Heavy rain terminal early warning method and system based on deep learning

Abstract

The invention relates to the technical field of rainstorm early warning, in particular to a deep learning-based rainstorm terminal early warning method and system, which introduces an image perception and semantic recognition mechanism to carry out consistency judgment on an overturning state of a rain gauge and image features such as a rain curtain and a water mark so as to realize effective elimination of unnatural trigger factors (such as manual water pouring and the like). And the reliability of rainfall sensing data is improved. A double-branch deep learning model capable of being deployed locally is adopted, multi-source data such as rainfall, temperature and humidity and topographic factors are fused, feature extraction and recurrence period prediction are performed based on a cross attention mechanism, and a threshold correction coefficient is dynamically generated in combination with historical events, so that self-adaptive adjustment of an early warning standard is realized, and misinformation and missing report risks are effectively reduced. The system supports low-power-consumption local operation, early warning decision is executed, and timeliness and accuracy of local early warning response are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of heavy rain early warning, in particular to a heavy rain terminal early warning method and system based on deep learning. BACKGROUND

[0002] Heavy rain is easy to cause disasters such as mountain torrents, landslides, and urban waterlogging, especially in areas with complex terrain and weak infrastructure, which are more destructive. The existing heavy rain early warning means mainly relies on meteorological big models, radar echo monitoring, and satellite remote sensing deduction. Most of the existing early warning systems rely on central meteorological platforms and cloud servers to perform model operation and decision logic execution. This results in the inability to operate normally in scenarios with poor network, weak network, and poor power conditions, and high response delay.

[0003] The existing technology realizes a certain degree of localized data collection and early warning through the local deployment of rain gauges in the early warning area, but still has the following problems: the rain gauge is easy to be affected by factors such as artificial water diversion and non-natural rainfall during the collection process, resulting in a high false alarm rate and affecting the credibility of the early warning and the accuracy of the public response. The current local early warning uses fixed rainfall threshold judgment logic, without combining regional differences, historical event statistical characteristics, and real-time weather conditions to dynamically adjust the early warning standard, which is easy to cause false alarms or missed alarms.

[0004] Therefore, it is necessary to provide a heavy rain terminal early warning method and system suitable for local deployment and high precision. SUMMARY

[0005] To solve the above problems, the present application provides a heavy rain terminal early warning method and system based on deep learning.

[0006] To achieve the above purpose, the technical solution adopted by the present application is: A heavy rain terminal early warning method based on deep learning, comprising: S1, collecting rainfall perception data and image data in a target area, the rainfall perception data including rainfall data and temperature and humidity data; S2, performing rainfall state recognition based on the image data, outputting a rainfall scene verification result, and filtering effective rainfall perception data according to the rainfall scene verification result; S3, inputting the effective rainfall perception data into a double-branch deep learning model based on cross-attention, performing feature extraction and pattern matching, and outputting a heavy rain return period corresponding to the target area and a dynamic threshold correction coefficient; S4, performing parameter adjustment on the engineering threshold of the current area based on the dynamic threshold correction coefficient to obtain a dynamic threshold, comparing the real-time rainfall data with the dynamic threshold, generating heavy rain early warning information according to the comparison result, and controlling the terminal display device to complete the early warning action output according to the heavy rain early warning information.

[0007] Further, the rainfall data is obtained by setting tipping bucket rain gauges in the target area, and the temperature and humidity data is obtained by setting temperature and humidity sensors in the target area.

[0008] Further, the image data is obtained by setting cameras in the target area, and the image data includes images of the setting area of the tipping bucket rain gauges.

[0009] Further, the rainfall state recognition based on the image data includes the following steps: Performing structured detection on the original image data, identifying and extracting the spatial position of the tipping bucket rain gauges in the image, and obtaining the corresponding tipping bucket image area; Based on the tipping bucket image area, constructing an image sequence, performing tipping bucket state recognition on consecutive image frames through inter-frame pixel difference and edge contour extraction algorithm, extracting tipping bucket overturn start and end frames, overturn amplitude and overturn frequency, and generating structured motion feature data representing the tipping bucket motion state; Based on the background image area in the original image data except the tipping bucket area, extracting rainfall semantic feature vectors of rain curtain texture, water droplet trajectory and ground water trace diffusion through a convolutional neural network model; Conducting consistency determination on the structured motion feature data and the rainfall semantic feature vector, if there are semantic feature distributions consistent with natural rainfall in the images in the corresponding time period of the tipping bucket overturn behavior, generating a rainfall scene verification result as natural rainfall triggering; if there is deviation between the structured motion feature and the semantic feature distribution, generating a rainfall scene verification result as non-natural rainfall triggering.

