Large-scale fish resource detection method based on multi-frequency sound wave and AI algorithm

Abstract

The invention discloses a large-scale fish resource detection method based on multi-frequency sound waves and an AI algorithm. The method comprises the following steps: correcting an original echo signal into a signal under a standard environment characteristic in combination with the environment characteristic; extracting and normalizing frequency spectrum characteristics and time sequence distribution information of the echo signals; obtaining frequency response characteristic data of fish types and associating spatial position information of the fish types; obtaining a fish swarm proportion through three-dimensional point cloud clustering analysis; identifying a fish school abnormal behavior type and an area range, and formulating a fish school abnormal intervention strategy; predicting a time point, an area range, an abnormality type and a reason of the abnormal behavior of the fish school in a future time period; and in combination with Mahalanobis distance and fish school behavior prediction, evaluating the effect of the fish school abnormal intervention strategy, and dynamically adjusting the fish school abnormal intervention strategy. According to the invention, the precision and efficiency of fish resource detection can be improved, the requirements of dynamic monitoring and precise management of fish schools in a large-range water area are met, and the healthy and sustainable development of a water area ecological system is further promoted.

Description

Technical Field

[0001] The present invention relates to the field of aquaculture monitoring, and in particular to a large-scale fish resource detection method using multi-frequency sound waves and AI algorithms. Background Art

[0002] With the changes in the ecological environment of global waters and the increasing scarcity of fishery resources, scientifically and efficiently monitoring the dynamics of fish resources has become an important task for ecological protection and fishery management. However, the current fish resource detection and monitoring technology still faces many problems that need to be solved in practical applications. Traditional monitoring technology has limited detection capabilities for large areas of water, and it is difficult to meet the needs of modern fishery management for real-time monitoring of large-scale fish populations. Secondly, as a non-contact means of fish resource detection, although multi-frequency sonar technology can use sound waves to obtain fish echo signals in waters, there are still many problems in practical applications. The propagation of sound wave signals will be significantly affected by the complex characteristics of the water environment, such as temperature, salinity, flow rate, turbidity, etc., resulting in noise, distortion, and even partial signal loss in the echo signal. The interference of these environmental variables makes the quality of the raw data collected by multi-frequency sonar unstable, which directly affects the accuracy of subsequent fish identification. In addition, the behavior of fish in natural waters is highly dynamic and complex, including group aggregation, dispersion, and changes in movement direction. These behavioral changes are usually driven by multiple factors, such as predator threats, environmental disturbances, or human activities. However, traditional fish monitoring methods have limited ability to capture these dynamic behavior characteristics, especially the lack of real-time and refined analysis capabilities, making it difficult to achieve timely monitoring and accurate identification of fish behavior. In addition, the complexity of the spatiotemporal distribution of fish behavior data makes it more difficult to predict its future dynamic changes. At present, there is a lack of accurate prediction methods for the future behavior change trends of fish schools, which leads to fishery management often being in a lagging and passive state when dealing with abnormal behavior. In addition, there are also great technical challenges for the identification and intervention of abnormal behavior. At present, the monitoring of abnormal fish behavior mostly relies on manual experience or the judgment of a single indicator, lacking a scientific behavioral baseline model as a reference, which is prone to omissions or misjudgments. At the same time, in terms of monitoring the expansion trend of abnormal behavior and adjusting the scope of intervention, the current means also have great deficiencies, especially for the situation where abnormal behavior may spread to unintervention areas, and existing technologies are difficult to respond in a timely and effective manner. Therefore, the above problems seriously restrict the improvement of fish resource management efficiency and the scientific protection of aquatic ecosystems, and more accurate and efficient technical means are urgently needed to solve them. Summary of the invention

[0003] The present invention solves the problems existing in the above-mentioned prior art and provides a large-scale fish resource detection method using multi-frequency sound waves and AI algorithms, which mainly includes:

[0004] Adjust the sound wave emission parameters through multi-frequency sonar equipment, and correct the original echo signal to the signal under standard environmental characteristics based on environmental characteristics;

[0005] According to the corrected echo signal, the spectrum characteristics and time series distribution information of the echo signal are extracted and normalized through time window segmentation, signal arrival time calculation and denoising processing;

[0006] Through time series analysis of multi-frequency signals, frequency response gradient calculation and envelope curve extraction, the frequency response characteristic data of fish species are generated and associated with their spatial position information;

[0007] Based on the frequency response characteristic data of multi-frequency sound waves in the fish sound wave monitoring database, the spatial distribution, population ratio and mixed fish marking information of fish schools are obtained through signal classification and three-dimensional point cloud clustering analysis;

[0008] Based on the fish school echo signals and dynamic data obtained in real time by the multi-frequency sonar receiver, the parameter range of the normal behavior baseline of the fish school is determined. By calculating the Mahalanobis distance between the fish school dynamic data and the mean of the normal behavior baseline of the fish school, the type and area of ​​abnormal behavior of the fish school are identified, and an intervention strategy for abnormal fish school behavior is formulated;

[0009] Based on the historical dynamic data of the fish school behavior pattern database and the baseline average of normal fish school behavior, the time point, area range, abnormal type and cause of abnormal fish school behavior in the future time period are predicted, and the fish school abnormality intervention strategy is implemented in advance;

[0010] Based on the fish school dynamics data obtained by the multi-frequency sonar receiver, combined with the Mahalanobis distance and fish school behavior prediction, the effectiveness of the fish school abnormality intervention strategy is evaluated, and the fish school abnormality intervention strategy and intervention range are dynamically adjusted.

[0011] Furthermore, the method of adjusting the sound wave emission parameters by the multi-frequency sonar device and correcting the original echo signal to a signal under standard environmental characteristics in combination with environmental characteristics includes:

[0012] According to the preset detection waters and the possible distribution range of target fish species, a multi-frequency sonar device is used to configure multiple frequency ranges, and the order of low-frequency, medium-frequency and high-frequency transmission signals is set; through the sound wave transmission system of the multi-frequency sonar device, the sound wave transmission power, beam width and transmission angle are adjusted to transmit the sound wave signals to the designated waters one by one, and the original echo signals of the multi-frequency sound waves are obtained through the multi-frequency sonar receiver, and the spatial position information of the sound wave transmission is marked and stored in the fish sound wave monitoring database; the echo signals under standard environmental characteristics of the same distance, azimuth and depth and the echo signals under non-standard environmental characteristics are obtained through the fish sound wave monitoring database, and the recurrent neural network is used for model training to construct an echo signal correction model, and the environmental characteristics include water temperature, salinity, flow rate and turbidity; according to the original echo signals and environmental characteristics obtained in real time, the echo signal correction model is used to correct the original echo signals to echo signals under standard environmental characteristics.

[0013] Furthermore, the spectral characteristics and time series distribution information of the echo signal are extracted and normalized through time window segmentation, signal arrival time calculation and denoising processing according to the corrected echo signal, including:

[0014] According to the corrected echo signal, the echo signal is separated based on time window segmentation and signal arrival time calculation, and the echo intensity, scattering pattern and arrival time of sound waves of different frequencies are recorded; according to the time delay of the received signal, the sound wave propagation path in the water area is calculated, and the distance, azimuth, depth and position coordinates corresponding to the echo signal are recorded; the acquired echo signal is denoised by a bandpass filter to remove environmental noise and equipment interference in the non-target frequency range; the time series signal is converted into a time-frequency domain signal by a short-time Fourier transform algorithm, and the spectral feature data of the echo is extracted; the echo signal intensity data is amplitude normalized by the normalization method to unify the power amplitude of signals of different frequencies, and the frequency distribution and time change pattern are recorded to obtain the normalized spectral feature data and its time series distribution information.

