A water conservancy and hydropower engineering fishway fish passing monitoring method and system coupling AI identification and sonar

Abstract

The application discloses a kind of coupling AI identification and water conservancy and hydropower engineering fishway fish monitoring method and system of sounder, it is related to aquatic ecological monitoring technical field, including: sounder intelligent imaging module, based on the intelligent sensing system of multiple element sounder array, combined with grouping time emission strategy, generate high-density parallel beam, carry out fine scanning to fishway section, obtain target boundary clear fishway section acoustic image.The present application realizes high-precision, non-invasive monitoring to fishway fish target by the deep integration of sounder intelligent imaging and AI identification technology, sounder array generates high-density parallel beam using grouping time emission strategy, can clearly image under the condition that water is turbid or light is insufficient, effectively overcome the limitations of traditional optical monitoring method in low visibility environment, combined with improved ant colony algorithm to intelligently separate overlapping targets, significantly improve the individual identification accuracy under the condition of high-density fish school or target overlap.

Description

Technical Field

[0001] This invention relates to the field of aquatic ecological monitoring technology, specifically to a method and system for monitoring fish passage in water conservancy and hydropower projects that couples AI recognition and sonar. Background Technology

[0002] Hydropower projects, in the process of developing and utilizing water energy and water resources, will have adverse effects on aquatic ecosystems, especially on the habitat, reproduction and migration of aquatic organisms such as fish. Fishways are specially designed to solve this problem, aiming to provide fish with safe migration channels and ensure their free movement between facilities such as hydropower stations and dams. Therefore, the design and management of fishways in hydropower projects are of paramount importance.

[0003] For example, a fish passage monitoring device and method based on acoustic imaging, which is published in Chinese Patent Publication No. CN115079148A, is suitable for counting fish passing through fish passages in turbid water environments. It can not only count fish passing through fish passages but also generate three-dimensional images of fish passing through fish passages, which are intuitive to display and accurate in counting.

[0004] However, while traditional acoustic imaging-based methods can achieve basic counting and imaging, under high-density fish passage, sonar signals are easily affected by clutter and target overlap, making it difficult for the system to accurately separate and track the movement trajectory of individual fish, resulting in significant deviations in the counting results. Secondly, even if preliminary target separation is achieved, there is still a lack of ability to identify fish species and their growth stages, making it difficult to distinguish between fish fry and adult fish, and between different fish species, thus limiting the in-depth application of fish passages in fish resource conservation effectiveness assessment. To address these issues, a fish passage monitoring system for hydraulic and hydropower projects that couples AI recognition and sonar is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring fish passage in water conservancy and hydropower projects that couples AI recognition and sonar, in order to solve the problems in the existing technology that the counting results have large deviations, lack the ability to identify fish species and their growth stages, and have difficulty distinguishing between fry and adult fish, as well as the differences in passage between different fish species.

[0006] To achieve the above objectives, this invention provides a method for monitoring fish passage in fishways of hydraulic and hydropower projects that couples AI recognition and sonar, comprising the following steps:

[0007] S1. Intelligent sonar imaging: An intelligent sensing system that deploys a multi-element sonar array in the fishway uses a group time-division transmission strategy to generate a high-density parallel beam to scan the fishway cross-section. After envelope detection and digital processing of the received echo signals, pixel mapping is used to synthesize an acoustic image of the fishway cross-section. S2. Target separation calculation: Identify fish targets in the acoustic image. If there is a high-density passage of fish or target overlap, the mixed echo signal is separated by an improved ant colony algorithm combined with a physical acoustic model. If there is no high-density passage of fish or target overlap, the fish is directly extracted to obtain individual fish acoustic morphology data. S3. Morphological depth recognition: After preprocessing the acoustic morphological data of the individual fish, the YOLOv7 target detection algorithm is used to locate the fish target and output the accurate contour bounding box. Then, the ResNet50 deep residual network is used to extract the fine-grained morphological feature vector of the fish. S4. Adaptive classification decision: Combining the fine-grained morphological feature vectors, a deep reinforcement learning strategy network is used to complete fish species identification and growth stage assessment, and fish common behavior patterns are analyzed based on the feature and behavior association rule knowledge base. S5. Monitoring Data Extraction: Integrates information on fish species, growth stages, behavioral patterns, and corresponding spatiotemporal stamps, calculates core statistical indicators in a rolling manner according to preset time windows, and automatically generates multi-dimensional fish monitoring reports by combining spatial distribution heatmaps.

[0008] Based on the above method, a fish passage monitoring system for hydraulic and hydropower projects that couples AI recognition and sonar is provided, including: The sonar intelligent imaging module, based on an intelligent sensing system covering a multi-element sonar array, combined with a group time-division transmission strategy, generates a high-density parallel beam to perform fine scanning of the fishway cross section, obtains an acoustic image of the fishway cross section with clear target boundaries, ensures imaging quality from the source, effectively improves image spatial resolution and signal-to-noise ratio, and enhances the ability to identify target contours. The target separation and calculation module is used to introduce an improved ant colony algorithm to intelligently calculate and simulate the mixed echo signals generated by the high-density passage of fish or target overlap in the acoustic image of the fishway cross section. It effectively separates overlapping individuals, obtains the acoustic morphology data of individual fish, improves the accuracy of individual target identification, significantly reduces the false detection and false detection rate of targets in high-density fish scene, and achieves accurate individual separation. The morphological depth recognition module is used to receive the calculated acoustic morphological data of individual fish bodies. It uses the YOLOv7 target detection algorithm to quickly locate and frame fish targets, and uses the ResNet50 deep residual network to perform fine-grained pattern recognition on fish targets to extract fine-grained morphological features required for species identification. This enables real-time localization and high-precision morphological feature extraction of fish targets, supporting subsequent classification and recognition. The adaptive classification decision module is used to combine extracted fine-grained morphological features and reinforcement learning strategy networks to dynamically learn and associate behavioral patterns and morphological association rules of different fish species and different growth stages, construct a classification decision model, identify fish species in fish passages, assess growth stages and analyze passage behavior, realize intelligent classification of fish species and growth stages, associate behavioral patterns, and improve the depth of fish monitoring data analysis. The fish monitoring data extraction module integrates classification results, quantity statistics, spatiotemporal distribution, and behavioral pattern information to automatically generate multi-dimensional fish monitoring reports, including the composition of fish species, size distribution, peak passage periods, and behavioral preferences. This provides direct data support for fishway effectiveness assessment and fish conservation strategy formulation.

[0009] The execution steps of the sonar intelligent imaging module include: By deploying an intelligent sensing system covering a multi-element sonar array in the target fishway, the array elements of the multi-element sonar array are divided into M equal number of array element groups, and the array elements in each array element group are numbered sequentially. The array elements with the same number in all array element groups are controlled to simultaneously transmit short pulse high-frequency acoustic signals. After each group of array elements completes its transmission, the receiving process of the corresponding array element is immediately started, so that each array element independently generates a highly directional parallel beam, ensuring that the beam direction is consistent and improving the spatial directivity and anti-interference capability of the signal. Repeat the group time-division transmission and reception process until the transmission and reception of all array elements in all array groups are completed, forming a high-density parallel beam scanning network covering the entire monitoring area within the fishway cross section, achieving full cross section coverage scanning, avoiding monitoring blind spots, and improving target acquisition rate; Each array element receives an echo signal that undergoes independent envelope detection and digitization. The echo energy data of all parallel beams are spatially arranged and pixel-mapped according to the physical deployment order of the array elements. This process synthesizes a clear acoustic image of the fishway cross-section, generating a high-resolution acoustic image that clearly presents the fish's outline and spatial distribution.

