A shrimp breeding risk early warning method and system based on water body index time sequence learning

Abstract

The application discloses a shrimp breeding risk early warning method and system based on water body index time sequence learning, and relates to the technical field of aquaculture data processing. The method obtains a water body index sequence and the latest fry state detection data, generates a breeding risk benchmark value according to the fry state detection data and determines an evaluation time; identifies a water body index state segment according to a fixed monitoring cycle, connects a cross-cycle water body index evolution process according to a stage connection relationship, calibrates a process position and generates a water body process record; calls a pre-trained cumulative risk change amount prediction model to generate a start-stop cumulative risk change amount, obtains an interval risk change amount through difference calculation and collection, generates a predicted risk value in combination with the breeding risk benchmark value, determines a risk grade and an abnormal water body index and outputs an early warning result. The application reduces early warning deviations caused by the fact that a water body abnormal process is cut by a cycle and multiple independent abnormal processes share a whole breeding risk value.

Description

Technical Field

[0001] This invention relates to the field of aquatic seedling data processing technology, specifically to a shrimp seedling risk early warning method and system based on time-series learning of water body indicators. Background Technology

[0002] Shrimp larvae are highly sensitive to changes in the aquatic environment during the larval stage. Continuous deviations in water indicators such as water temperature, dissolved oxygen, pH, salinity, ammonia nitrogen, nitrite, and turbidity can affect the larvae's feeding, molting, and survival. By continuously collecting water indicators and learning the correlation between changes in water indicators and larval risks, data can be provided for risk warning in the shrimp larval process.

[0003] Existing shrimp breeding risk early warning systems collect water body indicators according to fixed monitoring cycles, combine the water body indicator data from each fixed monitoring cycle into time-series learning samples, and generate breeding risk values ​​based on the proportion of dead seedlings, the proportion of abnormally active seedlings, and the proportion of decreased feeding response detected at the end of the cycle. The breeding risk values ​​are then used as sample labels to train risk prediction models. However, abnormal water body indicators have a continuous evolutionary process of deviation, development, and gradual recovery. A complete abnormal water body process may span adjacent fixed monitoring cycles, and multiple independent abnormal water body indicator processes may occur successively within the same fixed monitoring cycle.

[0004] The limitations of existing technologies include at least the following problems: When water body index data is correlated with seedling risk values ​​according to a fixed monitoring cycle, the same water body abnormal process spanning adjacent fixed monitoring cycles will be split and correspond to different seedling risk values. Multiple independent water body abnormal processes within the same fixed monitoring cycle will share a single overall seedling risk value as the learning basis, making the correspondence between water body index changes and seedling risk values ​​inaccurate and affecting the accuracy of shrimp seedling risk early warning. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a shrimp seedling risk early warning method and system based on time-series learning of water body indicators. This method solves the problem that the complete abnormal process of a water body is fragmented under a fixed monitoring period, and multiple independent abnormal processes of water bodies share the whole seedling risk value, resulting in inaccurate correspondence between changes in water body indicators and seedling risk values.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a shrimp seedling risk early warning method based on time-series learning of water body indicators, comprising the following steps: acquiring the current water body indicator sequence and the most recent seedling status detection data, the seedling status detection data including the proportion of dead seedlings, the proportion of abnormally active seedlings, and the proportion of decreased feeding response; generating a current seedling risk benchmark value by weighting the data according to preset weights, and determining the time of completion of the detection as the assessment time; based on the changing trend of water body indicators relative to the benchmark range, identifying current water body indicator status segments according to a fixed monitoring cycle, and connecting time-adjacent current water body indicator status segments according to the succession relationship of the benchmark stable stage, deviation development stage, deviation persistence stage, and recovery stage to generate the current water body... The system generates a set of indicator evolution processes; it calibrates the process position of each current water body indicator evolution process, and generates a current water body process record set based on the overlap between the current water body indicator evolution process and the current risk prediction interval formed from the assessment time to the current time; it calls a pre-trained cumulative risk change prediction model, and generates the current initial cumulative risk change and the current final cumulative risk change based on the initial process prefix features, final process prefix features, initial process position, and final process position of each current water body process record, and determines the process interval risk change; it aggregates the risk change of each process interval to generate the current interval risk change, generates the predicted risk value based on the current seedling risk benchmark value and the current interval risk change, and outputs the shrimp seedling risk warning result.

[0007] Further, the identification steps for the current water body indicator status segment are as follows: Read the water body indicator vectors in the current water body indicator sequence according to the acquisition time; establish candidate stable windows in the current water body indicator sequence; generate an indicator stability score based on the ratio obtained by dividing the standard deviation of each water body indicator by the sum of its absolute mean and a preset smoothing parameter, and the inversion ratio of adjacent value changes; aggregate the water body indicator vectors corresponding to the candidate stable windows that continuously meet the indicator stability conditions into indicator benchmark segments; generate indicator benchmark ranges based on the mean and standard deviation of each water body indicator; determine the deviation amplitude by the ratio of the magnitude of each water body indicator exceeding the indicator benchmark range to the width of the corresponding range, and determine the deviation direction based on the position of the upper and lower boundaries of the relative range of the water body indicator; mark the benchmark stable stage, deviation development stage, deviation persistence stage, and recovery stage according to the changing trend of the deviation amplitude; aggregate water body indicator vectors with continuous time and the same stage markings within each fixed monitoring cycle to generate the current water body indicator status segment.

[0008] Furthermore, the steps for generating the current water body indicator evolution process set are as follows: Read the stage markers and end water body indicator vectors of time-adjacent current water body indicator state segments; determine time-adjacent segments whose stage markers are sequentially connected as stage connection candidate connections, and determine segments with the same stage markers on both sides of adjacent fixed monitoring cycle boundaries as cross-cycle continuation candidate connections; generate a comprehensive connection distance based on the ratio of the absolute difference of the values ​​of the common effective water body indicators at both ends of the candidate connection to the width of the corresponding indicator benchmark range; determine candidate connections whose comprehensive connection distance is not higher than the connection threshold and whose stage markers conform to the corresponding stage relationship as effective connections; aggregate current water body indicator state segments along the effective connections; when water body indicators after the end of the recovery stage are marked as benchmark stable stages, the end of the recovery stage is taken as the endpoint of the current water body indicator evolution process; determine connection paths that have not returned to the benchmark stable stage by the end of the current water body indicator sequence as incomplete current water body indicator evolution processes; state segments containing only benchmark stable stages are not used to establish current water body indicator evolution processes.