[0015] Furthermore, the method of generating frequency response characteristic data of fish species and associating their spatial position information through time series analysis of multi-frequency signals, frequency response gradient calculation and envelope curve extraction includes:

[0016] According to the normalized spectrum feature data, including the time series of multi-frequency signals and their spectrum features, the amplitude and phase information of signals in different frequency bands are extracted using fast Fourier transform, and the time series data at each frequency is recorded; according to the time series signals of each frequency, the formula I = 20·log 10(|S|) Convert the time series signal into intensity value, and calculate the mean and variance of the echo signal intensity in each frequency range to obtain the scattering intensity characteristics corresponding to different frequencies, where S is the amplitude of the normalized signal; according to the scattering intensity characteristic data corresponding to different frequencies, perform fitting analysis on the scattering intensity data of different frequencies, establish a curve of scattering intensity changing with frequency, and calculate the gradient value of the curve Extract the intensity change rate between different frequencies to determine the sensitivity of fish to frequency; group the features according to the range of frequency change gradient to form a characteristic template of fish frequency response, and obtain the frequency sensitivity characteristics of fish species, including frequency response change gradient and characteristic template; extract the frequency response data of various fish at each frequency, calculate the characteristic separation index of different fish at each frequency band, and perform frequency band scaling on the frequency band signal intensity in the overlapping area of ​​signal characteristics; extract the envelope curve of the signal using Hilbert transform according to the time series signal of each frequency to obtain the shape characteristics of the reflected signal; according to the descending part of the echo signal envelope, use the exponential fitting model S(t)=S 0 ·e -R·t Calculate the signal decay rate R to obtain the shape characteristics and decay rate characteristics of the fish echo signal, where S(t) is the envelope amplitude, S 0 is the starting value of the envelope curve; the scattering intensity characteristics of sound waves of each frequency, the frequency sensitivity characteristics of fish species, the shape characteristics and attenuation rate characteristics of fish echo signals are combined according to the frequency range to generate a frequency response characteristic data table of fish species, and the frequency response characteristic data of fish species are associated with the distance, azimuth, depth and position coordinates of the waters where the fish are located.

[0017] It also includes extracting the frequency response data of various fish species at various frequencies, calculating the characteristic separation index of different fish species at various frequency bands, and performing frequency band scaling on the frequency band signal strength in the signal feature overlapping area, specifically including:

[0018] Extract various fish species at various frequencies from the echo signal, according to the formula Calculate its frequency response gradient G(f), record the G(f) value of each fish in the full frequency band and establish a frequency response gradient database, where S(f) represents the normalized signal intensity at frequency f, and Δf represents the frequency increment; according to the frequency response gradient of each fish, use the characteristic separation index formula Calculate the characteristic separation index I(f) of each fish in each frequency band, where G 1 (f) and G 2 (f) is the frequency response gradient of the two fish at frequency f, σ 1 (f) and σ 2(f) are the standard deviations of the signal strength of the two fish species at frequency f; based on the preset separation index threshold, the frequency region where the characteristic separation index I(f) is less than the preset separation index threshold is marked as the signal feature overlapping region; through the frequency scaling formula f′=ξ·(ff 0 )+f 0 , frequency band scaling is performed on the frequency range of the signal feature overlap area, where f′ is the scaled frequency, ξ is the scaling factor, which is used to amplify or compress the frequency interval and is obtained by fitting historical data, f 0 is the center point of the frequency, maintaining the consistency of the frequency band translation; for the scaled signal, the frequency response gradient G′(f) and the separation index I′(f) are calculated again. If I′(f) of all frequency bands is greater than or equal to the preset separation index threshold, there is no signal feature overlap problem. Otherwise, the scaling factor is adjusted until the separation index of all frequency bands reaches or exceeds the threshold.

[0019] Furthermore, the spatial distribution, population ratio and mixed fish marking information of fish schools are obtained by signal classification and three-dimensional point cloud cluster analysis based on the frequency response characteristic data of multi-frequency sound waves in the fish sound wave monitoring database, including:

[0020] The frequency response characteristic data of different fish species under multi-frequency sound waves are obtained through the fish sound wave monitoring database, and the recurrent neural network is used for model training to build a fish recognition model; based on the frequency response characteristic data of fish species obtained in real time, the fish recognition model is used to determine the fish label and classification confidence of each echo signal; if the classification confidence is lower than the preset confidence threshold, it is marked as an invalid signal and removed; the classification results of all echo signals are statistically analyzed to obtain the overall distribution of fish species in the water area, determine the proportion and number of different fish populations, and map the signal points based on the distance, azimuth, depth and position coordinates corresponding to the echo signal. The data are projected into a three-dimensional spatial coordinate system to form point cloud data; based on the point cloud data, the K-means clustering algorithm or the DBSCAN clustering algorithm is used for model training, and the signal points belonging to the same fish species are input into the clustering algorithm to obtain the spatial boundaries of each fish school, the size of the fish school in each clustering area, the center position of the fish school and the boundary range, where the size of the fish school is the number of signal points in the area; the clustering results of different fish in the space are overlapped to determine whether there are two or more types of fish overlapping in the same space. If an overlapping area is detected, it is marked as a mixed fish school; if a mixed fish school exists, the area is marked separately and the fish species and quantity statistics are output.

[0021] Furthermore, the method determines the parameter range of the baseline of normal behavior of the fish school based on the fish school echo signal and dynamic data obtained in real time by the multi-frequency sonar receiver, identifies the type and area range of abnormal behavior of the fish school by calculating the Mahalanobis distance between the dynamic data of the fish school and the mean of the baseline of normal behavior of the fish school, and formulates an abnormal fish school intervention strategy, including:

[0022] The multi-frequency sonar receiver receives the fish school echo signal in real time, obtains the timestamp, distance, azimuth, depth and position coordinates of each signal point, and determines the fish school dynamic data, including the fish school's position coordinates, group density, movement direction and movement rate; according to the fish school dynamic data, the fish school behavior characteristics are determined, and the fish school behavior characteristics include the fish school density change rate, spatial aggregation characteristics, movement direction consistency and speed change rate. The spatial aggregation characteristics are the degree of aggregation of the fish school in a specific area, and the movement direction consistency is the similarity of the movement direction of individuals in the group; according to the fish school behavior characteristics, the K-means clustering algorithm is used for clustering analysis, and a baseline model of normal fish school behavior is established to determine the parameter range of the normal fish school behavior baseline, including the normal group density change range, aggregation radius and spatial distribution characteristics, average movement direction and speed range; the Mahalanobis distance calculation formula is used Calculate the mean Mahalanobis distance D between the fish school dynamic data and the baseline of the fish school's normal behavior M , determine the degree of deviation between the real-time behavior of the fish school and the normal behavior baseline, where ∑ is the covariance matrix; if D M If the confidence interval of the normal behavior of the fish school is exceeded, the current fish school behavior is considered abnormal. M and preset abnormality assessment thresholds to determine the degree of abnormality of the fish school, and mark the type and degree of abnormal behavior of the fish school. The types of abnormal behavior include but are not limited to aggregation, dispersion or high-speed movement, and the degrees of abnormality include mild, moderate and severe. According to the location coordinates of the fish school, determine the location coordinates and regional scope of the abnormal behavior of the fish school, and count the group characteristics of the area where the abnormal behavior of the fish school occurs, including the number and density of the fish school; store the fish school behavior characteristics of the area where the abnormal behavior of the fish school occurs in the fish school behavior pattern database; obtain the fish school behavior characteristics of the historical area where the abnormal behavior of the fish school occurs through the fish school behavior pattern database, and mark the abnormal reasons, use the random forest algorithm to train the model, and build a fish school abnormal cause identification model. The abnormal reasons include but are not limited to predator threats, environmental disturbances or human intervention; according to the real-time acquired fish school behavior characteristics, use the fish school abnormal cause identification model to determine the cause of the fish school abnormality; based on the time of occurrence, regional scope, cause of the fish school abnormality and the degree of abnormality of the fish school abnormal behavior, formulate and implement fish school abnormality intervention strategies, including but not limited to adjusting water quality parameters, reducing pollution load, increasing dissolved oxygen concentration or optimizing water flow rate.

[0023] Furthermore, the historical dynamic data of the fish school behavior pattern database is combined with the baseline mean of the normal behavior of the fish school to predict the time point, area range, abnormal type and cause of the abnormal behavior of the fish school in the future time period, and implement the fish school abnormality intervention strategy in advance, including:

[0024] Through the fish school behavior pattern database, historical fish school dynamic data is obtained, and the model is trained using the long short-term memory network to build a fish school behavior prediction model to predict the fish school dynamic data within a preset time period; based on the predicted fish school dynamic data within the preset time period, the Mahalanobis distance between the fish school dynamic data and the mean baseline of the fish school's normal behavior is calculated to determine the time point, regional range, abnormal type and abnormal cause of the abnormal behavior of the fish school within the future preset time period; according to the predicted time point when the fish school will have abnormal behavior, the fish school abnormality intervention strategy is implemented in advance to intervene.

