Tunnel intelligent monitoring system and method based on data acquisition stage processing

Abstract

The invention discloses a tunnel intelligent monitoring system and method based on data acquisition and grading processing, and the method comprises the steps: carrying out the preprocessing of visual data, and obtaining clear structural data; the data size of the clear structured data is judged based on the grading threshold value, the stage of the data size of the current structured data is judged, and the structured data is processed according to the stage of the data size of the current structured data; when the stage is a large sample stage, clear structured data is transmitted to a large model data processing module for deep analysis; and processing clear structured data by adopting a multi-level and multi-dimensional data analysis framework constructed based on a large model, analyzing the data to obtain a maximum deformation condition in a maximum strain region, and performing risk early warning after comparing the obtained deformation condition with an early warning threshold. A hierarchical data layout transmission mode and a trigger type small sample incremental learning mechanism are adopted, and a mode of combining machine vision and a large model is adopted, so that dynamic learning of the model is carried out, and the limitation in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel engineering safety monitoring, and in particular to a tunnel intelligent monitoring system and method based on data acquisition and hierarchical processing. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] In the construction and operation of tunnels in karst terrain, due to factors such as unstable geological structures, the development of karst caves, groundwater, and surrounding rock deformation, traditional detection methods rely heavily on manual empirical analysis and simple threshold alarms. These methods are unable to meet the dynamic changes in the complex terrain of karst tunnels and enable real-time data transmission and analysis. The rapid development of artificial intelligence and machine vision technologies has provided new options for automated monitoring systems for tunnel in-hole engineering. In particular, considering the synchronization and accuracy of data collection and processing during the construction of complex karst tunnels, the application of machine vision data acquisition and large-scale model data processing in engineering construction has become a trend.

[0004] Traditional monitoring systems struggle to effectively integrate data collected by various monitoring equipment modules (such as displacement, stress, temperature and humidity, and gas concentration sensors), hindering comprehensive risk analysis and resulting in low data processing efficiency. Traditional monitoring system algorithms are inadequate for modeling complex nonlinear relationships (such as the interplay between surrounding rock and multiple influencing factors in complex karst tunnels), resulting in high rates of errors and omissions. Traditional intelligent detection systems rely on network transmission for data transmission, making data loss prone to interruptions in tunnel network signals. Furthermore, model updates are often delayed, requiring manual retraining and preventing the ability to dynamically adjust early warning strategies based on accumulated data. Summary of the Invention

[0005] In order to overcome the shortcomings of the above-mentioned existing technologies, the present invention provides a tunnel intelligent monitoring system and method based on hierarchical data acquisition and processing, adopts a hierarchical data layout and transmission method and a triggered small sample incremental learning mechanism, and adopts a combination of machine vision and large models to perform dynamic learning of the model, thereby solving the limitations of the existing technology.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides an intelligent tunnel monitoring system based on data acquisition and hierarchical processing, comprising: A data acquisition module is used to obtain visual data inside the tunnel; the visual data includes original RGB images and real-time video streams; A data preprocessing module, used to preprocess the visual data to obtain clear structured data; A data cache transmission module is used to determine the data volume of the clear structured data based on a classification threshold, determine the stage to which the current data volume of the structured data belongs, and process the structured data according to the stage to which it belongs; when the stage is the large sample stage, the clear structured data is transmitted to the large model data processing module for in-depth analysis; A large model data processing module is used to process the clear structured data using a multi-level and multi-dimensional data analysis framework built based on the large model, analyze the data to obtain the maximum deformation in the maximum strain area, and issue a risk warning based on the comparison of the obtained deformation with the warning threshold; The RISC-V-based chip architecture module is used to integrate the data acquisition module, data preprocessing module, data cache transmission module and large model data processing module architecture into the chip.

[0007] A further technical solution is to pre-process the visual data as follows: filtering the visual data based on the CycleGAN dehazing network, and using an algorithm to correct blur in combination with the inertial measurement unit data, filtering the visual data, and finally converting the filtered data based on the spatial scale of the target to obtain structured data.

[0008] A further technical solution is that the stages include a small sample stage, a medium sample stage and a large sample stage. When the amount of structured data is less than a first threshold, it belongs to the small sample stage; when the amount of structured data is greater than or equal to the first threshold and less than a second threshold, it belongs to the medium sample stage; when the amount of structured data is greater than or equal to the second threshold, it belongs to the large sample stage.

