Unmanned aerial vehicle water depth monitoring method for channel inspection

Abstract

The invention relates to the technical field of channel monitoring, and discloses an unmanned aerial vehicle water depth monitoring method for channel inspection. The method comprises the following steps: deploying an unmanned aerial vehicle to a target channel area, initializing a carried sonar sensor, a camera and a positioning module, and planning a flight path to cover a to-be-detected point according to a preset algorithm; water depth and image data are collected in real time during flight, position information is obtained, multi-modal monitoring data are formed through space-time alignment, a water depth fusion model is input to generate a channel water depth distribution diagram, an abnormal area is recognized through an abnormal detection model, and finally the position and water depth data of the abnormal area are encrypted and transmitted to a ground control terminal. The method improves the efficiency and accuracy of channel water depth monitoring, can timely discover the abnormality, guarantees the safety and smoothness of the channel, and provides an effective technical means for channel management.

Description

Technical Field

[0001] The present invention relates to the technical field of waterway monitoring, and in particular to a method for water depth monitoring using an unmanned aerial vehicle (UAV) for waterway inspection. Background Art

[0002] In the modern water transport system, the safety and smooth flow of waterways play a vital role in economic development. As a key indicator reflecting the navigation conditions of waterways, accurate monitoring of waterway depth has always been an important topic in the field of water transport. Traditional waterway depth monitoring methods mainly rely on manual measurement and fixed monitoring stations. Manual measurement usually relies on survey ships, and staff use depth sounders to measure the water depth point by point in the waterway. This method is not only inefficient, but also greatly restricted by weather and hydrological conditions. In severe weather such as heavy rain and strong winds and waves, measurement work is difficult to carry out. In addition, the labor cost is high, requiring the participation of a large number of professionals, and the measurement cycle is long, and it is impossible to obtain real-time water depth information of large waterways in a timely manner.

[0003] Although fixed monitoring stations can achieve continuous monitoring of local waterway depths to a certain extent, there is a problem of limited monitoring range. Waterways are widely distributed and the terrain is complex and changeable. It is difficult to fully cover all important areas by relying solely on fixed monitoring stations, and there are a large number of monitoring blind spots. For example, in some remote or complex waterway sections, it is difficult to set up fixed monitoring stations due to factors such as construction and maintenance costs, resulting in the inability to timely grasp the changes in water depth in these areas.

[0004] With the rapid development of drone technology, its application in various fields is becoming more and more extensive. In terms of waterway monitoring, drones have gradually become a potential monitoring method due to their advantages such as strong maneuverability, high flexibility and rapid deployment. However, the application of drones in waterway depth monitoring still faces many challenges. First of all, how to plan the flight path of drones in the waterway area to ensure that all test points can be covered efficiently and comprehensively while taking into account energy consumption and flight efficiency is a key issue. Secondly, the various monitoring equipment carried by drones, such as sonar sensors, cameras and positioning modules, will produce temporal and spatial differences during the data collection process. How to accurately align these multimodal data in time and space for subsequent effective fusion analysis is an important link in achieving accurate water depth monitoring. Furthermore, due to the complexity of the measurement environment, the data collected by the sonar sensor inevitably has noise interference. How to build an effective model to fuse the water depth data at different locations and eliminate the measurement noise, so as to generate an accurate waterway depth distribution map, is a technical problem that needs to be solved urgently. In addition, after obtaining the water depth distribution map, how to quickly and accurately identify abnormal areas, such as siltation areas and underwater obstacles, and transmit relevant information to the ground control terminal in a timely and reliably manner is also an important factor restricting the development of UAV waterway depth monitoring technology.