[0010] The target separation and calculation module includes an intelligent trajectory simulation unit and an overlapping signal decoupling unit; The intelligent trajectory simulation unit applies an improved ant colony algorithm to simulate the movement behavior and trajectory distribution of fish in a limited space. It intelligently segments and predicts the trajectory of the echo regions that are stuck or partially overlapping in the acoustic image of the fish passage section, separates the movement contours of suspected independent individuals, effectively tracks the movement paths of each fish in the overlapping area, and improves the continuity of the trajectory and the accuracy of the segmentation. The overlapping signal decoupling unit is used to combine the physical acoustic model with the motion contours of the separated suspected independent individuals to perform inverse calculations on the echo energy and time delay information of the overlapping area, decompose the mixed echo signal into multiple independent target signals that conform to acoustic scattering characteristics, and output the separated individual fish body acoustic morphology data with clear contours, thereby realizing the physical-level decomposition of complex echo signals and restoring the complete acoustic scattering characteristics of a single fish body.

[0011] The execution steps of the intelligent trajectory simulation unit include: Based on the acoustic image of the fishway section output by the sonar intelligent imaging module, the echo energy adhesion area caused by the high-density passage of fish or target overlap in the image is identified, and the pixel set of the area is initialized as the search space to be explored by artificial ants in the ant colony algorithm, which effectively focuses on high-density areas and improves the processing efficiency of the algorithm. Running the ant colony algorithm within the search space, by introducing local gradient pheromones and dynamic heuristic functions, simulates the possible movement paths of fish in the limited space of the fishway, guiding artificial ants to iteratively traverse and simulate the echo energy distribution of the overlapping area, significantly enhancing the accuracy of path search and target tracking capability. Based on the path trajectory clusters formed after the ant colony algorithm converges, the initially contiguous echo region is intelligently segmented to separate the motion contours of multiple suspected independent individuals. A unique trajectory identifier is assigned to each motion contour to achieve effective separation of overlapping individuals, which facilitates subsequent accurate identification and tracking.

[0012] The execution steps of the overlapping signal decoupling unit include: For each suspected independent individual's motion profile, a corresponding acoustic body scattering physical model is established and matched with the sound wave propagation characteristics of the fishway water area. A theoretical echo generation model based on the profile shape and spatial position is constructed to achieve high-fidelity acoustic feature modeling, enhance the consistency between the model and the actual fish body scattering, and improve the subsequent separation accuracy. Within the spatial region corresponding to the motion contour, the energy distribution and time delay information of the mixed echo signal in the original acoustic image are extracted. Combined with the theoretical echo generation model, non-negative matrix decomposition is used to perform inverse calculation of the mixed echo, estimate the energy contribution ratio of each independent target to the mixed echo, effectively decompose the overlapping signal, accurately quantify the energy ratio of each target, and significantly improve the objectivity and interpretability of target separation. Based on the estimated energy contribution ratio, the clear echo signal corresponding to each individual fish target is separated and reconstructed from the mixed echo signal, and a set of individual fish acoustic morphology data with complete acoustic scattering characteristics and clear contour boundaries is output, resulting in high-quality individual acoustic morphology data, which provides clear and reliable input for target recognition and classification, and supports subsequent intelligent analysis.

[0013] The morphological depth recognition module includes a target detection unit and a fine-grained feature extraction unit; The target detection unit, based on the YOLOv7 target detection algorithm, processes the calculated acoustic morphology data of individual fish bodies in real time, locates and selects the contour boundaries of all fish targets in the image, quickly and accurately locates the contours of each fish body, and provides reliable boundary information for feature extraction. The fine-grained feature extraction unit is used to perform fine-grained pattern recognition on the contour boundary of the detected fish target using a ResNet50 deep residual network, and extract fine-grained morphological features including fish contour, body size, aspect ratio and local scattering intensity distribution to form a depth feature vector that can characterize different types of fish species.

[0014] The execution steps of the target detection unit include: The acoustic morphology data of individual fish bodies is received, normalized into grayscale image format with uniform size and intensity, and data augmentation operations are performed to construct a standardized dataset suitable for deep learning model input, which ensures data consistency and improves the stability and generalization ability of model training. By combining preprocessed individual fish acoustic morphology data with the YOLOv7 target detection algorithm, and through the backbone feature extraction layer and multi-scale prediction head of the YOLOv7 target detection algorithm, fast forward reasoning is performed on fish targets in the image, generating multiple candidate bounding boxes and their corresponding confidence scores in real time, achieving efficient target localization and significantly improving the real-time performance and coverage of detection. The nonmaximum suppression algorithm is applied to filter the candidate bounding boxes output by the YOLOv7 object detection algorithm, eliminating redundant and low-confidence detection boxes, determining and outputting the precise contour bounding box coordinates of each fish target in the image, effectively eliminating overlapping box interference, and ensuring the accuracy and uniqueness of the output bounding boxes.

[0015] The execution steps of the fine-grained feature extraction unit include: Based on the output precise contour bounding box coordinates, the local image region corresponding to each fish target is extracted from the original individual fish acoustic morphology data. The extracted target region image is then scaled to the standard input size of the ResNet50 deep residual network to ensure that the size of each fish image is uniform, avoid deformation, and provide regular input for subsequent feature extraction. The cropped target region image is input into a pre-trained ResNet50 deep residual network. The network uses multiple residual block-level convolution and pooling operations to perform layer-by-layer abstraction and feature encoding on the target region image, extracting deep feature maps containing multi-level semantic information. The deep network automatically learns and encodes the details and structural features of the fish body morphology, thereby improving the feature expression capability. A global average pooling operation is performed on the depth feature map to compress it into a fixed-dimensional feature vector. This feature vector integrates the outline shape, body structure, aspect ratio, and acoustic scattering intensity distribution pattern of the fish target in the local area of ​​the image to form a fine-grained morphological feature descriptor for subsequent classification. This generates a compact feature vector with strong discriminative power, which effectively supports high-precision classification of fish species and growth stages.

[0016] The execution steps of the adaptive classification decision module include: Collect historical monitoring data, construct the correspondence between fine-grained morphological feature vectors of fish at different growth stages and their typical common behavior patterns, form an initial feature-behavior association rule knowledge base, and form an interpretable behavior rule base covering multiple fish species and multiple stages to improve classification credibility and system interpretability. Using fine-grained morphological feature vectors as state input and classification labels of fish species and growth stages as action space, a deep reinforcement learning policy network is constructed. Through interaction with the environment (real-time monitoring data) and a reward feedback mechanism, the network is dynamically trained, enabling it to adaptively learn and optimize classification decision strategies. This allows the model to autonomously optimize classification strategies in complex environments, significantly improving the recognition accuracy of different fish species and growth stages. For a newly input fine-grained morphological feature vector of a fish target, the trained deep reinforcement learning policy network outputs the fish species and growth stage assessment results of the target based on the currently learned optimal policy, and analyzes and infers its common behavior patterns based on the feature-behavior association rule knowledge base, thereby completing target classification and behavior analysis in real time, quickly identifying abnormal behavior, and providing an immediate decision-making basis for fish passage warning.

