Tunnel bridge culvert water level detection alarm system and method based on deep learning

Abstract

The invention discloses a tunnel, bridge and culvert water level detection and alarm system and method based on deep learning, and the method comprises the following steps: S1, collecting an image and a capacitance water level signal at the lowest point of a tunnel or an easy water accumulation region, and generating standardized multi-source data; s2, inputting the image into a SegmentAnything model, extracting a water body mask, and calculating an image water level value in combination with scale mapping; s3, carrying out confidence weighted fusion on the image water level and the capacitance signal to obtain a fused water level value; s4, constructing a fused water level time sequence, constructing an SCI Net model, and outputting a predicted water level sequence; s5, generating an alarm label according to the fused water level and the prediction result; s6, LED display and acousto-optic response are controlled according to the alarm label, and an early warning prompt is started when a serious alarm is given; s7, capturing an image in an early warning or serious alarm state, and generating an alarm data packet; and S8, uploading to an urban water level monitoring platform through the Internet of Things module. According to the invention, intelligent identification and remote early warning of the ponding state of the tunnel bridge and culvert are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of water level monitoring, and in particular to a tunnel, bridge and culvert water level detection and alarm system and method based on deep learning. Background Art

[0002] The widespread distribution of tunnels, bridges, and culverts in urban road traffic infrastructure plays a vital role in alleviating ground traffic congestion and improving traffic efficiency. However, these structures are often located in relatively low-lying areas and are highly susceptible to extreme weather events such as heavy rain, typhoons, and flooding, which can lead to the accumulation of rainwater and severe waterlogging. Failure to quickly detect and respond to rising water levels can not only disrupt traffic operations but, in severe cases, endanger human life and property. Therefore, tunnel, bridge, and culvert water level detection and early warning control have become important research areas in the interdisciplinary fields of urban smart transportation, water security, and disaster emergency management.

[0003] In existing technologies, some cities or engineering units have deployed preliminary water level monitoring systems. These systems typically rely on a single physical water level sensor, such as an ultrasonic sensor, an electrode-type liquid level switch, or a capacitive liquid level gauge, installed at the lowest point of a tunnel or near a drainage outlet. These sensors measure the water level by sensing the reflection of the signal or the change in the dielectric constant of the water body, and trigger a simple alarm module, such as an audible buzzer or a red indicator light, when the water level exceeds a preset threshold. However, due to the single location of the sensor, the equipment is susceptible to problems such as silt blockage, corrosion, vibration, or electromagnetic interference. Monitoring accuracy and long-term stability are difficult to guarantee, and false alarms and missed alarms are prone to occur. At the same time, traditional physical sensors lack adaptive data correction capabilities and are unable to determine the validity and confidence level of the measured signals.

[0004] Furthermore, current systems have significant shortcomings in information output and response methods. Most systems only provide local audio and visual alerts, lacking effective visual cues for passing drivers. This is especially true at night or in inclement weather, where a simple flashing red light often fails to attract attention, creating traffic hazards. Furthermore, some systems lack remote transmission and platform-based access capabilities, making it difficult for management units to monitor on-site conditions and implement preventative dispatch and response measures.

[0005] More critically, traditional water-level monitoring systems are primarily passive response systems, issuing alarms only when the set water level is reached, failing to effectively incorporate trend prediction mechanisms. If waterlogging develops rapidly, the alarm may lag behind the actual occurrence of the risk, hindering the development of a smart prevention and control system that combines proactive early warning and proactive risk avoidance. Furthermore, existing systems generally lack image-assisted recording and visual confirmation capabilities, preventing on-site managers from obtaining immediate visual images of flooded areas. This results in insufficient basis for decision-making and hinders the timeliness and relevance of emergency response.

[0006] When it comes to information fusion, current technologies primarily rely on single-modal signal sources, making it impossible to combine image and physical sensor information for multi-source cross-validation. A single sensor misjudgment directly impacts alarm accuracy, reducing system reliability and public trust. Furthermore, the lack of a quality assessment mechanism prevents the system from dynamically adjusting the weighting of various data types in the final judgment based on the reliability of real-time data, resulting in a mechanical and rigid fusion process lacking intelligence.

[0007] The rapid development of artificial intelligence and deep learning technologies, particularly the widespread application of models such as image segmentation and time series prediction in scenarios such as transportation, security, and water services, has opened up new possibilities for building more intelligent water level detection and alarm systems. Some cutting-edge research has attempted to use computer vision methods to automatically identify water levels or analyze flooded areas from video images, but most are limited to static image processing and lack mechanisms for deep integration with real-time sensor data. Furthermore, some studies have used traditional time series models (such as ARIMA and LSTM) for water level trend forecasting. However, these models have limitations in capturing non-stationary and multi-scale variations, and particularly face challenges in real-time inference and deployment efficiency.

[0008] Therefore, how to provide a tunnel and culvert water level detection alarm system and method based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0009] One objective of the present invention is to propose a deep learning-based tunnel and culvert water level detection and alarm system and method. This system leverages the intelligent analysis capabilities of the image segmentation model SegmentAnything and the time series prediction model SCINet, integrating video images with capacitive water level sensor data to construct a confidence-weighted water level fusion mechanism, enabling accurate extraction and dynamic prediction of multi-source water level data. By establishing multidimensional alarm rules and linkage control logic, the system supports graded judgment and response to water level risks. Combining image capture with IoT remote transmission capabilities, it forms a visual, predictive, and networked integrated intelligent monitoring system with the advantages of rapid response, accurate identification, strong adaptability, and remote controllability.

[0010] The deep learning-based tunnel and culvert water level detection and alarm method according to an embodiment of the present invention includes the following steps:

[0011] S1. Install equipment at the lowest point of the tunnel or in areas prone to water accumulation to collect video image data from tunnel bridges and culverts and capacitive water level sensor signals, perform preprocessing, and generate standardized multi-source data.

[0012] S2. Input each frame of the standardized multi-source data into the SegmentAnything model to generate an image mask set, filter out the water area mask, and calculate the image recognition water level value by combining the pixel-elevation mapping relationship of the LED water level scale;

[0013] S3, performing confidence-weighted fusion of the image water level value and the capacitive water level sensor signal to obtain a fused water level value;

[0014] S4. Construct a time series input vector based on the fused water level value and build an SCINet model. Perform trend decomposition and structure reconstruction through the recursive decomposition module and the multi-scale connection module to output the predicted water level value sequence.