[0025] Furthermore, the fish school dynamic data obtained by the multi-frequency sonar receiver is combined with the Mahalanobis distance and the fish school behavior prediction to evaluate the effect of the fish school abnormality intervention strategy, and dynamically adjust the fish school abnormality intervention strategy and intervention range, including:

[0026] Through the multi-frequency sonar receiver, the fish school dynamic data after the implementation of the fish school abnormal intervention strategy is obtained, and the Mahalanobis distance between the fish school dynamic data and the fish school normal behavior baseline is calculated. If the Mahalanobis distance is lower than the preset fish school normal behavior confidence interval, the fish school abnormal intervention strategy is judged to be effective, and the fish school abnormal behavior area returns to normal; if the Mahalanobis distance still exceeds the preset fish school normal behavior confidence interval, the fish school abnormal intervention strategy is judged to be ineffective, and based on the fish school dynamic data, it is judged whether the abnormal area has expanded to the non-intervention area, and the fish school abnormal intervention strategy and the area where the strategy is implemented are adjusted until the fish school abnormal behavior area returns to normal; obtain the fish school abnormal intervention strategy implemented in advance The fish school dynamics data of the area are re-predicted based on the fish school behavior prediction model, and the fish school dynamics data of the fish school within the preset time period are re-predicted, and the Mahalanobis distance between the predicted fish school dynamics data of the fish school within the preset time period and the baseline mean of the normal behavior of the fish school is determined; based on the Mahalanobis distance, it is determined whether the early implementation of the fish school abnormality intervention strategy can eliminate the abnormal risk of the abnormal behavior in the future; if the Mahalanobis distance still exceeds the preset fish school normal behavior confidence interval, it is determined that the early implementation of the fish school abnormality intervention strategy is ineffective, and based on the predicted fish school dynamics data of the fish school within the preset time period, it is determined whether the abnormal area expands to the non-intervention area, and the fish school abnormality intervention strategy and the regional scope of the strategy implementation are adjusted.

[0027] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0028] The present invention provides a large-scale fish resource detection method using multi-frequency sound waves and AI algorithms. The present invention combines multi-frequency sonar equipment with environmental feature correction technology to achieve high-precision correction of fish echo signals, effectively eliminate the influence of environmental variables on sound wave propagation, ensure the quality and stability of echo signals, and comprehensively obtain frequency response feature data of fish species through normalized spectrum feature extraction, multi-frequency signal time series analysis and frequency response characteristic generation, and associate it with spatial position information, so as to accurately identify fish species and distribution characteristics. The present invention accurately identifies the spatial distribution, population proportion and mixed fish characteristics of fish schools through three-dimensional point cloud cluster analysis, and provides high-precision data support for dynamic behavior research and ecological monitoring of fish schools. At the same time, the present invention can quickly identify abnormal behavior of fish schools and their regional scope through real-time monitoring and dynamic data analysis, combined with the baseline model of normal behavior of fish schools, and provide a scientific basis for timely intervention of abnormal behavior. In addition, the present invention uses historical dynamic data combined with a prediction model to accurately predict abnormal behavior of fish schools and their causes in future time periods, thereby realizing early intervention and dynamic management of abnormal behavior of fish schools. By dynamically adjusting the intervention strategy and regional scope, abnormal behavior can be effectively eliminated and prevented from spreading. The present invention can not only improve the accuracy and efficiency of fish resource detection and meet the needs of dynamic monitoring and precise management of fish populations in large waters, but also quickly identify the source of the problem and conduct effective intervention in the event of environmental disturbance or ecological anomaly, providing a reliable technical basis and scientific support for fishery resource protection, aquatic ecosystem restoration and environmental management, and further promoting the health and sustainable development of aquatic ecosystems. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flow chart of a large-scale fish resource detection method using multi-frequency sound waves and AI algorithms of the present invention;

[0030] Figure 2 A schematic diagram of a large-scale fish resource detection method using multi-frequency sound waves and AI algorithms of the present invention;

[0031] Figure 3 This is another schematic diagram of a large-scale fish resource detection method using multi-frequency sound waves and AI algorithms according to the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] like Figure 1-3 In this embodiment, a large-scale fish resource detection method using multi-frequency sound waves and AI algorithms may specifically include:

[0034] Step S101, adjusting the sound wave emission parameters through the multi-frequency sonar device, and correcting the original echo signal to a signal under standard environmental characteristics in combination with environmental characteristics.

[0035] According to the preset detection waters and the possible distribution range of target fish species, multi-frequency sonar equipment is used to configure multiple frequency ranges and set the order of low-frequency, medium-frequency and high-frequency transmission signals. Through the sound wave transmission system of the multi-frequency sonar equipment, the sound wave transmission power, beam width and transmission angle are adjusted to transmit the sound wave signals to the designated waters one by one. The original echo signal of the multi-frequency sound wave is obtained through the multi-frequency sonar receiver, and the spatial position information of the sound wave transmission is marked and stored in the fish sound wave monitoring database. The echo signal under the standard environmental characteristics of the same distance, azimuth and depth and the echo signal under non-standard environmental characteristics are obtained through the fish sound wave monitoring database. The recurrent neural network is used for model training to construct an echo signal correction model. The environmental characteristics include water temperature, salinity, flow rate and turbidity. According to the original echo signal obtained in real time and the environmental characteristics, the echo signal correction model is used to correct the original echo signal to the echo signal under the standard environmental characteristics.

[0036] For example, in an aquaculture farm with an area of ​​1000 square meters and a water depth of 10 meters, a variety of fish species including carp and perch are cultivated. A multi-frequency sonar device is configured with a low-frequency signal of 30-50kHz, an intermediate-frequency signal of 50-100kHz, and a high-frequency signal of 100-300kHz, and the signal is transmitted in the order of low frequency to high frequency. The sound wave transmission power is adjusted to 150W by the sound wave transmission system, the beam width is 30°, and the transmission angle is set to 45° below the horizontal plane, and the sound wave signal is transmitted to different areas of the designated waters one by one. The echo signal is received by a multi-frequency sonar receiver and the spatial position information of the sound wave transmission is recorded. The spatial coordinates of the transmission point are (0, 0, 0), and the spatial position of the target fish is (5m, 3m, -3m), where x is the horizontal distance, y is the longitudinal distance, and z is the depth. The received echo signal strength is -40dB, and the echo delay time is 0.02s, and these data are stored in the fish sound wave monitoring database. The echo signal of the same target distance (5m, 3m, -3m) under standard environmental characteristics (temperature 20℃, salinity 30‰, flow velocity 0.1m / s, turbidity 1NTU) is obtained from the database with an intensity of -35dB and a delay time of 0.018s, as well as non-standard environmental characteristics, temperature 25℃, salinity 35‰, flow velocity 0.3m / s, turbidity 5NTU, and an intensity of -50dB and a delay time of 0.025s, and a recurrent neural network model is used for training to build an echo signal correction model. According to the real-time acquired original echo signal with an intensity of -50dB and a delay time of 0.025s and the current environmental characteristics, temperature 25℃, salinity 35‰, flow velocity 0.3m / s, turbidity 5NTU, the echo signal correction model is input. After model correction, the original echo signal is adjusted to the correction signal under standard environmental characteristics, with an intensity of -40dB and a delay time of 0.02s.

[0037] Step S102 , based on the corrected echo signal, the spectrum characteristics and time series distribution information of the echo signal are extracted and normalized through time window segmentation, signal arrival time calculation and denoising processing.

[0038] According to the corrected echo signal, the echo signal is separated based on time window segmentation and signal arrival time calculation, and the echo intensity, scattering pattern and arrival time of sound waves of different frequencies are recorded. According to the time delay of the received signal, the sound wave propagation path in the water area is calculated, and the distance, azimuth, depth and position coordinates corresponding to the echo signal are recorded. The acquired echo signal is denoised by a bandpass filter to remove environmental noise and equipment interference in the non-target frequency range. The time series signal is converted into a time-frequency domain signal by the short-time Fourier transform algorithm, and the spectral feature data of the echo is extracted. The echo signal intensity data is amplitude normalized by the normalization method to unify the power amplitude of signals of different frequencies, and the frequency distribution and time change pattern are recorded to obtain the normalized spectral feature data and its time series distribution information.