[0005] During channel inspection, real-time and accurate water depth monitoring can effectively evaluate the navigability of the channel, timely detect potential safety hazards such as siltation and underwater obstacles, provide data support for dredging operations and navigation warnings, thus ensuring the safety of ship navigation and reducing grounding or reef hitting accidents caused by insufficient water depth. Dynamically obtaining high-resolution water depth data through drones can significantly improve the inspection efficiency, make up for the spatio-temporal limitations of traditional monitoring methods, provide a scientific basis for channel maintenance decision-making, and have important application value for optimizing the navigation environment and reducing operating costs. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for drone water depth monitoring for channel inspection to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above purpose, the present invention provides the following technical solution: A method for drone water depth monitoring for channel inspection, the method includes: Deploy the drone to the target channel area and initialize the monitoring equipment carried by the drone, wherein the monitoring equipment at least includes a sonar sensor, a camera and a positioning module; Generate the flight path of the drone based on a preset inspection path planning algorithm, wherein the flight path covers all the measurement points within the target channel area; During the flight of the drone along the flight path, use the sonar sensor to collect water depth data in real time and use the camera to synchronously obtain the image data of the channel surface; Obtain the position information of the drone through the positioning module, and align the water depth data, the image data and the position information in space and time to form multi-modal monitoring data; Input the multi-modal monitoring data into a preset water depth fusion model to generate a water depth distribution map of the channel, wherein the water depth fusion model is used to fuse the water depth data at different positions and eliminate measurement noise; Use a preset anomaly detection model to analyze the water depth distribution map of the channel and identify the abnormal areas in the channel, wherein the abnormal areas include siltation areas or underwater obstacles; Transmit the position information of the abnormal area and the corresponding water depth data to the ground control terminal.

[0008] Preferably, generating the flight path of the drone based on a preset inspection path planning algorithm includes: Obtain the geographical boundary information and historical water depth data of the target channel area; Divide the target channel area into multiple sub-areas according to the geographical boundary information and assign a priority weight to each sub-area, wherein the priority weight is determined based on the anomaly frequency in the historical water depth data; The path planning model is trained using a reinforcement learning algorithm, where the reward function of the path planning model is related to the priority weight, path length, and energy consumption; Input the geographical coordinates of the sub-region into the path planning model, and output the flight path sequence of the UAV.

[0009] Preferably, the training process of the reinforcement learning algorithm includes: Construct a state space, where the state space includes the current position of the UAV, remaining battery power, priority weights of the covered sub-regions and uncovered sub-regions; Define an action space, where the action space includes the UAV moving to an adjacent sub-region or returning to the charging station; Calculate the immediate reward for each state-action pair according to the reward function, and update the policy network parameters through the Q-learning algorithm; Stop training when the path planning model reaches the preset path coverage threshold in the simulation environment.

[0010] Preferably, the spatio-temporal alignment of the water depth data, the image data, and the position information includes: Extract the acquisition timestamp of the water depth data and the shooting timestamp of the image data; Perform interpolation processing on the position information according to the positioning frequency of the positioning module to generate a continuous position sequence matching the acquisition timestamp and the shooting timestamp; Map the water depth data and the image data to the corresponding time points of the continuous position sequence respectively to form a spatio-temporal correlation dataset.

[0011] Preferably, the interpolation processing for generating the continuous position sequence uses a cubic spline interpolation algorithm, which specifically includes: Construct a position-time curve based on the discrete position points recorded by the positioning module; Insert a cubic polynomial function between adjacent discrete position points to make the first and second derivatives of the continuous position sequence continuous.

[0012] Preferably, the construction process of the water depth fusion model includes: Collect historical water depth data of the waterway and corresponding multi-modal monitoring data as training samples; Perform normalization processing on the training samples and divide them into a training set and a validation set; Construct a deep neural network model, where the input layer of the model includes the water depth data, image feature vectors, and position coordinates, and the output layer is the predicted water depth value; The mean squared error is used as the loss function, and the deep neural network model is iteratively trained using the training set until the prediction error of the validation set is lower than a preset threshold.

[0013] Preferably, the extraction of the image feature vector uses a pre-trained convolutional neural network, specifically including: Input the waterway surface image data into the convolutional neural network, and extract the output feature map of the last convolutional layer; Perform global average pooling on the feature map to generate a feature vector with a fixed dimension.

[0014] Preferably, the construction of the anomaly detection model includes: Obtain samples of normal waterway depth distribution maps and anomaly samples, and label the anomaly types for the anomaly samples; Input the samples into an autoencoder network for unsupervised training, where the reconstruction error of the autoencoder network is used to distinguish normal and abnormal data; Set a dynamic threshold based on the reconstruction error. When the reconstruction error of the waterway depth distribution map exceeds the dynamic threshold, it is determined as an abnormal area.

