A multi-source information fusion water rescue task planning decision method and system

Abstract

The present application relates to the field of rescue equipment control, and provides a multi-source information fusion water rescue task planning decision method and system. The method comprises: obtaining original data collected by a multi-source sensor carried by a rescue robot, obtaining a multi-modal feature data set through data preprocessing and feature extraction; constructing a comprehensive information model based on the multi-modal feature data set, and outputting a fusion perception result; according to the fusion perception result, evaluating a rescue scheme through an auxiliary decision system, and generating a decision suggestion scheme; executing the decision suggestion scheme, and performing real-time tracking and dynamic adjustment on the rescue process to generate rescue execution feedback data. The present application realizes closed-loop optimization from perception to execution of water rescue tasks, and significantly improves rescue efficiency and success rate.

Description

Technical Field

[0001] This invention relates to the field of rescue equipment control technology, and in particular to a multi-source information fusion method and system for planning and decision-making in water rescue missions. Background Technology

[0002] In water rescue operations, accurately obtaining the location and status information of people in the water is a prerequisite for formulating effective rescue plans. Traditional water rescue systems mainly rely on single sensors for target detection and positioning, such as using only visible light cameras for visual recognition or relying solely on radar systems for distance measurement. Under good environmental conditions, a single sensor can provide basic target location information, supporting rescuers in making initial decision-making judgments. Existing rescue decision-making processes mainly rely on the experience and judgment of operators. Based on visual observation or information from a single sensor, operators manually assess the urgency and feasibility of the rescue, and then select rescue methods and plan rescue routes. In the selection of rescue methods, existing technologies mostly adopt fixed operating procedures, executing rescue tasks according to standard operating procedures, with little consideration for the actual status of the people in the water and the dynamic changes in environmental conditions. Regarding the monitoring of the rescue process, existing systems mainly track the rescue progress through manual observation or simple distance monitoring, lacking quantitative evaluation and real-time feedback mechanisms for key indicators of the rescue process. Some advanced rescue systems have begun to attempt to integrate data from multiple sensors, but these often remain at the data display level, failing to deeply integrate the complementary advantages of different modalities of information, nor have they established a systematic decision support model.

[0003] Existing rescue decision-making methods suffer from the following shortcomings: First, single sensors are prone to failure or accuracy degradation under specific environmental conditions, and the limitations of single sensors lead to a lack of reliable information support for rescue decisions. Second, traditional manual decision-making methods heavily rely on the experience and subjective judgment of operators. In high-pressure, high-time-sensitivity rescue scenarios, manual decision-making is prone to errors in judgment or delayed responses, lacking quantitative assessment of key factors such as environmental complexity and urgency, making it difficult to scientifically select the optimal rescue method and route. Third, existing rescue systems lack the ability to dynamically monitor and intelligently adjust the rescue process. Once a rescue plan is formulated, it is executed according to a fixed procedure. When abnormal situations such as sudden environmental changes, route deviations, or resource shortages occur during the actual rescue process, the system cannot identify the problem in time and dynamically adjust the rescue strategy, resulting in low rescue efficiency or even mission failure. Summary of the Invention

[0004] This invention provides a method and system for planning and making decisions on water rescue missions by fusing multi-source information, in order to overcome the shortcomings of existing technologies.

[0005] This invention provides a multi-source information fusion method for planning and decision-making in maritime rescue missions, comprising:

[0006] S1. Obtain raw data collected by the multi-source sensors on the rescue robot, and obtain a multimodal feature dataset through data preprocessing and feature extraction; S2. Construct a comprehensive information model based on the multimodal feature dataset and output the fusion perception results; S3. Based on the fusion perception results, the rescue plan is evaluated through the auxiliary decision-making system, and a decision-making suggestion plan is generated; S4. Execute the decision-making suggestion plan, track and dynamically adjust the rescue process in real time, and generate rescue execution feedback data.

[0007] According to the multi-source information fusion method for planning and decision-making in water rescue missions provided by the present invention, step S1 further includes: S11. Noise filtering and contrast enhancement are performed on the acquired image data. The contour features and geometric features of the target are extracted through edge detection and corner detection to obtain the image feature vector. S12. Perform temperature calibration and image sharpening on the acquired thermal imaging data, extract the regional and gradient features of the heat source distribution, and obtain the thermal feature vector. S13. Denoise and perform spectrum analysis on the collected audio data to extract the directional and intensity features of the sound source and obtain the acoustic feature vector. S14. Perform low-pass filtering and time synchronization on the collected navigation data, extract position coordinates and velocity information, and obtain navigation feature vectors; S15. Align the image feature vector, thermal feature vector, acoustic feature vector, and navigation feature vector according to the timestamp to form a multimodal feature dataset.

[0008] According to the multi-source information fusion method for planning and decision-making in water rescue missions provided by the present invention, step S2 further includes: S21. A Kalman filter is used to estimate the state of the navigation feature vectors in the multimodal feature dataset. The state equation of the Kalman filter is:

[0009] in, for The state vector at time t, Here is the state transition matrix. Assign a matrix to the noise. for Real-time system noise; S22. Using a deep neural network, perform feature-level fusion on the image feature vectors and hot feature vectors in the multimodal feature dataset, and output the target category probability and location confidence. S23. Cross-validate the orientation information provided by the acoustic feature vectors in the multimodal feature dataset with the position information provided by the visual features, and eliminate detection results with confidence levels lower than a preset threshold. S24. Integrate the fusion results, establish a comprehensive information model that includes the target's absolute position, relative velocity, and the distribution of environmental obstacles, and output the fusion perception results.

[0010] According to the multi-source information fusion method for planning and decision-making in water rescue missions provided by the present invention, step S22 further includes: S221. Input the image feature vector and the hot feature vector into two parallel convolutional neural network branches respectively to extract the high-level semantic features of the image and the hot high-level semantic features. S222. In the feature fusion layer, the high-level semantic features of the image and the hot high-level semantic features are added element-wise by channel concatenation to obtain a fused feature map. S223. Input the fused feature map into the fully connected layer, output the target class probability distribution and bounding box regression parameters, and obtain the fusion result including the target class probability and location confidence. S224. Perform non-maximum suppression processing on the fusion result to select the target detection box with the highest confidence.