[0039] For example, if there is a corrected echo signal, the intensity of the low frequency 30kHz is -40dB, the intensity of the medium frequency 80kHz is -45dB, the intensity of the high frequency 200kHz is -50dB, and the echo delay times are 0.02s, 0.018s and 0.015s respectively. The echo signals are separated based on time window segmentation and signal arrival time calculation, and the sound wave echo intensities, scattering patterns and arrival times of different frequencies are recorded. According to the delay of the echo signal, the delay of the low-frequency echo, medium-frequency echo and high-frequency echo are 0.02s, 0.018s and 0.015s respectively. The propagation path corresponding to the low-frequency echo is calculated to be 30m, the propagation path corresponding to the medium-frequency echo is 27m, and the propagation path corresponding to the high-frequency echo is 22.5m. Combined with the emission angle and the direction of the sound beam, the spatial coordinates corresponding to the recorded echo signal are (25m, 10m, -3m), (20m, 12m, -4m) and (18m, 15m, -5m). The echo signals of 30kHz, 80kHz and 200kHz are denoised by bandpass filters to remove environmental noise in the non-target frequency range, such as water flow sound, ship noise and equipment interference. The short-time Fourier transform algorithm is used to convert the time series signal into a time-frequency domain signal, and the spectrum feature data of each frequency is extracted. The spectrum peak of the 30kHz signal is around 31kHz, the spectrum peak of the 80kHz signal is around 79kHz, and the spectrum peak of the 200kHz signal is around 201kHz. The normalization algorithm is used to normalize the amplitude of the echo signal intensity, and the intensity of the low-frequency, medium-frequency and high-frequency signals is normalized to the amplitude range of 0.8, 0.7 and 0.6 respectively, unifying the power amplitude of signals of different frequencies, and recording the frequency distribution time variation pattern, and the intensity of 30-50kHz is concentrated above 0.75, and the signal intensity reaches the peak within 0.015-0.025s.

[0040] Step S103, generating frequency response characteristic data of fish species and associating their spatial position information through time series analysis of multi-frequency signals, frequency response gradient calculation and envelope curve extraction.

[0041] According to the normalized spectrum feature data, including the time series of multi-frequency signals and their spectrum features, the amplitude and phase information of signals in different frequency bands are extracted using fast Fourier transform, and the time series data at each frequency is recorded. According to the time series signals of each frequency, the formula I = 20·log 10 (|S|) Convert the time series signal into intensity value, and calculate the mean and variance of the echo signal intensity in each frequency range to obtain the scattering intensity characteristics corresponding to different frequencies, where S is the amplitude of the normalized signal. According to the scattering intensity characteristic data corresponding to different frequencies, perform fitting analysis on the scattering intensity data of different frequencies, establish a curve of scattering intensity changing with frequency, and calculate the gradient value of the curve Extract the intensity change rate between different frequencies to determine the sensitivity of fish to frequency. According to the range of frequency change gradient, group the features to form a characteristic template of fish frequency response, and obtain the frequency sensitivity characteristics of fish species, including frequency response change gradient and characteristic template. Extract the frequency response data of various fish at each frequency, calculate the characteristic separation index of different fish at each frequency band, and perform frequency band scaling on the frequency band signal intensity in the overlapping area of ​​signal characteristics. According to the time series signal of each frequency, use Hilbert transform to extract the envelope curve of the signal to obtain the shape characteristics of the reflected signal. According to the descending part of the echo signal envelope, use the exponential fitting model S(t)=S 0 ·e -R·t Calculate the signal decay rate R to obtain the shape characteristics and decay rate characteristics of the fish echo signal, where S(t) is the envelope amplitude, S 0 is the starting value of the envelope curve. The scattering intensity characteristics of sound waves of each frequency, the frequency sensitivity characteristics of fish species, the shape characteristics and attenuation rate characteristics of fish echo signals are combined according to the frequency range to generate a frequency response characteristic data table of fish species, and the frequency response characteristic data of fish species are associated with the distance, azimuth, depth and position coordinates of the waters where the fish are located.

[0042] For example, according to the normalized spectrum feature data, including the time series of the multi-frequency signal and its spectrum features, in the three frequency bands of 30kHz, 80kHz and 200kHz, the amplitude of the signal is 0.8, 0.6 and 0.4 respectively. The signal amplitude and phase information of the three frequency bands are extracted by fast Fourier transform, and the time series signal data is recorded. For example, the signal amplitude in the 30kHz frequency band changes with time with a starting value of 0.8, drops to 0.6 after 0.01s, and drops to 0.4 at 0.02s. According to these time series signals, the formula I = 20·log 10 (|S|) Convert the amplitude to intensity value, and the intensity values ​​of the 30kHz signal are -1.94dB, corresponding to an amplitude of 0.8, -4.44dB, corresponding to an amplitude of 0.6, and -7.96dB, corresponding to an amplitude of 0.4. The average intensity of the frequency band is calculated to be -4.78dB, with a variance of 2.55dB2, where S is the amplitude of the normalized signal. By fitting and analyzing the scattering intensity characteristic data of the three frequency bands of 30kHz, 80kHz and 200kHz, a curve of scattering intensity changing with frequency is established. The average intensities corresponding to 30kHz, 80kHz and 200kHz are -4.78dB, -6.02dB and -8.12dB, respectively. The fitting curve shows that the intensity decreases linearly with the increase of frequency, and the gradient value of the curve is calculated. The sensitivity of fish to frequency was judged by the gradient value. It was found that fish responded more strongly to the low frequency band of 30-80kHz, while the response to the high frequency band of 80-200kHz gradually weakened. According to the range of frequency gradient, the frequency gradient was divided into a low frequency gradient of -1.68 to -1.00, a medium frequency gradient of -1.00 to -0.50, and a high frequency gradient less than -0.50, forming a characteristic template of fish frequency response, and generating a fish template containing gradient range and frequency response characteristics. According to the time series signal of each frequency band, the envelope curve of the signal was extracted using Hilbert transform. In the 30kHz frequency band, the starting amplitude S of the envelope curve was 0.0447 W, which was 0.033 W. 0 is 0.8, and then the signal decays with time, using the exponential fitting model S(t) = S 0 ·e -R·t Calculate the signal decay rate R, and get decay rate R = 50s -1 , where S(t) is the envelope amplitude, S 0The scattering intensity characteristics of sound waves of different frequencies, the frequency sensitivity characteristics of fish, such as the strong response to 30kHz and the weak response to 200kHz, and the shape characteristics of fish echo signals, such as the envelope curve and the attenuation rate, are combined to generate a frequency response characteristic data table of fish species. For example, if the target fish is marked as species A, its frequency response data table records the average intensity of 30kHz as -4.78dB, the frequency change gradient as -1.68dB / kHz, and the envelope attenuation rate as 50s. -1 The frequency response characteristics of the fish species are associated with the spatial information of the waters where it is located, such as distance 10m, azimuth 45°, depth 5m, and position coordinates (7.1m, 7.1m, -5m).

[0043] Among them, the frequency response data of various fish at each frequency are extracted, the characteristic separation index of different fish at each frequency band is calculated, and the frequency band signal strength in the signal feature overlapping area is scaled.

[0044] Extract various fish species at various frequencies from the echo signal, according to the formula Calculate the frequency response gradient G(f), record the G(f) value of each fish in the full frequency band and establish a frequency response gradient database, where S(f) represents the normalized signal intensity at frequency f, and Δf represents the frequency increment. According to the frequency response gradient of each fish, use the characteristic separation index formula Calculate the characteristic separation index I(f) of each fish in each frequency band, where G 1 (f) and G 2 (f) is the frequency response gradient of the two fish at frequency f, σ 1 (f) and σ 2 (f) are the standard deviations of the signal strength of the two fish species at frequency f. Based on the preset separation index threshold, the frequency region where the feature separation index I(f) is less than the preset separation index threshold is marked as the signal feature overlap region. 0 )+f 0 , frequency band scaling is performed on the frequency range of the signal feature overlap area, where f′ is the scaled frequency, ξ is the scaling factor, which is used to amplify or compress the frequency interval and is obtained by fitting historical data, f 0 It is the center point of the frequency, maintaining the consistency of the frequency band translation. For the scaled signal, the frequency response gradient G′(f) and the separation index I′(f) are calculated again. If the I′(f) of all frequency bands is greater than or equal to the preset separation index threshold, there is no signal feature overlap problem. Otherwise, the scaling factor is adjusted until the separation index of all frequency bands reaches or exceeds the threshold.