[0015] S5. Generate an alarm label based on the fused water level value and the predicted water level value sequence;

[0016] S6. Control the LED water level gauge inside the tunnel and the LED display screen at the tunnel entrance to display corresponding display modes according to the alarm tag. If the alarm tag is in a serious alarm state, activate the tunnel entrance early warning prompt mode;

[0017] S7. When the alarm tag is in the warning or serious alarm state, trigger the camera to capture the image, collect the current image frame, and construct an alarm information data packet;

[0018] S8. Upload the alarm information data packet to the remote city water level monitoring platform through the Internet of Things communication module.

[0019] Optionally, the standardized multi-source data includes normalized video image frames, denoised capacitive water level sensor signals, unified timestamps corresponding to image frames, image enhancement parameter recording information, and filter status identifiers of sensor signals.

[0020] Optionally, the S2 specifically includes:

[0021] S21. Input each image frame in the standardized multi-source data into the image encoder of the SegmentAnything model. The image encoder is an image embedding network built based on the visual Transformer structure, which is used to perform global semantic modeling on the input image frame and output an image embedding representation.

[0022] S22. Input the empty prompt information into the prompt encoder of the SegmentAnything model, configure the prompt encoder to be in an empty input state, adopt a zero-shot segmentation mode, and output an empty prompt embedding representation;

[0023] S23. Input the image embedding representation and the empty prompt embedding representation into the mask prediction head of the SegmentAnything model. The mask prediction head generates a semantic mask set based on the cross-attention mechanism.

[0024] S24. Calculate a color histogram for each candidate mask in the mask set and compare it with a preset water body color histogram:

[0025]

[0026] Among them, S i is the water body similarity score of the i-th candidate mask, i is the mask index, j is the color interval index, K is the number of color intervals, is the water color reference histogram value in the jth color interval, is a candidate mask, is the pixel distribution density of the i-th candidate mask in the j-th color interval, and t is the time index variable;

[0027] S25, extract the pixel coordinates of the lower boundary of the water body mask, and calculate the image recognition water level value by combining the preset pixel-physical height mapping function

[0028] Optionally, the confidence weighted fusion includes fusion input definition, fusion weight calculation, and fusion water level output; the fusion input definition obtains the image recognition water level value at time t and capacitance sensor water level value The fusion weight calculation calculates the image recognition quality score and the capacitive sensor quality score respectively, and determines the fusion weight of the image water level based on the two score values;

[0029] The fused water level output is weighted by the image water level value and the capacitive sensor water level value based on the fusion weight to calculate the final fused water level value:

[0030]

[0031] Among them, H (t) is the fusion water level value, α t is the confidence fusion weight of the image water level, Identify water level values ​​for images, is the water level value of the capacitance sensor, (1-α t ) is the confidence fusion weight of the capacitor water level, and t is the time index variable.

[0032] Optionally, the S4 specifically includes:

[0033] S41. Construct a prediction time series input vector, select the fused water level values ​​of the most recent N consecutive time steps, and record it as X = {x(t-N+1), x(t-N+2), ..., x(t)}, where x(t) is the fused water level value of the current time step, t is the time index variable, and N is the length of the time series;

[0034] S42: Input the time series input vector X into a structured connectivity inference network (SCINet) model. The SCINet model includes a recursive decomposition module and a multi-scale connection module, wherein each module includes multiple convolutional layers, residual connection paths, and an attention fusion structure.

[0035] S43. In the recursive decomposition module, the time series input vector X is decomposed layer by layer, with the input of each layer being the residual term of the previous layer, and the output being the trend term and the new residual term;

[0036] S44. In the multi-scale connection module, the trend components of each layer are input into the fusion structure. The fusion structure includes a channel attention mechanism and a global pooling process. First, the fusion weight is calculated through the channel attention mechanism, and then all trend components are weighted and fused to generate a reconstructed feature vector:

[0037]

[0038] Among them, α i is the attention fusion weight corresponding to the trend component of the i-th layer, W1 is the first set of weight parameter matrices for channel attention mapping, W2 is the second set of weight parameter matrices for channel attention mapping, ReLU(·) is the rectified linear unit activation function, T (i) is the trend component obtained by recursive decomposition at the i-th layer, GAP(·) is the global average pooling operation performed, n is the total number of layers of recursive decomposition, i is the index of the trend component currently being processed, and j is the traversal index used for normalized weight calculation;

[0039] S45, reconstruct the feature vector, perform linear mapping, and output the predicted water level value sequence for the next M time steps in, is the predicted water level value at step i, and M is the prediction step length.

[0040] Optionally, the layer-by-layer decomposition includes a first layer decomposition using the predicted time series input vector as the initial residual term to generate a first layer trend component and a first layer residual component; the first layer residual component is used as the input residual term of the second layer, and the decomposition operation is continued to generate a second layer trend component and a second layer residual component; it is performed iteratively, in the i-th layer, the input is the residual component generated by the i-1-th layer, and the output is the trend component and the residual component of the i-th layer until the preset total number of decomposition layers n is completed, and finally a set of trend components of multiple levels and the residual component of the deepest layer are obtained.

[0041] Optionally, the generation of alarm labels includes alarm rule setting, alarm condition judgment and alarm label output; the alarm rule setting includes static threshold rule, trend growth rule and extreme value prediction rule;

[0042] The alarm condition judgment is based on the alarm rules. If the current fused water level value is greater than the preset water level alarm trigger threshold, it is determined that a static water level alarm is triggered; if there are m consecutive time steps in the predicted water level value sequence that meet the difference between the water level value of the i-th predicted time step and the current fused water level value is greater than the trend increase threshold, it is determined that a trend rising warning is triggered; if any water level value in the predicted water level value sequence is greater than or equal to the future water level extreme risk threshold, it is determined that an extreme value warning mechanism is triggered;

[0043] The alarm tag output generates an alarm tag based on the judgment result. If any one of the alarm condition judgments meets the conditions for triggering the extreme value warning mechanism or triggering the trend rising warning and reaches the highest risk threshold, L(t)=2 is directly output, that is, a serious alarm; if no serious alarm is triggered, but any one of the warning level conditions for triggering the trend rising warning or triggering the static water level alarm is met, then L(t)=1 is output, that is, a warning; if all alarm rules are not triggered or do not meet any risk criteria, then L(t)=0 is output, that is, the current state is normal.

[0044] Optionally, the S6 specifically includes:

[0045] S61. Obtain the alarm tag value of the current time step, and control the display status of the LED water level gauge and the LED display screen at the tunnel entrance based on the alarm tag value;

[0046] S62. When the alarm tag value is L(t)=0, the display screen maintains the standard water level display mode. When the alarm tag value is L(t)=1, the LED display screen is controlled to flash orange or display dynamic graphics to indicate that the current water level is approaching the warning threshold. When the alarm tag value is L(t)=2, the display screen switches to high-frequency flashing red.