[0009] Furthermore, the calibration steps for the process position are as follows: set the stage position for the current water body indicator state segment according to the stage type; determine the position within the stage based on the positional relationship between the water body indicator vector acquisition time and the start and end times of the current water body indicator state segment; associate the stage position and the position within the stage to generate the process position of the corresponding water body indicator vector in the current water body indicator evolution process.

[0010] Furthermore, the steps for generating the current water body process record set are as follows: First, determine the overlapping intervals between the evolution processes of each current water body indicator and the current risk prediction interval. Then, define the process positions at both ends of the overlapping interval as the starting and ending process positions, respectively. Next, extract the stage markers from the starting point to the starting position and from the starting point to the ending position of the current water body indicator evolution process, as well as the deviation direction and magnitude of the water body indicator, to generate the starting and ending process prefix features. Finally, associate the current water body indicator evolution process, the overlapping intervals, the starting and ending process positions, the starting and ending process prefix features, and the ending process prefix features to generate the current water body process record set.

[0011] Furthermore, the steps for determining the starting and ending positions of the process are as follows: When the starting point of the current water body index evolution process is earlier than the starting time of the current risk prediction interval, the position of the first effective water body index vector that is no earlier than the starting time of the current risk prediction interval is determined as the starting position; when the starting point of the current water body index evolution process is within the current risk prediction interval, its starting point is determined as the starting position; when the current water body index evolution process completes recovery and returns to the baseline stable stage within the current risk prediction interval, the position of the last effective water body index vector in the corresponding recovery stage is determined as the ending position; when it has not yet completed recovery and returned to the baseline stable stage, the position of the last effective water body index vector that is no later than the end time of the current risk prediction interval is determined as the ending position.

[0012] Furthermore, the pre-trained cumulative risk change prediction model includes a temporal coding layer, a process position coding layer, a feature fusion layer, and a risk change regression layer. The temporal coding layer uses gated recurrent units to process the prefix features of the starting process and the prefix features of the ending process, respectively. The process position coding layer processes the starting process position and the ending process position, respectively. The feature fusion layer fuses the corresponding coding results. The risk change regression layer uses linear output to generate the current starting cumulative risk change and the current ending cumulative risk change, which can be positive, zero, or negative.

[0013] Furthermore, the steps for generating the current interval risk change are as follows: subtract the current start cumulative risk change from the current end cumulative risk change of the same current water body process record to obtain the process interval risk change; aggregate the process interval risk changes of each current water body process record to generate the current interval risk change.

[0014] Furthermore, the steps for generating the shrimp seedling risk warning result are as follows: The current seedling risk baseline value is added to the current interval risk change to generate an initial predicted risk value, and boundary correction is performed according to the preset effective range of the risk value; the corrected predicted risk value is matched with the preset risk level interval to determine the current risk level; the maximum deviation of each water body indicator in each current water body process record within the corresponding overlapping interval is determined, and water body indicators whose maximum deviation reaches the abnormal indicator threshold are identified as abnormal water body indicators, and a warning result containing the current risk level, abnormal water body indicators, and the corresponding water body indicator evolution process identifier is output.

[0015] A shrimp seedling risk early warning system based on time-series learning of water body indicators includes: a water body data acquisition module for acquiring the current water body indicator sequence and the most recent seedling status detection data, generating the current seedling risk benchmark value and determining the assessment time; a water body segment identification module for generating current water body indicator status segments; a water body process generation module for generating the current water body indicator evolution process set; a process location calibration module for calibrating the process location; a process record generation module for generating the current water body process record set; a model invocation module for generating the current initial cumulative risk change and the current final cumulative risk change; an interval risk change analysis module for generating the process interval risk change and the current interval risk change; and a risk early warning module for generating predicted risk values ​​and outputting early warning results.

[0016] The present invention has the following beneficial effects: (1) The shrimp seedling risk early warning method based on water body index time series learning identifies the current water body index state segments according to a fixed monitoring cycle, and connects the current water body index state segments with adjacent time according to the connection relationship of the baseline stable stage, deviation development stage, deviation continuous stage and recovery stage. This enables the abnormal changes of the same water body that span the fixed monitoring cycle to be classified into the same current water body index evolution process according to the original evolution sequence, reducing the abnormal process break caused by fixed cycle segmentation.

[0017] (2) The shrimp seedling risk early warning method based on water body index time series learning generates the starting process prefix feature and the ending process prefix feature by calibrating the stage position and the position within the stage, and generating the starting process prefix feature and the ending process prefix feature according to the overlap relationship between the current water body index evolution process and the current risk prediction interval. This allows the current water body process record to retain the stages that the water body abnormality has already gone through before entering the prediction interval and the change process within the interval, reducing the loss of time series context caused by only using the current monitoring cycle data.

[0018] (3) The shrimp seedling risk early warning method based on water body index time series learning generates the cumulative risk change corresponding to the start and end positions of the current water body process record through the pre-trained cumulative risk change prediction model. Based on the difference between the start and end cumulative risk changes, the risk change of each independent water body abnormal process in the current risk prediction interval is determined, so that multiple independent water body abnormal processes in the same monitoring period form corresponding risk results, reducing the learning bias caused by multiple abnormal processes sharing the whole seedling risk value.

[0019] (4) The shrimp seedling risk early warning system based on water body index time series learning, through the collaborative processing of water body data acquisition module, water body segment identification module, water body process generation module, process location calibration module, process record generation module, model calling module, interval risk change analysis module and risk early warning module, converts the current water body index sequence into the current water body index state segment, the current water body index evolution process, the current water body process record and the current interval risk change amount, and generates the predicted risk value and shrimp seedling risk early warning result by combining the current seedling risk benchmark value, which is convenient for continuously identifying the current risk level, abnormal water body index and the corresponding water body index evolution process.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of a shrimp seedling risk early warning method based on time-series learning of water body indicators according to the present invention.