[0045] Exemplarily, the frequency response characteristic data of two kinds of fish are extracted from the echo signal, which are respectively denoted as fish A and fish B. In the frequency range of f = 10kHz to f = 50kHz, the normalized signal strength S(f) is recorded with a frequency increment Δf of 2kHz. For example, when f = 10kHz and f = 12kHz, the normalized signal strength of fish A is S A (10) = 0.8 and S A (12)=0.6, then according to the formula Its frequency response gradient at 10kHz is -0.1kHz -1 The signal strengths of fish B in the same frequency range are 0.75 and 0.7 respectively, so the frequency response gradient of fish B is -0.025kHz -1 After recording the G(f) value of each fish species in the full frequency band, a frequency response gradient database was established. Using the characteristic separation index formula Calculate the characteristic separation index of the two fish in each frequency band. For example, at 10kHz, the frequency response gradients of fish A and B are G and G, respectively. A (10) = -0.1 and G B (10) = -0.025, and the standard deviations of the signal intensities are σ A (10) = 0.05 and σ B (10) = 0.03, the feature separation index is 0.9375. The preset feature separation index threshold is 1, then I(10) = 0.9375 < 1.0, indicating that there is a signal feature overlap problem in the 10kHz frequency region. For the frequency region marked as signal feature overlap, the frequency scaling formula f′ = ξ·(ff 0 )+f 0 The frequency band is scaled, where the scaling factor ξ is obtained by fitting historical data and is used to amplify the frequency interval. The frequency center point f 0 The frequency response gradient and feature separation index of the scaled signal are recalculated. For example, the gradient of fish A becomes G′A(9.8)=-0.12 after scaling, and the gradient of fish B becomes G′A(9.8)=-0.12 after scaling. B(9.8) = -0.02, the signal strength standard deviation remains unchanged, and the feature separation index is calculated to be 1.25, which is greater than the preset feature separation index threshold of 1 and meets the separation index threshold requirement. Therefore, it can be determined that the signal feature overlap problem has been solved. If the separation index still does not meet the threshold requirement, the optimization can be continued by adjusting the scaling factor ξ or the preprocessing parameters until the separation index of all frequency bands reaches or exceeds the threshold.

[0046] Step S104, based on the frequency response characteristic data under multi-frequency sound waves in the fish sound wave monitoring database, the spatial distribution, population ratio and mixed fish marking information of the fish school are obtained through signal classification and three-dimensional point cloud clustering analysis.

[0047] Through the fish acoustic wave monitoring database, the frequency response characteristic data of different fish species under multi-frequency sound waves are obtained, and the recurrent neural network is used for model training to build a fish recognition model. According to the frequency response characteristic data of fish species obtained in real time, the fish recognition model is used to determine the fish label and classification confidence of each echo signal. If the classification confidence is lower than the preset confidence threshold, it is marked as an invalid signal and eliminated. The classification results of all echo signals are statistically analyzed to obtain the overall distribution of fish species in the water area, determine the proportion and number of different fish populations, and map the signal points to the three-dimensional space coordinate system based on the distance, azimuth, depth and position coordinates corresponding to the echo signal to form point cloud data. According to the point cloud data, the K-means clustering algorithm or the DBSCAN clustering algorithm is used for model training, and the signal points belonging to the same fish species are input into the clustering algorithm to obtain the spatial boundaries of each fish school, the size of the fish school in each clustering area, the center position of the fish school and the boundary range, where the size of the fish school is the number of signal points in the area. The clustering results of different fish in the space are overlapped to determine whether two or more species of fish overlap in the same space. If an overlapping area is detected, it is marked as a mixed fish school. If a mixed fish school exists, the area is marked separately and the fish species and quantity statistics are output.

[0048] Exemplarily, the frequency response characteristic data of carp, perch and mandarin fish under multi-frequency sound waves of 30kHz, 80kHz and 200kHz are obtained through the fish sound wave monitoring database. The scattering intensity of carp at 30kHz is 0.8 and the variance is 0.1. The frequency sensitivity is manifested as a high low-frequency response intensity. The scattering intensity of perch at 80kHz is 0.75 and the variance is 0.08, and the medium-frequency response is strong. The scattering intensity of mandarin fish at 200kHz is 0.85 and the variance is 0.05, and the high-frequency response is most obvious. These frequency response characteristic data are trained by recurrent neural networks to construct a fish recognition model. When the echo signal of the multi-frequency sound wave is received in real time, each echo signal is classified using the fish recognition model. The classification results show that the fish labels corresponding to the echo signals are carp, perch and mandarin fish, respectively, and the classification confidences are 90%, 85% and 60%, respectively. Among them, the classification confidence of mandarin fish is lower than the preset threshold of 70%, so it is marked as an invalid signal and removed. The final classified echo signals are carp and perch. According to the classification results of all echo signals, the population proportion of carp in the water area is 60%, the number is 120, and the population proportion of perch is 40%, the number is 80. Combined with the spatial position information of the echo signal, the coordinates corresponding to the carp signal are (10m, 5m, -2m), and the coordinates corresponding to the perch signal are (20m, 8m, -3m). These coordinate points are mapped to the three-dimensional space coordinate system to form point cloud data. The K-means clustering algorithm was used to analyze the point cloud data, and the cluster center position of carp was (12m, 6m, -2.5m), the boundary range of the cluster area was (8-16m, 4-8m, -3--2m), and the cluster center position of perch was (22m, 10m, -3.5m), and the boundary range of the cluster area was (18-26m, 7-13m, -4--3m). The clustering results of different fish were overlapped and it was found that the cluster boundaries of carp and perch overlapped in some areas. The number of signal points in this area was 30, which was marked as a mixed fish area. The number of carp and perch in the mixed area was counted as 20, and the number of perch was 10. Finally, the mixed fish area was marked as a special area.

[0049] Step S105, based on the fish school echo signals and dynamic data acquired in real time by the multi-frequency sonar receiver, the parameter range of the normal behavior baseline of the fish school is determined, and by calculating the Mahalanobis distance between the fish school dynamic data and the mean of the normal behavior baseline of the fish school, the type and regional range of abnormal behavior of the fish school are identified, and an abnormal fish school intervention strategy is formulated.

[0050] The multi-frequency sonar receiver receives the fish school echo signal in real time, obtains the timestamp, distance, azimuth, depth and position coordinates of each signal point, and determines the fish school dynamic data, including the fish school's position coordinates, group density, movement direction and movement rate. According to the fish school dynamic data, the fish school behavior characteristics are determined. The fish school behavior characteristics include the fish school density change rate, spatial aggregation characteristics, movement direction consistency and speed change rate. The spatial aggregation characteristics are the degree of aggregation of the fish school in a specific area, and the movement direction consistency is the similarity of the movement direction of individuals in the group. According to the fish school behavior characteristics, the K-means clustering algorithm is used for clustering analysis to establish a baseline model for normal fish school behavior and determine the parameter range of the normal fish school behavior baseline, including the normal group density change range, aggregation radius and spatial distribution characteristics, average movement direction and speed range. The Mahalanobis distance calculation formula Calculate the mean Mahalanobis distance D between the fish school dynamic data and the baseline of the fish school's normal behavior M , determine the degree of deviation between the real-time behavior of the fish school and the normal behavior baseline, where ∑ is the covariance matrix. If D M If the confidence interval of the normal behavior of the fish school is exceeded, the current fish school behavior is considered abnormal. M The abnormal degree of the fish school is judged by the preset abnormal degree assessment threshold, and the abnormal behavior type and abnormal degree of the fish school are marked. The abnormal behavior type includes but is not limited to aggregation, dispersion or high-speed movement, and the abnormal degree includes mild, moderate and severe. According to the location coordinates of the fish school, the location coordinates and regional scope of the abnormal behavior of the fish school are determined, and the group characteristics of the area where the abnormal behavior of the fish school occurs are counted, including the number and density of the fish school. The fish school behavior characteristics of the area where the fish school has abnormal behavior are stored in the fish school behavior pattern database. The fish school behavior characteristics of the historical fish school abnormal behavior area are obtained through the fish school behavior pattern database, and the abnormal reasons are marked. The model is trained using the random forest algorithm to construct a fish school abnormal cause identification model. The abnormal reasons include but are not limited to predator threats, environmental disturbances or human intervention. According to the fish school behavior characteristics obtained in real time, the fish school abnormal cause identification model is used to determine the abnormal cause of the fish school. Based on the time of occurrence, regional scope, abnormal cause of the fish school and the degree of abnormality of the fish school abnormal behavior, formulate and implement fish school abnormal intervention strategies, including but not limited to adjusting water quality parameters, reducing pollution load, increasing dissolved oxygen concentration or optimizing water flow rate.

