A method for passive tracking and fixed-point monitoring of rare fish in a protected area

Abstract

The present application relates to the technical field of aquatic organism monitoring, and discloses a rare fish passive tracking and fixed-point monitoring method for a protected area, which comprises the following steps: obtaining historical scheduling records and original environmental time series of each fixed-point monitoring station; calculating pivotal scheduling response signal time series of each station by using a one-dimensional hydrodynamic propagation delay model; separating residual environmental time series by means of hour-by-hour difference value operation; extracting residual environmental feature vectors and scheduling disturbance statistical feature vectors; training a two-channel gradient boosting decision tree ensemble model; and outputting cluster activity start probability in real time and generating a fixed-point monitoring early warning signal. The present application explicitly separates artificial disturbance from original environmental data to obtain residual environmental time series, eliminates the condition of false early warning caused by scheduling events from the data source, and makes the output of the early warning model more accurately reflect the behavior response of rare fish to natural hydrological rhythms.

Description

Technical Field

[0001] This invention relates to the field of aquatic organism monitoring technology, specifically to a method for passive tracking and fixed-point monitoring of rare fish species in protected areas. Background Technology

[0002] The schooling and migratory behavior of rare fish species in the protected area's river sections is closely related to natural hydrological rhythms. Existing fixed-point monitoring methods typically use the raw water temperature, water level, and flow velocity time series collected from each station as input features for the early warning model, and use event data detected by acoustic receivers as supervisory signals for model training.

[0003] However, when there are cascade hydropower projects upstream of the protected area, the sudden changes in water temperature, water level jumps, and pulsed flow rates caused by the project's release and scheduling will be superimposed on the original environmental time series, becoming intertwined with the gradual trends of natural hydrological changes. Existing methods do not explicitly separate artificial scheduling disturbance signals from natural hydrological signals, causing monitoring models to misinterpret scheduling events as precursors to the migration of rare fish populations, resulting in high-frequency false alarms. Furthermore, the scheduling patterns of these projects vary annually due to power generation and flood control demands, leading to unstable temporal distribution of artificial disturbances. This significantly reduces the generalization ability of existing monitoring models in cross-year applications, making it difficult to support the long-term tracking and monitoring needs of rare fish species. Summary of the Invention

[0004] This invention provides a method for passive tracking and fixed-point monitoring of rare fish species in protected areas, solving the technical problem in related technologies that cannot effectively isolate the interference of water conservancy hub scheduling disturbances on environmental signals, resulting in low accuracy of early warning of rare fish migration behavior.

[0005] This invention discloses a method for passive tracking and fixed-point monitoring of rare fish species in protected areas, comprising: acquiring upstream water conservancy hub scheduling records, environmental time series data of each fixed-point monitoring station, and acoustic detection event data of marked individuals; calculating the hub scheduling response signal of each station based on the scheduling records using a one-dimensional hydrodynamic propagation delay model, and performing time-by-time difference calculation between the environmental time series and the response signal to obtain a residual environmental time series; marking historical cluster migration initiation events based on the displacement direction and displacement rate of marked individuals in the detection event data; extracting residual environmental feature vectors based on the residual environmental time series, and extracting scheduling disturbance statistical feature vectors based on the response signals; training a dual-channel gradient boosting decision tree ensemble model with the residual environmental feature vectors as the main channel input and the scheduling disturbance statistical feature vectors as the auxiliary channel input, and outputting an early warning model; inputting real-time data into the early warning model after residual separation and feature extraction, and generating an early warning signal when the probability value of cluster activity initiation exceeds a preset threshold for several consecutive days.

[0006] Furthermore, the inputs of the one-dimensional hydrodynamic propagation delay model are the time of change of the discharge flow of the hub, the magnitude of the change, the cross-sectional parameters of the river channel, and the distance between each station and the hub. The outputs are the arrival time, peak amplitude, attenuation time of the water level response waveform of each station, and the corresponding parameters of the water temperature disturbance waveform.