[0011] According to the multi-source information fusion method for planning and decision-making in water rescue missions provided by the present invention, step S3 further includes: S31. Calculate the environmental complexity index based on the distribution of environmental obstacles and sea state parameters in the comprehensive information model. The expression for the environmental complexity index is:

[0012] in, This is an index representing environmental complexity. For the number of obstacles, The maximum obstacle threshold, Sea state rating The maximum sea state level, For wind speed, The maximum wind speed threshold, As the first weighting coefficient, This is the second weighting coefficient. This is the third weighting coefficient; S32. Calculate the rescue urgency index based on the target location and target status in the comprehensive information model; The expression for the urgency index of the rescue is:

[0013] in, As an indicator of the urgency of the rescue, For minimum safe distance, Current distance, This is the drift time. This is the critical time. As an indicator of physical activity, For the fourth weighting coefficient, The fifth weighting coefficient This is the sixth weighting coefficient; S33. Based on the environmental complexity index and the rescue urgency index, an initial rescue method is matched using a decision matrix; S34. Perform path feasibility analysis and time cost estimation on the initial rescue method to generate a path plan that includes execution steps and expected duration; S35. Evaluate the success probability of the proposed path scheme under different environmental disturbances through Monte Carlo simulation, obtain risk assessment data, and formulate a decision-making recommendation scheme.

[0014] According to the multi-source information fusion method for planning and decision-making in water rescue missions provided by the present invention, the matching rules for the decision matrix in step S33 specifically include: When the urgency index score is greater than 0.7 and the environmental complexity index is less than 0.5, the first rescue method is selected, which is the rapid approach rescue method. When the rescue urgency index score is greater than 0.7 and the environmental complexity index is greater than or equal to 0.5, the second rescue method is selected, which is the remote throwing of a life-saving device. When the rescue urgency index score is greater than or equal to 0.4 and less than or equal to 0.7, the third rescue method is selected, which is the standard approach and throwing of a life-saving device method; When the rescue urgency index score is less than 0.4 and there are multiple targets to be rescued, the fourth rescue method is selected. The fourth rescue method is a sequential rescue method that is executed in order of urgency.

[0015] According to the multi-source information fusion method for planning and decision-making in water rescue missions provided by the present invention, step S34 further includes: S341. Based on the obstacle distribution in the comprehensive information model, calculate the set of candidate paths from the current position to the target position using an optimization algorithm; the optimization algorithm includes A Algorithms and Dijkstra's algorithm; S342. For each candidate path in the candidate path set, calculate the path length, estimated travel time, and energy consumption to obtain the time constraint result; S343. Based on the remaining battery power of the rescue robot and the energy consumption, determine the energy feasibility of the path and obtain the energy constraint result; S344. Based on the time constraint results and the energy constraint results, a recommended path scheme is obtained by screening.

[0016] According to the multi-source information fusion method for planning and decision-making in water rescue missions provided by the present invention, step S4 further includes: S41. Establish a key performance indicator monitoring system, which includes multiple key performance indicators, specifically including path tracking deviation, rescue progress deviation, and resource consumption deviation. S42. Collect real-time operational data of the rescue robot and calculate the actual values ​​of multiple key performance indicators; S43. Compare the actual values ​​of multiple key performance indicators with the expected values ​​in the proposed decision-making scheme, and calculate the deviation rate; S44. When the deviation rate of any key performance indicator exceeds the preset threshold, the adjustment mechanism is triggered, and the auxiliary decision-making system is called again to generate an updated decision-making suggestion scheme. S45. Record all key events and decision adjustments during the rescue process to generate rescue execution feedback data.

[0017] According to the multi-source information fusion method for planning and decision-making in water rescue missions provided by the present invention, the triggering rules and processing strategies of the adjustment mechanism in step S44 specifically include: When the path tracking deviation rate is greater than 25%, it is judged as an abnormal navigation deviation, and the handling strategy is to replan the local path or activate the obstacle avoidance mode. When the rescue progress deviation rate is greater than 30%, it is judged as an abnormal progress delay, and the handling strategy is to increase the sailing speed or simplify the rescue process. When the resource consumption deviation rate is greater than 35%, it is judged as an energy shortage anomaly, and the handling strategy is to deploy backup rescue robots or optimize the energy allocation strategy. When multiple indicators are abnormal at the same time, the priority of the rescue targets is reordered according to the urgency of the rescue, and the rescue of targets with high urgency is given priority.

[0018] This invention also provides a multi-source information fusion system for planning and decision-making in water rescue missions, comprising: Extraction module: Used to acquire raw data collected by the multi-source sensors on the rescue robot, and obtain a multimodal feature dataset through data preprocessing and feature extraction; Construction module: used to build a comprehensive information model based on the multimodal feature dataset and output fusion perception results; Evaluation module: Used to evaluate rescue plans through the auxiliary decision-making system based on the fused perception results, and generate decision-making suggestions; Adjustment module: Used to execute the decision-making suggestions, track and dynamically adjust the rescue process in real time, and generate rescue execution feedback data.

[0019] This invention provides a multi-source information fusion method and system for planning and decision-making in water rescue missions. By integrating multi-source sensor information and intelligent decision-making algorithms, it achieves closed-loop optimization from perception to execution in water rescue missions, significantly improving rescue efficiency and success rate.

[0020] At the perception level, by building a multimodal feature dataset and integrating technologies such as Kalman filtering and deep neural networks, the system can effectively overcome the perception limitations of a single sensor in complex aquatic environments. It not only significantly improves the detection accuracy and robustness of targets and obstacles falling into the water, but also enhances the environmental adaptability under conditions such as severe sea conditions and low visibility, thus providing a more comprehensive and reliable information foundation for subsequent decision-making.

[0021] At the decision-making level, by introducing multi-dimensional evaluation factors such as environmental complexity index and rescue urgency index, the system can dynamically quantify the risks and urgency of rescue scenarios. With the help of decision matrix and Monte Carlo simulation, it can achieve intelligent matching and risk assessment of rescue plans, so that the generated paths and strategies not only have strong feasibility, but also achieve multi-objective trade-offs under resource and time constraints, effectively improving the response speed of rescue and the rationality of task planning.

[0022] In the execution and control phase, relying on the key performance indicator monitoring system and real-time feedback mechanism, this invention can continuously track and dynamically adjust the entire rescue process. Once there is a path deviation, progress delay or resource anomaly, it can quickly trigger adjustment strategies and regenerate optimized solutions, greatly enhancing the system's anti-interference ability and task fault tolerance in a dynamically changing environment.