[0051] For example, the echo signal of the school of fish is received in real time by a multi-frequency sonar receiver, and the timestamp, distance, azimuth, depth and position coordinates of each signal point are obtained. For example, the echo signal of a school of fish shows that its position coordinates are (10m, 15m, -5m) and the timestamp is 10s. According to the echo data of multiple points, the school density of the school of fish is calculated to be every 10m 3There are 20 fish in the school, and the movement direction of the school is 45°, relative to the north direction, and the movement rate is 0.5m / s. According to the dynamic data of the school of fish, the behavioral characteristics of the school of fish are determined, including the rate of change of the school of fish density is 2 fish less per second, and the spatial aggregation characteristics show that the school of fish is concentrated in an area with a radius of 5m centered on the center point (10m, 15m, -5m), and the consistency of the movement direction is 85%, indicating that most individuals in the group have the same movement direction, and the speed change rate increases by 0.1m / s per second. According to the behavioral characteristics of the school of fish, the K-means clustering algorithm is used to perform cluster analysis on the school of fish, establish a baseline model of normal behavior of the school of fish, and determine the normal behavioral parameter range of the school of fish. For example, the normal group density change range is ±3 fish per second, the aggregation radius range is 3-6m, the movement direction consistency range is 80%-95%, and the speed range is 0.3-0.7m / s. The degree of deviation of the current behavior of the school of fish from the baseline mean is calculated by the Mahalanobis distance calculation formula. If the Mahalanobis distance D M =4.5, which exceeds the preset normal behavior confidence interval of 95%. The corresponding D M ≤3, the fish school behavior is judged to be abnormal. Based on the degree of deviation D M =4.5 and the preset abnormality assessment threshold, the abnormality is judged as a serious abnormality and marked as a high-speed movement behavior type. According to the location coordinates of the fish school, the area where the abnormal behavior occurs is determined to be a space with a radius of 5m centered on the coordinates (10m, 15m, -5m), and the fish school characteristics in the area are counted, including the number of fish is 100 and the density is 10m per 10m 3 Contains 25 fish. The fish behavior characteristics in the abnormal behavior area are stored in the fish behavior pattern database, and combined with the historical data in the database, the abnormal behavior records in similar areas are extracted and the abnormal causes are marked. For example, similar situations in historical data are often related to predator threats. After training the model with the random forest algorithm, a fish school abnormal cause identification model is constructed, and the real-time data is input into the fish school abnormal cause identification model to determine that the current abnormal cause is a predator threat. Combined with the occurrence time of the fish school abnormal behavior 10s, the center point (10m, 15m, -5m), the area with a radius of 5m, the abnormal cause is a predator threat, and the degree of abnormality is a serious abnormality, a fish school abnormality intervention strategy is formulated. The intervention measures include deploying high-frequency sound wave equipment to drive away predators, increasing the dissolved oxygen concentration in the water to relieve the pressure on the fish school, and continuously monitoring the environmental parameters of the area, including water temperature, dissolved oxygen, and pollutant concentration, to confirm that the abnormality has been resolved.

[0052] Step S106, based on the historical dynamic data of the fish school behavior pattern database and the baseline mean of the normal behavior of the fish school, predict the time point, area range, abnormal type and cause of the abnormal behavior of the fish school in the future time period, and implement the fish school abnormality intervention strategy in advance.

[0053] Through the fish school behavior pattern database, historical fish school dynamic data is obtained, and the model is trained using the long short-term memory network to build a fish school behavior prediction model to predict the fish school dynamic data within a preset time period. Based on the predicted fish school dynamic data within the preset time period, the Mahalanobis distance between the fish school dynamic data and the baseline mean of the normal behavior of the fish school is calculated to determine the time point, area range, abnormal type and abnormal cause of abnormal behavior of the fish school within the future preset time period. According to the predicted time point of abnormal behavior of the fish school, the fish school abnormality intervention strategy is implemented in advance to intervene.

[0054] Exemplarily, through the fish school behavior pattern database, the fish school dynamic data of the past week is obtained, including the spatial position, movement direction, density change, speed change and other characteristics of the fish school. The historical data records show that under the low-frequency 30kHz, medium-frequency 80kHz and high-frequency 200kHz sound waves, the echo signals of different fish schools show that their average movement speed is 0.4m / s, the density change rate is within the range of ±5 fish per second, and the consistency of movement direction is 80%-95%. Using these data, the long short-term memory network is used for model training to construct a fish school behavior prediction model. The model predicts the fish school dynamic data in the next 24 hours, and the prediction for the next 10th hour shows that the aggregation radius of the fish school at the position (15m, 30m, -4m) will be reduced to less than 5m, and the group density will increase significantly to every 10m 3 Including 30 fish, the movement speed will increase to 1.2m / s, and the consistency of movement direction will drop to 60%. The Mahalanobis distance D is calculated by calculating the predicted dynamic data and the mean baseline of the normal behavior of the fish school. M =4.8, which is beyond the 95% confidence interval of the normal behavior baseline, corresponding to D M =3, it is determined that the fish will behave abnormally in the area at the 10th hour. Based on the classification output of the model, the abnormal behavior type is predicted to be high-speed movement with aggregation, and the abnormal cause is predicted to be predator threat. According to the prediction results, an intervention strategy is formulated and implemented in advance. Before the predicted time arrives, high-frequency sound wave equipment is deployed in the area of ​​​​the position (15m, 30m, -4m) to drive away possible predators, and oxygenation equipment is added in the surrounding area to increase the dissolved oxygen concentration to alleviate the stress response of the fish.

[0055] Step S107, based on the fish school dynamic data acquired by the multi-frequency sonar receiver, combined with the Mahalanobis distance and the fish school behavior prediction, the effect of the fish school abnormality intervention strategy is evaluated, and the fish school abnormality intervention strategy and intervention range are dynamically adjusted.

[0056] Through the multi-frequency sonar receiver, the fish school dynamic data after the implementation of the fish school abnormal intervention strategy is obtained, and the Mahalanobis distance between the fish school dynamic data and the fish school normal behavior baseline is calculated. If the Mahalanobis distance is lower than the preset fish school normal behavior confidence interval, the fish school abnormal intervention strategy is judged to be effective, and the fish school abnormal behavior area has returned to normal. If the Mahalanobis distance still exceeds the preset fish school normal behavior confidence interval, the fish school abnormal intervention strategy is judged to be ineffective, and based on the fish school dynamic data, it is judged whether the abnormal area has expanded to the non-intervention area, and the fish school abnormal intervention strategy and the area where the strategy is implemented are adjusted until the fish school abnormal behavior area returns to normal. The fish school dynamic data of the area where the fish school abnormal intervention strategy is implemented in advance is obtained, and based on the fish school behavior prediction model, the fish school dynamic data of the fish school within the preset time period is re-predicted, and the fish school dynamic data of the fish school within the preset time period and the average Mahalanobis distance of the fish school normal behavior baseline obtained by prediction are used. Based on the Mahalanobis distance, it is judged whether the implementation of the fish school abnormal intervention strategy in advance eliminates the abnormal risk of the abnormal behavior in the future. If the Mahalanobis distance still exceeds the preset confidence interval of the normal behavior of the fish school, it is judged that the fish school abnormality intervention strategy implemented in advance is ineffective, and based on the predicted fish school dynamic data within the preset time period, it is judged whether the abnormal area has expanded to the non-intervention area, and the fish school abnormality intervention strategy and the regional scope of the strategy implementation are adjusted.