[0010] Further, the cross-attention-based double-branch deep learning model is constructed by the following steps: Based on the effective rainfall perception data of the rainfall scene verification in historical heavy rain events, performing coefficient scanning on the rainfall sequence of each event and the corresponding engineering threshold, and statistically obtaining the false alarm rate and the false alarm rate of each candidate coefficient, obtaining the optimal threshold correction coefficient with the lowest false alarm rate and the false alarm rate not higher than the preset upper limit, and the optimal threshold correction coefficient and the actual heavy rain return period grade of the corresponding heavy rain event jointly constitute the supervision label set; Based on the terrain type, altitude and environment category of the heavy rain event occurrence place, constructing a terrain factor, and based on the effective rainfall perception data of the heavy rain event occurrence place, constructing a rainfall sequence and a temperature and humidity sequence; Constructing a multi-source training sample set through the terrain factor, the rainfall sequence and the temperature and humidity sequence, and associating the multi-source training sample set with the supervision label set; Based on the multi-source training sample set, temporal position embedding and cross-channel attention fusion are performed on the rainfall sequence, temperature and humidity sequence and terrain factor encoding to obtain a spatiotemporal joint feature sequence; Based on the spatiotemporal joint feature sequence, a backbone network including residual connections is input to obtain a prediction vector for return period prediction and a correction vector for dynamic threshold regression, and the return period error and threshold error are calculated according to the supervision label; Gradient optimization is performed on the network parameters through the loss function to obtain a dual-branch deep learning model.

[0011] Furthermore, the loss function is as follows: ; in, is the loss value; is the number of training samples; For the The predicted return period level of the samples; For the The true return period level label of each sample; For the The prediction threshold correction coefficient of the sample; For the The optimal threshold correction coefficient label for each sample; For the The historical false alarm rate weight corresponding to the sample; is the weight factor.

[0012] Furthermore, generating rainstorm warning information according to the comparison result includes: Real-time rainfall data is obtained and compared with the dynamic threshold. If the real-time rainfall data is greater than or equal to the dynamic threshold, a rainstorm warning message is generated; if the rainfall data in the previous time period is less than the dynamic threshold, no rainstorm warning message is generated.

[0013] Furthermore, the rainstorm warning information includes: target area identification, warning trigger time, real-time rainfall data, dynamic threshold, predicted rainstorm recurrence period level and warning level identification.

[0014] Furthermore, the warning actions include: visual interface color change, pop-up window prompt, voice broadcast and buzzer sounding.

[0015] A deep learning-based rainstorm terminal warning system, applied to any of the aforementioned deep learning-based rainstorm terminal warning methods, comprising: a data perception layer, an operation and maintenance layer, a core computing layer, and an early warning decision layer connected in sequence; The data perception layer is used to collect rainfall perception data and image data in the target area, wherein the rainfall perception data includes rainfall data and temperature and humidity data; An operation and maintenance layer, configured to identify rainfall status based on the image data, output a rainfall scene verification result, and filter effective rainfall perception data according to the rainfall scene verification result; A core computing layer, configured to input the effective rainfall perception data into a two-branch deep learning model based on cross-attention, perform feature extraction and pattern matching, and output a rainstorm return period corresponding to the target area and a dynamic threshold correction coefficient for correcting the engineering threshold; The early warning decision layer is used to perform parameter adjustment on the engineering threshold of the current area based on the dynamic threshold correction coefficient to obtain the dynamic threshold, and compare the real-time rainfall data with the dynamic threshold, generate heavy rain warning information according to the comparison results, and control the terminal display device to complete the warning action output according to the warning information.

[0016] The beneficial effects of the present invention include: by incorporating image perception and semantic recognition mechanisms, the system visually verifies the actual rainfall triggering of rain gauges within a designated area. By extracting image features such as rain curtains, water marks, and bucket flipping, and comparing them with the rain gauge data for semantic consistency, it significantly reduces false alarms caused by non-natural factors (such as human spillage and accidental touch), thereby enhancing the credibility of the perceived data. A branched deep learning model, deployed locally, is constructed based on a cross-attention architecture. This model integrates features and patterns from multiple sources of rainfall perception data (such as rainfall, temperature, humidity, and topographic factors). It also generates dynamic threshold correction coefficients based on historical event samples. This adaptively adjusts the warning threshold based on regional characteristics and recurrence levels, thereby avoiding the false or missed warnings associated with traditional static threshold methods. The system can operate autonomously in extreme environments, such as network disconnection, weak network connectivity, and power outages, using low-power LoRa technology. This ensures timely generation and on-site output of warning information, significantly enhancing response speed and deployment flexibility at the onset of disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Fig. 1 This is a step flow chart of a rainstorm terminal warning method based on deep learning in the present invention.