[0022] Figure 2 This is an architecture diagram of a shrimp seedling risk early warning system based on time-series learning of water body indicators according to the present invention. Detailed Implementation

[0023] Please see Figure 1 This invention provides a technical solution: a shrimp seedling risk early warning method based on time-series learning of water body indicators, comprising the following steps: acquiring the current water body indicator sequence and the most recent seedling status detection data, the seedling status detection data including the proportion of dead seedlings, the proportion of abnormally active seedlings, and the proportion of decreased feeding response; generating a current seedling risk benchmark value by weighting the data according to preset weights, and determining the detection completion time as the assessment time; based on the changing trend of water body indicators relative to the benchmark range, identifying current water body indicator state segments according to a fixed monitoring cycle, and connecting time-adjacent current water body indicator state segments according to the succession relationship of the benchmark stable stage, deviation development stage, deviation persistence stage, and recovery stage to generate the current water body indicator evolution. The process set is defined; the process position of the evolution process of each current water body indicator is identified; based on the overlap between the evolution process of each current water body indicator and the current risk prediction interval formed from the assessment time to the current time, a current water body process record set is generated; a pre-trained cumulative risk change prediction model is invoked; based on the start process prefix features, end process prefix features, start process position, and end process position of each current water body process record, the current starting cumulative risk change and the current ending cumulative risk change are generated, and the process interval risk change is determined; the risk change of each process interval is aggregated to generate the current interval risk change; based on the current seedling risk benchmark value and the current interval risk change, a predicted risk value is generated, and the shrimp seedling risk warning result is output.

[0024] Among them, the preset collection cycle is the set time interval between two adjacent collections by the water quality sensor, the fixed monitoring cycle is a continuous time interval set according to the seedling management requirements, and its duration is an integer multiple of the preset collection cycle. The benchmark observation interval ends at the assessment time and extends forward by a preset duration. The current risk prediction interval extends from the assessment time to the current time. When a water body index exceeds the range of the most recent effective index at the starting point of the baseline observation interval, the baseline observation interval is extended forward to the most recent historical stable segment, so that the current water body index sequence covers the starting point of the corresponding water body index evolution process.

[0025] Water quality indicators include water temperature, dissolved oxygen, pH, salinity, ammonia nitrogen, nitrite, and turbidity. Each water quality indicator is collected by the corresponding water quality sensor in the seedling pond. The controller combines the indicator values ​​and their valid markers at the same collection time into a water quality indicator vector. The water quality indicator vector is then associated with the seedling pond identifier, seedling batch identifier, and collection time and arranged according to the collection time to form the current water quality indicator sequence.

[0026] The detection area is the pre-marked observation area of ​​the seedling pond in the image acquired by the seedling status detection terminal. The effective seedling trajectory is the trajectory formed by the association of continuous seedling target positions within the corresponding detection time window for a duration that reaches the effective trajectory duration. The number of dead seedlings is the number of seedling targets in the detection area whose seedling contour centroid displacement is lower than the death displacement threshold, whose average optical flow amplitude of adjacent frames within the seedling contour is lower than the local motion threshold, and whose death confirmation duration is reached. The most recent seedling status detection data refers to the latest detection record where both the total number of identified seedlings and the number of valid seedling trajectories have reached the corresponding detection threshold, and both the swimming speed response and the feeding area aggregation response have valid calculation results. The proportion of dead seedlings is the ratio of the number of dead seedlings to the total number of identified seedlings. The proportion of abnormally active seedlings is the ratio of the number of seedling trajectories with an average swimming speed lower than the corresponding seedling age speed threshold and an actual distance from the start point to the end point of the trajectory lower than the displacement threshold to the total number of valid seedling trajectories.

[0027] The method for obtaining the decrease in feeding response is as follows: the feed area is the pre-marked area in the seedling status detection image where the feeder's discharge position is located, and the detection period is from the start of the detection time window before feeding to the end of the detection time window after feeding. Calculate the difference between the average swimming speed within the detection time window after feeding and the average swimming speed within the detection time window before feeding. Also calculate the ratio of the number of effective seedling tracks entering the feeding area after feeding to the total number of effective seedling tracks after feeding. This ratio is based on the most recent data from the same seedling batch before the current detection time. The median of the previous normal feeding record is used as the swimming speed response benchmark and the food area aggregation response benchmark. The swimming speed response decline rate and the food area aggregation response decline rate are calculated by dividing the difference between the response benchmark and the current response by the larger of the absolute value of the response benchmark and the corresponding lower limit of the response. The result is recorded as 0 when it is less than 0 and as 1 when it is greater than 1. The average of the two decline rates is the feeding response decline ratio.

[0028] Normal feeding records refer to feeding records where all water indicators before and after feeding do not exceed the corresponding benchmark range, the proportion of dead seedlings and the proportion of abnormally active seedlings in the seedling status monitoring data do not reach the corresponding abnormal thresholds, and there are no feeding records of changing ponds, merging ponds, separating seedlings, or concentrated harvesting during the monitoring period. Insufficient normal feeding records for the same seedling batch are also considered abnormal. Next, supplementary data was collected by reading historical normal feeding records of the same shrimp species and the same seedling age, arranged from most recent to oldest testing time. However, this still did not achieve the desired result. The feeding response baseline was determined from historical normal feeding records prior to model deployment.

[0029] The current seedling risk benchmark value is calculated using the following formula: ; In the formula, This serves as the current baseline value for seedling cultivation risk. , and In order of frequency, the proportion of dead seedlings, the proportion of seedlings exhibiting abnormal activity, and the proportion of seedlings with decreased feeding response were ranked as follows: , and These are the risk item weights, where each risk item weight is a non-negative number and their sum is 1. In one implementation, the risk component weights corresponding to the proportion of dead seedlings, the proportion of abnormally active seedlings, and the proportion of decreased feeding response are 0.5, 0.3, and 0.2, respectively, and the historical seedling risk value and the current seedling risk benchmark value use the same risk component weights.