[0009] A further technical solution is to cache the structured data locally during the small sample stage and use the structured data to optimize the local monitoring model.

[0010] A further technical solution is to use a rule engine to determine whether there are any anomalies in the structured data during the sample stage, and trigger an anomaly mark if there are any anomalies.

[0011] As a further technical solution, the multi-level and multi-dimensional data analysis framework divides structured data into original signal layer, feature abstraction layer and semantic understanding layer, and each layer is further divided into three dimensions: space, time and attributes.

[0012] A further technical solution is to divide the structured data into different levels and dimensions, and then establish the deformation propagation dynamics equation, which can be expressed as:

[0013] in, 、 Indicates target and its adjacent targets The displacement, Indicates time, Indicates target and target The mechanical coupling coefficient at position, is the external load influencing factor, represents the time-dependent external activation function.

[0014] In a second aspect, the present invention provides a tunnel intelligent monitoring method based on data acquisition and hierarchical processing, comprising: Acquire visual data inside the tunnel; the visual data includes original RGB images and real-time video streams; Preprocessing the visual data to obtain clear structured data; The data volume of the clear structured data is judged based on the classification threshold, the stage to which the current data volume of the structured data belongs is judged, and the structured data is processed according to the stage to which it belongs; when the stage is the large sample stage, the clear structured data is transmitted to the large model data processing module for in-depth analysis; A multi-level and multi-dimensional data analysis framework based on a large model is used to process the clear structured data. The data is analyzed to obtain the maximum deformation in the maximum strain area. After comparing the obtained deformation with the warning threshold, a risk warning is issued.

[0015] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a tunnel intelligent monitoring method based on data acquisition and hierarchical processing as described in the second aspect.

[0016] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the tunnel intelligent monitoring method based on data acquisition and hierarchical processing as described in the second aspect are implemented.

[0017] One or more of the above technical solutions have the following beneficial effects: The intelligent tunnel monitoring system proposed in this invention processes collected machine vision data in a hierarchical manner, dividing it into three levels based on data volume. When the data volume is less than threshold A, the data is cached locally, and the limited data is used to optimize the local monitoring model to avoid overfitting. When threshold A ≤ data volume < threshold B, a rules engine is integrated to achieve rapid local warnings and reduce computational latency. When the data volume is greater than or equal to threshold B, a remote transmission module is triggered, uploading the data via the network to a large-scale data processing platform for in-depth analysis. As machine vision data is collected, the volume of data is constantly changing. The data grading proposed in this invention improves data processing efficiency and effectively accounts for memory usage at different data volumes, effectively reducing memory usage when data is high.

[0018] The present invention proposes a large model data processing module. The construction of the large model data processing module is based on the open source large model on the market to build a multi-level and multi-dimensional data analysis framework. After dividing the data into different levels and dimensions, the collected multi-source data are coupled to avoid independent analysis of the data and the error rate of risk prediction.

[0019] The large-scale model data processing module in this invention generates a 3D tunnel model based on existing data and dynamically marks outliers based on machine vision data. Clicking on anomalies allows users to view historical data curves and associated parameters. It also includes a future risk visualization module, which displays historical data curves for that location (e.g., displacement over time). This module supports zooming and panning, allowing users to quickly locate data segments within specific time periods. It also supports exporting reports in CSV or PDF format, including graphs and key statistics (maximum, mean, and variance). Risk levels (red / yellow / green) can be annotated within the 3D tunnel model, indicating future risk hotspots. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0021] Figure 1 It is a flow chart of the tunnel intelligent monitoring method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0023] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0024] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0025] Example 1 like Figure 1 As shown, this embodiment discloses an intelligent tunnel monitoring system based on data collection and hierarchical processing, including: A data acquisition module is used to obtain visual data inside the tunnel; the visual data includes original RGB images and real-time video streams; A data preprocessing module, used to preprocess the visual data to obtain clear structured data; A data cache transmission module is used to determine the data volume of the clear structured data based on a classification threshold, determine the stage to which the current data volume of the structured data belongs, and process the structured data according to the stage to which it belongs; when the stage is the large sample stage, the clear structured data is transmitted to the large model data processing module for in-depth analysis; A large model data processing module is used to process the clear structured data using a multi-level and multi-dimensional data analysis framework built based on the large model, analyze the data to obtain the maximum deformation in the maximum strain area, and issue a risk warning based on the comparison of the obtained deformation with the warning threshold; The RISC-V-based chip architecture module is used to integrate the data acquisition module, data preprocessing module, data cache transmission module and large model data processing module architecture into the chip.