[0018] Fig. 2 This is a flowchart of the steps for constructing a dual-branch deep learning model based on cross-attention in the present invention. DETAILED DESCRIPTION

[0019] See also Figs. 1-2 As shown, the present invention relates to a rainstorm terminal early warning method based on deep learning, comprising: S1, collect rainfall perception data and image data in a target area, the rainfall perception data including rainfall data and temperature and humidity data; S2, perform rainfall state recognition based on the image data, output rainfall scene verification results, and filter effective rainfall perception data according to the rainfall scene verification results; S3, input the effective rainfall perception data into a double-branch deep learning model based on cross attention, perform feature extraction and pattern matching, and output a corresponding storm return period and dynamic threshold correction coefficient of the target area; S4, perform parameter adjustment on the engineering threshold of the current area based on the dynamic threshold correction coefficient, obtain a dynamic threshold, compare real-time rainfall data with the dynamic threshold, generate storm warning information according to the comparison result, and control a terminal display device to complete a warning action output according to the storm warning information.

[0020] In some embodiments, a tipping bucket rain gauge and a temperature and humidity sensor are deployed inside the target area to obtain raw rainfall and temperature and humidity parameters, and a camera is set up to capture a continuous image sequence covering the rain gauge area and its surrounding environment. In step S2, through image structural processing, the area of the rain gauge in the image is automatically identified and its tipping bucket action features (such as tipping start and end frames, tipping amplitude, frequency, etc.) are extracted, and at the same time, the rain curtain, water trace diffusion texture, etc. in the background image area are semantically encoded using a convolutional neural network to form a visual semantic feature vector. Then, the physical tipping bucket structure features and the visual semantic features are matched, and when they are consistent, it is determined as natural rainfall triggering, otherwise it is recognized as non-natural disturbance (such as artificial pouring water), which can significantly reduce false alarms caused by false triggering of the rain gauge and improve the local perception credibility. In step S3, the system inputs the effective rainfall perception data filtered by semantic verification into a pre-trained deep learning model based on cross-attention with double branches. The model consists of a return period branch and a threshold correction branch, which respectively predict the heavy rain return period grade in the future time period and the adjustment coefficient required to correct the current regional static engineering threshold. Among them, the model input not only includes the current rainfall and temperature and humidity sequence, but also includes regional background factors such as terrain type, elevation, and environment category, through time sequence position embedding and inter-channel attention fusion mechanism, to construct deep feature expression with spatiotemporal consistency, significantly improving the adaptability and generalization ability of the model to complex regional differences. In step S4, the dynamic threshold correction coefficient obtained by the above prediction is used to adjust the engineering preset threshold, and the rainfall data in the current period is compared in real time. If the current rainfall intensity reaches or exceeds the dynamically corrected threshold, a heavy rain warning information containing the target area, real-time rainfall value, return period grade and warning level is generated locally, and the local linkage response of the terminal device is triggered, such as screen highlighting, voice broadcast, and buzzer prompt. The local decision-making capability significantly enhances the independent operation capability of the system in the case of network interruption or extreme disaster environment, avoiding the response delay of traditional cloud computing architecture and the short board of relying on network.

[0021] In a preferred embodiment, the rainfall perception data is not limited to rainfall data and temperature and humidity data, but can also include air pressure, wind force and direction, and water level. The device specifically includes: a rain gauge for obtaining ground instantaneous and cumulative rainfall, a digital temperature and humidity sensor for collecting atmospheric temperature and relative humidity, a micro-pressure sensor for recording real-time air pressure change trend, an ultrasonic wind speed and direction instrument for collecting wind strength and direction change, and an ultrasonic or radar water level sensor for measuring surface or river water level change. The data collected by these sensors is preliminarily cleaned and time-stamped by the edge computing node, and constitutes a multi-dimensional rainfall perception data set. For multi-dimensional rainfall perception data, in model construction, all dimensions of rainfall perception variables are normalized by a unified encoder structure, and are constructed into a unified multi-source input tensor according to time alignment, and are input into a double-branch deep learning model. Not only does it provide more rich context information for heavy rain return period grade discrimination, but it also provides more comprehensive environmental parameter support for dynamic threshold correction, which can improve the discrimination ability of the system in the early stage of extreme weather evolution.