[0015] Preferably, the method for setting the dynamic threshold includes: Statistically analyze the distribution of the reconstruction errors of normal samples, and calculate their mean and standard deviation; Set the dynamic threshold to the mean plus three times the standard deviation.

[0016] Preferably, the location information of the abnormal area and the corresponding water depth data are transmitted to the ground control terminal through an encrypted communication module; the data transmission process of the encrypted communication module includes: Perform block processing on the location information of the abnormal area and the water depth data to generate multiple data packets; Generate a unique identifier for each data packet, and encrypt the data packet based on the national cryptographic algorithm; Send the encrypted data packets to the ground control terminal through a multi-hop transmission protocol, where the multi-hop transmission protocol dynamically selects relay nodes according to the channel quality.

[0017] Compared with the prior art, the beneficial effects of the present invention are: The method for monitoring water depth of an unmanned aerial vehicle for waterway inspection proposed by the present invention solves the problems existing in traditional waterway depth monitoring and existing unmanned aerial vehicle monitoring applications from multiple aspects, showing significant beneficial effects.

[0018] In terms of monitoring efficiency, the traditional manual measurement method relies on a survey ship to measure point by point, which is slow and easily restricted by weather and hydrological conditions. Fixed monitoring stations also have limitations in the monitoring range. Through the preset inspection path planning algorithm, the present invention can divide the area into multiple sub-areas and assign priority weights according to the geographical boundary information and historical water depth data of the target waterway area, and use the reinforcement learning algorithm to train the path planning model. This enables the unmanned aerial vehicle (UAV) to fly along the optimal path, comprehensively covering all points to be measured, greatly improving the monitoring efficiency. Taking a 100-kilometer waterway as an example, traditional manual measurement may take several days or even a week, and it needs to be interrupted in case of bad weather. However, the UAV monitoring method of the present invention may only take a few hours to complete the preliminary monitoring in good weather, significantly shortening the monitoring cycle and enabling timely acquisition of the water depth information of a large area of the waterway.

[0019] In terms of data accuracy, the water depth data collected by sonar sensors is vulnerable to noise interference, affecting the reliability of monitoring results. The water depth fusion model constructed by the present invention uses the historical water depth data of the waterway and the corresponding multi-modal monitoring data as training samples, performs normalization processing and divides the training set and the validation set, and uses a deep neural network model for training. The input layer of the model comprehensively considers the water depth data, image feature vectors, and position coordinates, and the output layer predicts the water depth value. The mean square error is used as the loss function for iterative training until the prediction error of the validation set is lower than the preset threshold. This method effectively fuses the water depth data at different positions and eliminates measurement noise, generating a more accurate water depth distribution map of the waterway. For example, in an area of a certain waterway with complex water flow and underwater terrain, the water depth data obtained by traditional measurement methods has a large error, while the method of the present invention can accurately present the true water depth situation in this area, with the error controlled within a very small range, providing reliable data support for waterway management and ship navigation.

[0020] In terms of anomaly detection capabilities, previous monitoring methods often required manual post-analysis of a large amount of data when identifying abnormal areas in waterways, resulting in low efficiency and easy omission. The present invention utilizes a preset anomaly detection model. By obtaining samples of normal and abnormal water depth distribution maps of waterways, after annotating the types of abnormal samples, they are input into an autoencoder network for unsupervised training, and a dynamic threshold is set based on the reconstruction error. When the reconstruction error of the water depth distribution map of the waterway exceeds the threshold, it can be quickly and accurately determined as an abnormal area, including siltation areas or underwater obstacles. This greatly improves the recognition efficiency and accuracy of abnormal areas, and timely discovers potential waterway safety hazards. For example, in a waterway near a certain port, due to frequent loading and unloading of ships, siltation is likely to occur. The anomaly detection model of the present invention can timely discover these siltation areas, providing a basis for the waterway management department to arrange dredging work in a timely manner and ensuring the normal navigation of the waterway. The present invention dynamically optimizes the UAV inspection path through a reinforcement learning algorithm, preferentially covering areas with a high incidence of historical siltation, combines the spatio-temporal alignment technology of sonar and visual data, and effectively improves the accuracy and coverage density of water depth data collection; the water depth fusion model constructed by a deep neural network can eliminate noise interference in a complex hydrological environment, generate a high-resolution water depth distribution map of the waterway, and quickly locate siltation areas through an anomaly detection model driven by an autoencoder, providing centimeter-level accuracy siltation coordinates and depth data for the waterway management department. This technology can achieve fully automated dynamic monitoring of waterway water depth, shorten the identification of abnormal areas and data transmission delay, assist the waterway management department in accurately formulating dredging plans, avoid the problem of lagging waterway maintenance caused by the long traditional manual inspection cycle, and ensure that the navigable water depth index continuously meets the standards.