[0026] In this embodiment, the data acquisition module is used to acquire visual data from within the tunnel. It includes a target (made of a specially designed, highly reflective, waterproof material), a machine vision measuring instrument, and an electrical cabinet. The visual data includes high-resolution raw RGB image data and real-time video streams.

[0027] This static machine vision measuring instrument performs multi-scenario monitoring, powered by PoE (PoE), and has a measurement range of 0-400m. The sampling rate is less than 1Hz, and measurement is available in both horizontal and vertical directions. Calibration uses an AI algorithm to automatically correct for rotational angles and distance effects, eliminating the need for distance measurement and leveling, significantly improving monitoring efficiency and reducing human resource waste. The static machine vision measuring instrument achieves millimeter-level accuracy and operates in a temperature range of -40°C to 80°C, ensuring high precision and multi-scenario applications. The instrument incorporates specialized sensors to generate IMU (Inertial Measurement Unit) data, including an accelerometer, gyroscope, and magnetometer. The accelerometer uses a microelectromechanical system (MEMS) to detect the displacement of a mass under acceleration. The gyroscope uses the Coriolis effect or a vibrating structure to measure angular velocity. The magnetometer uses the Hall effect to sense magnetic field strength and determine the device's orientation relative to magnetic north.

[0028] The target is an ordinary target that can be attached to the object to be measured. To reduce the interference of ambient light, a ring-shaped LED array or a near-infrared light source with resistance to dust scattering is used for active fill light. The light intensity is adjusted according to the dust concentration in the tunnel to avoid overexposure or underexposure.

[0029] Several targets are placed on the tunnel structure, and a machine vision measuring instrument is installed in a relatively stable position. The machine vision measuring instrument identifies the target image on the structure. When the structure undergoes planar displacement, the target coordinates change accordingly, thereby measuring the horizontal and vertical displacement of the object. The machine vision measuring instrument inputs 4K@30fps RAW12 data via the MIPI-CSI2 interface, facilitating subsequent chip storage.

[0030] An electrical cabinet is set up to protect the electrical equipment at the storage tunnel monitoring site, including the power supply system (including air switches, leakage protection, and power converters), routers, switches, lightning protection system (optional), and other accessories (terminal blocks, antenna brackets, etc.).

[0031] The present invention is applicable to tunnels constructed using various methods, enabling visual acquisition and hardware module deployment based on the tunnel construction method and site conditions. For example, when using traditional drill-and-blast methods, a static machine vision measuring instrument can be installed on the drilling rig, and the face camera can be retracted to the explosion-proof cabin before blasting. This provides both real-time monitoring and protection from damage during blasting. When using a TBM (full-face tunnel boring machine) for construction, a track-mounted slide can be installed inside the tunnel, equipped with a machine vision measuring instrument for movement, covering the entire tunnel section. A mobile platform can also be installed at the rear of the TBM for real-time monitoring.

[0032] In this embodiment, the data preprocessing module is used to preprocess the visual data to obtain clear structured data. Before caching and transmitting the collected visual data, it is first filtered based on the CycleGAN dehazing network to remove dust or fog and restore a clear image. The module also combines IMU (Inertial Measurement Unit) data obtained by sensors installed within the machine vision measuring instrument. This data collects device motion status information, including real-time measurement of the object's acceleration, angular velocity, and attitude changes. A specialized algorithm is used to correct blur caused by mechanical vibration. The visual data is then filtered locally. Finally, the filtered data is converted based on the spatial scale of the target to obtain preprocessed structured data such as coordinates, dimensions, and shape descriptions, which are then cached and transmitted. In this module, the visual data is filtered based on the CycleGAN dehazing network, and the VPU performs hardware-level filtering on the visual data, generating a YUV420 format data stream for storage, enabling chip integration through a RISC-V architecture.

[0033] In this embodiment, the data cache transmission module is used to judge the data volume of the clear structured data based on the grading threshold, judge the stage to which the data volume of the current structured data belongs, and process the structured data according to the stage to which it belongs; when the stage is the large sample stage, the clear structured data is transmitted to the large model data processing module for in-depth analysis.