[0022] Further, the rainfall data is obtained by setting a tipping bucket rain gauge in the target area, and the temperature and humidity data is obtained by distributing temperature and humidity sensors in the target area.

[0023] Specifically, the collection of rainfall data is realized by using a tipping bucket rain gauge, which has the advantages of simple structure, sensitive response and low power consumption, and is suitable for long-term deployment in remote areas or areas with weak infrastructure. The tipping bucket rain gauge quantitatively records the rainfall by driving the tipping bucket mechanism to turn over after collecting rainwater, and each turn represents a fixed volume of rainfall, thereby realizing accurate rainfall data collection. To further enhance the multi-dimensional coverage of rainfall perception, an array of temperature and humidity sensors is distributed in the target area to monitor real-time changes in local air temperature and humidity, and to compensate for possible point perception blind spots of a single rain gauge. By fusing temperature and humidity data with rainfall information, it is helpful to more comprehensively perceive the meteorological dynamics in the region, and to provide a richer data foundation for subsequent model temporal feature extraction, spatial attribution, and semantic consistency judgment.

[0024] Further, the image data is obtained by setting a camera in the target area, and the image data includes images of the setting area of the tipping bucket rain gauge.

[0025] It should be noted that the image data is obtained by deploying a camera device in the target area, and the camera is installed at a reasonable position to ensure that its field of view covers the installation area of the tipping bucket rain gauge, so as to collect image sequences containing the rain gauge and its surrounding environment in real time. This setting not only obtains the physical state change of the rain gauge, but also captures visual information features in the rainfall process, including rain curtain texture, water droplet motion trajectory, and ground water spreading phenomena. By processing the image data obtained by the camera, a continuous image sequence can be constructed for analyzing the semantic consistency between the tipping behavior of the rain gauge and the external rainfall characteristics. This image fusion method effectively makes up for the shortcoming of traditional rain gauges that cannot self-identify when subjected to human interference (such as artificial watering, accidental touching, etc.). The image data provided by the camera provides the necessary input basis for subsequent image recognition algorithms and is the core information source for building a rainfall authenticity recognition mechanism based on image verification.

[0026] Further, the rainfall state recognition based on the image data includes the following steps: Performing structured detection on the original image data, identifying and extracting the spatial position of the tipping bucket rain gauge in the image to obtain the corresponding tipping bucket image region; Based on the tipping bucket image region, an image sequence is constructed, and by using frame-to-frame pixel difference and edge contour extraction algorithms, tipping state recognition is performed on continuous image frames to extract tipping start and end frames, tipping amplitude, and tipping frequency, and generate structure motion feature data representing the tipping action state; Based on the background image region in the original image data excluding the tipping bucket region, a convolutional neural network model is used to extract rainfall semantic feature vectors of rain curtain texture, water droplet trajectory, and ground water trace diffusion; The structure motion feature data and the rainfall semantic feature vector are consistent, and if the image in the corresponding time period of the tipping bucket tipping behavior exists consistent semantic feature distribution with natural rainfall, the rainfall scene verification result is generated as natural rainfall trigger; if the structure motion feature of the tipping bucket and the semantic feature distribution exist deviation, the rainfall scene verification result is generated as non-natural rainfall trigger.

[0027] In some embodiments, the original image data is first subjected to structured detection, specifically using a target detection model based on YOLOv5 or Faster R-CNN to locate the tipping bucket rain gauge in the image frame. Through this detection process, the bounding box coordinates of the rain gauge in the image are obtained, so as to crop out a stable and target device-containing tipping bucket image region, providing a focused area for subsequent action recognition processing. After completing the spatial positioning, the system constructs an image sequence based on the tipping bucket image region. For this image sequence, the inter-frame pixel difference algorithm is applied to extract the gray scale change area in the continuous frames, so as to capture the dynamic change process of the tipping bucket device, and the edge contour extraction algorithm based on Canny or Sobel is combined to further enhance the extraction effect of the tipping bucket edge structure. By setting the flip angle change threshold and the continuous frame number condition, the starting frame and the ending frame of the tipping bucket flip can be identified, and the flip amplitude and the number of flips per unit time are counted, so as to finally generate a sequence of structural motion feature data for quantitatively representing the behavior characteristics of the tipping bucket action in this time period. In order to enhance the robustness of scene recognition, the system further performs semantic feature extraction on the background image region in the original image data except the tipping bucket region. In this step, a convolutional neural network is introduced as a basic feature encoder to extract key semantic features related to natural rainfall in the image, such as the flowing texture of the rain curtain, the track formed by water droplets on the lens surface, and the direction and density of water trace diffusion on the ground. The high-dimensional feature vector output by the network is mapped to a standardized semantic representation through a fully connected layer. Finally, the structural motion feature data and the rainfall semantic feature vector are input into a similarity matching module, and a cosine similarity or Mahalanobis distance metric function is used for consistency determination. If the flip behavior of the tipping bucket and the semantic image features in this time period have a high consistency, it is inferred that it is caused by natural rainfall, and the output of the rainfall scene verification result is natural rainfall trigger; if the two features are significantly different, it is determined that there is an abnormal trigger factor (such as human interference), and the verification result output is non-natural rainfall trigger.