[0007] Furthermore, when multiple water conservancy projects release water in a coordinated manner, the one-dimensional hydrodynamic propagation delay model calculates the response waveforms of each project independently and then superimposes them at the same station. When the response waveforms of multiple projects overlap in time, the time-by-time linear superposition value of each waveform is taken as the synthetic scheduling response signal of that station.

[0008] Furthermore, the marker historical cluster migration initiation event includes: sorting the acoustic detection event data of each marker individual by time, calculating the spatial displacement direction and displacement rate between adjacent fixed-point monitoring stations; when the displacement direction continuously points to the functional area for no less than a preset number of days and the displacement rate exceeds the daily activity rate threshold, the corresponding cluster migration initiation time point is marked; wherein, the daily activity rate threshold is the upper quantile value of the statistical distribution of displacement rates of each marker individual during non-migration periods.

[0009] Furthermore, the extraction of the residual environmental feature vector includes: for each event in the historical cluster migration initiation event, extracting the residual daily average water temperature sequence, residual daily water temperature change rate sequence, residual cumulative water level change, residual flow velocity mean, and dissolved oxygen mean within a set number of days before the initiation time point, and encoding each component into the residual environmental feature vector after Z-score standardization.

[0010] Furthermore, the extraction of the scheduling disturbance statistical feature vector includes: extracting the scheduling disturbance frequency, maximum single disturbance amplitude, and cumulative disturbance energy within the same window based on the hub scheduling response signal, and encoding each component into the scheduling disturbance statistical feature vector after Z-score standardization; wherein, the cumulative disturbance energy is the total amount obtained by accumulating the square values ​​of the water level response waveform amplitude at each time point within a set number of days according to the time step.

[0011] Furthermore, the dual-channel gradient boosting decision tree ensemble model includes a main channel decision tree group, an auxiliary channel decision tree group, and a weighted fusion unit. The main channel decision tree group takes the residual environmental feature vector as input, the start-up labels of historical start-up events as supervision signals, and the logarithmic loss function as the optimization objective for gradient boosting training, outputting the main channel start-up probability prediction value. The auxiliary channel decision tree group takes the scheduling perturbation statistical feature vector as input, uses gradient boosting training to learn the adjustment effect of scheduling perturbation intensity on the behavioral response threshold, and outputs the auxiliary channel adjustment coefficient. The weighted fusion unit performs a weighted operation on the main channel start-up probability prediction value and the auxiliary channel adjustment coefficient, outputting the final cluster activity start-up probability prediction value. The main channel decision tree group and the auxiliary channel decision tree group are trained and updated independently, and the two sets of outputs are only merged in the weighted fusion unit.

[0012] Further, the weighted fusion unit multiplies the predicted value of the main channel start probability by the main channel fusion weight, multiplies the adjustment coefficient of the auxiliary channel by the auxiliary channel fusion weight, and sums the two products to obtain the final predicted value of the cluster activity start probability; wherein the sum of the main channel fusion weight and the auxiliary channel fusion weight is one, and the fusion weight is determined by minimizing the logarithmic loss function between the predicted value of the cluster activity start probability and the actual start tag on the validation set.

[0013] Furthermore, it also includes an incremental update step: tracking the acoustic detection event data of the marked individuals in the subsequent period after the warning, statistically analyzing the actual cluster response ratio, and calculating the prediction deviation value between the cluster activity initiation probability value and the actual cluster response ratio; feeding the prediction deviation value and the actual hub scheduling mode type of the corresponding period as new samples back to the training dataset; incrementally training the main channel decision tree group and the auxiliary channel decision tree group according to a set period, with the incremental updates of the two channels being independent of each other, and keeping the parameters of the channel unchanged when the new samples of a certain channel are insufficient.