[0034] To achieve low-latency, high-reliability data transmission, and taking into account signal blind spots within tunnels, this paper divides data volume into small, medium, and large sample stages, employing different storage and transmission strategies for each stage. Specifically, two thresholds, Threshold A and Threshold B, are set to categorize the three data volume stages.

[0035] Threshold A: The minimum data threshold that triggers the switch from small sample incremental learning (small sample phase) to the medium sample local rule engine (medium sample phase). The minimum sample size required for incremental learning is determined based on preliminary experiments, and the model is verified by cross-validation in terms of sample size. The prediction error is: , ,in For the allowable error.

[0036] Threshold B: The critical data volume threshold that triggers the switch from the local rule engine for medium samples to remote cloud analysis (large sample phase). This threshold is determined based on the single data transmission volume, which should match the network bandwidth and the allowed transmission time:

[0037] in, Indicates the initial critical data volume threshold, which indicates the minimum data volume requirement for triggering data transmission; Indicates network bandwidth, which refers to the maximum rate of the data transmission channel; Indicates the maximum upload time allowed (Upload Time), the system's time limit for a single data transmission; Indicates the single sample transmission cost (Sample Cost), including comprehensive costs such as data compression rate and network resource usage.

[0038] It can be dynamically adjusted according to the network conditions in different tunnels and the allowed transmission time, which is more conducive to applications in different situations.

[0039] (1) When the amount of collected data is less than the minimum sample size required for fine-tuning the local monitoring model, that is, the amount of collected data is less than the threshold A, the structured data is cached locally, and the local monitoring model is optimized using limited data to avoid overfitting. The local monitoring model adopts a multi-source migration-multi-task network (MT-MTNet) and is pre-trained with monitoring data from three similar karst tunnels. The pre-trained model is loaded and the locally pre-processed structured data, such as the displacement value (μ), surrounding rock grade (I~V), groundwater pressure (MPa), and crack parameters calculated by the sliding window, are input. The displacement statistics (μ, σ), geological parameters, and crack parameters are spliced ​​into a mixed feature vector and the proportion of each parameter is adjusted in real time through a dynamic weighting mechanism to train the local monitoring model. This can improve the accuracy of the local monitoring model in locating the displacement-settlement correlation curve and high-risk areas when the sample is medium.

[0040] The monitoring data collected in the early stages of the tunnel was insufficient to directly train a high-precision model. Therefore, the local monitoring model is a pre-trained model based on common monitoring tasks (such as crack identification and settlement prediction). This model is pre-trained using monitoring data from other tunnels (e.g., tunnels with similar geological conditions or structural types) and then transferred to the target tunnel. Based on the collected data, meta-learning is combined with rapid adaptation using a small amount of local data, achieving the goal of rapid adaptation and optimization of the pre-trained model and locally collected data.

[0041] New local data ,in Indicates the number of samples of newly added data, Indicates the The input features (or independent variables) of the data points, Indicates the The local monitoring model selects the algorithm based on fine-tuning of transfer learning and incremental learning (Online Gradient Descent), and defines a weighted loss for the newly added local data, expressed as:

[0042] in, 、 represents the weight coefficient, balancing the new data fitting and the stability of the local monitoring model; Indicates that the model has parameters Lower pair input The predicted output of represents the optimized model parameters, Represents the parameters of the pre-trained base model.

[0043] (2) When threshold A ≤ data volume < threshold B, this is the mid-sample stage. When the amount of collected data reaches the mid-sample stage, it is time to issue an early warning. The rule engine can be combined to achieve fast local early warnings and reduce computing delays. Early warnings in the mid-sample stage are preliminary warnings. Moment (displacement value), dynamically calculate the sliding window mean and variance: ,

[0044] in, represents the sliding window mean, represents the size of the sliding window, Indicates the index of the current time point, Represents the sum index variable ( as the end point), Indicates at a point in time The displacement values ​​collected are represents the sliding window variance, represents the sliding window standard deviation.

[0045] when When the abnormal mark is triggered, a preliminary warning (triggering abnormal mark) can be issued based on the existing data volume, and special attention can be paid after a large sample data volume is reached.