[0023] Overall, this invention constructs a highly adaptable, reliable, and responsive maritime rescue mission planning and execution system through deep collaboration of perception fusion, intelligent decision-making, and closed-loop control, providing important technical support for improving rescue success rates, reducing human decision-making risks, and optimizing the allocation of rescue resources. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 A schematic diagram of a multi-source information fusion method for planning and decision-making in water rescue missions, provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of a multi-source information fusion-based water rescue mission planning and decision-making system provided in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0027] The embodiments of the present invention are described below with reference to the figures.

[0028] like Figure 1 As shown, this invention provides a multi-source information fusion method for planning and decision-making in water rescue missions, comprising: S1. Obtain raw data collected by the multi-source sensors on the rescue robot, and obtain a multimodal feature dataset through data preprocessing and feature extraction.

[0029] Step S1 further includes: S11. Noise filtering and contrast enhancement are performed on the acquired image data. The contour features and geometric features of the target are extracted through edge detection and corner detection to obtain the image feature vector.

[0030] This invention first acquires image data streams from a binocular camera and an infrared camera, and applies a Gaussian filter to each frame of the image for noise filtering. The filter kernel size is 5×5, and the standard deviation is... Then, the filtered image matrix is ​​read, and each pixel value is linearly stretched to enhance contrast. The stretching formula extends the original grayscale range. Mapped to Subsequently, this invention applies the Canny edge detection operator to the enhanced image, setting a high threshold of 150 and a low threshold of 50 to extract edge pixels with drastic gray-level gradient changes, forming an edge binary map. Then, connected component analysis is performed on the edge binary map to extract the complete target contour, calculating its perimeter, area, and centroid coordinates, among other geometric parameters. Next, the Harris corner detection operator is applied to the original image, setting a response function threshold of 0.01 to detect corner locations with significant gray-level changes, recording the pixel coordinates of the corners. Finally, the contour geometric parameters and corner coordinate sequences are packaged into a one-dimensional vector to form an image feature vector, including perimeter, area, centroid coordinates, corner coordinates, and the number of corners.

[0031] S12. Perform temperature calibration and image sharpening on the acquired thermal imaging data, extract the regional features and gradient features of the heat source distribution, and obtain the thermal feature vector.

[0032] In step S12, the present invention acquires a thermal radiation image from an infrared thermal imaging sensor, with each pixel value corresponding to the temperature value at that location. Sensor calibration parameters are read, including the temperature-to-radiation conversion coefficient and the ambient temperature compensation coefficient. Temperature calibration is performed on each pixel value of the thermal radiation image, converting the original radiation value to a Celsius temperature value according to the calibration parameters. Subsequently, the present invention applies a Laplacian sharpening operator to the calibrated temperature image to enhance the gradient information of the temperature boundaries, making the temperature difference between the heat source and the background more obvious. Threshold segmentation is performed on the sharpened temperature image, setting the temperature threshold to 25°C. Pixels above this threshold are marked as foreground, and pixels below this threshold are marked as background, obtaining a heat source segmentation mask. Then, connected component analysis is performed on the heat source segmentation mask to extract each independent heat source region, calculating the centroid coordinates, area, average temperature, and maximum temperature of each region. The gradient of the temperature image is calculated, calculating the temperature difference between each pixel location and its adjacent pixels. The gradient in the X direction is the temperature difference between the current pixel and the pixel to its right, and the gradient in the Y direction is the temperature difference between the current pixel and the pixel below it. In addition, this invention also calculates the average gradient magnitude within each heat source region as an indicator of the clarity of the heat source boundary. Finally, the centroid coordinates, area, average temperature, maximum temperature, and average gradient of each heat source region are packaged into a thermal feature vector.

[0033] S13. Denoise and perform spectrum analysis on the collected audio data to extract the directional and intensity features of the sound source and obtain the acoustic feature vector.

[0034] In step S13, the present invention acquires multi-channel audio signals from a microphone array. Each channel outputs a discrete time-domain signal sequence. After receiving the signals, a bandpass filter is applied to each channel signal for noise reduction. The passband frequency range is set to 200Hz-4000Hz, which covers the main frequency components of human distress cries, filtering out low-frequency environmental noise and high-frequency electronic noise. Subsequently, the present invention performs a short-time Fourier transform on the filtered signal, converting the time-domain signal into a time-frequency domain representation. Then, the window length is set to 1024 sampling points, corresponding to a time length of 21ms. The power spectrum is calculated for each time-frequency window, and the power spectrum is a frequency-amplitude squared relationship curve. After calculation, a peak value is searched in the power spectrum. The frequency corresponding to the peak value is the main frequency component of the sound source, and the peak frequency is recorded. and corresponding power Then, the cross-correlation function of each channel signal is calculated, and the time delay between adjacent microphones is estimated using the generalized cross-correlation method. Simultaneously, based on the geometric arrangement and time delay information of the microphone array, the azimuth and elevation angles of the sound source are calculated. Beamforming is then performed on the multi-channel signals, and the power of the beamout signal is calculated as a sound source intensity index. Finally, the peak frequency, source azimuth, elevation angle, and source intensity are packaged into an acoustic feature vector.

[0035] S14. Perform low-pass filtering and time synchronization on the collected navigation data, extract position coordinates and velocity information, and obtain navigation feature vectors.

[0036] In step S14, this invention acquires navigation data from the BeiDou positioning module and the inertial navigation unit. The BeiDou positioning module outputs latitude and longitude coordinates and altitude, while the inertial navigation unit outputs three-axis acceleration and three-axis angular velocity. After acquisition, this invention performs low-pass filtering on the BeiDou positioning coordinates, specifically using a first-order low-pass filter to remove coordinate jump noise; similarly, longitude and altitude are filtered. Subsequently, this invention performs time integration on the acceleration data from the inertial navigation unit to calculate the three-axis velocities; and performs time integration on the angular velocity data to calculate the attitude angle change; in addition, this invention converts the latitude and longitude coordinates of the BeiDou positioning to X and Y coordinates in the Cartesian coordinate system, using the conversion reference point as the reference point of the rescue area. Finally, this invention synchronizes the BeiDou positioning data and the inertial navigation data in time, with the BeiDou data timestamp recorded as follows: Inertial data timestamp is .when The two data points are considered synchronized at the second. When synchronization is confirmed, the position coordinates and velocity vector at the synchronization moment are combined into a navigation feature vector.

[0037] S15. Align the image feature vector, thermal feature vector, acoustic feature vector, and navigation feature vector according to the timestamp to form a multimodal feature dataset.