[0028] Further, the cross-attention-based double-branch deep learning model is constructed by the following steps: Based on the effective rainfall perception data verified by the rainfall scene in the historical heavy rain events, the rainfall sequence of each event and the corresponding engineering threshold are scanned, the false alarm rate and the missed alarm rate of each candidate coefficient are counted, the optimal threshold correction coefficient that makes the false alarm rate the lowest and the missed alarm rate not higher than the preset upper limit is obtained, and the optimal threshold correction coefficient and the actual heavy rain return period grade of the corresponding heavy rain event jointly constitute the supervision label set; Based on the terrain type, altitude and environment category of the heavy rain event occurrence place, a terrain factor is constructed, and based on the effective rainfall perception data of the heavy rain event occurrence place, a rainfall sequence and a temperature and humidity sequence are constructed; A multi-source training sample set is constructed by terrain factors, rainfall sequences and temperature and humidity sequences, and the multi-source training sample set is associated with a supervision label set; Based on the multi-source training sample set, time sequence position embedding and cross-channel attention fusion are performed on rainfall sequence, temperature and humidity sequence and terrain factor coding to obtain a spatio-temporal joint feature sequence; Based on the spatio-temporal joint feature sequence, a main network containing residual connection is input to obtain a prediction vector for return period prediction and a correction vector for dynamic threshold regression, and return period error and threshold error are calculated according to the supervision label; Gradient optimization of network parameters is performed by a loss function to obtain a double-branch deep learning model.

[0029] In some embodiments, a supervised label is first constructed based on valid rainfall sensing data from historical rainstorm events that has been verified as natural rainfall through image verification. A set of candidate correction coefficients is set for each event's rainfall variation curve. The original engineering warning threshold is weighted and adjusted accordingly, and the false alarm rate and missed alarm rate under these correction conditions are calculated. A heuristic search strategy is employed to select the correction coefficient with the lowest false alarm rate as the optimal threshold correction coefficient for the event, provided that the missed alarm rate does not exceed a preset tolerance. Simultaneously, a corresponding dual-label supervision structure is formed based on the actual rainstorm return period level annotated in meteorological data for subsequent model training. During feature construction, the system constructs a terrain factor based on factors such as the terrain type (e.g., hilly, mountainous, plain), altitude, and environmental category (e.g., urban-rural distribution) of the rainstorm event region. This factor is combined with rainfall and temperature and humidity sequences collected by sensing terminals in the same region to generate three types of heterogeneous input data. To ensure the model's ability to perceive temporal evolution, positional encoding is performed on each sequence, and time step identifiers are introduced to enhance the ability to capture periodic feature changes. Subsequently, a structured terrain factor is constructed by combining the terrain type, altitude, and environmental category of the event site. Combined with the corresponding rainfall and temperature and humidity series, these factors together form multi-source input features, forming a unified multimodal training sample set. In the feature encoding stage, the model first embeds the time series data of different modalities in a temporal positional manner to preserve their temporal evolution during the rainstorm formation process. A cross-channel attention mechanism is then introduced to model the information flow and dependencies between modalities. Specifically, the cross-channel attention module generates statistical descriptions by globally pooling the three input channels of rainfall, temperature and humidity, and terrain factors. Channel attention weights are then calculated through a shared weight network. This adaptively adjusts the importance of each channel feature during fusion, enhancing the model's ability to focus on key influencing factors. Residual connections are used to preserve the features of the underlying perceptual layer, ultimately generating a structurally stable spatiotemporal joint feature representation. At the output, the model has two branches: one for predicting the rainstorm return period level and the other for regressing the dynamic threshold correction coefficient. During the training phase, a joint loss function is defined using the constructed supervised label set as the target. Finally, the parameters are iteratively updated through the Adam optimizer to complete the training convergence of the two-branch model.