[0030] Specifically, the steps for identifying the current water body indicator status segment are as follows: The water body indicator vector is read from the current water body indicator sequence according to the collection time, specifically as follows: An invalid flag is written when the index value exceeds the sensor's effective range, the sensor self-test fails, or the change relative to the previous valid value exceeds the device's allowable transition threshold. Linear interpolation is used when the missing duration does not exceed the missing duration threshold and valid index values ​​exist before and after the missing interval. An invalid flag is retained when only one valid index value exists in the missing interval, the missing duration exceeds the missing duration threshold, or the interpolation result exceeds the sensor's effective range. Water index vectors with fewer valid water body indexes than the minimum number of valid indexes do not participate in stage marking.

[0031] Candidate stable windows are established in the current water body indicator sequence. A stability score is generated based on the ratio of the standard deviation of each water body indicator divided by the sum of its absolute mean and a preset smoothing parameter, and the inversion ratio of adjacent values. Specifically: An initial candidate stable window is established within the baseline observation interval, and updated candidate stable windows are established in the water body index sequence after the completion of the water body index evolution process and before the next deviation development stage. Each candidate stable window contains a preset number of window points. A continuous water body index vector is generated and moved at time intervals according to a preset collection cycle. The index stability score is calculated according to the following formula: ; In the formula, To ensure stable scoring of indicators, For water body indicators, For the first The weight of each water body indicator and , To participate in the calculation of the label, the first The number of effective values ​​of a water body indicator within the candidate stable window reaches the minimum number of effective values ​​in the window. If the value is 1, then the value is 0; otherwise, the value is 0. , , and The order is number 1 The mean, standard deviation, and proportion of valid data for each water body indicator within the candidate stability window. To be a smoothing parameter with the same dimensions as the corresponding water body index. This is the direction flip ratio. , When the denominator of the formula is 0, the corresponding candidate stable window is marked as an invalid window.

[0032] The direction reversal ratio is calculated as follows: Calculate the difference between adjacent valid index values ​​according to the collection order, delete the positions with a difference of 0, divide the number of times the sign of two adjacent non-zero difference values ​​changes by the number of comparable adjacent non-zero difference value pairs, and record the direction flipping ratio as 0 when no comparable sign pairs can be formed. The effective data ratio is the ratio of the number of water body indicator values ​​with valid labels within the candidate stable window to the total number of water body indicator values ​​to be collected. The total number of water body indicator values ​​to be collected is... and The product of.

[0033] The water body index vectors corresponding to candidate stability windows that continuously meet the index stability conditions are aggregated into index benchmark segments. The index benchmark range is then generated based on the mean and standard deviation of each water body index, specifically as follows: Candidate stable windows with an indicator stability score not higher than the stability score threshold, a window effective data ratio reaching the window effective ratio threshold, and the window mean of each water body indicator involved in the calculation within the range of the corresponding target water body indicator are marked as stable windows. Stable windows with continuous time and overlap are merged to form an indicator benchmark segment. No. The initial lower and upper bounds of the water body index are respectively the baseline mean minus and plus the range coefficient. The product of the initial range width and the baseline standard deviation results in an initial range width that is less than the minimum range width. Expanding to the baseline mean When the extended end exceeds the sensor's effective range, the excess width will be compensated to the other end, resulting in both the lower and upper bounds being limited within the sensor's effective range. And it should not exceed the effective range width of the corresponding sensor.

[0034] When no stable window is detected in the current baseline observation interval, the most recent effective indicator baseline segment before the evaluation time of the same seedling batch, the historical indicator baseline segment of the same shrimp species and the same seedling age, and the target water body indicator range pre-recorded by the seedling process are used in sequence. When using the target water body indicator range, its upper and lower limits are limited to the effective range of the corresponding sensor. Arrange the benchmark segments of each indicator according to their end time, select the benchmark segment with the closest end time before the evaluation time as the initial benchmark, and identify the first deviation from the starting point of the development stage based on the initial benchmark. Whenever the evolution process of a current water body index ends, the most recently formed index benchmark segment is prioritized to be updated as the new initial index benchmark. If no new index benchmark segment has been formed, the previous initial index benchmark is used, and the starting point of the next deviation development stage is identified. The initial index benchmarks adopted at each deviation from the starting point of the development stage are determined as the index benchmarks corresponding to the current water body index evolution process, and remain unchanged until the end of the process; When using the seedling cultivation process to pre-record the target water body index range, the midpoint between its upper and lower boundaries is determined as the benchmark mean value of the index.

[0035] The deviation range is determined by the ratio of the extent to which each water body indicator exceeds the benchmark range to the width of the corresponding range. The direction of deviation is then determined based on the position of the water body indicator relative to the upper and lower boundaries of the range. Specifically: ; ; ; In the formula, For the first The effective range of the water body indicators and These are the lower and upper bounds of the indicator's benchmark range, respectively. For the first Water body indicators at the time of collection The index value, To correspond to the deviation range, For the time of data collection The overall deviation, The value is 1 if the flag is valid and 0 if it is invalid. When the index value is below the lower limit of the index benchmark range, within the index benchmark range, or above the upper limit of the index benchmark range, the deviation direction is coded as -1, 0, and 1 respectively. When the number of effective water body indicators is lower than the minimum number of effective indicators, the comprehensive deviation is not calculated.

[0036] Based on the changing trend of the deviation magnitude, the baseline stabilization stage, the deviation development stage, the deviation persistence stage, and the recovery stage are marked. Within each fixed monitoring cycle, water body indicator vectors with continuous time and identical stage markings are collected to generate the current water body indicator state segment, specifically: When the overall deviation is continuously lower than the stable deviation threshold and the number of continuous sampling points reaches the number of stable confirmation points, the first sampling point among the continuous sampling points is determined as the starting point of the baseline stable stage. When no stage marker has been formed and the overall deviation reaches the stable deviation threshold, the corresponding sampling point is marked as the deviation development stage. When the overall deviation reaches the stable deviation threshold for the first time in the baseline stable stage, it transitions to the deviation development stage. When the overall deviation reaches the continuous deviation threshold in the deviation development stage and the range within the continuous judgment window is not higher than the continuous fluctuation threshold, it transitions to the deviation continuity stage. When the average reduction of the overall deviation in the recovery judgment window in the deviation development stage or the deviation continuity stage reaches the recovery reduction threshold, it transitions to the recovery stage. When the overall deviation reaches the stable deviation threshold in the recovery stage and the average increment in the development judgment window reaches the development increment threshold, it returns to the deviation development stage. When the overall deviation is continuously lower than the stable deviation threshold in the recovery stage and the number of continuous sampling points reaches the number of stable confirmation points, the first sampling point among the continuous sampling points is determined as the new starting point of the baseline stable stage, and the last valid sampling point before this starting point is determined as the end of the recovery stage. If the stage transition conditions are not met, the current stage marker is maintained.