[0017] The execution steps of the fish monitoring data extraction module include: The system receives and integrates the species, growth stage, and behavioral tag information of all fish targets in real time, and associates them with the spatiotemporal stamps of their passage to form a structured fish passage event record stream, ensuring data integrity and temporal consistency, and providing a reliable basis for subsequent analysis. Based on the fish passage event record stream, rolling calculations are performed according to a preset time window to generate core statistical indicators including the total number of fish passing through, the number and proportion of various fish species, the size distribution of each growth stage, and the passage frequency time series, so as to realize dynamic monitoring and trend analysis and support real-time decision-making and fish passage assessment. By combining core statistical indicators with spatial distribution heat maps and behavioral pattern analysis conclusions, and automatically formatting according to preset templates, a multi-dimensional fish monitoring report is generated, which includes the composition of fish species, size distribution, peak passage periods, and behavioral preferences. The report outputs an intuitive and comprehensive fish monitoring report, improving management efficiency and scientific rigor.

[0018] Compared with the prior art, the beneficial effects of the present invention include: (1) This system achieves high-precision, non-invasive monitoring of fish passing through fish passages by deeply integrating sonar intelligent imaging and AI recognition technology. The sonar array adopts a group-based time-division transmission strategy to generate high-density parallel beams, which can clearly image under conditions of turbid water or insufficient light, effectively overcoming the limitations of traditional optical monitoring methods in low visibility environments. Combined with an improved ant colony algorithm, it can intelligently separate overlapping targets, significantly improving the accuracy of individual identification in cases of high-density fish groups or overlapping targets.

[0019] (2) By introducing fine-grained feature extraction based on deep residual networks and a classification decision model driven by reinforcement learning, this system can adaptively learn the morphological and behavioral characteristics of different fish species, realize the full-stage identification from juvenile fish to adult fish, break through the technical bottleneck of traditional acoustic monitoring that can only count but not classify, and provide richer biological parameters for fish passage effect evaluation.

[0020] (3) By constructing a feature-behavior association rule knowledge base and an adaptive classification decision mechanism, this system can continuously learn and optimize the recognition model, adapt to changes in different aquatic environments and fish behavior, integrate multi-source information in real time, and automatically generate fish monitoring reports containing multiple dimensions such as species composition, size distribution, and spatiotemporal patterns, which significantly improves the scientificity and timeliness of fishway operation effect evaluation and fish resource protection strategy formulation. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This is a flowchart illustrating a method for monitoring fish passage in water conservancy and hydropower projects that couples AI recognition and sonar.

[0023] Figure 2 This is a data flow diagram of a fish passage monitoring system for hydraulic and hydropower projects that couples AI recognition and sonar. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1 A method for monitoring fish passage in fishways of hydraulic and hydropower projects that combines AI recognition and sonar, such as Figure 1 As shown, it includes the following steps: S1. Intelligent sonar imaging: An intelligent sensing system that deploys a multi-element sonar array in the fishway uses a group time-division transmission strategy to generate a high-density parallel beam to scan the fishway cross-section. After envelope detection and digital processing of the received echo signals, pixel mapping is used to synthesize an acoustic image of the fishway cross-section. S2. Target separation calculation: Identify fish targets in the acoustic image. If there is a high-density passage of fish or target overlap, the mixed echo signal is separated by an improved ant colony algorithm combined with a physical acoustic model. If there is no high-density passage of fish or target overlap, the fish is directly extracted to obtain individual fish acoustic morphology data. S3. Morphological depth recognition: After preprocessing the acoustic morphological data of the individual fish, the YOLOv7 target detection algorithm is used to locate the fish target and output the accurate contour bounding box. Then, the ResNet50 deep residual network is used to extract the fine-grained morphological feature vector of the fish. S4. Adaptive classification decision: Combining the fine-grained morphological feature vectors, a deep reinforcement learning strategy network is used to complete fish species identification and growth stage assessment, and fish common behavior patterns are analyzed based on the feature and behavior association rule knowledge base. S5. Monitoring Data Extraction: Integrates information on fish species, growth stages, behavioral patterns, and corresponding spatiotemporal stamps, calculates core statistical indicators in a rolling manner according to preset time windows, and automatically generates multi-dimensional fish monitoring reports by combining spatial distribution heatmaps.