[0038] In step S15, the present invention adds a timestamp to each modal feature vector, and the image feature vector timestamp The hot feature vector timestamp represents the time of image acquisition. The acoustic feature vector timestamp represents the time of thermal imaging acquisition. The navigation feature vector timestamp is the center time of the audio frame. This is the data synchronization time. Subsequently, this invention sets a time alignment window. The timestamp sequence of navigation feature vectors is used as the baseline timeline, traversing seconds. The present invention searches for satisfaction The image feature vectors are searched to satisfy... The hot feature vectors are searched to satisfy... The acoustic feature vector.

[0039] When all four modal feature vectors exist at a certain moment, the four feature vectors are concatenated into a joint feature vector. Finally, all joint feature vectors are arranged in chronological order to form a time series. This time series is stored as a multimodal feature dataset with a two-dimensional matrix structure, where rows correspond to time frames and columns correspond to feature dimensions.

[0040] S2. Construct a comprehensive information model based on the multimodal feature dataset and output the fusion perception results.

[0041] Step S2 further includes: S21. A Kalman filter is used to estimate the state of the navigation feature vectors in the multimodal feature dataset. The state equation of the Kalman filter is:

[0042] in, for The state vector at time t, Here is the state transition matrix. Assign a matrix to the noise. for Time-based system noise.

[0043] In step S21, the present invention aims to construct a Kalman filter to estimate the state of the navigation feature vector. First, the state vector is defined. The system contains 19 state variables, including position coordinates, velocity components, position error, velocity error, attitude error, gyroscope error, accelerometer error, spatial lever error, and time asynchrony error. A 19×19 dimensional state transition matrix is ​​then established to describe the propagation relationship of state variables between adjacent time steps; simultaneously, a 19×6 dimensional noise allocation matrix is ​​established to allocate 6-dimensional system noise. Mapping to the state space; assume the system noise includes gyroscope random walk error and accelerometer random walk error. Subsequently, initialize the state vector, setting the position coordinates to initial values ​​from BeiDou positioning, setting the velocity component to zero, and setting all error terms to zero. After setting, establish an initial state covariance matrix, which is a 19×19 dimensional diagonal matrix, with the diagonal elements representing the initial uncertainties of each state variable.

[0044] At time k, the navigation feature vector is read, the position coordinates and velocity components are extracted as observations, and the Kalman filter prediction step is executed to calculate the prior state estimate and prior covariance. Then, the update step is executed to calculate the observation residuals. Subsequently, the Kalman gain is calculated. Then, the posterior state estimate and posterior covariance are updated. Finally, the corrected position coordinates and velocity vectors are extracted from the posterior state estimate, and the output is the filtered navigation feature vector.

[0045] S22. Using a deep neural network, perform feature-level fusion on the image feature vectors and hot feature vectors in the multimodal feature dataset, and output the target category probability and location confidence.

[0046] Step S22 further includes: S221. Input the image feature vector and the hot feature vector into two parallel convolutional neural network branches respectively to extract high-level semantic features of the image and high-level semantic features of the hot feature vector; S222. In the feature fusion layer, perform element-wise addition of the high-level semantic features of the image and the high-level semantic features of the hot feature vector through channel concatenation to obtain a fused feature map; S223. Input the fused feature map into a fully connected layer to output the target category probability distribution and bounding box regression parameters, and obtain a fusion result including target category probability and location confidence; S224. Perform non-maximum suppression processing on the fusion result to select the target detection box with the highest confidence.

[0047] In step S22, the present invention aims to construct a deep neural network to fuse image features and thermal features. Specifically, the image feature vector is input into the first convolutional neural network branch, which contains three convolutional layers, each with a kernel size of 3×3 and channel numbers of 64, 128, and 256 respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. After convolution processing, the image feature map has a dimension of H×W×256, where H and W are the height and width of the feature map. Similarly, the thermal feature vector is input into the second convolutional neural network branch, which has the same structure as the first branch, and outputs a thermal feature map with a dimension of H×W×256.

[0048] In the feature fusion layer, this invention concatenates the feature maps output from the two branches along the channel dimension, resulting in a concatenated feature map with dimensions H×W×512. After concatenation, a 1×1 convolution is applied to the concatenated feature map for channel compression, outputting a dimension H×W×256. The compressed feature map is then input into a global average pooling layer, where the spatial dimensions are averaged to output a 256-dimensional feature vector. After average pooling, this feature vector is input into a fully connected layer containing two linear layers with dimensions 256→128→C, where C is the number of target classes plus 4. The first C-4 elements of the output vector represent the class probability distribution, and the last 4 elements represent the center coordinate offset and width / height scaling parameters of the bounding boxes. Finally, softmax normalization is applied to the class probabilities, selecting the class with the highest probability as the target class, and the corresponding probability value is used as the confidence level. Non-maximum suppression is applied to the multiple output bounding boxes, retaining the bounding boxes with the highest confidence and an overlap of less than 0.5 with other boxes. The final output fusion result includes the target class, confidence level, and bounding box coordinates.

[0049] S23. Cross-validate the orientation information provided by the acoustic feature vectors in the multimodal feature dataset with the location information provided by the visual features, and eliminate detection results with confidence levels lower than a preset threshold.

[0050] In step S23, the present invention reads the azimuth and elevation angles of the sound source from the acoustic feature vector, reads the center coordinates of the target bounding box from the fusion result, and then converts the center coordinates of the bounding box into azimuth angles relative to the camera. After obtaining the azimuth angle, the present invention calculates the difference between the sound source azimuth angle and the visual azimuth angle, and sets the azimuth angle consistency threshold to [value missing]. The system determines whether the difference is less than a threshold. If the azimuth difference is less than the threshold, the acoustic and visual localization are considered consistent, and the detection result is retained. If the azimuth difference is greater than the threshold, the two modal information is considered to conflict, and the confidence level of the fusion result is checked. Subsequently, the invention reads the confidence level value, sets the confidence level threshold to 0.7, retains the detection result when the confidence level is higher than the threshold, and discards the detection result when the confidence level is lower than the threshold. Finally, for the retained detection results, the visual and acoustic azimuth angles are weighted and averaged, with the weights determined according to their respective signal-to-noise ratios, and the fused azimuth angle is output.

[0051] S24. Integrate the fusion results, establish a comprehensive information model that includes the target's absolute position, relative velocity, and the distribution of environmental obstacles, and output the fusion perception results.