[0014] This invention discloses a passive tracking and fixed-point monitoring system for rare fish species in a protected area, used to execute the aforementioned method, comprising: a data acquisition module for acquiring upstream water conservancy hub scheduling records, environmental time series data of each fixed-point monitoring station, and acoustic detection event data of marked individuals; a scheduling response calculation module for calculating the hub scheduling response signal of each station based on the scheduling records using a one-dimensional hydrodynamic propagation delay model; a residual separation module for performing time-by-time difference calculation between the environmental time series and the response signal to obtain a residual environmental time series; and an event marking module for marking events based on the marked individuals in the detection event data. The displacement direction and displacement rate of the body mark historical cluster migration initiation events; the feature extraction module is used to extract residual environment feature vectors based on the residual environment time series and extract scheduling disturbance statistical feature vectors based on the response signal; the model training module is used to train a dual-channel gradient boosting decision tree ensemble model with the residual environment feature vector as the main channel input and the scheduling disturbance statistical feature vector as the auxiliary channel input and output an early warning model; the early warning output module is used to input real-time data into the early warning model after residual separation and feature extraction, and generate an early warning signal when the cluster activity initiation probability value exceeds a preset threshold for several consecutive days.

[0015] This invention introduces hub scheduling records and utilizes a one-dimensional hydrodynamic propagation delay model to calculate the time series of hub scheduling response signals at each station. This explicitly separates artificial disturbances from the raw environmental data, obtaining a residual environmental time series. The residual environmental time series no longer contains false precursor features caused by scheduling, such as sudden changes in water temperature and water level jumps. This eliminates the conditions for false warnings caused by scheduling events at the data source, enabling the warning model's output to more accurately reflect the behavioral responses of rare fish species to natural hydrological rhythms.

[0016] The dual-channel gradient boosting decision tree ensemble model learns the correlation between fish behavior triggered by natural rhythms and the moderating effect of scheduling disturbances on behavior thresholds in independent channels. When the hub scheduling strategy changes in different years, only the parameters of the auxiliary channel decision tree group need to be updated, while the natural rhythm correlations learned by the main channel decision tree group remain stable, ensuring that the fixed-point monitoring and early warning capability does not become completely ineffective due to changes in scheduling strategy during cross-year operation.

[0017] The incremental update step periodically adds new samples, enabling the dual-channel gradient boosting decision tree ensemble model to continuously track the long-term drift of natural rhythms and changes in scheduling strategies. The incremental updates of the two channels are independent of each other. When there are insufficient new samples in a certain channel, the parameters of that channel remain unchanged, avoiding improper modification of model parameters due to sample imbalance, thereby supporting the long-term tracking and monitoring needs of rare fish species. Attached Figure Description

[0018] Figure 1This is a flowchart of the passive tracking and fixed-point monitoring method for rare fish species in protected areas provided in this embodiment of the invention; Figure 2 This is a schematic diagram illustrating the changes in the original environmental parameters of station 3 during a typical time period, as provided in this embodiment of the invention. Figure 3 This is a schematic diagram comparing the superposition effect of hub scheduling response signals provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the residual sequence separation effect at station 3 provided in this embodiment of the invention: water temperature comparison; Figure 5 This is a schematic diagram of the residual sequence separation effect at station 3 provided in this embodiment of the invention: water level comparison; Figure 6 This is a schematic diagram comparing the average displacement rate of a typical marked individual five days before the start of its migration with a threshold, provided in an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the comparison of output probabilities of the dual-channel model validation set provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the continuous warning probability change from day 074 to 076 of year B, provided in an embodiment of the present invention. Figure 9 This is a schematic diagram of the distribution of the Z-score standardized value of the F007 dual-channel feature vector provided in an embodiment of the present invention. Detailed Implementation

[0019] The protected river section is affected by the upstream cascade hydropower project's scheduling. The original environmental time series data collected by various fixed monitoring stations, including water temperature, water level, and flow velocity, are superimposed with artificial disturbance signals generated by the release scheduling. Sudden changes in water temperature, abrupt changes in water level, pulsed flow variations, and gradual trends in natural hydrology are mixed together, causing the monitoring system to misinterpret scheduling events as precursors to rare fish colonies or migrations, resulting in high-frequency false alarms. Simultaneously, the scheduling patterns of the hydropower projects change annually due to power generation and flood control needs, leading to unstable temporal distribution of artificial disturbances and a significant decrease in the generalization ability of the monitoring model in cross-year applications.