[0047] S63, controlling the operating state of the sound and light alarm device based on the alarm tag value, when the alarm tag value L(t) = 0, all the sound and light devices remain off, when the alarm tag value L(t) = 1, the sound and light devices are controlled to work in an intermittent ringing mode, and when the alarm tag value L(t) = 2, the sound and light devices are controlled to ring continuously;

[0048] S64. When the alarm tag value is L(t)=2, the tunnel entrance warning prompt module is activated in conjunction with the tunnel entrance, and the early warning mechanism is turned on at the tunnel entrance to output a prompt of "water accumulation ahead" or "detour recommended".

[0049] Optionally, the S7 specifically includes:

[0050] S71, determine whether the image capture trigger condition is met, when the alarm tag value L(t)∈{1,2}, meets the capture condition, start the capture process;

[0051] S72. Calling a surveillance camera installed inside a tunnel or bridge to collect an image frame at the current time step, where the image frame is denoted as I(t), where t is a time index variable when the alarm occurs;

[0052] S73. Extract the fused water level value corresponding to the current time step and read the alarm tag value;

[0053] S74. Encapsulate the image frame, fused water level value, alarm tag and time index into a structured alarm information data packet.

[0054] The deep learning-based tunnel and culvert water level detection and alarm system according to an embodiment of the present invention includes the following modules:

[0055] The image and sensor acquisition module is used to deploy surveillance cameras and capacitive water level sensors in low-lying areas of tunnels or bridges and culverts, collect video image frames and capacitive water level signals, and generate raw multi-source observation data streams;

[0056] A multi-source data standardization processing module is used to process the original image data and generate standardized multi-source input data;

[0057] The image water level extraction module is used to input the standardized image frame into the SegmentAnything model, generate a semantic mask set, extract the water area mask by filtering the water body features, obtain the water body edge position and combine it with the scale mapping relationship to output the image recognition water level value;

[0058] Confidence-weighted water level fusion module, used to obtain the image recognition water level value and the capacitive sensor water level value to generate a fused water level value;

[0059] The water level trend prediction module is used to construct a time series input vector and input it into the SCINet model, and output the predicted water level value sequence;

[0060] The alarm level determination module is used to output the alarm level label based on the fused water level value and the water level prediction sequence;

[0061] The LED display and sound and light control module is used to control the display status of the LED water level gauge and the LED display screen at the tunnel entrance according to the alarm label, and control the local sound and light alarm device to achieve a graded response. In the event of a serious alarm, the entrance early warning prompt mode is activated;

[0062] The image capture and alarm information packaging module is used to trigger the camera to capture the current image frame when the alarm tag is in the warning or serious alarm state, extract and fuse the water level value and alarm tag information, and combine them to generate a structured alarm information data packet;

[0063] The IoT remote transmission module is used to upload structured alarm information data packets to the urban water level monitoring platform through the IoT communication module.

[0064] The beneficial effects of the present invention are:

[0065] This invention significantly enhances the intelligence and practical value of tunnel and bridge water level detection and early warning response by introducing deep learning technology and a multi-source data fusion mechanism. Compared with traditional water level monitoring systems that rely on a single physical sensor, have a delayed response, and lack image verification capabilities, this invention fuses video images with capacitive water level sensor signals to construct a confidence-weighted fusion model. While ensuring real-time data, it also introduces a quality scoring mechanism that enables the system to dynamically adjust the contribution weight of each source data based on its reliability. This effectively mitigates the interference of single data anomalies on the overall judgment accuracy of the system and improves the accuracy and stability of the fused water level value.

[0066] This invention, for the first time, introduces the Structured Connectivity Inference Network (SCINet) into the field of urban drainage, enabling multi-scale modeling and recursive prediction of water accumulation trends in tunnels, bridges, and culverts. This system not only reflects current water levels but also issues early risk warnings based on trend predictions, establishing a technological transition from "passive detection" to "active prediction." A multidimensional alarm rule system, combining static threshold determination, trend growth analysis, and extreme value prediction mechanisms, makes alarm decisions more comprehensive and sensitive, effectively responding to sudden and rapidly evolving heavy rainfall events.

[0067] In terms of alarm response control, this invention implements hierarchical response logic to achieve coordinated control of the LED water level gauge, the tunnel entrance LED display, and the local sound and light alarm device. Different levels of alarm tags will trigger different levels of prompts and guidance actions, ensuring that on-site personnel and passing vehicles receive effective early warning information in the first place. At the same time, the system automatically activates the image capture mechanism in high-risk situations and constructs a structured alarm information data packet, establishing a connection path between image, water level, and tag data, forming a complete chain of evidence for the event.

[0068] Furthermore, the present invention integrates an Internet of Things (IoT) communication module, enabling remote upload of alarm information and connection to the city's monitoring platform. This supports remote data viewing, analysis, and dispatch, providing accurate, timely, and visual decision-making support for back-end smart traffic management, water regulation, and emergency response. Overall, the present invention boasts outstanding advantages such as intelligent data fusion, forward-looking trend analysis, a hierarchical response mechanism, and practical system deployment. It significantly improves the water level monitoring accuracy and response efficiency of urban drainage channels under extreme weather conditions, and has promising engineering application prospects and social security value. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0070] Figure 1 This is a flow chart of the tunnel and culvert water level detection and alarm method based on deep learning proposed by the present invention;

[0071] Figure 2 This is a schematic diagram of the structure of SCINet trend prediction in the tunnel and culvert water level detection and alarm method based on deep learning proposed in the present invention;

[0072] Figure 3 This is a schematic diagram of the tunnel and culvert water level detection and alarm system based on deep learning proposed in this invention. DETAILED DESCRIPTION

[0073] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0074] refer to Figure 1 and Figure 2 , a tunnel and culvert water level detection and alarm method based on deep learning includes the following steps:

[0075] S1. Install equipment at the lowest point of the tunnel or in areas prone to water accumulation to collect video image data from tunnel bridges and culverts and capacitive water level sensor signals, perform preprocessing, and generate standardized multi-source data.

[0076] S2. Input each frame of the standardized multi-source data into the SegmentAnything model to generate an image mask set, filter out the water area mask, and calculate the image recognition water level value by combining the pixel-elevation mapping relationship of the LED water level scale;

[0077] S3, performing confidence-weighted fusion of the image water level value and the capacitive water level sensor signal to obtain a fused water level value;

[0078] S4. Construct a time series input vector based on the fused water level value and build an SCINet model. Perform trend decomposition and structure reconstruction through the recursive decomposition module and the multi-scale connection module to output the predicted water level value sequence.