[0026] Example 2 A fish passage monitoring system for water conservancy and hydropower projects that couples AI recognition and sonar, such as Figure 2 As shown, it includes: The intelligent sonar imaging module, based on an intelligent sensing system encompassing a multi-element sonar array, combines a grouped time-division transmission strategy to generate high-density parallel beams for fine scanning of the fishway cross-section. This yields acoustic images of the fishway cross-section with clear target boundaries, ensuring imaging quality from the source and effectively improving image spatial resolution and signal-to-noise ratio, thus enhancing target contour recognition capabilities. By deploying the intelligent sensing system encompassing a multi-element sonar array within the target fishway, the array elements are divided into eight equal-numbered groups. Each group's elements are sequentially numbered, and all elements with the same number in each group simultaneously transmit short-pulse high-frequency acoustic signals. Immediately after each group's transmission, the corresponding element's receiving process begins, ensuring that each element... A highly directional parallel beam is independently generated to ensure beam direction consistency, improve signal spatial directivity and anti-interference capability, and repeat the group time-division transmission and reception process until the transmission and reception of all array elements in all array groups are completed. A high-density parallel beam scanning network covering the entire monitoring area is formed in the fishway cross section to achieve full cross section coverage scanning, avoid monitoring blind spots, and improve target acquisition rate. The echo signal received by each array element is independently envelope detected and digitized, and the echo energy data of all parallel beams are spatially arranged and pixel mapped according to the physical deployment order of the array elements to synthesize a fishway cross section acoustic image with clear target boundaries, generating a high-resolution acoustic image that clearly presents the fish body outline and spatial distribution. The specific work involves: selecting the physical dimensions and number of elements of a multi-element sonar array based on the specific width and depth of the target fishway. The array is installed vertically from top to bottom along one side wall of the fishway, ensuring its central axis is perpendicular to the direction of water flow. Each element in the array is an independent transceiver, and its operating frequency is selected based on a combination of resolution and water attenuation characteristics. The transmit beamwidth of a single element is designed to be 1.5° to 2.5° to ensure high directivity. During system initialization, the control unit divides all 64 elements into 8 equal groups and sequentially numbers the elements within each group along the vertical direction, driving all 8 groups sequentially. Array elements with the same number synchronously transmit short-pulse acoustic signals, with pulse widths set to 50 to 100 microseconds. The pulse repetition interval is determined based on the fishway width W (set to 10) and the speed of sound C (approximately 1480 m / s in freshwater). This interval must be slightly greater than 2W / C to avoid multiple echo interference. The transmission control logic follows a grouped time-division sequence, controlling the first array element in all eight groups to transmit simultaneously. Immediately after transmission, the system switches to receiving mode, initiating synchronous signal acquisition for the corresponding eight array elements. Each receiving channel pre-amplifies, bandpass filters, and dynamically adjusts the gain of the echo signal, followed by analog-to-digital conversion. After this process, the second array element in each group is controlled sequentially. The transmit-receive cycle is repeated for array elements 3 through 8, thus forming 64 independent highly directional parallel beams within a single complete scan cycle. These beams form a dense scanning network within the fishway cross-section, with an angular interval of approximately 1.4° between beams, achieving seamless coverage of the entire monitoring section. All digitized echo data are buffered according to the physical position (i.e., vertical sequence) of the array elements, with each data point containing a timestamp and amplitude information. Envelope detection (using Hilbert transform) is performed on the discrete echo sequence acquired by each array element to extract its energy-time domain curve. Subsequently, based on the known spatial position (vertical coordinates) and beam pointing angle of each array element... (Fixed to be perpendicular to the array plane normal), each energy curve is converted into radial distance information along the beam direction with time delay. Through coordinate transformation, the 64 radial energy distributions are mapped to a unified two-dimensional rectangular coordinate system (with the array plane as the reference plane), forming an initial energy distribution matrix with distance as the horizontal axis and depth as the vertical axis. Pixel interpolation and dynamic range compression are performed on this matrix to finally synthesize a fishway cross-section acoustic image. The horizontal spatial resolution of this fishway cross-section acoustic image is determined by the pulse length, and the vertical resolution is determined by the array element spacing and beamwidth. It can clearly present the outline boundary of the fish target and the water structure within the fishway cross-section. The target separation and calculation module is used to introduce an improved ant colony algorithm to intelligently calculate and simulate the mixed echo signals generated by high-density fish passage or target overlap in the acoustic image of the fishway cross section. It effectively separates overlapping individuals, obtains acoustic morphology data of individual fish, improves the accuracy of individual target identification, significantly reduces the false detection and false detection rate of targets in high-density fish scene, and achieves accurate individual separation. The target separation and calculation module includes an intelligent trajectory simulation unit and an overlapping signal decoupling unit. The intelligent trajectory simulation unit applies an improved ant colony algorithm to simulate the movement behavior and trajectory distribution of fish in a confined space. It intelligently segments and predicts the trajectories of overlapping or partially overlapping echo regions in the acoustic image of the fishway cross-section, separating the movement contours of suspected independent individuals. This effectively tracks the movement paths of fish within overlapping areas, improving trajectory continuity and segmentation accuracy. Based on the acoustic image of the fishway cross-section output by the sonar intelligent imaging module, it identifies echo energy adhesion regions caused by high-density fish passage or target overlap, and initializes the pixel set of these regions as the search space to be explored by artificial ants in the ant colony algorithm, effectively clustering... Focusing on high-density areas improves algorithm processing efficiency. The ant colony algorithm is run in the search space. By introducing local gradient pheromones and dynamic heuristic functions, the possible movement paths of fish in the limited space of the fishway are simulated. Artificial ants are guided to iteratively traverse and simulate the echo energy distribution of overlapping areas, which significantly enhances the accuracy of path search and target tracking ability. Based on the path trajectory clusters formed after the convergence of the ant colony algorithm, the initially stuck echo area is intelligently segmented to separate the movement contours of multiple suspected independent individuals. A unique trajectory identifier is assigned to each movement contour to achieve effective separation of overlapping individuals, which facilitates subsequent accurate identification and tracking. The specific work involves: after generating the acoustic image of the fishway cross-section, preprocessing the image to convert the original acoustic intensity data into a normalized grayscale image with a grayscale range of 0-255. By calculating the energy gradient features of local areas of the image, automatically identifying areas where the echo energy value exceeds a preset threshold (set to 1.5-2.0 times the average background energy) and is spatially continuous, and marking them as potential target adhesion areas. Each adhesion area consists of a set of spatially adjacent pixels. The coordinates of the pixels are extracted and constructed into a two-dimensional point set, which serves as the search space for the subsequent ant colony algorithm. The boundary of the search space is determined by the minimum bounding rectangle of the adhesion area. The size of the rectangle does not exceed 15% of the effective monitoring range of the fishway cross-section to ensure that the algorithm focuses on local high-density targets. At the same time, the energy gradient amplitude and direction of each pixel in the search space are calculated as the basic input of heuristic information. During this process, the image resolution is 2-5 cm per pixel corresponding to the actual distance. The target detection sensitivity is dynamically adjusted according to the turbidity of the water. When the turbidity is high, the threshold is appropriately reduced to improve the recall rate. Within the initialized search space, a set number of artificial ants (50-100) are deployed. Each ant starts from a randomly selected boundary pixel and moves iteratively according to the improved ant colony algorithm rules. The state transition probability of the algorithm is determined by two parts: first, the local gradient pheromone, whose concentration is proportional to the echo energy gradient amplitude of the pixel; the higher the energy, the higher the pheromone concentration. Second, the dynamic heuristic function, which comprehensively considers the continuity of the gradient direction of the pixel, the proximity to the adjacent explored regions, and the path smoothness constraint. When the ant moves in each step, it selects the position with the highest state transition probability among the 8 neighboring connected pixels as the next target, with the forward direction change not exceeding ±30° as a constraint. After each iteration, the pheromone concentration on the path is updated according to the total energy gradient amplitude of the pixels covered by the ant's path. Paths with higher energy receive more pheromone enhancement. The number of algorithm iterations is set to 100-200. When the optimal path no longer shows significant improvement (path energy change less than 5%) in 20 