[0021] In terms of data transmission, the present invention uses an encrypted communication module to transmit the location information and water depth data of the abnormal area to the ground control terminal. The data is block-processed, a unique identifier is generated, and encrypted based on the national cryptographic algorithm, and relay nodes are dynamically selected according to the channel quality through a multi-hop transmission protocol. This method not only ensures the security and integrity of the data during transmission, prevents the data from being stolen or tampered with, but also improves the reliability of data transmission, ensuring that the ground control terminal can obtain abnormal information in a timely and accurate manner, facilitating relevant personnel to quickly take measures to respond, and ensuring the safety and unobstructedness of the waterway. For example, in some important waterways, near military-sensitive areas or waterways with busy commercial transportation, data security is crucial. The encrypted transmission method of the present invention can effectively protect data security and avoid safety accidents caused by data leakage or transmission errors. Brief Description of the Drawings

[0022] Figure 1 It is a processing step diagram of the UAV water depth monitoring method for waterway inspection described in the present invention; Figure 2 It is a flowchart of the flight path planning method; Figure 3 It is a diagram of the data spatio-temporal alignment method; Figure 4 It is a diagram for constructing an anomaly detection model. Specific implementation manners

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0024] Please refer to Figures 1 - 4 , the present invention provides a technical solution: an unmanned aerial vehicle water depth monitoring method for channel inspection, and its specific implementation manners are elaborated in detail below.

[0025] Before channel inspection, the unmanned aerial vehicle is deployed to the target channel area. The unmanned aerial vehicle is equipped with a variety of key monitoring devices, including a sonar sensor, a camera, and a positioning module. After being deployed in place, initialization operations are performed on these monitoring devices to ensure that each device can work normally and prepare for subsequent data collection. For example, the sonar sensor needs to calibrate the transmission and reception frequencies, the camera needs to adjust the focal length and exposure parameters, and the positioning module needs to search for and lock the satellite signal to obtain accurate positioning information.

[0026] According to the preset inspection path planning algorithm, the flight path of the unmanned aerial vehicle is generated. This algorithm will fully consider the characteristics of the target channel area and plan a flight path that can cover all the points to be measured in the target channel area. This can ensure comprehensive and non-missing monitoring of the entire channel and lay a foundation for subsequent acquisition of accurate water depth data and channel surface image data.

[0027] When the unmanned aerial vehicle flies along the planned flight path, the sonar sensor starts to play a role and collects water depth data in real time. The sonar sensor calculates the water depth by emitting sound waves underwater and receiving the reflected sound wave signals and based on the time of sound wave propagation. At the same time, the camera works synchronously to obtain channel surface image data for recording the conditions of the channel surface, such as whether there are floating objects and the ship navigation situation, etc.

[0028] During the data collection process, the positioning module will obtain the position information of the unmanned aerial vehicle. In order to make the collected water depth data, image data, and position information correspond to each other, spatio-temporal alignment is required. Through specific processing methods, the data collected at different times is matched with the corresponding positions to form multi-modal monitoring data, which is convenient for subsequent analysis and processing.

[0029] Input the multi-modal monitoring data that has been spatio-temporally aligned into a preset water depth fusion model. This model can fuse water depth data from different positions and effectively eliminate the noise generated during the measurement process, thereby generating an accurate water depth distribution map of the waterway, visually showing the water depth conditions at different positions of the waterway.