[0079] S5. Generate an alarm label based on the fused water level value and the predicted water level value sequence;

[0080] S6. Control the LED water level gauge inside the tunnel and the LED display screen at the tunnel entrance to display corresponding display modes according to the alarm tag. If the alarm tag is in a serious alarm state, activate the tunnel entrance early warning prompt mode;

[0081] S7. When the alarm tag is in the warning or serious alarm state, trigger the camera to capture the image, collect the current image frame, and construct an alarm information data packet;

[0082] S8. Upload the alarm information data packet to the remote city water level monitoring platform through the Internet of Things communication module.

[0083] The deep learning-based tunnel and culvert water level detection and alarm method provided by the present invention can significantly improve the intelligence level and response efficiency of waterlogging monitoring in low-lying areas of urban roads. By collecting images and capacitive water level signals at the lowest point of the tunnel, the system integrates the SegmentAnything model to extract the image water level value, and generates high-precision water level results based on the confidence weighting mechanism and sensor data fusion. The SCINet model further realizes water level trend prediction and identifies risk changes in advance. The system drives the LED display and sound and light device to respond according to the alarm tag, and automatically captures images and uploads them remotely through the Internet of Things module when a serious alarm occurs. It has the advantages of high prediction accuracy, fast response time, and strong information integrity. It is suitable for the prevention and control of waterlogging risks in key flood-prone areas such as tunnels and culverts.

[0084] In this embodiment, the standardized multi-source data includes normalized video image frames, denoised capacitive water level sensor signals, unified timestamps corresponding to image frames, image enhancement parameter recording information, and filter status identifiers of sensor signals.

[0085] This invention uses a standardized multi-source data processing method to fuse image frames, capacitance signals, and their quality information to improve the accuracy and stability of water level detection. Through normalization, denoising, and state labeling, it achieves efficient fusion of multimodal data, providing a reliable foundation for subsequent identification and prediction.

[0086] In this embodiment, S2 specifically includes:

[0087] S21. Input each image frame in the standardized multi-source data into the image encoder of the SegmentAnything model. The image encoder is an image embedding network built based on the visual Transformer structure, which is used to perform global semantic modeling on the input image frame and output an image embedding representation.

[0088] S22. Input the empty prompt information into the prompt encoder of the SegmentAnything model, configure the prompt encoder to be in an empty input state, adopt a zero-shot segmentation mode, and output an empty prompt embedding representation;

[0089] S23. Input the image embedding representation and the empty prompt embedding representation into the mask prediction head of the SegmentAnything model. The mask prediction head generates a semantic mask set based on the cross-attention mechanism.

[0090] S24. Calculate a color histogram for each candidate mask in the mask set and compare it with a preset water body color histogram:

[0091]

[0092] Among them, S i is the water body similarity score of the i-th candidate mask, i is the mask index, j is the color interval index, K is the number of color intervals, is the water color reference histogram value in the jth color interval, is a candidate mask, is the pixel distribution density of the i-th candidate mask in the j-th color interval, and t is the time index variable;

[0093] S25, extract the pixel coordinates of the lower boundary of the water body mask, and calculate the image recognition water level value by combining the preset pixel-physical height mapping function

[0094] The present invention introduces the SegmentAnything model based on the visual Transformer structure to achieve accurate extraction of water areas in images. It adopts a zero-sample segmentation method and can output a high-quality semantic mask set without manual labeling. By jointly modeling the image encoder and the prompt encoder, and combining the color histogram similarity judgment, the real water area is effectively screened out. The image water level value is calculated using the mapping relationship between the lower boundary position of the mask and the scale, realizing stable water level recognition in complex lighting and occlusion environments. This method has the advantages of strong adaptability, high segmentation accuracy, and no need for training, which significantly improves the accuracy and practicality of image water level recognition.

[0095] In this embodiment, the confidence weighted fusion includes fusion input definition, fusion weight calculation, and fusion water level output; the fusion input definition obtains the image recognition water level value at time t and capacitance sensor water level value The fusion weight calculation calculates the image recognition quality score and the capacitive sensor quality score respectively, and determines the fusion weight of the image water level based on the two score values;

[0096] The fused water level output is weighted by the image water level value and the capacitive sensor water level value based on the fusion weight to calculate the final fused water level value:

[0097]

[0098] Among them, H (t) is the fusion water level value, α t is the confidence fusion weight of the image water level, Identify water level values ​​for images, is the water level value of the capacitance sensor, (1-α t ) is the confidence fusion weight of the capacitor water level, and t is the time index variable.

[0099] The present invention uses a confidence-weighted fusion mechanism to effectively combine image recognition water level values ​​with capacitive sensor water level values, dynamically adjusting the fusion weights based on the real-time quality of different data sources. By separately calculating the image recognition quality score and the capacitive signal stability score, the adaptability and robustness of the fusion process are improved, preventing single data anomalies from interfering with the overall judgment. This fusion method not only improves the accuracy of water level measurement, but also enhances the system's reliability in complex environments. It is particularly suitable for working conditions such as rain reflection, sediment obstruction, or signal drift, significantly improving the accuracy and stable operation of the water level detection system.

[0100] In this embodiment, the S4 specifically includes:

[0101] S41. Construct a prediction time series input vector, select the fused water level values ​​of the most recent N consecutive time steps, and record it as X = {x(t-N+1), x(t-N+2), ..., x(t)}, where x(t) is the fused water level value of the current time step, t is the time index variable, and N is the length of the time series;

[0102] S42: Input the time series input vector X into a structured connectivity inference network (SCINet) model. The SCINet model includes a recursive decomposition module and a multi-scale connection module, wherein each module includes multiple convolutional layers, residual connection paths, and an attention fusion structure.

[0103] S43. In the recursive decomposition module, the time series input vector X is decomposed layer by layer, with the input of each layer being the residual term of the previous layer, and the output being the trend term and the new residual term;

[0104] S44. In the multi-scale connection module, the trend components of each layer are input into the fusion structure. The fusion structure includes a channel attention mechanism and a global pooling process. First, the fusion weight is calculated through the channel attention mechanism, and then all trend components are weighted and fused to generate a reconstructed feature vector:

[0105]

[0106] Among them, α i is the attention fusion weight corresponding to the trend component of the i-th layer, W1 is the first set of weight parameter matrices for channel attention mapping, W2 is the second set of weight parameter matrices for channel attention mapping, ReLU(·) is the rectified linear unit activation function, T (i) is the trend component obtained by recursive decomposition at the i-th layer, GAP(·) is the global average pooling operation performed, n is the total number of layers of recursive decomposition, i is the index of the trend component currently being processed, and j is the traversal index used for normalized weight calculation;

[0107] S45, reconstruct the feature vector, perform linear mapping, and output the predicted water level value sequence for the next M time steps in, is the predicted water level value at step i, and M is the prediction step length.