[0052] In step S24, the present invention first reads the target bounding box coordinates and category from the fusion result, and calculates the target's three-dimensional coordinates by combining the depth information from binocular vision; it then reads the filtered navigation feature vector to obtain the rescue robot's position, and calculates the target's position vector relative to the rescue robot; the present invention also reads the target's position coordinate sequence in consecutive frames, calculates the position difference between adjacent frames divided by the time interval, and obtains the target's velocity vector; in addition, it reads the edge detection results from the image data and counts the number and location distribution of obstacles in the image; simultaneously, it performs cluster analysis on the obstacle grid in the grid map, extracts the centroid coordinates and size parameters of each obstacle region; and it reads environmental sensor data to obtain environmental parameters such as sea state level, wind speed, and visibility. Finally, the target's absolute position, target's relative position, target velocity vector, obstacle distribution, and environmental parameters are packaged into a comprehensive information model data structure, and the output is the fusion perception result.

[0053] S3. Based on the fusion perception results, the rescue plan is evaluated through the auxiliary decision-making system, and a decision-making suggestion plan is generated.

[0054] Step S3 further includes: S31. Calculate the environmental complexity index based on the distribution of environmental obstacles and sea state parameters in the comprehensive information model. The expression for the environmental complexity index is:

[0055] in, This is an index representing environmental complexity. For the number of obstacles, The maximum obstacle threshold, Sea state rating The maximum sea state level, For wind speed, The maximum wind speed threshold, As the first weighting coefficient, This is the second weighting coefficient. This is the third weighting coefficient.

[0056] In step S31, the present invention reads obstacle distribution data from the integrated information model, counts the number of grid cells marked as obstacles in the grid map, traverses the grid matrix, counts the number of grid cells with an element value of 1, and accumulates the total number of obstacles. After obtaining the total number of obstacles, a maximum obstacle threshold is set as the normalization benchmark, and the obstacle density ratio is calculated. Next, the present invention reads sea state parameters, the sea state level is determined by wave height and period, ranging from level 1 to 9, and then sets a maximum sea state level to calculate the sea state ratio. Third, the present invention also reads meteorological sensor data to obtain wind speed, sets a maximum wind speed threshold, and calculates the wind speed ratio. Finally, a first weighting coefficient is set. Emphasizing the impact of obstacles, the second weighting coefficient Taking sea state into account, the third weighting coefficient Taking wind speed into account, the three ratios are multiplied by their respective weights and then summed to obtain the environmental complexity index. The value ranges from 0 to 1, with a larger value indicating a more complex environment.

[0057] S32. Calculate the rescue urgency index based on the target location and target status in the comprehensive information model; The expression for the urgency index of the rescue is:

[0058] in, As an indicator of the urgency of the rescue, For minimum safe distance, Current distance, This is the drift time. This is the critical time. As an indicator of physical activity, For the fourth weighting coefficient, The fifth weighting coefficient It is the sixth weighting coefficient.

[0059] In step S32, the present invention reads the target relative position vector from the integrated information model. The modulus is calculated to obtain the current distance. Then, the present invention sets a minimum safe distance. Using meters as a benchmark for throwing distance, the distance ratio is calculated; simultaneously, this invention reads the timestamp of the first target detection. and current timestamp Calculate the drift time Then set the critical time. The time frame is calculated based on the number of seconds elapsed after a fall into the water, beyond which the person's body temperature drops, threatening their life. The time ratio is then calculated. Additionally, this invention reads motion amplitude data from the image feature vector, extracts the target centroid displacement sequence from 10 consecutive frames, and then calculates the standard deviation of the displacement sequence. A larger standard deviation indicates greater physical activity. The standard deviation will then be normalized to a physical activity index. The pixel value is used as the reference standard deviation. After calculating the above values, this invention sets a fourth weighting coefficient. Emphasizing distance factors, the fifth weighting coefficient Taking time factors into account, the sixth weighting coefficient Taking into account the state of motion, the three evaluation metrics are multiplied by their respective weights and then summed to obtain the rescue urgency index. The value ranges from 0 to above 1, with a larger value indicating a more urgent need for rescue.

[0060] S33. Based on the environmental complexity index and the rescue urgency index, the initial rescue method is matched using a decision matrix.

[0061] Specifically, the matching rules for the decision matrix in step S33 include: when the rescue urgency index score is greater than 0.7 and the environmental complexity index is less than 0.5, a first rescue method is selected, which is a rapid approach rescue method; when the rescue urgency index score is greater than 0.7 and the environmental complexity index is greater than or equal to 0.5, a second rescue method is selected, which is a remote throwing rescue device method; when the rescue urgency index score is greater than or equal to 0.4 and less than or equal to 0.7, a third rescue method is selected, which is a standard approach and throwing rescue device method; when the rescue urgency index score is less than 0.4 and there are multiple targets to be rescued, a fourth rescue method is selected, which is a sequential rescue method executed in order of urgency.

[0062] Specifically, in step S33, the present invention reads the rescue urgency index. and environmental complexity index Construct a two-dimensional decision space. In the decision space, first determine... and Whether this condition is met indicates that the rescue is extremely urgent and the environment is relatively simple. The first rescue method, i.e., the rapid approach rescue method, is selected. The rescue robot approaches the person in the water at its maximum speed of 5 meters per second, and performs the pick-up operation after approaching to a distance of 2.5 meters. Next, a judgment is made... and The condition indicates an extremely urgent but complex rescue situation. Therefore, the second rescue method, remotely throwing a life-saving device, is chosen. The rescue robot stops at a distance of 5-10 meters, triggering the throwing device to release a lifebuoy, thus avoiding the risks of close proximity in a complex environment. A second assessment is then required. Whether this condition is met indicates a moderate level of urgency in the rescue. The third rescue method, the standard approach and deployment of the rescue device, is selected. The rescue robot approaches at a moderate speed of 3 meters per second and performs the deployment operation upon reaching a distance of 2.5 meters. Final judgment... To determine if the conditions are met, the list of targets to be rescued in the integrated information model is read, and the number of targets is counted. When there are multiple targets to be rescued, the fourth rescue method, namely the sequential rescue method, is selected. The targets are sorted from highest to lowest according to their urgency index and the rescue tasks are executed sequentially.

[0063] S34. Perform path feasibility analysis and time cost estimation on the initial rescue method to generate a path plan that includes execution steps and expected duration.

[0064] Step S34 further includes: S341. Based on the obstacle distribution in the comprehensive information model, calculate the set of candidate paths from the current position to the target position using an optimization algorithm; the optimization algorithm includes A Algorithms and Dijkstra's algorithm.