[0030] Furthermore, the loss function is as follows: ; in, is the loss value; is the number of training samples; For the The predicted return period level of the samples; For the a true return period rank label of the sample; a predicted threshold correction coefficient of the first sample; an optimal threshold correction coefficient label of the first sample; a historical false alarm rate weight corresponding to the first sample; a weight factor.

[0031] It should be noted that the loss function is composed of two sub-items, corresponding to the return period rank prediction error and the threshold correction error respectively. The first term is used to measure the regression accuracy of the model on the storm return period rank, and its value is the squared error between the model predicted rank and the true rank; this part ensures that the model can effectively distinguish different levels of rainfall events and provide hierarchical risk prompts. The second term is weighted processing for the prediction error of the dynamic threshold correction coefficient, and the historical false alarm rate is introduced as a weight factor to give higher training punishment intensity to the samples prone to false alarm, significantly reducing the probability of false alarm in actual application. At the same time, in order to balance the relative importance of the two prediction targets in the training process, an adjustable weight coefficient is set in the loss function to control the contribution ratio between the return period error term and the threshold correction error term. Through this structural design, the loss function not only improves the multi-objective learning ability of the model, but also enhances its robustness and adaptability to abnormal rainfall events in edge scenarios, effectively making up for the problems of high false alarm rate and inability to dynamically adjust the threshold in traditional warning methods.

[0032] Further, the generating the storm warning information according to the comparison result comprises: obtaining real-time rainfall data and comparing it with the dynamic threshold, if the real-time rainfall data is greater than or equal to the dynamic threshold, generating storm warning information; if the rainfall data in the previous time period is less than the dynamic threshold, not generating storm warning information.

[0033] Specifically, the process of generating the heavy rain warning information according to the comparison result relies on the accurate comparison logic of the real-time rainfall and the dynamic threshold. The system first continuously receives the real-time rainfall data collected by the tipping bucket rain gauge in the target area, and compares it with the threshold dynamically calculated by the deep learning model at each time period. The dynamic threshold has taken into account the regional terrain features, the return period level of historical rainfall events, and the false alarm control factor, and can flexibly adjust the warning standard for different scenarios. In the comparison process, it is determined whether the actual rainfall in the current time period has reached or exceeded the dynamic threshold. If the comparison result is "satisfies" the condition, that is, the real-time rainfall intensity has reached the dynamically set critical degree, the corresponding heavy rain warning information is immediately generated as the trigger signal for disaster response. At the same time, the warning information will carry structured fields including regional identification, trigger time, current rainfall and threshold, warning level, etc. for subsequent decision-making and terminal display. On the contrary, if the rainfall in the current time period has not reached the dynamic threshold, it is determined that it does not constitute a heavy rain trigger situation, and the system does not generate a warning information to avoid false alarms due to slight rainfall, thereby improving the reliability of the overall warning and the accuracy of public response. This mechanism is significantly different from the traditional static threshold method, with stronger environmental adaptability and error control capability, and is suitable for local autonomous warning applications in complex weather backgrounds.

[0034] Further, the heavy rain warning information includes: target area identification, warning trigger time, real-time rainfall data, dynamic threshold, predicted heavy rain return period level and warning level identification.

[0035] Further, the warning action includes: visual interface color change, pop-up prompt, voice broadcast and buzzer ringing.

[0036] In some embodiments, the visual interface color change refers to the automatic switching of the terminal interface background color or border highlight color under different warning levels, such as yellow, orange, red, etc., corresponding to the national meteorological warning levels, to enhance risk perception; the pop-up prompt is to pop up a warning window on the terminal screen to display key information such as current rainfall, dynamic threshold, return period level and warning level for on-site managers to quickly understand the current risk status; the voice broadcast is to broadcast the current warning information content in real time through the built-in or external speaker, such as "the current area has triggered a ten-year return period heavy rain warning, please take immediate precautions", to ensure that it can still be effectively received when the vision is limited or the operator is not in front of the terminal; the buzzer ringing is a high-intensity sound alarm device that starts continuous or intermittent ringing when reaching the preset level, which is used to strengthen the warning signal in noisy or visually impaired environments.