[0108] This paper introduces the structured connection inference network SCINet to model the fused water level value sequence, extracts multi-layer trend information through a recursive decomposition module, and fuses trend features of different scales in a multi-scale connection module to achieve high-precision prediction of future water level changes. The fusion structure adopts a channel attention mechanism and a global pooling method to automatically assign weights to each trend component, effectively enhancing the expressive power of feature reconstruction. The prediction results output the water level values ​​at multiple moments in the future through linear mapping. It has the advantages of high sequence prediction accuracy, fast response, and strong adaptability. It can identify risk trends in advance and provide forward-looking decision support for waterlogging prevention and control.

[0109] In this embodiment, the layer-by-layer decomposition includes a first layer decomposition using the predicted time series input vector as the initial residual term to generate a first layer trend component and a first layer residual component; the first layer residual component is used as the input residual term of the second layer, and the decomposition operation is continued to generate a second layer trend component and a second layer residual component; it is performed iteratively, in the i-th layer, the input is the residual component generated by the i-1-th layer, and the output is the trend component and the residual component of the i-th layer until the preset total number of decomposition layers n is completed, and finally a set of trend components of multiple levels and the residual component of the deepest layer are obtained.

[0110] The present invention decomposes the water level time series in a layer-by-layer recursive manner, extracting the corresponding trend components at each layer using the residual as input, effectively achieving a multi-level deconstruction of the water level change pattern. This method can capture the change characteristics of different scales from coarse to fine, enhancing the model's ability to model multiple types of trends such as slow rise, periodic fluctuations, and sudden anomalies in complex water level series. The resulting multi-layer trend component set provides a rich and differentiated input basis for subsequent feature fusion and prediction, significantly improving the accuracy of trend prediction and the generalization ability of the model, and is suitable for risk assessment and early warning response in a variety of waterlogging evolution scenarios.

[0111] In this embodiment, the generation of alarm tags includes alarm rule setting, alarm condition judgment and alarm tag output; the alarm rule setting includes static threshold rule, trend growth rule and extreme value prediction rule;

[0112] The alarm condition judgment is based on the alarm rules. If the current fused water level value is greater than the preset water level alarm trigger threshold, it is determined that a static water level alarm is triggered; if there are m consecutive time steps in the predicted water level value sequence that meet the difference between the water level value of the i-th predicted time step and the current fused water level value is greater than the trend increase threshold, it is determined that a trend rising warning is triggered; if any water level value in the predicted water level value sequence is greater than or equal to the future water level extreme risk threshold, it is determined that an extreme value warning mechanism is triggered;

[0113] The alarm tag output generates an alarm tag based on the judgment result. If any one of the alarm condition judgments meets the conditions for triggering the extreme value warning mechanism or triggering the trend rising warning and reaches the highest risk threshold, L(t)=2 is directly output, that is, a serious alarm; if no serious alarm is triggered, but any one of the warning level conditions for triggering the trend rising warning or triggering the static water level alarm is met, then L(t)=1 is output, that is, a warning; if all alarm rules are not triggered or do not meet any risk criteria, then L(t)=0 is output, that is, the current state is normal.

[0114] The present invention establishes a multi-dimensional water level risk assessment mechanism by setting three types of alarm rules: static threshold, trend growth, and extreme value prediction. This mechanism can comprehensively identify different types of water accumulation evolution situations. The system makes a joint judgment based on the current fused water level value and the prediction sequence, and outputs graded alarm labels to achieve precise classification from normal, early warning to severe alarm. This method avoids the false alarm and lag problems caused by traditional single threshold triggering, improves the sensitivity and foresight of the alarm response, and is particularly suitable for dynamic early warning needs in sudden and fast-changing heavy rainfall environments, providing efficient and reliable intelligent decision-making support for urban drainage emergency response.

[0115] In this embodiment, S6 specifically includes:

[0116] S61. Obtain the alarm tag value of the current time step, and control the display status of the LED water level gauge and the LED display screen at the tunnel entrance based on the alarm tag value;

[0117] S62. When the alarm tag value is L(t)=0, the display screen maintains the standard water level display mode. When the alarm tag value is L(t)=1, the LED display screen is controlled to flash orange or display dynamic graphics to indicate that the current water level is approaching the warning threshold. When the alarm tag value is L(t)=2, the display screen switches to high-frequency flashing red.

[0118] S63, controlling the operating state of the sound and light alarm device based on the alarm tag value, when the alarm tag value L(t) = 0, all the sound and light devices remain off, when the alarm tag value L(t) = 1, the sound and light devices are controlled to work in an intermittent ringing mode, and when the alarm tag value L(t) = 2, the sound and light devices are controlled to ring continuously;

[0119] S64. When the alarm tag value is L(t)=2, the tunnel entrance warning prompt module is activated in conjunction with the tunnel entrance, and the early warning mechanism is turned on at the tunnel entrance to output a prompt of "water accumulation ahead" or "detour recommended".

[0120] The present invention realizes hierarchical linkage control of the LED water level gauge, entrance display screen and sound and light alarm device by obtaining the alarm tag value, forming a complete visual and auditory response system. The display mode and sound and light status are dynamically adjusted according to the alarm level to ensure that the prompt means under different risk levels are both clear and targeted, thereby improving the efficiency of information transmission and the intensity of on-site perception. When in a serious alarm state, the system can link and activate the entrance early warning prompt module to realize early notification of water accumulation risks and vehicle guidance. This mechanism has the advantages of rapid response, clear control, and strong adaptability, and effectively improves the initiative and safety of water accumulation treatment.

[0121] In this embodiment, the S7 specifically includes:

[0122] S71, determine whether the image capture trigger condition is met, when the alarm tag value L(t)∈{1,2}, meets the capture condition, start the capture process;

[0123] S72. Calling a surveillance camera installed inside a tunnel or bridge to collect an image frame at the current time step, where the image frame is denoted as I(t), where t is a time index variable when the alarm occurs;

[0124] S73. Extract the fused water level value corresponding to the current time step and read the alarm tag value;

[0125] S74. Encapsulate the image frame, fused water level value, alarm tag and time index into a structured alarm information data packet.

[0126] This invention intelligently triggers the image capture process by determining whether the alarm tag value meets preset conditions, ensuring complete recording of on-site image information at critical warning moments. The system uses surveillance cameras to capture image frames in real time and encapsulates these images, along with the corresponding fused water level value, alarm tag, and time index, into a structured alarm information data packet. This mechanism synchronizes image and water level data, providing authentic, intuitive, and traceable data support for subsequent event tracing, risk analysis, and remote decision-making, effectively enhancing the system's overall effectiveness in intelligent identification and information integrity.