consecutive iterations, it is considered converged. After convergence, the algorithm... The effective paths explored by ants are clustered into several trajectory clusters. After the algorithm converges, the trajectory clusters represent the most likely fish movement path hypotheses within the adhesion region. Each trajectory cluster is post-processed, and morphological operations are used to fill in the small discontinuities in the trajectory to ensure path continuity. Then, the center line of each complete trajectory is extracted, and the trajectory is extended to both sides based on the center line. The extension width is adaptively determined according to the local curvature and energy distribution of the trajectory, and controlled within the range of 4-10 pixels (corresponding to the actual fish width). The strip-shaped region formed after the extension is then initially segmented into an independent motion contour. Subsequently, the geometric features of each contour are calculated, including contour area, aspect ratio, and average energy intensity within the contour. For contours that are too close in space (center distance less than 30% of the shorter contour length) and have the same direction of movement, a second merging and verification is performed to exclude over-segmentation caused by partial occlusion of a single fish body. Finally, a unique trajectory identifier is assigned to each independent contour that passes the verification. This trajectory identifier is bound to its spatial location, timestamp, and extracted feature vector, and serves as the output of the intelligent trajectory simulation unit. The overlapping signal decoupling unit combines the physical acoustic model with the motion contours of the separated suspected independent individuals to perform inverse calculations on the echo energy and time delay information of the overlapping area. This decomposes the mixed echo signal into multiple independent target signals conforming to acoustic scattering characteristics, outputting clearly defined acoustic morphological data of the separated individual fish bodies. This achieves physical-level decomposition of complex echo signals, restoring the complete acoustic scattering characteristics of a single fish body. For each output motion contour of a suspected independent individual, a corresponding acoustic body scattering physical model is established and matched with the sound wave propagation characteristics of the fishway waters. A theoretical echo generation model based on the contour shape and spatial position is constructed, achieving high-fidelity acoustic feature modeling, enhancing the consistency between the model and actual fish scattering, and improving subsequent separation accuracy. To ensure accuracy, within the spatial region corresponding to the motion contour, the energy distribution and time delay information of the mixed echo signal in the original acoustic image are extracted. Combined with the theoretical echo generation model, non-negative matrix decomposition is used to inversely solve the mixed echo, estimating the energy contribution ratio of each independent target to the mixed echo. This effectively decomposes overlapping signals, accurately quantifies the energy proportion of each target, and significantly improves the objectivity and interpretability of target separation. Based on the estimated energy contribution ratio, the clear echo signal corresponding to each independent fish target is separated and reconstructed from the mixed echo signal. A set of individual fish acoustic morphology data with complete acoustic scattering characteristics and clear contour boundaries is output, resulting in high-quality individual acoustic morphology data. This provides clear and reliable input for target recognition and classification, supporting subsequent intelligent analysis. The specific work involves: establishing a corresponding acoustic scattering physical model for the motion profile of each suspected independent individual. This model is based on the theory of elastic body scattering and fish morphological parameters (including profile aspect ratio, equivalent diameter, and local curvature). A layered medium scattering model is used to describe the acoustic impedance difference between the fish tissue and the water. The model inputs include the geometric center coordinates of the profile, the principal axis direction angle of the profile, the equivalent acoustic scattering cross section (calibrated based on measured data in the 1.5-2.5MHz frequency band), and the local water sound velocity profile (measured in layers every 0.5 meters). Simultaneously, combined with the measured sound wave propagation characteristics of the fishway waters (including sound attenuation coefficient α = 0.08-0.15 dB / m·MHz and sound velocity gradient 0.5-1.2 m / s·m), a sound wave... The propagation path compensation function discretizes the contour into 200-500 scattering units, assigning each unit a time delay compensation value and amplitude attenuation factor corresponding to its spatial location. This ultimately generates a normalized theoretical echo sequence based on the geometric center of the contour, with the sequence length matching the transmitted pulse width and a sampling interval of 0.1 μs. Within the spatial region corresponding to the moving contour, the energy distribution matrix and time delay vector of the mixed echo signal in the original acoustic image are extracted. The energy distribution matrix covers a monitoring area centered on the contour's circumscribed rectangle, extending by 15-25%, with a spatial sampling interval of 2-5 cm and a temporal sampling rate of 10-20 MHz. The time delay vector is calculated based on the sound wave's round-trip path, including time difference compensation for direct waves and boundary reflected waves, based on theoretical echo... A dictionary matrix is ​​constructed for the wave generation model, where each column corresponds to the theoretical echo sequence of a scattering unit. A non-negative matrix factorization algorithm with sparsity constraints is employed. The L1 regularization coefficient in the objective function is set between 0.05 and 0.12 to control the uniqueness of the solution. The number of iterations is set to 50-80, and the convergence threshold is a residual change rate ≤ 0.5%. By decomposing the mixed echo matrix, an energy contribution ratio matrix is ​​obtained. Each row of this energy contribution ratio matrix corresponds to the energy weight of a scattering unit, and each column corresponds to the energy distribution ratio of a time sampling point, accurately quantifying the energy proportion of each independent target in the mixed signal. Based on the energy contribution ratio matrix, the mixed echo signal is component-separated and reconstructed. First, the theoretical echo sequence of each scattering unit is... The corresponding energy weights are applied, and the contributions of each unit are superimposed along the time dimension to generate the initial reconstructed signal for each independent target. Then, phase consistency correction is performed, and a time delay fine-tuning algorithm based on cross-correlation function is adopted (adjustment step size 0.05μs, search window ±1μs) to ensure that the phase deviation between the reconstructed signal and the original signal at the main scattering points is ≤5°. During the reconstruction process, bandpass filtering and adaptive noise suppression are implemented simultaneously (signal-to-noise ratio improvement ≥6dB). The final output of individual fish acoustic morphology data includes the following structured information: reconstructed echo sequence, scattering intensity distribution map, effective scattering center coordinates and contour boundary confidence coefficient. All data are stored indexed by timestamp and trajectory identifier to form a standardized acoustic morphology dataset that can be used for classification and recognition. Let the energy distribution matrix of the hybrid echo signal be... ,in The number of time sampling points, The number of spatial sampling points is ; the dictionary matrix is ,in The number of scattering units per column Indicates the first The theoretical echo sequence of scattering units; the energy contribution ratio matrix is Its elements Indicates the first The scattering unit at the th ... Energy allocation ratio for each time sampling point; Through nonnegative matrix decomposition with sparse constraints, the energy contribution ratio matrix The estimation formula is: ; In the formula: This represents the energy distribution matrix of the mixed echo signal; Represents a dictionary matrix; This represents the energy contribution ratio matrix to be solved; The Frobenius norm is used to measure the performance of a matrix. and Reconstruction error between; Represents the L1 regularization coefficient, used to control the matrix. The sparsity of the solution ensures its uniqueness and stability. L1 norm represents the sum of the absolute values ​​of all elements in a matrix. It is used to penalize the number of non-zero elements and influence the energy contribution ratio matrix. It has sparsity; The morphological depth recognition module is used to receive the calculated acoustic morphological data of individual fish bodies. It uses the YOLOv7 target detection algorithm to quickly locate and frame fish targets, and uses the ResNet50 deep residual network to perform fine-grained pattern recognition on fish targets to extract fine-grained morphological features required for species identification. This enables real-time localization and high-precision morphological feature extraction of fish targets, supporting subsequent classification and recognition. The adaptive classification decision module combines extracted fine-grained morphological features and reinforcement learning strategy networks to dynamically learn and associate behavioral patterns and morphological association rules of different fish species and different growth stages, constructing a classification decision model to identify fish species, assess growth stages, and analyze passage behavior in fish passages. This enables intelligent classification of fish species and growth stages, as well as association with behavioral patterns, thereby improving the depth of target analysis. The fish monitoring data extraction module integrates classification results, quantity statistics, spatiotemporal distribution, and behavioral pattern information to automatically generate multi-dimensional fish monitoring reports, including the composition of fish species, size distribution, peak passage periods, and behavioral preferences. This provides direct data support for fishway effectiveness assessment and fish resource protection strategy formulation. The automated generation of multi-dimensional fish monitoring reports provides data basis for fishway operation effectiveness and fish resource protection.