[0020] In this embodiment, an acoustic receiver network is deployed within the protected area, and each fixed monitoring station is equipped with sensors for water temperature, water level, and flow velocity. Each acoustic receiver is used to receive coded pulse signals emitted by rare fish individuals carrying acoustic tags, recording the detection time and the corresponding receiver number. Each fixed monitoring station synchronously collects environmental parameters at fixed time intervals and stores them as a continuous time series. The scheduling and management system of each upstream water conservancy project records the discharge flow, gate opening, and the time and duration of their changes.

[0021] Historical dispatch records of each water conservancy hub upstream of the protected area were acquired. These records include the time series of discharge flows, the timing and duration of gate opening changes, and other relevant data. Simultaneously, continuous time series data of water temperature, water level, and flow velocity were acquired from each fixed monitoring station within the protected area. Acoustic receiver detection event data for all tagged rare fish individuals were also acquired. Each detection event data point includes the tagged individual number, detection time, and detection station number.

[0022] Based on the timing and magnitude of discharge changes recorded in the hub's scheduling records, a one-dimensional hydrodynamic propagation delay model is used to calculate the water level response waveforms and water temperature disturbance waveforms generated at downstream fixed monitoring stations for each scheduling event, generating a time series of hub scheduling response signals for each station. The inputs of the one-dimensional hydrodynamic propagation delay model are the timing and magnitude of the hub discharge changes, river cross-sectional parameters, and the river distance between each station and the hub. The outputs are the arrival time, peak amplitude, attenuation duration of the water level response waveform, and the corresponding parameters of the water temperature disturbance waveform for each station.

[0023] It should be noted that the above one-dimensional hydrodynamic propagation delay model is for the case of joint release of multiple hubs. The response waveforms of each hub are calculated independently and then superimposed at the same station. When the response waveforms of multiple hubs overlap in time, the time-by-time linear superposition value of each waveform is taken as the synthetic scheduling response signal of that station.

[0024] The residual water temperature time series is obtained by performing time-by-time difference calculations between the original water temperature time series from each fixed monitoring station and the corresponding water temperature disturbance waveform from the key project's scheduling. Similarly, the residual water level time series is obtained by performing time-by-time difference calculations between the original water level time series and the corresponding water level response waveform from the key project's scheduling. Finally, the residual flow velocity time series is obtained by performing time-by-time difference calculations between the original flow velocity time series and the corresponding flow velocity response component from the key project's scheduling. These residual series reflect the gradual environmental change trend driven by the natural hydrological rhythms of the protected river section.

[0025] The acoustic receiver-detected event data for each tagged rare fish individual were sorted chronologically, and the spatial displacement direction and velocity between adjacent fixed-point monitoring stations were calculated. Based on the condition that the displacement direction consistently points towards the functional area and the displacement velocity exceeds the daily activity rate threshold, the initiation time point of the swarm migration was marked. The initiation time points of the swarm migration of all tagged individuals were summarized to form a historical initiation event set.

[0026] It should be noted that the aforementioned daily activity rate threshold refers to the upper bound value determined based on the statistical distribution of the displacement rate of each marked individual during non-migratory periods. Specifically, the daily displacement rate of each marked individual during non-migratory periods in historical records is statistically analyzed, and the upper quantile value of its distribution is taken as the daily activity rate threshold. The criterion for determining whether the displacement direction continuously points towards the functional area is that the displacement direction points towards the location of the functional area for no less than a preset number of consecutive days.

[0027] Step 5 may specifically include: Step 501: For each event in the historical start-up event set, extract the residual daily average water temperature sequence, residual daily water temperature change rate sequence, residual cumulative water level change, residual flow velocity mean, and dissolved oxygen mean within a set number of days before the cluster migration start time point based on the residual environment time series. After performing Z-score standardization on each of the above components, encode them into a residual environment feature vector.