[0127] refer to Figure 3 The tunnel and culvert water level detection and alarm system based on deep learning includes the following modules:

[0128] The image and sensor acquisition module is used to deploy surveillance cameras and capacitive water level sensors in low-lying areas of tunnels or bridges and culverts, collect video image frames and capacitive water level signals, and generate raw multi-source observation data streams;

[0129] A multi-source data standardization processing module is used to process the original image data and generate standardized multi-source input data;

[0130] The image water level extraction module is used to input the standardized image frame into the SegmentAnything model, generate a semantic mask set, extract the water area mask by filtering the water body features, obtain the water body edge position and combine it with the scale mapping relationship to output the image recognition water level value;

[0131] Confidence-weighted water level fusion module, used to obtain the image recognition water level value and the capacitive sensor water level value to generate a fused water level value;

[0132] The water level trend prediction module is used to construct a time series input vector and input it into the SCINet model, and output the predicted water level value sequence;

[0133] The alarm level determination module is used to output the alarm level label based on the fused water level value and the water level prediction sequence;

[0134] The LED display and sound and light control module is used to control the display status of the LED water level gauge and the LED display screen at the tunnel entrance according to the alarm label, and control the local sound and light alarm device to achieve a graded response. In the event of a serious alarm, the entrance early warning prompt mode is activated;

[0135] The image capture and alarm information packaging module is used to trigger the camera to capture the current image frame when the alarm tag is in the warning or serious alarm state, extract and fuse the water level value and alarm tag information, and combine them to generate a structured alarm information data packet;

[0136] The IoT remote transmission module is used to upload structured alarm information data packets to the urban water level monitoring platform through the IoT communication module.

[0137] The deep learning-based tunnel and culvert water level detection and alarm system provided by the present invention realizes a closed loop of the entire process from data acquisition, intelligent identification, fusion prediction to hierarchical response and remote transmission through the collaborative operation of multiple modules. The system integrates image segmentation models and capacitive sensors to achieve high-precision water level identification, and combines the SCINet model to perform multi-scale prediction of water level trends, effectively improving the foresight and accuracy of early warning. The alarm tag drives the LED display and sound and light device to respond in a hierarchical manner, and automatically captures images and uploads structured data in high-risk conditions to ensure that the information is fully retained. The system has the advantages of accurate identification, rapid response, data synchronization and strong remote linkage capabilities. It is particularly suitable for water accumulation risk monitoring and intelligent management and control of low-lying tunnels and culverts in cities, significantly enhancing the city's drainage and flood prevention capabilities and emergency response levels.

[0138] Example 1:

[0139] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a water accumulation monitoring and intelligent early warning scenario in a typical low-lying road tunnel. This area is in a high-risk area prone to flooding all year round. The terrain is lower than the drainage system of the adjacent road section, and water accumulation problems occur frequently. Once heavy rainfall occurs, water accumulation is very likely to form at the low point in the middle of the tunnel. In severe cases, it can reach more than 30 centimeters, which not only affects the traffic efficiency, but also causes dangerous situations such as vehicle stalling and personnel being stranded. The existing water level detection device is a single electrode or capacitive sensor, which judges the water level by setting a height threshold. This type of system has problems such as measurement delay, lack of anti-interference, inability to feedback images in real time, and lack of trend prediction capabilities. It often triggers sound and light alarms only when water accumulation has become a fact, making it difficult to intervene in advance, resulting in missing the best window for disposal.

[0140] In this embodiment, a video surveillance unit and distributed capacitive water level sensors were deployed on-site. Image frames captured by the camera undergo image enhancement and size normalization before being fed into a deep image segmentation model. The system automatically identifies water areas within the image and calculates the corresponding image water level. The capacitive signals are filtered and time-aligned, and then combined with the image to identify the water level. The system then scores the image recognition results and sensor signals separately, dynamically assigning weights based on the quality scores. This achieves confidence-weighted fusion of water level values, significantly improving the accuracy and robustness of overall water level determination.

[0141] Based on the fused water level values, the system constructs a time series input and inputs it into a structured connection inference network model for trend prediction. The model outputs water level predictions for multiple future time points and compares them with the current water level. Combining the static water level threshold, the trend rise amplitude, and the predicted extreme value conditions for multi-dimensional judgment, the system automatically generates alarm labels, which are divided into three states: normal, warning, and severe alarm. When the system determines that it is a warning or severe alarm, the local response module is automatically started. The LED water level gauge and the tunnel entrance display screen switch to flashing warning mode, and the sound and light alarms sound intermittently or continuously according to the level. At the same time, the system captures the current image, encapsulates the alarm data, and uploads it to the remote monitoring platform through the Internet of Things module.

[0142] During the trial operation, which lasted a full flood season, the system covered multiple rainfall scenarios. During several significant rainfall events, the system successfully predicted the rapid rise in water levels and issued early warning signals, with an average lead time of 3 to 5 minutes, ensuring that management personnel could take countermeasures before waterlogging became a disaster. During the trial operation, the system's alarm accuracy reached 100%, with no false alarms or missed alarms. Compared with the results of on-site manual inspections, the water level error output after image and capacitor fusion was controlled within ±1.5 cm, far superior to traditional solutions where a single sensor on-site had a fluctuation error of more than ±5 cm. The system's image capture function accurately recorded the on-site footage in all alarm events, with a 100% information upload success rate and a 100% image integrity rate, providing a clear basis for risk tracing for the management backend.

[0143] Furthermore, the system demonstrates excellent stability during 24-hour continuous operation. The average response time for key modules is less than 2 seconds, and the latency for fusion and prediction is less than 1 second, fully meeting the real-time requirements of complex urban road environments. Upon receiving an alarm message, the platform instantly displays images, alarm levels, and water level data, significantly improving data transparency and command efficiency.

[0144] Table 1 Statistical table of measured data on the operating performance of the system of the present invention under typical heavy rainfall scenarios

[0145]

[0146] Experimental results demonstrate that the proposed deep learning-based tunnel and culvert water level detection and alarm system significantly outperforms traditional single-sensor waterlogging monitoring solutions in terms of water level sensing accuracy, trend prediction capabilities, and multi-level response linkage. Test data shows that in typical heavy rainfall scenarios, the confidence-weighted water level values ​​constructed by the system, which integrate image recognition and capacitive signals, can control the average measurement error to within ±1.2 centimeters. This significantly improves the accuracy and stability of monitoring results compared to conventional capacitive sensors, which are prone to drift in high-humidity environments and can experience errors exceeding ±5 centimeters.