[0057] For example, the center point of a certain abnormal area is located at (20m, 30m, -3m). Through the multi-frequency sonar receiver, the dynamic data of the fish school after the implementation of the fish school abnormal intervention strategy is obtained in real time. For example, the density of the fish school is every 10m 3 There are 25 fish in the school, with a movement speed of 0.8m / s, a movement direction consistency of 70%, and a gathering radius of 6m. These dynamic data are compared with the normal behavior baseline of the school of fish, and the D is calculated using the Mahalanobis distance formula. M =2.5, which is lower than the 95% confidence interval threshold of the preset normal behavior confidence interval, corresponding to D M ≤3, so it is judged that the current fish school abnormality intervention strategy is effective, and the abnormal behavior in the area has returned to normal. Then continue to monitor the dynamic data of the area intervened in advance, and record the latest fish school dynamic data through the sonar receiver. In the area intervened in advance, the density is maintained at every 10m 3 There are 20 fish in the area, with a movement speed of 0.6m / s, a directional consistency of 85%, and an aggregation radius of 8m. Based on these data, the fish school behavior prediction model is used to re-predict the dynamic data of the fish school in the next 12 hours. The prediction results show that in the next 6 hours, the dynamics of the fish school in this area will basically tend to be normal, with a Mahalanobis distance of D M=2.8, which is still within the preset confidence interval of normal behavior, further confirming that the early intervention strategy is effective in eliminating future risks of abnormal behavior. By monitoring other unintervention areas, it was found that the fish density in a new area outside the abnormal area (25m, 35m, -4m) increased to every 10m 3 There are 30 fish in the pool, the movement speed increases to 1.0 m / s, the directional consistency decreases to 60%, the aggregation radius shrinks to 4 m, and the Mahalanobis distance is calculated as D M =4.2, which is beyond the confidence interval of normal behavior, indicating that abnormal behavior begins to expand to the unintervention area. Therefore, the intervention strategy is adjusted to expand the coverage of the intervention equipment to the new abnormal area, such as deploying high-frequency sound wave equipment at (25m, 35m, -4m) to drive away predators, and increasing the dissolved oxygen concentration around the area to reduce the stress response of the fish. After the expansion of the intervention measures, real-time monitoring shows that the fish density in the new area has dropped to 1.0 per 10m 3 There are 22 fish in the pool, the speed is reduced to 0.7m / s, the directional consistency rises back to 80%, and the Mahalanobis distance is recalculated as D M =2.9, which is lower than the preset confidence interval of normal behavior. It is judged that the expanded fish school abnormal intervention strategy successfully controls the abnormal behavior and effectively prevents the abnormal area from further expanding.

[0058] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, the above features are replaced with the technical features with similar functions disclosed in the present application (but not limited to) to form a technical solution.

Claims

1. A large-scale fish resource detection method using multi-frequency sound waves and AI algorithms, characterized in that: The method comprises: Adjust the sound wave emission parameters through multi-frequency sonar equipment, and correct the original echo signal to the signal under standard environmental characteristics based on environmental characteristics; According to the corrected echo signal, the spectrum characteristics and time series distribution information of the echo signal are extracted and normalized through time window segmentation, signal arrival time calculation and denoising processing; Through time series analysis of multi-frequency signals, frequency response gradient calculation and envelope curve extraction, the frequency response characteristic data of fish species are generated and associated with their spatial position information; Based on the frequency response characteristic data of multi-frequency sound waves in the fish sound wave monitoring database, the spatial distribution, population ratio and mixed fish marking information of fish schools are obtained through signal classification and three-dimensional point cloud clustering analysis; Based on the fish school echo signals and dynamic data obtained in real time by the multi-frequency sonar receiver, the parameter range of the normal behavior baseline of the fish school is determined. By calculating the Mahalanobis distance between the fish school dynamic data and the mean of the normal behavior baseline of the fish school, the type and area of ​​abnormal behavior of the fish school are identified, and an intervention strategy for abnormal fish school behavior is formulated; Based on the historical dynamic data of the fish school behavior pattern database and the baseline average of normal fish school behavior, the time point, area range, abnormal type and cause of abnormal fish school behavior in the future time period are predicted, and the fish school abnormality intervention strategy is implemented in advance; Based on the fish school dynamics data obtained by the multi-frequency sonar receiver, combined with the Mahalanobis distance and fish school behavior prediction, the effectiveness of the fish school abnormality intervention strategy is evaluated, and the fish school abnormality intervention strategy and intervention range are dynamically adjusted.

2. The method according to claim 1, wherein: The method of adjusting the acoustic wave emission parameters by the multi-frequency sonar device and correcting the original echo signal to a signal under standard environmental characteristics in combination with environmental characteristics includes: According to the preset detection waters and the possible distribution range of target fish species, a multi-frequency sonar device is used to configure multiple frequency ranges, and the order of low-frequency, medium-frequency and high-frequency transmission signals is set; through the sound wave transmission system of the multi-frequency sonar device, the sound wave transmission power, beam width and transmission angle are adjusted to transmit the sound wave signals to the designated waters one by one, and the original echo signals of the multi-frequency sound waves are obtained through the multi-frequency sonar receiver, and the spatial position information of the sound wave transmission is marked and stored in the fish sound wave monitoring database; the echo signals under standard environmental characteristics of the same distance, azimuth and depth and the echo signals under non-standard environmental characteristics are obtained through the fish sound wave monitoring database, and the recurrent neural network is used for model training to construct an echo signal correction model, and the environmental characteristics include water temperature, salinity, flow rate and turbidity; according to the original echo signals and environmental characteristics obtained in real time, the echo signal correction model is used to correct the original echo signals to echo signals under standard environmental characteristics.

3. The method according to claim 1, wherein: The method extracts and normalizes the frequency spectrum characteristics and time series distribution information of the echo signal through time window segmentation, signal arrival time calculation and denoising processing according to the corrected echo signal, including: According to the corrected echo signal, the echo signal is separated based on time window segmentation and signal arrival time calculation, and the echo intensity, scattering pattern and arrival time of sound waves of different frequencies are recorded; according to the time delay of the received signal, the sound wave propagation path in the water area is calculated, and the distance, azimuth, depth and position coordinates corresponding to the echo signal are recorded; the acquired echo signal is denoised by a bandpass filter to remove environmental noise and equipment interference in the non-target frequency range; the time series signal is converted into a time-frequency domain signal by a short-time Fourier transform algorithm, and the spectral feature data of the echo is extracted; the echo signal intensity data is amplitude normalized by a normalization method to unify the power amplitude of signals of different frequencies, and the frequency distribution and time change pattern are recorded to obtain the normalized spectral feature data and its time series distribution information.

4. The method according to claim 1, wherein: The method generates frequency response characteristic data of fish species and associates their spatial position information through time series analysis of multi-frequency signals, frequency response gradient calculation and envelope curve extraction, including: According to the normalized spectrum feature data, including the time series of multi-frequency signals and their spectrum features, the amplitude and phase information of signals in different frequency bands are extracted using fast Fourier transform, and the time series data at each frequency is recorded; according to the time series signals of each frequency, the formula I = 20·log 10 (|S|) Convert the time series signal into intensity value, and calculate the mean and variance of the echo signal intensity in each frequency range to obtain the scattering intensity characteristics corresponding to different frequencies, where S is the amplitude of the normalized signal; according to the scattering intensity characteristic data corresponding to different frequencies, perform fitting analysis on the scattering intensity data of different frequencies, establish a curve of scattering intensity changing with frequency, and calculate the gradient value of the curve Extract the intensity change rate between different frequencies to determine the sensitivity of fish to frequency; group the features according to the range of frequency change gradient to form a characteristic template of fish frequency response, and obtain the frequency sensitivity characteristics of fish species, including frequency response change gradient and characteristic template; extract the frequency response data of various fish at each frequency, calculate the characteristic separation index of different fish at each frequency band, and perform frequency band scaling on the frequency band signal intensity in the overlapping area of ​​signal characteristics; extract the envelope curve of the signal using Hilbert transform according to the time series signal of each frequency to obtain the shape characteristics of the reflected signal; according to the descending part of the echo signal envelope, use the exponential fitting model S(t)=S0·e -R·t The attenuation rate R of the signal is calculated to obtain the shape characteristics and attenuation rate characteristics of the fish echo signal, where S(t) is the envelope amplitude and S0 is the starting value of the envelope curve; the scattering intensity characteristics of sound waves of each frequency, the frequency sensitivity characteristics of the fish species, the shape characteristics and attenuation rate characteristics of the fish echo signal are combined according to the frequency range to generate a frequency response characteristic data table of the fish species, and the frequency response characteristic data of the fish species are associated with the distance, azimuth, depth and position coordinates of the waters where the fish are located.