[0028] Step 502: Based on the time series of the hub scheduling response signal, extract the frequency of scheduling disturbances, the maximum single disturbance amplitude, and the cumulative disturbance energy within the same window. After performing Z-score standardization on each of the above components, encode them into a scheduling disturbance statistical feature vector.

[0029] It should be noted that the aforementioned cumulative disturbance energy refers to the total amount obtained by summing the squares of the water level response waveform amplitudes at each moment within a set number of days window, according to the time step, reflecting the overall intensity of the impact of scheduling disturbances on the hydrological status of the station within this window.

[0030] The aforementioned residual environmental feature vector and scheduling disturbance statistical feature vector are two independent feature vectors, which respectively describe the gradual change of natural hydrology and the intensity of artificial scheduling disturbance.

[0031] Using residual environmental feature vectors as the main input channel and scheduling disturbance statistical feature vectors as auxiliary adjustment channels, a dual-channel gradient boosting decision tree ensemble model is trained, and a scheduling perception fixed-point monitoring and early warning model is output.

[0032] It should be noted that the dual-channel gradient boosting decision tree ensemble model comprises three components: a main channel decision tree group, an auxiliary channel decision tree group, and a weighted fusion unit. The main channel decision tree group takes the residual environmental feature vector as input, is trained using the gradient boosting algorithm, uses the activation labels from the historical activation event set as supervision signals, and uses the logarithmic loss function as the optimization objective to iteratively fit the correlation between natural hydrological rhythms and the triggering of rare fish colony activities, outputting the main channel activation probability prediction value. The auxiliary channel decision tree group takes the scheduling perturbation statistical feature vector as input, is also trained using the gradient boosting algorithm, learns the moderating effect of scheduling perturbation intensity on the fish behavioral response threshold, and outputs the auxiliary channel modulation coefficient. The weighted fusion unit receives the main channel activation probability prediction value and the auxiliary channel modulation coefficient, performs a weighted calculation according to preset fusion weights, and generates the final colony activity activation probability prediction value.

[0033] It should be noted that the weighting calculation method of the above weighted fusion unit is as follows: Let the predicted probability of the main channel starting be... The auxiliary channel adjustment coefficient is The main channel fusion weight is The auxiliary channel fusion weight is The final predicted cluster activity initiation probability value for: ; in . The predicted initiation probability of the main channel output by the main channel decision tree group, with a value range of [value missing]. ; The auxiliary channel adjustment coefficient is the output of the auxiliary channel decision tree group, and its value range is [value range missing]. This reflects the amount by which the current scheduling disturbance intensity corrects the behavior threshold. Main channel weighting For auxiliary channel fusion weights. Fusion weights and By minimizing on the validation set The logarithmic loss function between the actual start tag and the actual start tag is determined.

[0034] The main channel decision tree group and the auxiliary channel decision tree group are trained and updated independently, and the outputs of the two groups are only combined in the weighted fusion unit.

[0035] In real-time passive tracking operation, the latest hub scheduling records and raw environmental data of each fixed-point monitoring station are acquired daily. Hub scheduling response signal calculation and residual separation are performed to obtain a real-time residual environmental time series. Based on the real-time residual environmental time series and the hub scheduling response signal time series, the daily residual environmental feature vector and scheduling disturbance statistical feature vector are extracted, and each component is standardized according to the Z-score standardization parameters in steps 501 and 502, respectively. The residual environmental feature vector and scheduling disturbance statistical feature vector are input into the main channel and auxiliary channel of the scheduling perception fixed-point monitoring and early warning model, respectively, and the probability value of rare fish cluster activity initiation is output. When the probability value of rare fish cluster activity initiation exceeds a preset threshold for several consecutive days, a fixed-point monitoring and early warning signal is generated, simultaneously indicating the dominant natural environmental factor and the intensity level of hub scheduling disturbance in the current warning.