[0147] By introducing a structured connection reasoning network model, the system can construct a future trend prediction sequence based on the current fused water level value. In actual tests, the system successfully predicted multiple rapid water level rise events and triggered an alarm response an average of 3 to 5 minutes in advance before the actual formation of the waterlogging risk, greatly extending the management intervention time window. The prediction error was less than 1.3 cm in all test samples, demonstrating excellent sequence modeling and practical adaptability. In a test scenario where the maximum rainfall intensity exceeded 50 mm per hour, the system triggered a serious alarm in advance and prompted passing vehicles to avoid it in time through the LED display and sound and light alarm. No traffic interruption or people were trapped at the scene, verifying the effectiveness of the system's early warning-driven linkage.

[0148] The image capture and information encapsulation modules successfully executed in all alarm events, achieving 100% image integrity and upload success rates. This ensures that every alarm record contains real-time images and data, providing a crucial basis for subsequent analysis and traceability. The platform's communication module achieves stable, low-latency transmission, with an average upload delay of less than 2.1 seconds for alarm data, meeting the timeliness of event processing required by the city-level monitoring platform. Furthermore, the system experienced no false alarms or missed alarms during continuous operation, demonstrating its robustness in dynamically integrating multi-source information and adapting to complex environments.

[0149] Of particular note is the stable operation of the system's graded alarm response mechanism during testing. When the system determines the alarm level is at a warning level, the LED water level gauge automatically switches to orange and provides intermittent audio and visual alerts. When the system enters a severe alarm state, it automatically switches to high-frequency red flashing and captures the current image, completing a complete linkage path from local alerts to remote upload. The system's response delay is kept within 2 seconds for each module, supporting cascade control within seconds, avoiding the response delays often associated with traditional systems due to the fragmented alarm drive chain.

[0150] In summary, the system of the present invention demonstrates a high degree of integration and intelligence in terms of water level data fusion accuracy, trend prediction advance time, alarm response closed-loop and image recording integrity. It has obvious technological progress compared with existing single-point or distributed water level monitoring solutions. It is particularly suitable for key waterlogging risk points such as urban tunnels, bridges, culverts, and underpasses that have high requirements for early warning response and strong demand for visual monitoring. It has good engineering feasibility and wide promotion and application value.

[0151] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A tunnel and culvert water level detection and alarm method based on deep learning, characterized in that: The steps include: S1. Install equipment at the lowest point of the tunnel or in areas prone to water accumulation to collect video image data from tunnel bridges and culverts and capacitive water level sensor signals, perform preprocessing, and generate standardized multi-source data. S2. Input each frame of the standardized multi-source data into the SegmentAnything model to generate an image mask set, filter out the water area mask, and calculate the image recognition water level value by combining the pixel-elevation mapping relationship of the LED water level scale; S3, performing confidence-weighted fusion of the image water level value and the capacitive water level sensor signal to obtain a fused water level value; S4. Construct a time series input vector based on the fused water level value and build an SCINet model. Perform trend decomposition and structure reconstruction through the recursive decomposition module and the multi-scale connection module to output the predicted water level value sequence. S5. Generate an alarm label based on the fused water level value and the predicted water level value sequence; S6. Control the LED water level gauge inside the tunnel and the LED display screen at the tunnel entrance to display corresponding display modes according to the alarm tag. If the alarm tag is in a serious alarm state, activate the tunnel entrance early warning prompt mode; S7. When the alarm tag is in the warning or serious alarm state, trigger the camera to capture the image, collect the current image frame, and construct an alarm information data packet; S8. Upload the alarm information data packet to the remote city water level monitoring platform through the Internet of Things communication module.

2. The tunnel bridge and culvert water level detection and alarm method based on deep learning according to claim 1 is characterized in that: The standardized multi-source data includes normalized video image frames, denoised capacitive water level sensor signals, unified timestamps corresponding to image frames, image enhancement parameter recording information, and filter status identifiers of sensor signals.

3. The tunnel bridge and culvert water level detection and alarm method based on deep learning according to claim 1 is characterized in that: The S2 specifically includes: S21. Input each image frame in the standardized multi-source data into the image encoder of the SegmentAnything model. The image encoder is an image embedding network built based on the visual Transformer structure, which is used to perform global semantic modeling on the input image frame and output an image embedding representation. S22. Input the empty prompt information into the prompt encoder of the SegmentAnything model, configure the prompt encoder to be in an empty input state, adopt a zero-shot segmentation mode, and output an empty prompt embedding representation; S23. Input the image embedding representation and the empty prompt embedding representation into the mask prediction head of the SegmentAnything model. The mask prediction head generates a semantic mask set based on the cross-attention mechanism. S24. Calculate a color histogram for each candidate mask in the mask set and compare it with a preset water body color histogram: Among them, S i is the water body similarity score of the i-th candidate mask, i is the mask index, j is the color interval index, K is the number of color intervals, is the water color reference histogram value in the jth color interval, is the candidate mask, is the pixel distribution density of the i-th candidate mask in the j-th color interval, and t is the time index variable; S25, extract the pixel coordinates of the lower boundary of the water body mask, and calculate the image recognition water level value by combining the preset pixel-physical height mapping function 4. The tunnel bridge and culvert water level detection and alarm method based on deep learning according to claim 1 is characterized in that: The confidence weighted fusion includes fusion input definition, fusion weight calculation, and fusion water level output; the fusion input definition obtains the image recognition water level value at time t and capacitance sensor water level value The fusion weight calculation calculates the image recognition quality score and the capacitive sensor quality score respectively, and determines the fusion weight of the image water level based on the two score values; The fused water level output is weighted by the image water level value and the capacitive sensor water level value based on the fusion weight to calculate the final fused water level value: Among them, H (t) is the fusion water level value, α t is the confidence fusion weight of the image water level, Identify water level values ​​for images, is the water level value of the capacitance sensor, (1-α t ) is the confidence fusion weight of the capacitor water level, and t is the time index variable.