[0037] The average increment is the difference between the last value and the first value of the development judgment window divided by the number of collection intervals within the window; the average decrease is the difference between the first value and the last value of the recovery judgment window divided by the number of collection intervals within the window; the range within the continuous judgment window is the difference between the maximum and minimum values ​​of the comprehensive deviation within the window; and water body index vectors with the same stage marker and continuous collection time are grouped into the same current water body index state segment within each fixed monitoring cycle.

[0038] Specifically, the steps for generating the current water body index evolution process set are as follows: The stage marker and end water body index vector of the current water body index state segment with adjacent time are read, specifically as follows: Arrange the current water body index state segments according to the start time, and read the stage mark, start and end time, first effective water body index vector and last effective water body index vector of each segment. Segments that lack effective end water body index vectors are not included in the candidate connection.

[0039] The time-adjacent segments whose stage markers are sequentially connected are identified as candidate links for stage connection, and segments with the same stage markers on both sides of the boundary of adjacent fixed monitoring periods are identified as candidate links for cross-period continuation. Specifically: The benchmark stabilization phase is allowed to take over the deviation development phase; the deviation development phase is allowed to take over the deviation development phase, the deviation persistence phase, or the recovery phase; the deviation persistence phase is allowed to take over the deviation persistence phase or the recovery phase; and the recovery phase is allowed to take over the recovery phase or the deviation development phase. Only state segments with adjacent start and end times and a time interval not exceeding the maximum connection interval are established as phase succession candidate connections. For state segments located on both sides of adjacent fixed monitoring cycle boundaries and with the same phase label, cross-cycle continuation candidate connections are established. Each state segment corresponds to at most one candidate predecessor and one candidate successor.

[0040] The comprehensive connection distance is generated based on the ratio of the absolute difference of the values ​​of the common effective water body indicators at both ends of the candidate connection to the width of the corresponding indicator's baseline range. Candidate connections whose comprehensive connection distance is not higher than the connection threshold and whose stage markers conform to the corresponding stage relationships are determined as valid connections. Specifically: When the number of common effective water indicators at both ends of the candidate connection reaches the minimum number of common effective indicators, the comprehensive connection distance is calculated according to the following formula: ; In the formula, For fragments To fragment The overall connection distance and They are respectively the first two ends connected Water body index values, The value is 1 when both ends of the marker are valid and the corresponding water body indicators are valid; otherwise, the value is 0. It is the ratio of the number of common effective water body indicators to the total number of water body indicators. The time interval between two state segments. The maximum connection interval, For common effective proportion correction factor, This is a time interval correction factor; Insufficient number of common effective water body indicators If the overall connection distance exceeds the connection threshold, a valid connection will not be established.

[0041] The current water body index state segments are collected along valid connections. When a water body index after the end of the recovery phase is marked as a baseline stable phase, the end of the recovery phase is taken as the endpoint of the current water body index evolution process. Connection paths that have not returned to the baseline stable phase by the end of the current water body index sequence are determined as incomplete current water body index evolution processes. State segments that only contain the baseline stable phase are not used to establish current water body index evolution processes. Specifically: The evolution process of the current water body index is established starting from the baseline stable stage segment with a subsequent segment that deviates from the development stage. When there is no effective predecessor connection for the abnormal stage segment, the evolution process of the current water body index is established starting from the abnormal stage segment. The subsequent segments are collected along the effective connection and written with the same process identifier. When a water quality indicator after the end of the recovery phase is marked as the baseline stable phase, the end of the recovery phase is taken as the endpoint of the current water quality indicator evolution process. If the current water quality indicator sequence has not returned to the baseline stable phase by the end of the current phase, the collected connection paths are retained as incomplete processes. Only state fragments containing the baseline stable phase are not used to establish the current water quality indicator evolution process.

[0042] Specifically, the calibration steps for the process location are as follows: The stage position is set according to the stage type for the current water body indicator state segment, specifically as follows: The baseline stabilization stage, deviation development stage, deviation persistence stage, and recovery stage are coded as 0, 1, 2, and 3, respectively. The stage position indicates the stage to which the corresponding water body indicator vector belongs, but does not indicate the direction of risk change.

[0043] The position within a stage is determined based on the positional relationship between the water body index vector acquisition time and the start and end times of the corresponding current water body index state segment. Specifically: The time difference between the water body index vector acquisition time and the start time of the corresponding state segment is divided by the duration of the state segment. If the state segment contains only one valid water body index vector, the position within the segment is recorded as 0.

[0044] By associating the stage position and the position within the stage, the corresponding water body index vector is generated as the process position in the current water body index evolution process, specifically as follows: The process location includes the stage location, the location within the stage, the segment number, and the cumulative duration. The segment number is the sequence number of the state segment in the evolution process of the current water body index according to the start time and increases from 0. The cumulative duration is the time difference between the start time of the current water body index evolution process and the time of acquisition of the corresponding water body index vector.

[0045] Specifically, the steps for generating the current water body process record set are as follows: The overlapping interval between the current water body index evolution process and the current risk prediction interval is determined, and the process positions corresponding to the two ends of the overlapping interval are respectively determined as the start process position and the end process position, as follows: The start time of the overlapping interval is the later of the start time of the current water body index evolution process and the evaluation time. The end time of the overlapping interval is the earlier of the end time of the current water body index evolution process and the current time. For incomplete processes, the acquisition time of the last valid water body index vector at the end of the aggregated connection path is used as the process end time. If the start time of the overlapping interval is later than the end time of the overlapping interval, no current water body process record is generated.