[0030] Use a preset anomaly detection model to analyze the generated water depth distribution map of the waterway. Through this model, abnormal areas in the waterway can be identified, and these abnormal areas include siltation areas or underwater obstacles, etc., to timely detect potential safety hazards in the waterway.

[0031] Once an abnormal area is identified, transmit the location information of the abnormal area and the corresponding water depth data to the ground control terminal. Ground control personnel can take corresponding measures in a timely manner based on this data, such as arranging dredging work or setting warning signs to ensure the safe and unobstructed passage of the waterway.

[0032] The present invention will be further described below in conjunction with Embodiments 1 to 5:

[0033] Embodiment 1:

[0034] This embodiment mainly focuses on the flight path planning algorithm. In practical applications, obtaining the geographical boundary information and historical water depth data of the target waterway area is the basis for flight path planning. The geographical boundary information can be obtained through a Geographic Information System (GIS), which defines the flight range of the unmanned aerial vehicle, avoiding inaccurate data collection or safety accidents caused by the unmanned aerial vehicle flying out of the waterway area. The historical water depth data records the past water depth conditions of the waterway, and these data are stored in a database and can be called at any time.

[0035] Divide multiple sub-regions according to the geographical boundary information. This step is to more carefully plan the flight path. For example, for a long and irregularly shaped waterway, it can be divided into several rectangular or polygonal sub-regions according to factors such as its curvature and width change. Assign a priority weight to each sub-region, where the priority weight is determined based on the anomaly frequency in the historical water depth data. Assume that the frequency of anomalies (such as sudden shallowing of water depth or abnormal fluctuations) in sub-region A in the historical water depth data is , the total number of anomalies in this waterway area is , the area of sub-region A is , and the area of the entire target waterway area is , then the priority weight of sub-region A can be calculated through the formula . Here, represents the number of historical anomalies in sub-region A, represents the total number of historical anomalies in the entire waterway area, represents the area of sub-region A, represents the total area of the target waterway area. For the weights calculated by this formula, sub-regions with a high anomaly frequency and a large area will obtain higher priority weights.

[0036] The path planning model is trained using the reinforcement learning algorithm. The reward function of the path planning model is related to the priority weight, path length, and energy consumption. Let the reward function be , the sub-region has a priority weight of , the UAV moves from the sub-region to the sub-region with a path length of , and the energy consumption is , then the reward function can be expressed as . Among them, , , are weight coefficients used to adjust the importance of each factor in the reward function. For example, if currently more attention is paid to covering high-priority areas, the value of can be appropriately increased; if it is desired to reduce energy consumption, the weight of can be increased. By continuously adjusting the weight coefficients, the reward function can better meet the actual needs.

[0037] The geographical coordinates of the sub-regions are input into the path planning model, and the flight path sequence of the UAV is output. In actual operation, the geographical coordinates (such as longitude and latitude) of each divided sub-region are input into the trained path planning model in a certain format. The model will output an optimized flight path sequence according to the strategies learned during the previous training, guiding the UAV to efficiently complete the waterway inspection task.

[0038] Example 2: This example details the training process of the reinforcement learning algorithm. During the training process, constructing the state space is one of the key steps. The state space includes the current position of the UAV, the remaining battery power, the priority weights of the covered sub-regions and the uncovered sub-regions. The current position of the UAV can be obtained in real time through the positioning module, represented by longitude and latitude coordinates, denoted as . The remaining battery power can be obtained through the UAV battery management system, represented as a percentage of the battery power, denoted as . The covered sub-regions can be represented by a set , and the elements in the set are the numbers of the covered sub-regions. The priority weights of the uncovered sub-regions are a vector corresponding to the number of uncovered sub-regions, and each element in the vector represents the priority weight of the corresponding uncovered sub-region.

[0039] Define the action space, which includes the UAV moving to an adjacent sub-region or returning to the charging station. Assume the UAV is currently located in sub-region , and its set of adjacent sub-regions is . Then the action of moving to an adjacent sub-region can be represented as , where indicates that the UAV moves from sub-region to sub-region . The action of returning to the charging station is represented as .