5. The tunnel bridge and culvert water level detection and alarm method based on deep learning according to claim 1 is characterized in that: The S4 specifically includes: S41. Construct a prediction time series input vector, select the fused water level values ​​of the most recent N consecutive time steps, and record it as X = {x(t-N+1), x(t-N+2), ..., x(t)}, where x(t) is the fused water level value of the current time step, t is the time index variable, and N is the length of the time series; S42: Input the time series input vector X into a structured connectivity inference network (SCINet) model. The SCINet model includes a recursive decomposition module and a multi-scale connection module, wherein each module includes multiple convolutional layers, residual connection paths, and an attention fusion structure. S43. In the recursive decomposition module, the time series input vector X is decomposed layer by layer, with the input of each layer being the residual term of the previous layer, and the output being the trend term and the new residual term; S44. In the multi-scale connection module, the trend components of each layer are input into the fusion structure. The fusion structure includes a channel attention mechanism and a global pooling process. First, the fusion weight is calculated through the channel attention mechanism, and then all trend components are weighted and fused to generate a reconstructed feature vector: Among them, α i is the attention fusion weight corresponding to the trend component of the i-th layer, W1 is the first set of weight parameter matrices for channel attention mapping, W2 is the second set of weight parameter matrices for channel attention mapping, ReLU(·) is the rectified linear unit activation function, T (i) is the trend component obtained by recursive decomposition at the i-th layer, GAP(·) is the global average pooling operation performed, n is the total number of layers of recursive decomposition, i is the index of the trend component currently being processed, and j is the traversal index used for normalized weight calculation; S45, reconstruct the feature vector, perform linear mapping, and output the predicted water level value sequence for the next M time steps in, is the predicted water level value at step i, and M is the prediction step length.

6. The tunnel bridge and culvert water level detection and alarm method based on deep learning according to claim 5 is characterized in that: The layer-by-layer decomposition includes a first layer using the predicted time series input vector as the initial residual term, performing a decomposition to generate a first layer trend component and a first layer residual component; The first layer residual component is used as the input residual term of the second layer, and the decomposition operation is continued to generate the second layer trend component and the second layer residual component; The process is iterative. In the i-th layer, the input is the residual component generated by the i-1-th layer, and the output is the trend component and residual component of the i-th layer until the preset total number of decomposition layers n is completed, and finally a set of trend components of multiple levels and the residual component of the deepest layer are obtained.

7. The tunnel bridge and culvert water level detection and alarm method based on deep learning according to claim 1 is characterized in that: The generation of alarm tags includes alarm rule setting, alarm condition judgment and alarm tag output; the alarm rule setting includes static threshold rule, trend growth rule and extreme value prediction rule; The alarm condition judgment is based on the alarm rules. If the current fused water level value is greater than the preset water level alarm trigger threshold, it is determined that a static water level alarm is triggered; if there are m consecutive time steps in the predicted water level value sequence that meet the difference between the water level value of the i-th predicted time step and the current fused water level value is greater than the trend increase threshold, it is determined that a trend rising warning is triggered; if any water level value in the predicted water level value sequence is greater than or equal to the future water level extreme risk threshold, it is determined that an extreme value warning mechanism is triggered; The alarm tag output generates an alarm tag based on the judgment result. If any one of the alarm condition judgments meets the conditions for triggering the extreme value warning mechanism or triggering the trend rising warning and reaches the highest risk threshold, L(t)=2 is directly output, that is, a serious alarm; if no serious alarm is triggered, but any one of the warning level conditions for triggering the trend rising warning or triggering the static water level alarm is met, then L(t)=1 is output, that is, a warning; if all alarm rules are not triggered or do not meet any risk criteria, then L(t)=0 is output, that is, the current state is normal.

8. The tunnel, bridge and culvert water level detection and alarm method based on deep learning according to claim 1 is characterized in that: The S6 specifically includes: S61. Obtain the alarm tag value of the current time step, and control the display status of the LED water level gauge and the LED display screen at the tunnel entrance based on the alarm tag value; S62. When the alarm tag value is L(t)=0, the display screen maintains the standard water level display mode. When the alarm tag value is L(t)=1, the LED display screen is controlled to flash orange or display dynamic graphics to indicate that the current water level is approaching the warning threshold. When the alarm tag value is L(t)=2, the display screen switches to high-frequency flashing red. S63, controlling the operating state of the sound and light alarm device based on the alarm tag value, when the alarm tag value L(t) = 0, all the sound and light devices remain off, when the alarm tag value L(t) = 1, the sound and light devices are controlled to work in an intermittent ringing mode, and when the alarm tag value L(t) = 2, the sound and light devices are controlled to ring continuously; S64. When the alarm tag value is L(t)=2, the tunnel entrance warning prompt module is activated in conjunction with the tunnel entrance, and the early warning mechanism is turned on at the tunnel entrance, outputting a "water accumulation ahead" or "detour recommended" prompt.

9. The tunnel, bridge and culvert water level detection and alarm method based on deep learning according to claim 1 is characterized in that: The S7 specifically includes: S71, determine whether the image capture trigger condition is met, when the alarm tag value L(t)∈{1,2}, meets the capture condition, start the capture process; S72. Calling a surveillance camera installed inside a tunnel or bridge to collect an image frame at the current time step, where the image frame is denoted as I(t), where t is a time index variable when the alarm occurs; S73. Extract the fused water level value corresponding to the current time step and read the alarm tag value; S74. Encapsulate the image frame, fused water level value, alarm tag and time index into a structured alarm information data packet.

10. A tunnel bridge and culvert water level detection and alarm system based on deep learning, which implements the tunnel bridge and culvert water level detection and alarm method based on deep learning according to any one of claims 1 to 9, characterized in that: Includes the following modules: The image and sensor acquisition module is used to deploy surveillance cameras and capacitive water level sensors in low-lying areas of tunnels or bridges and culverts, collect video image frames and capacitive water level signals, and generate raw multi-source observation data streams; A multi-source data standardization processing module is used to process the original image data and generate standardized multi-source input data; The image water level extraction module is used to input the standardized image frame into the SegmentAnything model, generate a semantic mask set, extract the water area mask by filtering the water body features, obtain the water body edge position and combine it with the scale mapping relationship to output the image recognition water level value; Confidence-weighted water level fusion module, used to obtain the image recognition water level value and the capacitive sensor water level value to generate a fused water level value; The water level trend prediction module is used to construct a time series input vector and input it into the SCINet model, and output the predicted water level value sequence; The alarm level determination module is used to output the alarm level label based on the fused water level value and the water level prediction sequence; The LED display and sound and light control module is used to control the display status of the LED water level gauge and the LED display screen at the tunnel entrance according to the alarm label, and control the local sound and light alarm device to achieve a graded response. In the event of a serious alarm, the entrance early warning prompt mode is activated; The image capture and alarm information packaging module is used to trigger the camera to capture the current image frame when the alarm tag is in the warning or serious alarm state, extract and fuse the water level value and alarm tag information, and combine them to generate a structured alarm information data packet; The IoT remote transmission module is used to upload structured alarm information data packets to the urban water level monitoring platform through the IoT communication module.

Images