[0046] (3) When the data volume is ≥ threshold B, it is the large sample stage. At this time, the remote transmission module is triggered to upload the structured data to the large model data processing module through the network for in-depth analysis. The remote transmission module provides two transmission modes: wired transmission and wireless transmission. Wired transmission is optical fiber transmission. Anti-corrosion optical fiber is fixed along the side wall or top of the tunnel and connected to the sensor network (such as strain gauges, displacement gauges) through optical terminals. Optical fiber junction boxes are set every 500~1000 meters to facilitate segmented maintenance. This transmission mode is suitable for full-line monitoring of ultra-long tunnels (such as high-speed rail tunnels and cross-mountain tunnels) or far away from urban network signal quality. Wireless transmission is cellular network transmission. Data is directly uploaded to the network through 4G / 5G base stations deployed around the tunnel and underground sections through leaky cables or small base stations. This transmission mode is suitable for urban tunnels with good operator signal coverage. During remote transmission, the H.265 encoding bit rate (4Mbps~20Mbps) is adjusted according to the network status (such as 5G or Wi-Fi), and key area data is transmitted preferentially through the RISC-V NoC bus.

[0047] The large-scale model data processing module enables risk prediction for subsequent construction projects and automatically generates solutions based on the established solution library. Data transmission via signal base stations ensures high-definition video streaming (bandwidth > 100Mbps) with end-to-end latency < 50ms, enabling low-latency data upload to the large-scale model data processing module (large-scale model data processing platform).

[0048] In this embodiment, the large-scale model data processing module is used to process the clear structured data using a multi-level, multi-dimensional data analysis framework built based on the large-scale model to obtain analysis results. When the data volume is ≥ threshold B, the data is transmitted to the large-scale model data processing platform, namely the large-scale model data processing module. This platform directly connects to the data platform transmitted by the machine vision measuring instrument to avoid data loss or inefficiency during secondary data transmission.

[0049] The large-scale model data processing module builds a multi-level, multi-dimensional data analysis framework based on open-source large-scale models. It divides data into three distinct layers: the raw signal layer based on data acquisition source and physical signal form; the feature abstraction layer based on information abstraction level and task relevance; and the semantic understanding layer based on engineering semantics and decision-making requirements. A hierarchical processing chain is employed to process data at each layer. The raw signal layer uses a lightweight CNN to process images and a Kalman filter to process sensor data. The feature abstraction layer uses a 3D ResNet to extract spatial features of tunnel point clouds, and the semantic understanding layer uses a multimodal Transformer to fuse spatiotemporal features and output semantic labels. Each layer contains specific information. Furthermore, it not only focuses on a single dimension but also categorizes information across multiple dimensions, performing decoupled analysis based on the data's geographic structure, temporal evolution, and engineering characteristics into spatial, temporal, and attribute dimensions. After dividing the data into different layers and dimensions, the large-scale model data processing module couples the collected data based on weighted coefficients and establishes the deformation propagation dynamics equation:

[0050] in, 、 Indicates target and its adjacent targets The displacement, Indicates time, Indicates target and target The mechanical coupling coefficient at position, is the external load influencing factor, represents the time-dependent external activation function.

[0051] This module completes the INT8 quantization and layer fusion of YOLOv5-Tiny on the PC side, generating a binary model adapted to RISC-Tensor instructions.

[0052] In this embodiment, the hierarchical warning module establishes an abnormality scoring function based on all collected data in the large model data processing platform, which is expressed as:

[0053] in, Indicates time The abnormality score of , the larger the value, the more significant the system abnormality; represents the number of feature dimensions, Indicates the The predicted value of a feature, Indicates the The actual observed value of the feature, represents the attenuation coefficient, Indicates the current time The interval between the time when the data was generated.

[0054] Set the monitoring deformation threshold. When triggering graded warning ( Dynamic adjustment according to rock mass type) is beneficial to improving efficiency, reducing waste of human resources and improving the accuracy of the monitoring process.

[0055] Alerts are divided into three levels. Level 1 (emergency) is triggered when the anomaly score exceeds the preset safety threshold multiple times (e.g., displacement exceeds the upper limit three times within 10 minutes), or when the deviation between the large-scale model prediction and the measured value continues to widen. Triggering a level 1 alert immediately activates an audible and visual alarm, and the location of the anomaly is highlighted on the local display. If network connectivity is available, an emergency notification is also sent to the administrator's mobile phone and the monitoring platform.

[0056] A Level 2 alert (potential risk) is triggered when a single, but not sustained, occurrence of a comprehensive anomaly score exceeding the limit, or when trend analysis shows a parameter rate of change exceeding the historical average. A Level 2 alert pops up on the local screen, and the anomaly is logged and reported daily for subsequent manual review.