[0037] The application also comprises a heavy rain terminal early warning system based on deep learning, applied to the heavy rain terminal early warning method based on deep learning, comprising a data perception layer, an operation and maintenance layer, a core calculation layer and an early warning decision layer connected in turn. The data perception layer is used for collecting rainfall perception data and image data in a target area, wherein the rainfall perception data comprises rainfall amount data and temperature and humidity data. The operation and maintenance layer is used for identifying rainfall state based on the image data, outputting rainfall scene verification results, and screening effective rainfall perception data according to the rainfall scene verification results. The core calculation layer is used for inputting the effective rainfall perception data into a double-branch deep learning model based on cross attention, performing feature extraction and pattern matching, and outputting a heavy rain return period corresponding to the target area and a dynamic threshold correction coefficient for correcting the engineering threshold. The early warning decision layer is used for performing parameter adjustment on the engineering threshold of the current area based on the dynamic threshold correction coefficient, obtaining a dynamic threshold, comparing real-time rainfall amount data with the dynamic threshold, generating heavy rain early warning information according to the comparison result, and controlling a terminal display device to complete early warning action output according to the early warning information.

[0038] Specifically, the heavy rain terminal early warning system based on deep learning adopts a module integrated design, and the operation and maintenance layer, the core calculation layer and the early warning decision layer are integrated and deployed in a screen host device with embedded computing capability and graphical visualization output capability. The screen host can be an all-in-one machine device with an edge computing chip set, supporting Linux or a lightweight operating system environment, integrating a local neural network inference engine and a graphical user interface display module, which can not only perform inference calculation of the deep learning model locally, but also can present the early warning result and the data visualization interface in real time, meeting the requirements of independent operation and information prompt in a disconnected or weak network environment. The data perception layer and the screen host are connected through a low-power wireless communication protocol or a wired way to realize data access of multi-source perception devices. The rainfall amount data is collected through a tipping bucket rain gauge, the temperature and humidity data are collected by distributed temperature and humidity sensors deployed in the target area, and the image data are obtained by fixed cameras. All perception devices support communication with the screen host through LoRa, and the data is uniformly scheduled and cached by the host, and forms a standardized input format for deep learning processing. This integrated deployment method significantly reduces the dependence on cloud platforms and remote networks, improves the application accessibility and response timeliness of the system in remote and weak infrastructure areas, and has stronger autonomous early warning capability and on-site disaster prevention support capability.

[0039] The above embodiments are only used to describe the preferred embodiments of the present application, and are not used to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements made by those skilled in the art to the technical solutions of the present application shall fall within the protection scope of the claims of the present application.

Claims

1. A rainstorm terminal warning method based on deep learning, characterized in that: include: S1. Collecting rainfall perception data and image data in the target area, wherein the rainfall perception data includes rainfall data and temperature and humidity data; S2. Recognize rainfall status based on the image data, output a rainfall scene verification result, and filter valid rainfall perception data according to the rainfall scene verification result; S3. Input the effective rainfall perception data into a two-branch deep learning model based on cross-attention, perform feature extraction and pattern matching, and output the rainstorm return period and dynamic threshold correction coefficient corresponding to the target area; S4. Based on the dynamic threshold correction coefficient, the engineering threshold of the current area is adjusted to obtain a dynamic threshold, and the real-time rainfall data is compared with the dynamic threshold. According to the comparison result, heavy rain warning information is generated, and the terminal display device is controlled to complete the warning action output according to the heavy rain warning information.

2. The method for terminal early warning of rainstorm based on deep learning according to claim 1, characterized in that: The rainfall data is obtained by setting a tipping bucket rain gauge in the target area, and the temperature and humidity data is obtained by distributing temperature and humidity sensors in the target area.

3. The method for terminal early warning of rainstorm based on deep learning according to claim 2, characterized in that: The image data is acquired by setting a camera in the target area, and the image data includes an image of the setting area of ​​the tipping bucket rain gauge.