[0046] When the starting point of the current water body index evolution process is earlier than the starting time of the current risk prediction interval, the first valid water body index vector that is no earlier than the assessment time is read in ascending order of the collection time and its process position is determined as the starting process position. When the starting point of the current water body index evolution process is within the current risk prediction interval, the process starting point is determined as the starting process position. When the current water body index evolution process completes recovery and returns to the baseline stable stage within the current risk prediction interval, the position of the last effective water body index vector in the corresponding recovery stage is determined as the end of the process position. If the recovery has not yet been completed and the return to the baseline stable stage has not yet been completed, the last effective water body index vector in the current water body index evolution process is read in reverse order of the collection time, no later than the end time of the current risk prediction interval, and its position is determined as the end of the process position. If there is no effective water body index vector in the interval or the end of the process position is earlier than the start of the process position, no corresponding record is generated.

[0047] The stage markers, deviation directions, and deviation magnitudes of the current water body index evolution process from the starting point to the initial process position and from the starting point to the ending process position are extracted respectively, generating the starting process prefix features and the ending process prefix features, which are as follows: Extract all water body index vectors from the starting point of the current water body index evolution process to the corresponding process position according to the acquisition sequence. Subtract the mean value of the corresponding index used in the current water body index evolution process from the value of each water body index, divide it by the effective range width of the water body index, and form a single-step prefix feature with the effective marker, deviation direction, deviation magnitude, stage marker and adjacent acquisition time interval. When the water body index is invalid, the corresponding normalized index value, deviation magnitude and deviation direction are recorded as 0, the adjacent acquisition time interval at the process starting point is recorded as 0, the features at the sequence completion position are recorded as 0 and excluded by the sequence mask.

[0048] By associating the current water body index evolution process, overlapping intervals, starting process location, ending process location, starting process prefix features, and ending process prefix features, a current water body process record set is generated, which is specifically as follows: Record the process identifier, start and end times of overlapping intervals, start and end process positions, start and end process prefix features, sequence mask, and data validity ratio. Calculate the ratio of the number of validly labeled water body indicator values ​​in the start and end process prefix features to the total number of water body indicator values ​​to be collected. Use the smaller of the two ratios as the data validity ratio for the current water body process record. The total number of water body indicator values ​​to be collected for each of the start and end process prefix features is represented by the number of water body indicator vectors and the number of water body indicators contained in the corresponding process prefix, respectively. The product of the data and the data ratio is used to determine which records are retained in the current water process record set, and which are moved to the insufficient data record set.

[0049] Specifically, the pre-trained cumulative risk change prediction model includes a time-series encoding layer, a process location encoding layer, a feature fusion layer, and a risk change regression layer; The temporal coding layer uses gated recurrent units with shared parameters to process the prefix features of the start process and the prefix features of the end process respectively, and ignores the padding position according to the sequence mask; The process location encoding layer consists of a first fully connected layer, a rectified linear activation layer, and a second fully connected layer. The inputs are the stage position, the position within the stage, the segment number, and the cumulative duration normalized by a fixed monitoring period. The feature fusion layer concatenates the temporal coding results with the process position coding, and then generates fused features through a fully connected layer; The risk change regression layer consists of a linear neuron, which outputs the cumulative risk change from the starting point of the current water body index evolution process to the corresponding process position.

[0050] Pre-training data is collected according to seedling ponds and seedling batches. Historical seedling risk values ​​for each historical assessment time are generated using the same calculation method as the current seedling risk benchmark value. The same state segment identification, process connection, process position calibration, and prefix feature extraction are performed on the data between adjacent historical assessment times. The actual risk change in the historical interval is obtained by subtracting the previous historical seedling risk value from the subsequent historical seedling risk value. Historical risk intervals that involve pond replacement, pond merging, seedling separation, centralized harvesting, insufficient total number of identified seedlings, and insufficient number of effective seedling trajectories are removed. Historical water process records with a data effective ratio lower than the minimum input ratio of the model are deleted from the remaining historical risk intervals. Historical risk intervals without effective water process records after deletion are not included in the training.

[0051] Historical data is divided into training and validation sets according to seedling batches. Data from the same seedling batch are not included in the training and validation sets simultaneously. The cumulative risk change prediction model is trained according to the following formula: ; In the formula, For training loss, For model parameters, This refers to the number of historical risk intervals that include valid records of water body processes. For the first A collection of valid water body process records within a historical risk range. and The process is as follows The prefix features of the ending process and the prefix features of the starting process, and These are the end point and the start point of the process, respectively. This is a model for predicting cumulative risk changes. and These represent the historical seedling risk values ​​at the end and start times of the historical risk interval, respectively. The number of water body index evolution processes that achieve the minimum input ratio for the model by ensuring that the effective proportion of the data in the training set is reached. and The first The process prefix characteristics and process location at the starting point of the evolution process of each water body index. The process starting point constraint coefficient, These are the constraint coefficients for the model parameters, used during training. and When executing, the risk change regression layer does not have non-negative output constraints.

[0052] In one implementation, the temporal coding layer uses two gated recurrent units, each containing 64 hidden units, the process position coding layer outputs a 16-dimensional position code, the training batch size is 32, the initial learning rate is 0.001, the adaptive moment estimation optimization algorithm is used to update the model parameters, and training stops when the validation set loss does not decrease for 20 consecutive rounds. After training is completed, the model structure parameters and model weights are fixed.

[0053] Specifically, the steps for generating the current interval risk change are as follows: The process interval risk change is obtained by subtracting the current cumulative risk change from the current cumulative risk change recorded at the end of the current water body process. Specifically: Input the prefix features and position of the starting process, the prefix features and position of the ending process, and the position of the ending process into the same cumulative risk change prediction model. Subtract the current cumulative risk change of the starting process from the current cumulative risk change of the ending process to obtain the risk change of the process interval. A result greater than 0 indicates that the risk has increased, less than 0 indicates that the risk has decreased, and equal to 0 indicates that the model has not identified any risk change.

[0054] The risk change data for each current water body process is collected over the corresponding time interval to generate the current time interval risk change data, which is as follows: Sum the risk changes of each current water body process record interval. If the current water body process record set is empty, the risk change of the current interval is recorded as 0. If the data is insufficient and the record set is not empty, the corresponding process identifier is retained in the risk warning result.