[0057] A Level 3 alert (Long-Term Trend Anomaly) is triggered when the anomaly score fluctuates within a safe range but exhibits a slow, unidirectional change (e.g., sustained subsidence) over a long period (e.g., 72 hours). A Level 3 alert generates a periodic health assessment report, prompting operations personnel to pay close attention.

[0058] In addition, the large-model data processing platform has set up a false alarm suppression mechanism. Through time window verification, it performs repeated detection within the time window for instantaneous abnormal signals, that is, abnormal scores that instantaneously exceed the threshold (for example, confirmation is required only when the same abnormality occurs twice within 5 minutes). The manual confirmation interface supports on-site personnel to manually mark false alarms through the touch screen or mobile terminal, and the system automatically learns and corrects the rule weights.

[0059] Future Risk Visualization Module: The large-scale model data processing platform uses a lightweight 3D engine or professional BIM software (such as Revit) to analyze the tunnel structure design files provided by the project and convert them into an interactive OpenGL model. After generating the 3D tunnel model, abnormal points are dynamically marked on the 3D tunnel model based on abnormal signals obtained from the large-scale model. Click to view historical data curves and associated parameters. To visualize future risks, the visualization panel displays historical data curves for each point (such as displacement over time), supports zooming and panning, and can quickly locate data segments within specific time periods. Reports can be exported in CSV or PDF formats, including curve graphs and key statistics (maximum, mean, and variance). Risk levels (red / yellow / green) can also be annotated on the 3D tunnel model, indicating future risk hotspots.

[0060] In this embodiment, the large-scale model data processing platform integrates industry standards, engineering cases, and expert experience to build a solution library, formulating standardized response strategies (e.g., crack repair solutions, grouting reinforcement processes). Based on feedback from actual treatment results, the solution library is continuously optimized (e.g., prioritizing a leak treatment solution as its efficiency improves).

[0061] The system matches and recommends solutions, selecting the optimal solution from a library based on the risk type (e.g., structural deformation, water leakage), warning severity (Level 1 / 2 / 3), and site conditions (constructability, resource availability) indicated by the large-scale model data platform. It also supports manual intervention to adjust priorities (e.g., selecting temporary support solutions when the construction schedule is tight). The risk warning module exchanges data via shared memory, and the main core executes the decision rules.

[0062] In this embodiment, the tunnel intelligent monitoring system also includes a RISC-V-based chip architecture module. This module, based on the RISC-V reduced instruction set (RISC-V) instruction set, integrates the aforementioned module architecture into the chip. The core layer of the hardware utilizes a 64-bit C906-based RISC-V processor. The main core is responsible for scheduling and system management of each module, while the real-time core utilizes an E24 processor to handle sensor interrupts and provide low-latency responses. Specifically, the RISC-V-based chip design integrates the following modules: a data acquisition module, responsible for collecting raw visual data from external sensors or other data sources; a data preprocessing module, which preprocesses the collected visual data to improve data quality; a data cache and transmission module, which provides temporary storage for preprocessed data and is responsible for efficiently transmitting the preprocessed data to the large model data processing module; and a large model data processing module, which supports complex large model data processing tasks.

[0063] The visual processing unit and large model data processing acceleration module respectively use integrated SIMD vector instruction extensions and deployed Tensor instruction extensions to support operations such as image filtering and geometric correction. They directly connect to the camera's MIPI interface and support INT8 quantized inference to process crack recognition models. The interface layer architecture uses a dual-channel DMA controller to directly transfer camera data to memory, avoiding CPU load. The local cache occupies 2MB of shared L2 cache and is divided into an image preprocessing area and a model parameter area for local monitoring model pre-training. Based on the RISC-V reduced instruction set, the data acquisition module, data preprocessing module, data cache transmission module, large model data processing module, as well as the graded warning and future risk visualization modules and solution intelligent generation module are integrated into a chip architecture to achieve integrated optimization of the detection system.

[0064] Example 2 This embodiment discloses a tunnel intelligent monitoring method based on data collection and hierarchical processing, including: Acquire visual data from inside the tunnel; Preprocessing the visual data to obtain clear visual data; The data volume of the clear visual data is judged based on a grading threshold, the stage to which the current visual data volume belongs is judged, and the visual data is processed according to the stage to which it belongs; when the stage is a large sample stage, the clear visual data is transmitted to a large model data processing module for in-depth analysis; The clear visual data is processed using a multi-level and multi-dimensional data analysis framework constructed based on a large model to obtain analysis results.