[0040] Calculate the immediate reward for each state-action pair according to the reward function, and update the policy network parameters through the Q-learning algorithm. The core formula of the Q-learning algorithm is . Among them, represents the Q value of executing action in state , is the learning rate, which controls the step size of each update; is the immediate reward obtained after executing action , and the immediate reward is calculated according to the previously defined reward function ; is the discount factor, which is used to balance the importance of current rewards and future rewards, and its value range is between ; is the new state transferred to after executing action , is the optimal action in the new state .

[0041] During the training process, continuously repeat the above steps. When the path planning model reaches the preset path coverage threshold in the simulation environment, stop the training. For example, the preset path coverage threshold is , that is, when the path planned by the model can cover or more of the points to be measured in the target waterway area, it is considered that the training has achieved the expected effect and stop the training. At this time, the obtained policy network parameters can be used for the actual UAV flight path planning.

[0042] Example 3: This example focuses on the specific implementation of data spatio-temporal alignment and interpolation processing. When performing data spatio-temporal alignment, first extract the acquisition timestamp of the water depth data and the shooting timestamp of the image data. The timestamp is a mark that records the moment of data acquisition or shooting, usually accurate to the second or millisecond level. For example, the water depth data acquisition timestamp records the specific moment when the sonar sensor collects the water depth data, and the image data shooting timestamp records the moment when the camera takes the image.

[0043] Interpolate the position information according to the positioning frequency of the positioning module to generate a continuous position sequence that matches the acquisition timestamp and the shooting timestamp. The positioning frequency of the positioning module refers to the time interval at which it obtains the position information of the UAV. Assume the positioning frequency is , that is, obtain the position information once every time. The collected position information is discrete points. In order to make the position information accurately match the water depth data and the image data, interpolation processing is required.

[0044] Map the water depth data and the image data to the corresponding time points of the continuous position sequence respectively to form a spatio-temporal correlation dataset. For example, for the water depth data at a certain moment , find the time point closest to in the continuous position sequence, and associate with the position corresponding to this time point ; perform a similar operation on the image data, associate the image data with the corresponding position, and finally form a spatio-temporal correlation dataset for subsequent comprehensive analysis.

[0045] Example 4: This example focuses on the construction process of the water depth fusion model. When constructing the water depth fusion model, first collect historical water depth data of the waterway and the corresponding multi-modal monitoring data as training samples. The historical water depth data of the waterway can be obtained from past waterway monitoring records. The multi-modal monitoring data includes the waterway surface image data collected at that time, the UAV position information, etc. These data are stored in the data warehouse, providing rich materials for model training.

[0046] Normalize the training samples and divide them into a training set and a validation set. Normalization is to make data with different features have the same scale, which is convenient for model training and convergence. For example, for the water depth data , the formula can be used for normalization, where and are the minimum and maximum values in the water depth data respectively. Divide the processed data into a training set and a validation set according to a certain ratio, such as a ratio of 70% and 30%. The training set is used to train the model, and the validation set is used to evaluate the performance of the model.

[0047] Build a deep neural network model. Among them, the input layer of the model includes water depth data, image feature vectors, and position coordinates, and the output layer is the predicted water depth value. The extraction of image feature vectors uses a pre-trained convolutional neural network. Input the waterway surface image data into the convolutional neural network, and extract the output feature map of the last convolutional layer. Through multiple convolutional and pooling operations, the convolutional neural network can automatically extract features in the image. For example, for the commonly used VGG16 network, the output feature map of its last convolutional layer contains rich image feature information. Perform global average pooling on the feature map to generate a feature vector with a fixed dimension. Assume that the size of the output feature map of the last convolutional layer is , through global average pooling, average the feature maps of each channel to obtain a fixed-dimension feature vector with a length of , , where is the pixel value of the -th channel, the -th row, and the -th column in the feature map.

[0048] Use the mean squared error as the loss function, and use the training set to iteratively train the deep neural network model until the prediction error of the validation set is lower than the preset threshold. The formula for the mean squared error loss function is , where is the number of training samples, is the true water depth value, and is the water depth value predicted by the model. During the training process, continuously adjust the parameters of the model through the backpropagation algorithm to make the loss function value continuously decrease. When the prediction error of the validation set is lower than the preset threshold, such as when the mean squared error is less than 0.1, it is considered that the model training has achieved a good effect and can be used for actual water depth fusion and prediction.