[0027] Example 3 Based on Example 2, the present invention provides a technical solution: the morphological depth recognition module includes a target detection unit and a fine-grained feature extraction unit; The target detection unit, based on the YOLOv7 target detection algorithm, processes the calculated individual fish acoustic morphology data in real time, locates and outlines the contours of all fish targets in the image, quickly and accurately locates the contours of each fish, providing reliable boundary information for feature extraction. It receives the individual fish acoustic morphology data, normalizes it into a grayscale image format with uniform size and intensity, and performs data augmentation operations to construct a standardized dataset suitable for deep learning model input, ensuring data consistency and improving the stability and generalization ability of model training. This, combined with the preprocessed individual fish acoustic morphology data and YOLOv7... The target detection algorithm, through the backbone feature extraction layer and multi-scale prediction head of the YOLOv7 target detection algorithm, performs fast forward inference on fish targets in the image, generates multiple candidate bounding boxes and their corresponding confidence scores in real time, achieves efficient target localization, and significantly improves the real-time performance and coverage of detection. The non-maximum suppression algorithm is applied to filter the candidate bounding boxes output by the YOLOv7 target detection algorithm, eliminating redundant and low-confidence detection boxes, determining and outputting the precise contour bounding box coordinates of each fish target in the image, effectively eliminating overlapping box interference, and ensuring the accuracy and uniqueness of the output bounding boxes. The specific work involves: in the target detection unit, standardizing the input acoustic morphology data of individual fish to ensure data consistency and model training effectiveness. Specifically, the original acoustic morphology data is input in the form of an acoustic scattering intensity distribution map. The size of this map varies depending on the target size and imaging resolution. The input image is then uniformly adjusted to a fixed size, standardized at 256 pixels × 256 pixels. Bicubic interpolation is used to maintain the geometric features of the image without distortion. For intensity normalization, pixel values ​​are linearly mapped to a grayscale range of 0-255. The mapping formula is as follows: ; in, These are the normalized pixel values ​​(0-255). This is the original strength value. and The minimum and maximum intensities in the image are represented by these values. To further enhance data diversity and model robustness, data augmentation techniques are applied during the training phase, including random horizontal flipping (probability set to 0.5), random rotation within ±5°, and brightness adjustment (adjustment coefficient range 0.8-1.2). All augmentation operations are performed after normalization to ensure that the input data is standardized in size and intensity. After preprocessing, the standardized images are input into the YOLOv7 object detection network for forward inference. The network uses CSPDarknet53 as the backbone feature extraction... The network takes a layer and takes an input image size of 256×256. It outputs three feature maps at different scales: 32×32, 16×16, and 8×8, corresponding to the detection of fish targets of different sizes. In the prediction head at each scale, the network predicts bounding box coordinates, a confidence score, and a class probability for each grid cell. The bounding box coordinates are represented by an offset relative to the grid cell. The confidence score is normalized to between 0 and 1 using the sigmoid function, reflecting the probability of the target being present in the box. During forward inference, a confidence threshold of 0.4 is set. Only candidate boxes with scores higher than this confidence threshold are retained to initially filter out low-quality detections. Simultaneously, to avoid missed detections, the IoU threshold for non-maximum suppression is set to 0.3, but this is only used for internal filtering at this stage. The network inference speed is controlled to be no more than 20 milliseconds per frame to meet real-time processing requirements. After obtaining multiple candidate bounding boxes from forward inference, a non-maximum suppression algorithm is used for further filtering to eliminate redundant detection boxes and determine the final output. Specifically, all candidate boxes are sorted in descending order of confidence score, and the box with the highest score is selected as the benchmark. Its intersection with the remaining boxes is calculated. The intersection-union ratio (IU) of a bounding box with the reference bounding box exceeds a set threshold (standard setting is 0.50), and it is considered a redundant bounding box and is removed. This process is repeated until all bounding boxes have been processed. Then, the coordinates of the remaining bounding boxes are refined, the relative coordinates are converted back to absolute image coordinates, and it is ensured that they do not exceed the image boundary. Finally, each fish target bounding box is represented in the form of a quadruple, namely the x-coordinate of the upper left corner, the y-coordinate of the upper left corner, the width and the height, with the precision at the pixel level. All bounding box information and corresponding confidence scores are output together to complete the entire target detection process. The fine-grained feature extraction unit uses a ResNet50 deep residual network to perform fine-grained pattern recognition on the contour boundaries of detected fish targets. It extracts fine-grained morphological features, including fish contour, body size, aspect ratio, and local scattering intensity distribution, forming deep feature vectors that can characterize different fish species. This extraction of highly discriminative deep-level morphological features enhances the ability to distinguish between species and growth stages. Based on the output precise contour bounding box coordinates, it extracts the corresponding local image region for each fish target from the original individual fish acoustic morphological data. The extracted target region image is then scaled to the standard input size of the ResNet50 deep residual network to ensure uniform image size and avoid deformation, providing a regularized input for subsequent feature extraction. The target region image is then input into a pre-trained ResNet50 deep residual network. Using multiple residual block-level convolution and pooling operations, the network performs layer-by-layer abstraction and feature encoding on the target region image, extracting a deep feature map containing multi-level semantic information. The deep network automatically learns and encodes the details and structural features of the fish body morphology, improving feature representation capabilities. Global average pooling is performed on the deep feature map to compress it into a fixed-dimensional feature vector. This feature vector integrates the outline shape, body structure, aspect ratio, and acoustic scattering intensity distribution pattern of the local area in the image to form a fine-grained morphological feature descriptor for subsequent classification, generating a compact feature vector with strong discriminative power, effectively supporting high-precision classification of fish species and growth stages. The specific work involves: In actual operation, based on the precise contour bounding box coordinates output by the target detection unit, accurately extracting the local image region corresponding to each fish target from the original individual fish acoustic morphology data matrix. This original data is a two-dimensional acoustic scattering intensity distribution map, where the pixel values ​​represent the normalized acoustic echo intensity. The extraction operation starts from the upper left corner of the bounding box, with the width and height of the box as the range. The corresponding sub-matrix is ​​directly indexed from the original data matrix. To ensure the standardization of subsequent deep network processing, the extracted target region image is uniformly scaled to the standard input size of the ResNet50 deep residual network, i.e., 224 pixels × 224 pixels. The scaling process uses a bilinear interpolation algorithm to maximize the preservation of the continuity of its geometric structure and intensity distribution while adjusting the image size. Fish target images of different sizes and positions are uniformly transformed into standardized inputs with a fixed spatial dimension. All processing is performed in memory using matrix operations, and the processing time for a single image is controlled within 5 milliseconds to meet the system's real-time requirements. The target region image is input into a pre-trained ResNet50 deep residual network for forward propagation to complete deep feature encoding. The network weights are pre-trained on the ImageNet large visual dataset and then fine-tuned through transfer learning to adapt to the domain characteristics of the acoustic image. The image passes through the network's 7×7 convolutional layer, max pooling layer, and four stages of residual block groups (containing 3, 4, 6, and 3 residual blocks respectively). Each residual block contains a sequence of 1×1, 3×3, and 1×1 convolutional layers. With batch normalization and ReLU activation function, the network achieves layer-by-layer nonlinear transformation and abstraction of the input image. During this process, the network's receptive field expands step by step. Lower-level residual blocks capture local details such as the edge and texture of the target, while higher-level residual blocks integrate global contextual information to form a deep feature map containing rich multi-level semantic information. Finally, after passing through all convolutional layers, the network outputs a 7×7 deep feature map with 2048 channels. This deep feature map compactly encodes the morphological structure and scattering pattern of the fish target.Global average pooling is performed on the depth feature map (7×7×2048) generated by the ResNet50 deep residual network to generate a fixed-dimensional feature vector. Specifically, the arithmetic mean of all pixel values ​​is calculated for each of the 2048 feature maps in each channel along the spatial dimensions (height and width) of the feature map, thereby compressing the 7×7 feature matrix of each channel into a single scalar value, and finally generating a one-dimensional feature vector with a dimension of 2048. This vector integrates the global statistical information of the fish target in the entire image area, and integrates the macroscopic geometric characteristics of its contour shape, the three-dimensional spatial implications of its body structure, the morphological proportion reflected by the aspect ratio, and the distribution pattern and texture features of the sound wave scattering intensity in the local area of ​​the image. The resulting 2048-dimensional feature vector constitutes a fine-grained morphological feature descriptor with strong discriminative power, and at the same time, a structured data package containing the fine-grained morphological feature descriptor is generated. The execution steps of the adaptive classification decision module include: collecting historical monitoring data, constructing a correspondence between fine-grained morphological feature vectors of fish at different growth stages and their typical common behavior patterns, forming an initial feature-behavior association rule knowledge base, forming an interpretable behavior rule base covering multiple fish species and stages, improving classification credibility and system interpretability, using fine-grained morphological feature vectors as state input and fish species and growth stage classification labels as action space, constructing a deep reinforcement learning policy network, dynamically training the network through interaction with the environment (real-time monitoring data) and reward feedback mechanism, enabling it to adaptively learn and optimize classification decision strategies, realizing the model's autonomous optimization of classification strategies in complex environments, significantly improving the recognition accuracy of different fish species and growth stages, for a new input fish target fine-grained morphological feature vector, the trained deep reinforcement learning policy network outputs the fish species and growth stage evaluation results of the target based on the currently learned optimal strategy, and analyzes and infers its common behavior patterns based on the feature-behavior association rule knowledge base, completing target classification and behavior analysis in real time, quickly identifying abnormal behavior, and providing immediate decision-making basis for fish passage warning; The specific work involves: In the initial stage of system deployment, constructing a feature-behavior association rule knowledge base based on historical monitoring data. Specifically, this involves collecting at least 12 months of acoustic monitoring data on fish passages under different seasons and hydrological conditions. Fine-grained morphological feature vectors of individual fish at labeled species and growth stages are extracted. Each feature vector is a 2048-dimensional floating-point array, corresponding to a pre-processed acoustic morphological sample. Simultaneously, the corresponding passage behavior pattern label is recorded, including upstream direction, swimming depth, passage speed, and whether the fish are passing in groups. The knowledge base uses a graph-structured database for storage, where nodes represent feature vectors and edges represent behavior association rules. Rule weights are calculated using statistical co-occurrence frequency and mutual information. The update cycle is every 30 days. The initial knowledge base must contain at least 200 samples each of at least 5 common fish species and 3 growth stages (juvenile, subadult, and adult) to ensure the typicality and statistical significance of rule coverage. During system operation, the knowledge base supports real-time insertion of newly labeled samples and updates through an incremental learning mechanism. Association rules are used to maintain the timeliness and adaptability of the knowledge base. Based on the constructed knowledge base, a deep reinforcement learning policy network is designed and trained to achieve adaptive classification of fish species and growth stages. The policy network adopts an Actor-Critic architecture, where the state space is a 2048-dimensional feature vector, and the action space is a discrete-continuous hybrid form: the discrete part corresponds to the fish species category, and the continuous part corresponds to the growth stage score (0.0 to 1.0, divided into three intervals: juvenile, sub-adult, and adult). The network hidden layers are set to 3 fully connected layers, with 512, 256, and 128 neurons per layer, respectively. The ReLU activation function is used. The output layer outputs action probabilities through Softmax (category) and Sigmoid (stage) respectively. During training, the agent interacts with the environment: for each input feature vector, the network outputs a classification action, and calculates an immediate reward based on the difference between the action and the true label. In the reward function, the reward is +1 for correct category, +0.5 for stage error less than 0.1, and -0.3 otherwise. The PPO algorithm is used for training, with a learning rate of 3×10. -4The discount factor γ = 0.99, the batch size for each training round is 64, and the total number of training rounds is no less than 500 rounds until the accuracy of the validation set stabilizes at over 92% and the overfitting coefficient is less than 1.05. After training, the policy network is deployed in the online inference module to process the new input fish target feature vectors in real time. After the features of each frame of acoustic image are extracted by the aforementioned module, they are input into the policy network for forward inference. The network outputs the category probability distribution and stage score within 5 milliseconds, and takes the one with the highest probability as the final classification result. At the same time, the feature-behavior association rule knowledge base is retrieved according to the classification result, and the typical behavior pattern template of the fish species and growth stage is matched. Combined with the actual movement trajectory of the current target (obtained by association with the trajectory identifier), the deviation of its behavior parameters from the template is calculated. If the speed deviation exceeds 30% or the swimming layer deviation exceeds 50%, it is marked as a behavior abnormality and an early warning is triggered. All classification results, behavior analysis conclusions and abnormal markers are written into the fish monitoring database in time series. The execution steps of the fish monitoring data extraction module include: receiving and integrating all fish species, growth stages, and behavioral tag information in real time, while associating them with their spatiotemporal stamps to form a structured fish passage event record stream, ensuring data integrity and temporal consistency, providing a reliable basis for subsequent analysis; based on the fish passage event record stream, performing rolling calculations according to preset time windows to generate core statistical indicators including the total number of fish passing through, the number and proportion of each type of fish species, the size distribution of each growth stage, and the passage frequency time series, realizing dynamic monitoring and trend analysis, supporting real-time decision-making and fish passage assessment; combining the core statistical indicators with spatial distribution heat maps and behavioral pattern analysis conclusions, automatically formatting according to preset templates to generate a multi-dimensional fish monitoring report containing the composition of fish species passing through, size distribution, peak passage periods, and behavioral preferences, outputting an intuitive and comprehensive fish monitoring report, improving management efficiency and scientific rigor; The specific tasks are as follows: The data acquisition interface receives structured data packets from the adaptive classification decision module in real time. Each data packet corresponds to a detected fish target, containing the target type, growth stage, behavioral tags, target spatial coordinates, and a high-precision timestamp. The data receiving service is deployed on an edge server, using a message queue for asynchronous buffering and order guarantee to ensure no data loss or out-of-order processing under high-concurrency scenarios (peak processing capacity ≥ 20 events / second). The received raw event data is immediately injected into the pipeline processing engine. The engine parses, verifies, and cleans the data according to a preset data pattern, removing invalid records with format errors or missing key fields. Verified event records are appended to the time-series database in chronological order. Each record contains complete attribute fields, forming a fish passage event record stream with time as the unique primary key, allowing for high-speed querying along the time dimension. The built-in configurable statistical calculation engine performs rolling calculations based on the fish passage event record stream, according to preset fixed time windows. The calculation is initiated by a time trigger and automatically executes at the end of each window. The engine first extracts all valid event records within the corresponding time window from the time-series database, and then calculates four core statistical indicators according to a predetermined algorithm: (1) Total number of fish caught: directly count the number of event records; (2) Quantity and percentage of each type of fish: Group statistics by type field and calculate the percentage of each type of fish in the total; (3) Size distribution of each growth stage: Group statistics according to the growth stage field, and calculate the average equivalent body length of fish at each stage (based on acoustic morphology model inversion) and its distribution histogram (group interval 5cm). (4) Passage Frequency Time Series: Using 5-minute intervals as sub-intervals, the number of fish passing through each sub-interval is counted to generate time-frequency series data. All calculation results are output in structured JSON format, along with metadata such as the time window on which the calculation is based, the total number of records, and the data quality score (based on record completeness and validity), and stored in the analysis results database. Report generation is triggered periodically, and the latest core statistical indicator results are retrieved from the analysis results database. At the same time, the spatial coordinate data of fish targets within the same time period are obtained to generate two-dimensional spatial distribution heatmaps of the longitudinal and transverse sections of the fish passage. The heatmap pixel resolution is... The system measures 0.1m x 0.1m, uses color mapping to represent target density, and provides a summary of behavioral patterns for the given time period. Then, following a predefined, industry-standard Word / PDF template, it automatically fills in and formats three parts: core statistical indicator tables and charts, spatial distribution heatmap, and behavioral pattern analysis text. The resulting fishway fish passage monitoring report includes a summary, monitoring overview, detailed data analysis (category composition, size distribution, passage time patterns, spatial distribution characteristics, behavioral pattern analysis), main conclusions, and recommendations. This report is then automatically distributed to pre-defined recipients via a file server or email system.