[0055] Specifically, the steps for generating shrimp seedling risk warning results are as follows: The initial predicted risk value is generated by adding the current seedling risk benchmark value to the current interval risk change value, and then the boundary is corrected according to the preset effective range of the risk value. Specifically: When the calculation result is less than 0, the predicted risk value is recorded as 0; when the calculation result is greater than 1, the predicted risk value is recorded as 1; and the remaining calculation results remain unchanged.

[0056] The revised predicted risk value is matched with the preset risk level range to determine the current risk level, specifically as follows: The risk level boundaries are determined based on the historical distribution of seedling risk values ​​and the seedling farm disposal rules. Each risk level interval is continuous and does not overlap. When the predicted risk value is equal to the common boundary of adjacent risk levels, it is classified into the higher risk level. The risk level boundaries are fixed before the model is deployed.

[0057] Determine the maximum deviation of each water body indicator in the current water body process record within the corresponding overlapping interval. Identify water body indicators whose maximum deviation reaches the abnormal indicator threshold as abnormal water body indicators, and output an early warning result containing the current risk level, the abnormal water body indicator, and the corresponding water body indicator evolution process identifier. Specifically: Abnormal indicator thresholds are set separately for each water body indicator. The early warning results include the current risk level, predicted risk value, name of abnormal water body indicator, direction of deviation, maximum deviation, time when the abnormal indicator threshold is first reached, process identifier, current stage of the process, and insufficient data marker. The process identifier consists of the seedling pond identifier, process start time, and process number.

[0058] In this embodiment, the sensor's effective range, sensor resolution, device allowable transition threshold, and missing duration threshold are determined based on the corresponding sensor's calibration records and stable acquisition records. The lower limit of swimming speed response and the lower limit of bait area aggregation response are determined based on the measurement resolution of the corresponding image detection quantity and the positive feeding response distribution in normal feeding records, and both are positive numbers. The preset acquisition period, fixed monitoring period, baseline observation interval duration, pre-feeding detection time window, post-feeding detection time window, and number of normal feeding records are also included. Number of candidate stable window points Minimum number of valid windows The thresholds for the total number of seedlings, the number of effective seedling trajectories, the minimum number of effective indicators, the minimum number of common effective indicators, the number of stable confirmation points, the length of the development judgment window, the length of the continuity judgment window, the length of the recovery judgment window, and the maximum connection interval are determined based on the seedling management cycle and the duration of historical water body changes.

[0059] The effective trajectory duration, velocity threshold, displacement threshold, mortality displacement threshold, local movement threshold, mortality confirmation duration, abnormal mortality seedling ratio threshold, and abnormal activity seedling ratio threshold are determined based on normal seedling detection records and manually verified seedling status records of the same shrimp species and the same seedling age. Smoothing parameters of water body indicators The range coefficient is determined based on the sensor resolution and the mean of historical stable segments. The minimum range width is determined based on the coverage ratio of effective indicator values ​​in historical stable segments. The index weights are determined based on sensor resolution and the fluctuation range of historical stable segments. , , , , Stable scoring threshold, effective window ratio threshold, stable deviation threshold, development increment threshold, persistent deviation threshold, persistent fluctuation threshold, recovery reduction threshold, connectivity threshold, abnormal indicator threshold, minimum input ratio of the model, and Candidate values ​​are generated from training data and selected based on stage markers, process connectivity, and risk prediction errors verified manually in validation data. The selected parameters are fixed before model deployment, and the same parameter caliber is used for historical training data and current water body index sequences.

[0060] Please see Figure 2 This invention provides a technical solution: a shrimp seedling risk early warning system based on time-series learning of water body indicators, comprising: a water body data acquisition module for acquiring the current water body indicator sequence and the most recent seedling status detection data, generating the current seedling risk benchmark value and determining the assessment time; a water body segment identification module for generating current water body indicator status segments; a water body process generation module for generating the current water body indicator evolution process set; a process location calibration module for calibrating the process location; a process record generation module for generating the current water body process record set; a model invocation module for generating the current initial cumulative risk change and the current final cumulative risk change; an interval risk change analysis module for generating the process interval risk change and the current interval risk change; and a risk early warning module for generating predicted risk values ​​and outputting early warning results.

[0061] Each module is executed by the edge server of the seedling farm. The relevant data is stored according to the seedling pond identifier, seedling batch identifier and collection time. The cumulative risk change prediction model is trained and the model parameters are fixed before deployment. During the seedling process, water body index processing, model calling and risk warning are performed.

[0062] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0063] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for early warning of shrimp breeding risks based on time-series learning of water body indicators, characterized in that, Includes the following steps: The current water body indicator sequence and the latest seedling status monitoring data are obtained. The seedling status monitoring data includes the proportion of dead seedlings, the proportion of abnormally active seedlings, and the proportion of decreased feeding response. The current seedling risk benchmark value is generated by weighting according to preset weights, and the time of completion of the monitoring is determined as the assessment time. Based on the changing trend of water body indicators relative to the benchmark range, the current water body indicator state segments are identified according to a fixed monitoring cycle, and the current water body indicator state segments with adjacent time are connected according to the succession relationship of the benchmark stable stage, deviation development stage, deviation persistence stage and recovery stage to generate the current water body indicator evolution process set. The process position of the evolution of each current water body indicator is marked, and a current water body process record set is generated based on the overlap between the evolution of each current water body indicator and the current risk prediction interval formed from the assessment time to the current time. The pre-trained cumulative risk change prediction model is invoked to generate the current starting cumulative risk change and the current ending cumulative risk change based on the starting process prefix features, ending process prefix features, starting process position, and ending process position of each current water body process record, and to determine the risk change of the process interval. The system aggregates risk changes across all process intervals to generate the current interval risk change. Based on the current seedling risk baseline and the current interval risk change, it generates a predicted risk value and outputs the shrimp seedling risk warning result.

