stored image database using a deep learning image feature matching algorithm to confirm the anomaly type, and simultaneously generates a navigation path based on a simultaneous localization and mapping (SLAM) algorithm, driving the robot to move a second time to the anomaly point, including: The track robot receives the execution command, extracts relevant information about the abnormal location and station number from the execution command, generates task initialization data, and transmits it to the internal control module of the track robot to obtain the task start signal; Based on the mission start signal, the mobile track robot moves to the nearest station to the abnormal situation, uses built-in sensors to collect on-site environmental data, generates preliminary environmental perception data, and transmits it to the cloud server. An image acquisition module is used to take pictures of the abnormal location from multiple angles, generate a dataset of on-site images, upload it to the cloud storage unit, and determine whether the clarity of the image data reaches the preset clarity threshold. If the image data's clarity reaches a preset clarity threshold, then feature matching is performed with a pre-stored image database using image comparison technology to extract key features of the anomaly type and generate anomaly classification results. Based on the anomaly classification results, and combined with simultaneous localization and mapping (SLAM) technology, a navigation path is generated for the anomaly location. The path data is then transmitted to the track robot drive unit, which drives the robot to make a secondary movement to the anomaly point.

7. The tunnel anomaly detection and response method based on a track robot according to claim 1, characterized in that, After the orbital robot reaches the anomaly point, it activates the nearest broadcast node to play pre-recorded voice commands. Simultaneously, it connects to surveillance cameras around the anomaly point to acquire real-time video stream data, which is then transmitted to the command center via a real-time communication protocol, establishing a multi-party video communication channel, including: After the track robot arrives at the abnormal location, it automatically scans the surrounding environment, identifies the nearest broadcast node, and sends an activation signal through its built-in communication module to drive the broadcast node to play pre-recorded voice information, thus providing initial on-site prompts. Based on the activation status of the broadcast node, the track robot synchronously connects to the monitoring equipment around the abnormal point, obtains real-time video data, generates an initial video stream, and determines whether the video stream is complete. If the initial video stream is complete, the video data is compressed using a preset communication protocol to generate a compressed video stream, which is then uploaded to the command center. The command center receives compressed video streams, decompresses them to generate restored video data, connects to the video interaction system, and builds a multi-party video interaction channel. Based on the connection status of the video interaction channel, the command center generates instruction data according to the on-site coordination needs, and sends it to the track robot through the communication protocol to update the task execution parameters; After acquiring the instruction data, the track robot analyzes the task execution parameters, combines them with real-time video data from abnormal locations, adjusts its own position or acquisition angle, and generates an updated on-site information stream. Through the updated on-site information flow, the track robot continuously transmits information to the command center, completing the closed-loop processing of instruction transmission and on-site coordination.

8. The tunnel anomaly detection and response method based on a track robot according to claim 1, characterized in that, During the anomaly handling process, a message queue telemetry transmission protocol is used for instruction data transmission, and all message transmission records are stored in a preset data analysis module, including: The service quality level is determined based on the level of the abnormal status. When the level of the abnormal status is level 1 or 2, the corresponding service quality level is level 2. When the level of the abnormal status is level 3 or 4, the corresponding service quality level is level 1. Based on the quality of service level, command data is transmitted via a preset message queue telemetry transmission protocol, wherein: When the service quality level is 2, a dedicated delivery path is established. Based on the dedicated delivery path, instruction data for abnormal events is distributed, and distribution status information is obtained to determine whether the instruction data has been delivered to the target node. If the instruction data has not been delivered to the target node, a new delivery request is generated. For the undelivered instruction data, the instruction data is distributed through the backup transmission channel to obtain new distribution status information. When the service quality level is 1, a shared transmission channel is generated, and instruction data is distributed and distribution status information is obtained through preset message routing rules. The distribution status information is integrated into a transmission record, and the transmission record is stored in a preset data analysis module.

9. The tunnel anomaly detection and response method based on a track robot according to claim 1, characterized in that, The process involves acquiring metric data, including response time and success rate, from message transmission records, dynamically adjusting the anomaly level, statistically evaluating the effectiveness of alarm push notifications using log analysis stacks, optimizing the push strategy, and periodically updating the deep Q-learning network parameters based on data accumulation. This is achieved through offline training to continuously improve the response strategy, including: Retrieve response time and success rate data from message transmission records, calculate the average response time and success rate distribution, and obtain real-time performance indicators for abnormal events; If the response time in the real-time performance metrics exceeds the preset time threshold, the abnormal situation level will be adjusted based on the correlation between time and success rate, and new graded data will be generated. By analyzing the log stack, alarm push performance metrics are extracted from push log data. The coverage and response rate of the push strategy are analyzed to obtain the optimized push strategy configuration. Based on the optimized push strategy configuration, update the execution parameters of the alarm push and generate new push log data; Obtain push log data and tiered data from the past three months, extract abnormal event features, and obtain an offline training dataset; By using the deep Q-learning algorithm, network parameters are updated based on the offline training dataset to generate an optimized response strategy model; Based on the optimized response strategy model, the push strategy configuration is adjusted to obtain new business execution parameters.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

Images