[0028] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0029] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring fish passage in fishways of hydraulic and hydropower projects that couples AI recognition and sonar, characterized in that, Includes the following steps: S1. Intelligent sonar imaging: An intelligent sensing system that deploys a multi-element sonar array in the fishway uses a group time-division transmission strategy to generate a high-density parallel beam to scan the fishway cross-section. After envelope detection and digital processing of the received echo signals, pixel mapping is used to synthesize an acoustic image of the fishway cross-section. S2. Target separation calculation: Identify fish targets in the acoustic image. If there is a high-density passage of fish or target overlap, the mixed echo signal is separated by an improved ant colony algorithm combined with a physical acoustic model. If there is no high-density passage of fish or target overlap, the fish is directly extracted to obtain individual fish acoustic morphology data. S3. Morphological depth recognition: After preprocessing the acoustic morphological data of the individual fish, the YOLOv7 target detection algorithm is used to locate the fish target and output the accurate contour bounding box. Then, the ResNet50 deep residual network is used to extract the fine-grained morphological feature vector of the fish. S4. Adaptive classification decision: Combining the fine-grained morphological feature vectors, a deep reinforcement learning strategy network is used to complete fish species identification and growth stage assessment, and fish common behavior patterns are analyzed based on the feature and behavior association rule knowledge base. S5. Monitoring Data Extraction: Integrates information on fish species, growth stages, behavioral patterns, and corresponding spatiotemporal stamps, calculates core statistical indicators in a rolling manner according to preset time windows, and automatically generates multi-dimensional fish monitoring reports by combining spatial distribution heatmaps.