4. The method for terminal early warning of rainstorm based on deep learning according to claim 3, characterized in that: The rain state recognition based on the image data comprises the following steps: Perform structured detection on the original image data to identify and extract the spatial position of the tipping bucket rain gauge in the image and obtain the corresponding tipping bucket image area; An image sequence is constructed based on the dump bucket image area, and dump bucket state recognition is performed on continuous image frames through inter-frame pixel difference and edge contour extraction algorithms, and the dump bucket flip start and end frames, flip amplitude and flip frequency are extracted to generate structural motion feature data representing the dump bucket action state; Based on the background image area excluding the tipping bucket area in the original image data, a convolutional neural network model is used to extract rainfall semantic feature vectors of rain curtain texture, water droplet trajectory, and ground water mark diffusion; The consistency of the structural motion feature data and the rainfall semantic feature vector is judged. If the image within the time period corresponding to the tipping bucket flipping behavior contains a semantic feature distribution consistent with natural rainfall, the rainfall scene verification result is generated as natural rainfall trigger; if there is a deviation between the tipping bucket structural motion feature and the semantic feature distribution, the rainfall scene verification result is generated as unnatural rainfall trigger.

5. The method for terminal early warning of rainstorm based on deep learning according to claim 4, characterized in that: The dual-branch deep learning model based on cross-attention is constructed by the following steps: Based on effective rainfall perception data verified by rainfall scenarios in historical rainstorm events, a coefficient scan is performed on the rainfall sequence of each event and the corresponding engineering threshold. The false alarm rate and false negative rate of each candidate coefficient are calculated to obtain the optimal threshold correction coefficient that minimizes the false alarm rate and the false negative rate does not exceed the preset upper limit. The optimal threshold correction coefficient and the actual rainstorm return period level of the corresponding rainstorm event are combined to form a supervision label set. The terrain factor is constructed based on the terrain type, altitude and environmental category of the rainstorm event location, and the rainfall series and temperature and humidity series are constructed based on the effective rainfall perception data of the rainstorm event location. Construct a multi-source training sample set through terrain factors, rainfall series, and temperature and humidity series, and associate the multi-source training sample set with a supervision label set; Based on the multi-source training sample set, temporal position embedding and cross-channel attention fusion are performed on the rainfall sequence, temperature and humidity sequence and terrain factor encoding to obtain a spatiotemporal joint feature sequence; Based on the spatiotemporal joint feature sequence, a backbone network including residual connections is input to obtain a prediction vector for return period prediction and a correction vector for dynamic threshold regression, and the return period error and threshold error are calculated according to the supervision label; Gradient optimization is performed on the network parameters through the loss function to obtain a dual-branch deep learning model.

6. The method for terminal early warning of rainstorm based on deep learning according to claim 5, characterized in that: The loss function is as follows: ； in, is the loss value; is the number of training samples; For the The predicted return period level of the samples; For the The true return period level label of each sample; For the The prediction threshold correction coefficient of the sample; For the The optimal threshold correction coefficient label for each sample; For the The historical false alarm rate weight corresponding to the sample; is the weight factor.

7. The method for terminal early warning of rainstorm based on deep learning according to claim 1, characterized in that: Generating rainstorm warning information according to the comparison results includes: Real-time rainfall data is obtained and compared with the dynamic threshold. If the real-time rainfall data is greater than or equal to the dynamic threshold, a rainstorm warning message is generated; if the rainfall data in the previous time period is less than the dynamic threshold, no rainstorm warning message is generated.

8. The method for terminal early warning of rainstorm based on deep learning according to claim 7, characterized in that: The rainstorm warning information includes: target area identification, warning trigger time, real-time rainfall data, dynamic threshold, predicted rainstorm recurrence period level and warning level identification.

9. The method for terminal early warning of rainstorm based on deep learning according to claim 8, characterized in that: The warning actions include: visual interface color change, pop-up prompts, voice broadcasts and buzzer sounds.

10. A rainstorm terminal warning system based on deep learning, applied to a rainstorm terminal warning method based on deep learning according to any one of claims 1 to 9, characterized in that: include: The data perception layer, operation and maintenance layer, core computing layer and early warning decision layer are connected in sequence; The data perception layer is used to collect rainfall perception data and image data in the target area, wherein the rainfall perception data includes rainfall data and temperature and humidity data; An operation and maintenance layer, configured to identify rainfall status based on the image data, output a rainfall scene verification result, and filter effective rainfall perception data according to the rainfall scene verification result; A core computing layer, configured to input the effective rainfall perception data into a two-branch deep learning model based on cross-attention, perform feature extraction and pattern matching, and output a rainstorm return period corresponding to the target area and a dynamic threshold correction coefficient for correcting the engineering threshold; The early warning decision layer is used to perform parameter adjustment on the engineering threshold of the current area based on the dynamic threshold correction coefficient to obtain the dynamic threshold, and compare the real-time rainfall data with the dynamic threshold, generate heavy rain warning information according to the comparison results, and control the terminal display device to complete the warning action output according to the warning information.

Images