[0065] Example 3 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method of embodiment 2 when executing the program.

[0066] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium, a computer-readable storage medium having a computer program stored thereon, which executes the steps of the method of embodiment 2 when the program is executed by a processor.

[0067] The steps involved in the apparatuses of Examples 3 and 4 above correspond to those of Method Example 2. For detailed implementation, please refer to the relevant description of Example 2. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any of the methods of the present invention.

[0068] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0069] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

[0070] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. An intelligent tunnel monitoring system based on data acquisition and hierarchical processing, characterized in that: include: Data acquisition module, used to obtain visual data inside the tunnel; The visual data includes original RGB images and real-time video streams; A data preprocessing module, used to preprocess the visual data to obtain clear structured data; A data cache transmission module is used to determine the data volume of the clear structured data based on a classification threshold, determine the stage to which the data volume of the current structured data belongs, and process the structured data according to the stage to which it belongs; When the stage is a large sample stage, the clear structured data is transmitted to the large model data processing module for in-depth analysis; A large model data processing module is used to process the clear structured data using a multi-level and multi-dimensional data analysis framework built based on the large model, analyze the data to obtain the maximum deformation in the maximum strain area, and issue a risk warning based on the comparison of the obtained deformation with the warning threshold; The RISC-V-based chip architecture module is used to integrate the data acquisition module, data preprocessing module, data cache transmission module and large model data processing module architecture into the chip.

2. The intelligent tunnel monitoring system based on data acquisition and hierarchical processing according to claim 1, characterized in that: The preprocessing of the visual data is specifically as follows: filtering the visual data based on the CycleGAN dehazing network, correcting blur using an algorithm combined with inertial measurement unit data, filtering the visual data, and finally converting the filtered data based on the spatial scale of the target to obtain structured data.

3. The intelligent tunnel monitoring system based on data acquisition and hierarchical processing according to claim 1, characterized in that: The stages include a small sample stage, a medium sample stage, and a large sample stage. When the amount of structured data is less than a first threshold, it belongs to the small sample stage; when the amount of structured data is greater than or equal to the first threshold and less than a second threshold, it belongs to the medium sample stage; when the amount of structured data is greater than or equal to the second threshold, it belongs to the large sample stage.

4. The intelligent tunnel monitoring system based on data acquisition and hierarchical processing according to claim 3 is characterized in that: During the small sample stage, the structured data is cached locally and the local monitoring model is optimized using the structured data.

5. The intelligent tunnel monitoring system based on data acquisition and hierarchical processing according to claim 3 is characterized in that: During the sample stage, the rule engine is used to determine whether there are any anomalies in the structured data. If there are any anomalies, the anomaly mark is triggered.

6. The intelligent tunnel monitoring system based on data acquisition and hierarchical processing according to claim 3, characterized in that: The multi-level and multi-dimensional data analysis framework divides structured data into the original signal layer, the feature abstraction layer and the semantic understanding layer, and each layer is further divided into three dimensions: space, time and attributes.

7. The intelligent tunnel monitoring system based on data acquisition and hierarchical processing according to claim 6, characterized in that: After dividing the structured data into different levels and dimensions, the deformation propagation dynamics equation is established, which can be expressed as: in, 、 Indicates target and its adjacent targets The displacement, Indicates time, Indicates target and target The mechanical coupling coefficient at position, is the external load influencing factor, represents the time-dependent external activation function.

8. A tunnel intelligent monitoring method based on data acquisition and hierarchical processing, characterized in that: include: Acquire visual data from inside the tunnel; The visual data includes original RGB images and real-time video streams; Preprocessing the visual data to obtain clear structured data; Determining the data volume of the clear structured data based on a classification threshold, determining the stage to which the data volume of the current structured data belongs, and processing the structured data according to the stage to which it belongs; When the stage is a large sample stage, the clear structured data is transmitted to the large model data processing module for in-depth analysis; A multi-level and multi-dimensional data analysis framework based on a large model is used to process the clear structured data. The data is analyzed to obtain the maximum deformation in the maximum strain area. After comparing the obtained deformation with the warning threshold, a risk warning is issued.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the tunnel intelligent monitoring method based on data acquisition and hierarchical processing as described in claim 8 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the tunnel intelligent monitoring method based on data acquisition and hierarchical processing as described in claim 8 are implemented.

Images