2. The method for monitoring fish passage in water conservancy and hydropower projects that couples AI recognition and sonar as described in claim 1, characterized in that, The specific execution method of the grouped time-division transmission strategy in step S1 is as follows: the array elements of the multi-element sonar array are divided into multiple array element groups of equal number and the array elements in the group are numbered sequentially. The array elements with the same number in all array element groups are controlled to synchronously transmit short pulse high-frequency acoustic signals. After the transmission is completed, the receiving process of the corresponding array element is started immediately. This operation is repeated until the transmission and reception of all array elements are completed, forming a high-density parallel beam scanning network covering the fishway monitoring area.

3. The method for monitoring fish passage in water conservancy and hydropower projects that couples AI recognition and sonar as described in claim 1, characterized in that, The specific steps for solving and separating the mixed echo signal in step S2 are as follows: First, identify the echo energy adhesion region in the acoustic image as the search space of the improved ant colony algorithm, introduce local gradient pheromone and dynamic heuristic function to simulate the movement path of the fish, segment out the movement contour of suspected independent individuals and assign a unique trajectory identifier; then, construct a theoretical echo generation model for each movement contour, use non-negative matrix decomposition to reverse solve the mixed echo signal, and separate and reconstruct the acoustic morphological data of the individual fish with clear contours.

4. The method for monitoring fish passage in water conservancy and hydropower projects that couples AI recognition and sonar as described in claim 1, characterized in that, The preprocessing described in step S3 involves normalizing the acoustic morphology data of individual fish bodies into a grayscale image format with uniform size and intensity, and performing data enhancement operations such as random flipping, rotation, and brightness adjustment. The fine-grained morphological features include the fish's outline, body shape, aspect ratio, and local scattering intensity distribution.

5. The method for monitoring fish passage in water conservancy and hydropower projects by coupling AI recognition and sonar as described in claim 1, characterized in that, The construction and training method of the deep reinforcement learning policy network in step S4 is as follows: using fine-grained morphological feature vectors as state input and fish species and growth stage classification labels as action space, a deep reinforcement learning policy network with an Actor-Critic architecture is constructed. The network training is completed through interaction with the real-time monitoring data environment and a reward feedback mechanism to achieve adaptive classification of fish species and growth stages.

6. A fish passage monitoring system for hydraulic and hydropower projects that couples AI recognition and sonar, used to implement the fish passage monitoring method for hydraulic and hydropower projects that couples AI recognition and sonar as described in any one of claims 1-5, characterized in that, Includes the following modules: The sonar intelligent imaging module is deployed at the fishway monitoring section. Based on the intelligent sensing system of the multi-element sonar array and the group time-division transmission strategy, it generates a high-density parallel beam and outputs an acoustic image of the fishway section after processing the echo signal. The target separation and calculation module receives the acoustic image of the fishway cross section, identifies the fish target, calculates and separates the mixed echo signal, and outputs individual fish acoustic morphology data. The morphological depth recognition module incorporates the YOLOv7 target detection algorithm and the ResNet50 deep residual network to perform target detection and feature extraction on the acoustic morphological data of individual fish, and outputs fine-grained morphological feature vectors. The adaptive classification decision module constructs a deep reinforcement learning strategy network and a knowledge base of feature and behavior association rules. It receives the fine-grained morphological feature vector and outputs the evaluation results of fish species and growth stage, as well as the analysis conclusions of traffic behavior. The fish monitoring data extraction module integrates the above assessment results, behavioral analysis conclusions, and corresponding spatiotemporal information, calculates core statistical indicators, and automatically generates multi-dimensional fish monitoring reports.

7. The fish passage monitoring system for hydraulic and hydropower projects that couples AI recognition and sonar as described in claim 6, characterized in that, The target separation and calculation module includes an intelligent trajectory simulation unit and an overlapping signal decoupling unit; The intelligent trajectory simulation unit runs an improved ant colony algorithm to segment the echo energy adhesion region and outputs the motion contour and trajectory identifier of the suspected independent individual. The overlapping signal decoupling unit constructs a theoretical echo generation model for the motion contour, uses non-negative matrix decomposition to solve the mixed echo signal, and outputs individual fish acoustic morphology data.

8. The fish passage monitoring system for hydraulic and hydropower projects that couples AI recognition and sonar as described in claim 6, characterized in that, The morphological depth recognition module includes a target detection unit and a fine-grained feature extraction unit; The target detection unit preprocesses and detects the acoustic morphology data of individual fish bodies, and outputs the precise contour bounding box coordinates of the fish body target. The fine-grained feature extraction unit extracts a local image of the fish body based on the coordinates, extracts and outputs a fine-grained morphological feature vector through a ResNet50 deep residual network.

9. The fish passage monitoring system for hydraulic and hydropower projects that couples AI recognition and sonar according to claim 6, characterized in that, The adaptive classification decision module can update the feature and behavior association rule knowledge base based on historical monitoring data, and the deep reinforcement learning strategy network can adapt to changes in fish morphology and behavioral characteristics in different aquatic environments through incremental learning, thereby achieving autonomous optimization of the classification model.

10. The fish passage monitoring system for hydraulic and hydropower projects that couples AI recognition and sonar according to claim 6, characterized in that, The fish monitoring data extraction module can calculate core statistical indicators in a rolling manner according to a preset time window, including the total number of fish passing through, the number and proportion of various fish species, the size distribution of each growth stage, and the passage frequency time series. It can also format the core statistical indicators, spatial distribution heatmaps, and behavioral pattern analysis conclusions according to industry standard templates to generate standardized fish monitoring reports.

Images