[0065] In step S341, this invention reads obstacle distribution data from the integrated information model and maps obstacle coordinates to a grid map. First, the current position of the rescue robot is read and converted into a grid index, and the target position is also read and converted into a grid index. Then, this invention uses Dijkstra's algorithm for path search, expanding the search in four directions from the starting grid, calculating the distance from adjacent grids to the starting point, and continuing the search until the destination grid is reached. The parent node of each grid is recorded during the search process. The search backtracks from the destination to the starting point, extracting the path grid coordinate sequence to form the first candidate path. Subsequent searches backtrack to the destination to extract the path, forming the second candidate path. Finally, the grid map is dilated to increase the obstacle edge width. Dijkstra's algorithm is executed again on the dilated map to obtain a more conservative third candidate path. After obtaining the third candidate path, the grid coordinate sequences of the three candidate paths are stored as a candidate path set.

[0066] S342. For each candidate path in the candidate path set, calculate the path length, estimated travel time, and energy consumption to obtain the time constraint result.

[0067] This invention iterates through each candidate path in the candidate path set, reads the path grid coordinate sequence, calculates the distance between adjacent grids, and sums up the distances of all path segments to obtain the total path length. Subsequently, this invention sets the travel speed for each path segment according to the rescue method: 5 m / s for the rapid approach rescue method and 3 m / s for the standard approach method. The travel time for each path segment is then calculated and summed to obtain the base travel time. Additionally, this invention analyzes the number of obstacle avoidances along the path, checking for abrupt changes in the heading angle within the path grid sequence. A change in heading angle exceeding 45 degrees is counted as one obstacle avoidance. Each obstacle avoidance requires deceleration, turning, and then acceleration, with the extra time recorded as 2 seconds. All extra time for obstacle avoidances is accumulated, and the estimated travel time is the sum of all times. Furthermore, this invention also calculates the path energy consumption, reading the energy consumption per unit distance of the rescue robot. Path energy consumption Ultimately, this invention packages the path length, estimated travel time, and energy consumption into a time-constrained result.

[0068] S343. Based on the remaining battery power of the rescue robot and the energy consumption, determine the energy feasibility of the path and obtain the energy constraint result.

[0069] In step S343, the present invention reads the remaining battery power of the rescue robot. Get the current battery percentage from the power management module and multiply it by the maximum battery capacity. Then, read the path energy consumption from the time constraint results. Calculate round-trip energy consumption The energy requirements for returning to the mother ship are taken into account. Therefore, this invention sets a safety margin factor. The total energy demand is calculated as follows: Then judge The invention determines whether the energy feasibility flag is true or false. If true, the path satisfies the energy constraint, and the energy feasibility flag is set to true; otherwise, the path does not satisfy the energy constraint, and the energy feasibility flag is set to false. The invention also calculates energy sufficiency. An energy sufficiency score greater than 1 indicates sufficient energy, while a score less than 1 indicates insufficient energy. Ultimately, the energy feasibility indicator and the energy sufficiency score are packaged into an energy constraint result.

[0070] S344. Based on the time constraint results and the energy constraint results, a recommended path scheme is obtained by screening.

[0071] In step S344, the present invention traverses the candidate path set, reads the time constraint result and energy constraint result of each path, and then filters paths with the energy feasibility flag as true, eliminating paths with insufficient energy. Subsequently, the filtered paths are sorted in ascending order of expected travel time, and the path with the shortest travel time is selected. Then, the present invention reads the path length, expected travel time, obstacle avoidance attempts, and energy sufficiency of this path, converts the path grid coordinates into actual geographic coordinates, generates an execution step list containing waypoint coordinate sequences, travel speeds for each segment, and expected arrival times, and records the total path duration as the expected travel time plus the 3-second drop operation time. Finally, the execution step list and expected duration are combined to form a recommended path scheme.

[0072] S35. Evaluate the success probability of the proposed path scheme under different environmental disturbances through Monte Carlo simulation, obtain risk assessment data, and formulate a decision-making recommendation scheme.

[0073] Furthermore, this invention sets a random distribution for environmental disturbance parameters: wind speed fluctuations follow a normal distribution, ocean current changes follow a uniform distribution, and target drift velocity follows a normal distribution. Then, it sets the number of simulations, performs the k-th simulation, and samples disturbance samples from each random distribution to generate disturbance samples. Subsequently, it corrects the rescue robot's navigation speed and target position based on these disturbance samples. Specifically, it simulates the rescue robot navigating along a recommended path, updating its position and speed at 0.1-second intervals. Then, it calculates the distance between the rescue robot and the target; when the distance is less than 2.5 meters, it is considered a successful arrival, and the arrival time is recorded. The simulation is considered a failure if it takes more than twice the expected duration. The simulation is repeated N times, and the number of successful simulations is counted. Calculate the probability of a successful rescue. Subsequently, simulation records of successful arrivals were extracted, and the average rescue time was calculated. And the standard deviation of time consumption. Finally, the probability of successful rescue, the average rescue time, and the standard deviation of time consumption are used as risk assessment data. The initial rescue method, the recommended route plan, and the risk assessment data are combined to form a decision-making recommendation plan.

[0074] S4. Execute the decision-making suggestion plan, track and dynamically adjust the rescue process in real time, and generate rescue execution feedback data.

[0075] Step S4 further includes: S41. Establish a key performance indicator monitoring system, which includes multiple key performance indicators, specifically including path tracking deviation, rescue progress deviation, and resource consumption deviation.

[0076] In step S41, this invention establishes three categories of key performance indicators (KPIs): path tracking deviation is defined as the vertical distance between the actual position of the rescue robot and the recommended path; rescue progress deviation is defined as the time difference between the actual rescue progress and the expected progress; and resource consumption deviation is defined as the percentage difference between the actual power consumption and the expected power consumption. This invention sets a monitoring frequency for each type of indicator: path tracking deviation is sampled once per second, rescue progress deviation is calculated every 5 seconds, and resource consumption deviation is evaluated every 10 seconds, ultimately forming a KPI monitoring system.

[0077] S42. Collect real-time operational data of the rescue robot and calculate the actual values ​​of multiple key performance indicators.

[0078] In step S42, the present invention reads the position coordinates of the rescue robot in real time. The system obtains an update frequency of 10Hz from the navigation system, then reads the waypoint coordinate sequence from the recommended path and finds the waypoint closest to the actual location as the reference point. Then, the present invention calculates the distance from the actual position to the reference point, which is used as the actual value of the path tracking deviation. Secondly, this invention reads the start time of the rescue mission. and the current moment Calculate the execution time Based on the expected duration and current execution progress in the recommended path, calculate the expected completion rate. Secondly, this invention calculates the actual completion rate based on the current and initial distances between the rescue robot and the target. Then calculate the actual value of the rescue progress deviation. Additionally, this invention reads the battery level at the start of the rescue mission. and current battery level Calculate actual power consumption Calculate the expected power consumption based on the execution time and path energy consumption. Finally, calculate the actual value of resource consumption deviation. .