2. The shrimp seedling risk early warning method based on time-series learning of water body indicators according to claim 1, characterized in that, The steps for identifying the current water body indicator status segment are as follows: Read the water body indicator vector from the current water body indicator sequence according to the collection time; Candidate stable windows are established in the current water body index sequence. The index stability score is generated based on the ratio obtained by dividing the standard deviation of each water body index by the sum of its mean absolute value and the preset smoothing parameter, as well as the flip ratio of the change direction of adjacent values. The water body index vectors corresponding to the candidate stable windows that continuously meet the index stability conditions are collected into index benchmark segments, and the index benchmark range is generated based on the mean and standard deviation of each water body index. The deviation range is determined by the ratio of the extent to which each water body indicator exceeds the benchmark range to the width of the corresponding range, and the deviation direction is determined based on the position of the upper and lower boundaries of the relative range of the water body indicator. Based on the changing trend of the deviation magnitude, the baseline stabilization stage, the deviation development stage, the deviation persistence stage, and the recovery stage are marked. Within each fixed monitoring cycle, water body indicator vectors with continuous time and the same stage markings are collected to generate the current water body indicator status segment.

3. The shrimp seedling risk early warning method based on time-series learning of water body indicators according to claim 1, characterized in that, The steps for generating the current water body index evolution process set are as follows: Read the stage markers and end water body index vectors of the current water body index state segments that are adjacent in time; The time-adjacent segments that are sequentially connected by stage markers are identified as candidate links for stage connection, and the segments with the same stage markers on both sides of the boundary of adjacent fixed monitoring cycles are identified as candidate links for cross-cycle continuation. The comprehensive connection distance is generated by the ratio of the absolute difference of the values ​​of the common effective water body indicators at both ends of the candidate connection to the width of the corresponding indicator benchmark range. Candidate connections whose comprehensive connection distance is not higher than the connection threshold and whose stage markers conform to the corresponding stage relationships are determined as effective connections. The current water body index state segments are collected along the effective connection. When the water body index after the end of the recovery phase is marked as the baseline stable phase, the end of the recovery phase is taken as the endpoint of the current water body index evolution process. The connection path that has not yet returned to the baseline stable stage by the end of the current water body indicator sequence is identified as the incomplete current water body indicator evolution process. The current water body indicator evolution process is not established if it only contains the state segment of the baseline stable stage.

4. The shrimp seedling risk early warning method based on time-series learning of water body indicators according to claim 1, characterized in that, The calibration steps for the process position are as follows: Set the stage position for the current water body indicator status segment according to the stage type; The position within a stage is determined based on the positional relationship between the water body index vector acquisition time and the start and end times of the corresponding current water body index state segment. By associating the stage position and the position within the stage, the corresponding water body index vector is generated to represent the process position of the current water body index evolution process.

5. The shrimp seedling risk early warning method based on water body index time series learning according to claim 1, characterized in that, The steps for generating the current water body process record set are as follows: Determine the overlap between the current water body index evolution process and the current risk prediction interval, and determine the process positions corresponding to the two ends of the overlap interval as the start process position and the end process position, respectively; Extract the stage markers from the starting point to the initial process position and from the starting point to the ending process position of the current water body index evolution process, as well as the deviation direction and deviation magnitude of the water body index, and generate the starting process prefix features and the ending process prefix features respectively; By associating the current water body index evolution process, overlapping intervals, starting process position, ending process position, starting process prefix features, and ending process prefix features, a current water body process record set is generated.

6. The shrimp seedling risk early warning method based on time-series learning of water body indicators according to claim 5, characterized in that, The steps for determining the start and end positions of the process are as follows: When the starting point of the current water body index evolution process is earlier than the starting time of the current risk prediction interval, the position of the first valid water body index vector that is no earlier than the starting time of the current risk prediction interval is determined as the starting process position. When the starting point of the current water body index evolution process is within the current risk prediction interval, its starting point is determined as the starting process position. When the current water body index evolution process completes recovery and returns to the baseline stable stage within the current risk prediction interval, the position of the last effective water body index vector in the corresponding recovery stage is determined as the end of the process position. If the recovery has not yet been completed and the return to the baseline stable stage has not yet been completed, the position of the last effective water body index vector no later than the end of the current risk prediction interval is determined as the end of the process position.

7. The shrimp seedling risk early warning method based on time-series learning of water body indicators according to claim 1, characterized in that, The pre-trained cumulative risk change prediction model includes a time-series encoding layer, a process location encoding layer, a feature fusion layer, and a risk change regression layer; The temporal coding layer uses gated recurrent units to process the prefix features of the start process and the prefix features of the end process respectively. The process position coding layer processes the start process position and the end process position respectively. The feature fusion layer fuses the corresponding coding results. The risk change regression layer uses linear output to generate the current start cumulative risk change and the current end cumulative risk change.

8. The shrimp seedling risk early warning method based on time-series learning of water body indicators according to claim 1, characterized in that, The steps for generating the current interval risk change are as follows: Subtracting the current start of the cumulative risk change from the current end of the current water body process record, we obtain the risk change over the process interval. Collect the risk changes of each current water body process record for the process interval, and generate the risk change of the current interval.

9. The shrimp seedling risk early warning method based on time-series learning of water body indicators according to claim 5, characterized in that, The steps for generating shrimp seedling risk warning results are as follows: The current seedling risk benchmark value is added to the current interval risk change to generate an initial predicted risk value, and the boundary is corrected according to the preset effective range of the risk value; The revised predicted risk value is matched with the preset risk level range to determine the current risk level; Determine the maximum deviation of each water body indicator in the current water body process record within the corresponding overlapping interval, identify the water body indicator whose maximum deviation reaches the abnormal indicator threshold as an abnormal water body indicator, and output an early warning result that includes the current risk level, the abnormal water body indicator, and the corresponding water body indicator evolution process identifier.

10. A shrimp seedling risk early warning system based on time-series learning of water body indicators, employing the shrimp seedling risk early warning method based on time-series learning of water body indicators as described in any one of claims 1-9, characterized in that, include: The water body data acquisition module is used to acquire the current water body indicator sequence and the latest seedling status detection data, generate the current seedling risk benchmark value and determine the assessment time; The water body segment recognition module is used to generate segments of the current water body indicator status; The water process generation module is used to generate a set of current water body index evolution processes; Process position calibration module, used to calibrate process position; The process record generation module is used to generate the current water body process record set; The model call module is used to generate the current initial cumulative risk change and the current final cumulative risk change; The interval risk change analysis module is used to generate the process interval risk change and the current interval risk change. The risk warning module is used to generate predicted risk values ​​and output warning results.

Images