5. The method according to claim 4, wherein: The method of extracting frequency response data of various fish species at various frequencies, calculating feature separation indexes of different fish species at various frequency bands, and performing frequency band scaling on the signal strength of the frequency band in the area where signal features overlap, includes: Extract various fish species at various frequencies from the echo signal, according to the formula Calculate its frequency response gradient G(f), record the G(f) value of each fish in the full frequency band and establish a frequency response gradient database, where S(f) represents the normalized signal intensity at frequency f, and Δf represents the frequency increment; according to the frequency response gradient of each fish, use the characteristic separation index formula The characteristic separation index I(f) of each fish species in each frequency band is calculated, where G1(f) and G2(f) are the frequency response gradients of the two fish species at frequency f, and σ1(f) and σ2(f) are the standard deviations of the signal strength of the two fish species at frequency f, respectively. Based on the preset separation index threshold, the frequency region where the characteristic separation index I(f) is less than the preset separation index threshold is marked as the signal feature overlapping region. The frequency range of the signal feature overlapping region is frequency band scaled by the frequency scaling formula f′=ξ·(f-f0)+f0, where f′ is the scaled frequency, ξ is the scaling factor used to amplify or compress the frequency interval, obtained by fitting historical data, and f0 is the center point of the frequency to maintain the consistency of the frequency band shift. For the scaled signal, the frequency response gradient G′(f) and the separation index I′(f) are calculated again. If I′(f) of all frequency bands is greater than or equal to the preset separation index threshold, there is no signal feature overlap problem. Otherwise, the scaling factor is adjusted until the separation index of all frequency bands reaches or exceeds the threshold.

6. The method according to claim 1, wherein: The method of obtaining the spatial distribution, population ratio and mixed fish marking information of fish schools by signal classification and three-dimensional point cloud cluster analysis based on the frequency response characteristic data of multi-frequency sound waves in the fish sound wave monitoring database includes: The frequency response characteristic data of different fish species under multi-frequency sound waves are obtained through the fish sound wave monitoring database, and the recurrent neural network is used for model training to build a fish recognition model; based on the frequency response characteristic data of fish species obtained in real time, the fish recognition model is used to determine the fish label and classification confidence of each echo signal; if the classification confidence is lower than the preset confidence threshold, it is marked as an invalid signal and removed; the classification results of all echo signals are statistically analyzed to obtain the overall distribution of fish species in the water area, determine the proportion and number of different fish populations, and map the signal points based on the distance, azimuth, depth and position coordinates corresponding to the echo signal. The data are projected into a three-dimensional spatial coordinate system to form point cloud data; based on the point cloud data, the K-means clustering algorithm or the DBSCAN clustering algorithm is used for model training, and the signal points belonging to the same fish species are input into the clustering algorithm to obtain the spatial boundaries of each fish school, the size of the fish school in each clustering area, the center position of the fish school and the boundary range, where the size of the fish school is the number of signal points in the area; the clustering results of different fish in the space are overlapped to determine whether there are two or more types of fish overlapping in the same space. If an overlapping area is detected, it is marked as a mixed fish school; if a mixed fish school exists, the area is marked separately and the fish species and quantity statistics are output.

7. The method according to claim 1, wherein: The method determines the parameter range of the baseline of normal behavior of the fish school based on the fish school echo signal and dynamic data obtained in real time by the multi-frequency sonar receiver, identifies the type and area of ​​abnormal behavior of the fish school by calculating the Mahalanobis distance between the dynamic data of the fish school and the mean of the baseline of normal behavior of the fish school, and formulates an intervention strategy for abnormal fish school behavior, including: The multi-frequency sonar receiver receives the fish school echo signal in real time, obtains the timestamp, distance, azimuth, depth and position coordinates of each signal point, and determines the fish school dynamic data, including the fish school's position coordinates, group density, movement direction and movement rate; according to the fish school dynamic data, the fish school behavior characteristics are determined, and the fish school behavior characteristics include the fish school density change rate, spatial aggregation characteristics, movement direction consistency and speed change rate. The spatial aggregation characteristics are the degree of aggregation of the fish school in a specific area, and the movement direction consistency is the similarity of the movement direction of individuals in the group; according to the fish school behavior characteristics, the K-means clustering algorithm is used for clustering analysis, and a baseline model of normal fish school behavior is established to determine the parameter range of the normal fish school behavior baseline, including the normal group density change range, aggregation radius and spatial distribution characteristics, average movement direction and speed range; the Mahalanobis distance calculation formula is used Calculate the mean Mahalanobis distance D between the fish school dynamic data and the baseline of the fish school's normal behavior M , determine the degree of deviation between the real-time behavior of the fish school and the normal behavior baseline, where ∑ is the covariance matrix; if D M If the confidence interval of the normal behavior of the fish school is exceeded, the current fish school behavior is considered abnormal. M and preset abnormality assessment thresholds to determine the degree of abnormality of the fish school, and mark the type and degree of abnormal behavior of the fish school. The types of abnormal behavior include but are not limited to aggregation, dispersion or high-speed movement, and the degrees of abnormality include mild, moderate and severe. According to the location coordinates of the fish school, determine the location coordinates and regional scope of the abnormal behavior of the fish school, and count the group characteristics of the area where the abnormal behavior of the fish school occurs, including the number and density of the fish school; store the fish school behavior characteristics of the area where the abnormal behavior of the fish school occurs in the fish school behavior pattern database; obtain the fish school behavior characteristics of the historical area where the abnormal behavior of the fish school occurs through the fish school behavior pattern database, and mark the abnormal reasons, use the random forest algorithm to train the model, and build a fish school abnormal cause identification model. The abnormal reasons include but are not limited to predator threats, environmental disturbances or human intervention; according to the real-time acquired fish school behavior characteristics, use the fish school abnormal cause identification model to determine the cause of the fish school abnormality; based on the time of occurrence, regional scope, cause of the fish school abnormality and the degree of abnormality of the fish school abnormal behavior, formulate and implement fish school abnormality intervention strategies, including but not limited to adjusting water quality parameters, reducing pollution load, increasing dissolved oxygen concentration or optimizing water flow rate.

8. The method according to claim 1, wherein: The historical dynamic data of the fish school behavior pattern database, combined with the baseline mean of the normal behavior of the fish school, predicts the time point, area range, abnormal type and cause of the abnormal behavior of the fish school in the future time period, and implements the fish school abnormality intervention strategy in advance, including: Through the fish school behavior pattern database, historical fish school dynamic data is obtained, and the model is trained using the long short-term memory network to build a fish school behavior prediction model to predict the fish school dynamic data within a preset time period; based on the predicted fish school dynamic data within the preset time period, the Mahalanobis distance between the fish school dynamic data and the mean baseline of the fish school's normal behavior is calculated to determine the time point, regional range, abnormal type and abnormal cause of the abnormal behavior of the fish school within the future preset time period; according to the predicted time point when the fish school will have abnormal behavior, the fish school abnormality intervention strategy is implemented in advance to intervene.

9. The method according to claim 1, wherein: The fish school dynamic data obtained by the multi-frequency sonar receiver is combined with the Mahalanobis distance and the fish school behavior prediction to evaluate the effect of the fish school abnormality intervention strategy, and dynamically adjust the fish school abnormality intervention strategy and intervention range, including: Through the multi-frequency sonar receiver, the fish school dynamic data after the implementation of the fish school abnormal intervention strategy is obtained, and the Mahalanobis distance between the fish school dynamic data and the fish school normal behavior baseline is calculated. If the Mahalanobis distance is lower than the preset fish school normal behavior confidence interval, the fish school abnormal intervention strategy is judged to be effective, and the fish school abnormal behavior area returns to normal; if the Mahalanobis distance still exceeds the preset fish school normal behavior confidence interval, the fish school abnormal intervention strategy is judged to be ineffective, and based on the fish school dynamic data, it is judged whether the abnormal area has expanded to the non-intervention area, and the fish school abnormal intervention strategy and the area where the strategy is implemented are adjusted until the fish school abnormal behavior area returns to normal; obtain the fish school abnormal intervention strategy implemented in advance The fish school dynamics data of the area are re-predicted based on the fish school behavior prediction model, and the fish school dynamics data of the fish school within the preset time period are re-predicted, and the Mahalanobis distance between the predicted fish school dynamics data of the fish school within the preset time period and the baseline mean of the normal behavior of the fish school is determined; based on the Mahalanobis distance, it is determined whether the early implementation of the fish school abnormality intervention strategy can eliminate the abnormal risk of the abnormal behavior in the future; if the Mahalanobis distance still exceeds the preset fish school normal behavior confidence interval, it is determined that the early implementation of the fish school abnormality intervention strategy is ineffective, and based on the predicted fish school dynamics data of the fish school within the preset time period, it is determined whether the abnormal area expands to the non-intervention area, and the fish school abnormality intervention strategy and the regional scope of the strategy implementation are adjusted.

Images