[0079] S43. Compare the actual values ​​of multiple key performance indicators with the expected values ​​in the proposed decision-making scheme, and calculate the deviation rate.

[0080] In step S43, the present invention first reads the expected path tracking deviation in the decision suggestion scheme. Meters, this value is the allowable deviation set during path planning, used to calculate the path tracking deviation rate. Secondly, read the deviation from the expected rescue progress. Seconds, ideally the actual progress matches the expected progress, calculate the rescue progress deviation rate. Read the expected resource consumption deviation. Ah, calculate resource consumption deviation rate. .

[0081] S44. When the deviation rate of any key performance indicator exceeds the preset threshold, the adjustment mechanism is triggered, and the auxiliary decision-making system is called again to generate an updated decision-making suggestion.

[0082] Specifically, the triggering rules and processing strategies for the adjustment mechanism in step S44 include: When the path tracking deviation rate is greater than 25%, it is judged as a navigation deviation anomaly, and the handling strategy is to replan the local path or activate the obstacle avoidance mode; when the rescue progress deviation rate is greater than 30%, it is judged as a progress delay anomaly, and the handling strategy is to increase the navigation speed or simplify the rescue process; when the resource consumption deviation rate is greater than 35%, it is judged as an energy shortage anomaly, and the handling strategy is to allocate backup rescue robots or optimize the energy allocation strategy; when multiple indicators are abnormal at the same time, the priority of the rescue targets is reordered according to the rescue urgency index, and the rescue of high-urgency targets is given priority.

[0083] In step S44, the present invention first determines whether the path tracking deviation rate is greater than 25%. If the deviation is detected as abnormal, this invention reads the obstacle distribution from the integrated information model and checks whether there are obstacles within 3 meters in front of the rescue robot. If obstacles exist, the handling strategy is to activate obstacle avoidance mode and call the dynamic window method to calculate the obstacle avoidance velocity vector. If there are no obstacles, the handling strategy is to replan the local path, starting from the current position and ending at the original target position, and call A... The algorithm calculates a new path. Secondly, this invention determines whether the rescue progress deviation rate is greater than 30%. If a progress delay is detected as abnormal, this invention reads the current speed, calculates a new speed, and issues a speed adjustment command to the propulsion system. Secondly, this invention checks if the resource consumption deviation rate is greater than 35%. If an energy shortage is detected, the remaining energy is calculated to ensure a round trip. If insufficient energy is found, the strategy is to deploy a backup rescue robot, send a support request to the control console, and transmit the current and target location information. Finally, the invention checks for multiple simultaneous anomalies. If at least two deviation rates exceed the limit, the urgency indicators of all targets to be rescued are read, reordered from highest to lowest value, and priority is given to rescuing the target with the highest urgency, while suspending or postponing rescue missions for other targets. Finally, an updated decision-making suggestion is generated, including an adjusted route, speed command, and mission priority.

[0084] S45. Record all key events and decision adjustments during the rescue process to generate rescue execution feedback data.

[0085] In step S45, the present invention aims to establish an event log data structure, including timestamp, event type, event description, and processing result fields. First, it records the rescue mission initiation event (timestamp, event type: mission initiation, event description: initial rescue method and recommended route plan); it records key events during the route tracking process, recording a navigation deviation event when the route tracking deviation exceeds a threshold, including the deviation distance and trigger time; it records rescue progress assessment events, recording the current completion percentage and expected completion percentage every 30 seconds; it records resource consumption monitoring events, recording the remaining power and consumption rate every minute; it records decision adjustment events, recording the exception type, processing strategy, and adjustment result when the adjustment mechanism is triggered; and it records the rescue mission completion event (timestamp). The event type is "task completed," and the event description includes total time, total energy consumption, and rescue result. Finally, all event logs are arranged in chronological order to form rescue execution feedback data.

[0086] like Figure 2 As shown, the present invention also provides a multi-source information fusion system for planning and decision-making in water rescue missions, comprising: Extraction module 100: Used to acquire raw data collected by the multi-source sensors on the rescue robot, and obtain a multimodal feature dataset through data preprocessing and feature extraction; Construction module 200: used to construct a comprehensive information model based on the multimodal feature dataset and output fusion perception results; Evaluation module 300: Used to evaluate the rescue plan through the auxiliary decision-making system based on the fused perception results, and generate a decision suggestion plan; Adjustment module 400: Used to execute the decision suggestion plan, track and dynamically adjust the rescue process in real time, and generate rescue execution feedback data.

[0087] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0088] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the multi-source information fusion water rescue mission planning and decision-making method described in various embodiments or some parts of embodiments.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-source information fusion method for planning and decision-making in water rescue missions, characterized in that, include: S1. Obtain raw data collected by the multi-source sensors on the rescue robot, and obtain a multimodal feature dataset through data preprocessing and feature extraction; S2. Construct a comprehensive information model based on the multimodal feature dataset and output the fusion perception results; S3. Based on the fusion perception results, the rescue plan is evaluated through the auxiliary decision-making system, and a decision-making suggestion plan is generated; S4. Execute the decision-making suggestion plan, track and dynamically adjust the rescue process in real time, and generate rescue execution feedback data.

2. The multi-source information fusion method for planning and decision-making in water rescue missions according to claim 1, characterized in that, Step S1 further includes: S11. Noise filtering and contrast enhancement are performed on the acquired image data. The contour features and geometric features of the target are extracted through edge detection and corner detection to obtain the image feature vector. S12. Perform temperature calibration and image sharpening on the acquired thermal imaging data, extract the regional and gradient features of the heat source distribution, and obtain the thermal feature vector. S13. Denoise and perform spectrum analysis on the collected audio data to extract the directional and intensity features of the sound source and obtain the acoustic feature vector. S14. Perform low-pass filtering and time synchronization on the collected navigation data, extract position coordinates and velocity information, and obtain navigation feature vectors; S15. Align the image feature vector, thermal feature vector, acoustic feature vector, and navigation feature vector according to the timestamp to form a multimodal feature dataset.