[0049] Example 5: This example elaborates in detail on the construction of the anomaly detection model and the data transmission process. When constructing the anomaly detection model, first obtain samples of the normal waterway water depth distribution map and anomaly samples, and label the anomaly types for the anomaly samples. Normal samples can be obtained from the monitoring data when the waterway conditions were good in history, and anomaly samples are screened from the monitoring data where there have been situations such as siltation and underwater obstacles. For example, for a siltation sample, label its anomaly type as "siltation" and record information such as the location and degree of siltation.

[0050] Input the samples into the autoencoder network for unsupervised training. Among them, the reconstruction error of the autoencoder network is used to distinguish normal and abnormal data. The autoencoder consists of an encoder and a decoder. The encoder compresses the input data into a low-dimensional representation, and the decoder then reconstructs the low-dimensional representation into the original data. Let the input sample be , a low-dimensional representation is obtained through an encoder , and then it is reconstructed through a decoder to obtain , the reconstruction error can be calculated using the formula . Here represents the Euclidean norm. During the training process, the reconstruction error of normal samples is small, while that of abnormal samples is large.

[0051] Based on the reconstruction error, a dynamic threshold is set. When the reconstruction error of the water depth distribution map of the waterway exceeds the dynamic threshold, it is determined as an abnormal area. The method for setting the dynamic threshold includes: statistically analyzing the distribution of the reconstruction errors of normal samples and calculating their mean and standard deviation . Suppose the number of normal samples is , and the reconstruction error is , then the mean , and the standard deviation . The dynamic threshold is set to the mean plus three times the standard deviation, that is . When a new water depth distribution map of the waterway is input into the model, its reconstruction error is calculated. If the reconstruction error is greater than , then this area is determined as an abnormal area.

[0052] The location information of the abnormal area and the corresponding water depth data are transmitted to the ground control terminal through the encrypted communication module. The data transmission process of the encrypted communication module includes: performing block processing on the location information and water depth data of the abnormal area to generate multiple data packets. For example, the data is divided according to a certain length. Suppose the length of each data packet is , and the total length of the data is , then the number of data packets , where represents rounding up. A unique identifier is generated for each data packet, and the data packets are encrypted based on the national cryptographic algorithm. The national cryptographic algorithm such as the SM4 algorithm has high security and encryption efficiency. Taking the SM4 algorithm as an example, the data packets are encrypted with a specific key to ensure the security of the data during transmission. The encrypted data packets are sent to the ground control terminal through the multi-hop transmission protocol. Among them, the multi-hop transmission protocol dynamically selects relay nodes according to the channel quality. Suppose there are multiple relay nodes , the multi-hop transmission protocol will monitor the channel quality between each relay node and the sender and receiver in real time, such as signal strength, bit error rate and other indicators, and select the relay node with the best channel quality for data forwarding to ensure that the data can be accurately and quickly transmitted to the ground control terminal.

[0053] It should be noted that, in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0054] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An unmanned aerial vehicle water depth monitoring method for waterway inspection, characterized in that, Including: Deploying a drone to a target waterway area and initializing the monitoring equipment carried by the drone, where the monitoring equipment at least includes a sonar sensor, a camera, and a positioning module; Generating a flight path of the drone based on a preset inspection path planning algorithm, where the flight path covers all measurement points within the target waterway area; During the flight of the drone along the flight path, using the sonar sensor to collect water depth data in real time, and using the camera to synchronously obtain waterway surface image data; Obtaining the position information of the drone through the positioning module, and performing spatio-temporal alignment on the water depth data, the image data, and the position information to form multi-modal monitoring data; Inputting the multi-modal monitoring data into a preset water depth fusion model to generate a water depth distribution map of the waterway, where the water depth fusion model is used to fuse water depth data at different positions and eliminate measurement noise; Analyzing the water depth distribution map of the waterway using a preset anomaly detection model to identify abnormal areas in the waterway, where the abnormal areas include siltation areas or underwater obstacles; Transmitting the position information of the abnormal area and the corresponding water depth data to a ground control terminal.