3. The multi-source information fusion method for planning and decision-making in water rescue missions according to claim 2, characterized in that, Step S2 further includes: S21. A Kalman filter is used to estimate the state of the navigation feature vectors in the multimodal feature dataset. The state equation of the Kalman filter is: in, for The state vector at time t, Here is the state transition matrix. Assign a matrix to the noise. for Real-time system noise; S22. Using a deep neural network, perform feature-level fusion on the image feature vectors and hot feature vectors in the multimodal feature dataset, and output the target category probability and location confidence. S23. Cross-validate the orientation information provided by the acoustic feature vectors in the multimodal feature dataset with the position information provided by the visual features, and eliminate detection results with confidence levels lower than a preset threshold. S24. Integrate the fusion results, establish a comprehensive information model that includes the target's absolute position, relative velocity, and the distribution of environmental obstacles, and output the fusion perception results.

4. The multi-source information fusion method for planning and decision-making in water rescue missions according to claim 3, characterized in that, Step S22 further includes: S221. Input the image feature vector and the hot feature vector into two parallel convolutional neural network branches respectively to extract the high-level semantic features of the image and the hot high-level semantic features. S222. In the feature fusion layer, the high-level semantic features of the image and the hot high-level semantic features are added element-wise by channel concatenation to obtain a fused feature map. S223. Input the fused feature map into the fully connected layer, output the target class probability distribution and bounding box regression parameters, and obtain the fusion result including the target class probability and location confidence. S224. Perform non-maximum suppression processing on the fusion result to select the target detection box with the highest confidence.

5. The multi-source information fusion method for planning and decision-making in water rescue missions according to claim 1, characterized in that, Step S3 further includes: S31. Calculate the environmental complexity index based on the distribution of environmental obstacles and sea state parameters in the comprehensive information model. The expression for the environmental complexity index is: in, This is an index representing environmental complexity. For the number of obstacles, The maximum obstacle threshold, Sea state rating The maximum sea state level, For wind speed, The maximum wind speed threshold, As the first weighting coefficient, This is the second weighting coefficient. This is the third weighting coefficient; S32. Calculate the rescue urgency index based on the target location and target status in the comprehensive information model; The expression for the urgency index of the rescue is: in, As an indicator of the urgency of the rescue, For minimum safe distance, This is the current distance. This is the drift time. This is the critical time. As an indicator of physical activity, For the fourth weighting coefficient, The fifth weighting coefficient This is the sixth weighting coefficient; S33. Based on the environmental complexity index and the rescue urgency index, an initial rescue method is matched using a decision matrix; S34. Perform path feasibility analysis and time cost estimation on the initial rescue method to generate a path plan that includes execution steps and expected duration; S35. Evaluate the success probability of the proposed path scheme under different environmental disturbances through Monte Carlo simulation, obtain risk assessment data, and formulate a decision-making recommendation scheme.

6. The multi-source information fusion method for planning and decision-making in water rescue missions according to claim 5, characterized in that, The matching rules for the decision matrix in step S33 specifically include: When the urgency index score is greater than 0.7 and the environmental complexity index is less than 0.5, the first rescue method is selected, which is the rapid approach rescue method. When the rescue urgency index score is greater than 0.7 and the environmental complexity index is greater than or equal to 0.5, the second rescue method is selected, which is the remote throwing of a life-saving device. When the rescue urgency index score is greater than or equal to 0.4 and less than or equal to 0.7, the third rescue method is selected, which is the standard approach and throwing of a life-saving device method; When the rescue urgency index score is less than 0.4 and there are multiple targets to be rescued, the fourth rescue method is selected. The fourth rescue method is a sequential rescue method that is executed in order of urgency.

7. The multi-source information fusion method for planning and decision-making in water rescue missions according to claim 5, characterized in that, Step S34 further includes: S341. Based on the obstacle distribution in the comprehensive information model, calculate the set of candidate paths from the current position to the target position using an optimization algorithm; the optimization algorithm includes A Algorithms and Dijkstra's algorithm; S342. For each candidate path in the candidate path set, calculate the path length, estimated travel time, and energy consumption to obtain the time constraint result; S343. Based on the remaining battery power of the rescue robot and the energy consumption, determine the energy feasibility of the path and obtain the energy constraint result; S344. Based on the time constraint results and the energy constraint results, a recommended path scheme is obtained by screening.

8. The multi-source information fusion method for planning and decision-making in water rescue missions according to claim 1, characterized in that, Step S4 further includes: S41. Establish a key performance indicator monitoring system, which includes multiple key performance indicators, specifically including path tracking deviation, rescue progress deviation, and resource consumption deviation. S42. Collect real-time operational data of the rescue robot and calculate the actual values ​​of multiple key performance indicators; S43. Compare the actual values ​​of multiple key performance indicators with the expected values ​​in the proposed decision-making scheme, and calculate the deviation rate; S44. When the deviation rate of any key performance indicator exceeds the preset threshold, the adjustment mechanism is triggered, and the auxiliary decision-making system is called again to generate an updated decision-making suggestion scheme. S45. Record all key events and decision adjustments during the rescue process to generate rescue execution feedback data.

9. The multi-source information fusion method for planning and decision-making in water rescue missions according to claim 8, characterized in that, The triggering rules and processing strategies for the adjustment mechanism in step S44 specifically include: When the path tracking deviation rate is greater than 25%, it is judged as an abnormal navigation deviation, and the handling strategy is to replan the local path or activate the obstacle avoidance mode. When the rescue progress deviation rate is greater than 30%, it is judged as an abnormal progress delay, and the handling strategy is to increase the sailing speed or simplify the rescue process. When the resource consumption deviation rate is greater than 35%, it is judged as an energy shortage anomaly, and the handling strategy is to deploy backup rescue robots or optimize the energy allocation strategy. When multiple indicators are abnormal at the same time, the priority of the rescue targets is reordered according to the urgency of the rescue, and the rescue of targets with high urgency is given priority.

10. A multi-source information fusion system for planning and decision-making in water rescue missions, characterized in that, include: Extraction module: Used to acquire raw data collected by the multi-source sensors on the rescue robot, and obtain a multimodal feature dataset through data preprocessing and feature extraction; Construction module: used to build a comprehensive information model based on the multimodal feature dataset and output fusion perception results; Evaluation module: Used to evaluate rescue plans through the auxiliary decision-making system based on the fused perception results, and generate decision-making suggestions; Adjustment module: Used to execute the decision-making suggestions, track and dynamically adjust the rescue process in real time, and generate rescue execution feedback data.

Images