2. The method according to claim 1, characterized in that, The generating the flight path of the drone based on a preset inspection path planning algorithm includes: Obtaining the geographical boundary information and historical water depth data of the target waterway area; Dividing multiple sub-areas according to the geographical boundary information and assigning a priority weight to each sub-area, where the priority weight is determined based on the anomaly frequency in the historical water depth data; Training a path planning model using a reinforcement learning algorithm, where the reward function of the path planning model is related to the priority weight, path length, and energy consumption; Inputting the geographical coordinates of the sub-areas into the path planning model and outputting a flight path sequence of the drone.

3. The method according to claim 2, wherein The training process of the reinforcement learning algorithm includes: Constructing a state space, where the state space includes the current position of the drone, the remaining battery power, the covered sub-areas, and the priority weights of the uncovered sub-areas; Defining an action space, where the action space includes the drone moving to an adjacent sub-area or returning to a charging station; Calculating the immediate reward of each state-action pair according to the reward function and updating the policy network parameters through the Q-learning algorithm; Stopping the training when the path planning model reaches a preset path coverage threshold in the simulation environment.

4. The method according to claim 1, wherein The performing spatio-temporal alignment on the water depth data, the image data, and the position information includes: Extracting the acquisition timestamp of the water depth data and the shooting timestamp of the image data; Performing interpolation processing on the position information according to the positioning frequency of the positioning module to generate a continuous position sequence matching the acquisition timestamp and the shooting timestamp; Mapping the water depth data and the image data to the corresponding time points of the continuous position sequence respectively to form a spatio-temporal correlation data set.

5. The method according to claim 4, wherein The interpolation processing for generating the continuous position sequence adopts a cubic spline interpolation algorithm, specifically including: Constructing a position-time curve according to the discrete position points recorded by the positioning module; Insert a cubic polynomial function between adjacent discrete position points so that the first and second derivatives of the continuous position sequence are continuous.

6. The method according to claim 1, characterized in that, The construction process of the water depth fusion model includes: Collect historical water depth data of the waterway and corresponding multi-modal monitoring data as training samples; Normalize the training samples and divide them into a training set and a validation set; Construct a deep neural network model, where the input layer of the model includes the water depth data, image feature vectors, and position coordinates, and the output layer is the predicted water depth value; Use the mean squared error as the loss function and iteratively train the deep neural network model using the training set until the prediction error of the validation set is lower than a preset threshold.

7. The method according to claim 6, characterized in that The extraction of the image feature vectors uses a pre-trained convolutional neural network, specifically including: Input the waterway surface image data into the convolutional neural network and extract the output feature map of the last convolutional layer; Perform global average pooling on the feature map to generate a feature vector with a fixed dimension.

8. The method according to claim 1, characterized in that, The construction of the anomaly detection model includes: Obtain normal water depth distribution map samples and anomaly samples of the waterway, and label the anomaly types for the anomaly samples; Input the samples into an autoencoder network for unsupervised training, where the reconstruction error of the autoencoder network is used to distinguish normal and abnormal data; Set a dynamic threshold based on the reconstruction error, and determine an abnormal area when the reconstruction error of the water depth distribution map of the waterway exceeds the dynamic threshold.

9. The method according to claim 8, characterized in that The method for setting the dynamic threshold includes: Statistically analyze the distribution of the reconstruction errors of the normal samples and calculate their mean and standard deviation; Set the dynamic threshold to the mean plus three times the standard deviation.

10. The method according to claim 1, wherein Transmit the position information of the abnormal area and the corresponding water depth data to the ground control terminal through an encryption communication module; the data transmission process of the encryption communication module includes: Perform block processing on the position information and water depth data of the abnormal area to generate multiple data packets; Generate a unique identifier for each data packet and encrypt the data packet based on the national cryptographic algorithm; Send the encrypted data packets to the ground control terminal through a multi-hop transmission protocol, where the multi-hop transmission protocol dynamically selects relay nodes according to the channel quality.

Images