[0036] In this embodiment of the application, in order to improve the adaptability of the scheduling perception fixed-point monitoring and early warning model to changes in scheduling strategies during cross-year operation, the following steps are also included in step 7:

[0037] In the period following the early warning, acoustic receivers tracking and marking rare fish individuals detected event data, statistically analyzed the actual cluster response ratio, and calculated the prediction deviation between the probability of rare fish cluster activity initiation and the actual cluster response ratio. Simultaneously, the actual hub scheduling mode type during this period was recorded. The prediction deviation and hub scheduling mode type were used as additional labels for new samples and fed back into the training dataset. The dual-channel gradient boosting decision tree ensemble model was incrementally updated at set intervals. The main channel decision tree group was incrementally trained based on newly added residual environmental feature samples to adapt to long-term drift of natural rhythms, while the auxiliary channel decision tree group was incrementally trained based on newly added scheduling perturbation statistical feature samples to adapt to changes in scheduling strategies. The incremental updates of the two channels were independent of each other; when the number of new samples for a certain channel was insufficient, the parameters of that channel remained unchanged.

[0038] This implementation introduces hub scheduling records and utilizes a one-dimensional hydrodynamic propagation delay model to calculate the time series of hub scheduling response signals at each station. This explicitly separates artificial disturbances from the original environmental data, enabling the residual environmental feature vector to reflect the actual natural hydrological change trends perceived by rare fish species. Since the residual environmental time series no longer contains false precursor features caused by scheduling, such as sudden changes in water temperature and water level jumps, the cluster activity prediction based on the residual environmental feature vector eliminates the conditions for false early warnings caused by scheduling events from the data source. The dual-channel gradient boosting decision tree ensemble model learns the correlation between fish behavior triggered by natural rhythms and the moderating effect of scheduling disturbances on behavior thresholds in independent channels. When the hub scheduling strategy changes in different years, only the parameters of the auxiliary channel decision tree group need to be updated, while the natural rhythm correlation learned by the main channel decision tree group remains stable. This ensures that the passive tracking and fixed-point monitoring system maintains stable early warning capabilities during long-term operation across years. The incremental update step further supplements the data with new samples periodically, enabling the dual-channel gradient boosting decision tree ensemble model to continuously track natural rhythm drift and scheduling strategy changes, supporting the long-term tracking and monitoring needs of rare fish species.

[0039] A section of a protected river is located downstream of two cascade hydropower projects, Hub A and Hub B. Seven fixed monitoring stations (stations 1 to 7) and an acoustic receiver network are deployed within this section. There are currently 32 rare fish individuals (numbered F001 to F032) carrying acoustic tags. During year A, sudden changes in water temperature and water level jumps caused by the combined release of water from the cascade hydropower projects led to frequent false alarms. Starting in year B, historical data from year A was retrospectively processed, and the early warning model was reconstructed. The following section uses station 3 as the core station to illustrate the complete data flow process.

[0040] The system extracts the historical scheduling records of hub A and hub B in year A from the hub scheduling management system, synchronously reads the original time series of water temperature, water level and flow velocity at station 3, and extracts the acoustic receiver detection event records of individuals marked with F007, F012, F019 at station 3 and adjacent stations.

[0041] Table 1 Hub Dispatch Records and Original Environmental Data of Station 3

[0042] Taking the example of the discharge flow of hub A increasing from 1240 m³ / s to 2850 m³ / s at 14:00 on day 83. The distance between hub A and station 3 is 187 km, and the distance between hub B and station 3 is 94 km. A one-dimensional hydrodynamic propagation delay model is used to calculate the arrival time and peak amplitude of the water level response waveforms of the two hubs based on the distance from each hub to station 3, cross-sectional parameters, and the magnitude of the flow change. These waveforms are then linearly superimposed at station 3 time-by-time to obtain the composite dispatch response signal.

[0043] Table 2 Calculation Results of Dispatch Response Signals for Station 3 Hub

[0044] The original water temperature sequence at station 3 was subtracted hourly from the water temperature disturbance waveform caused by the hub's scheduling; the original water level sequence was subtracted hourly from the water level response waveform caused by the hub's scheduling; and the original flow velocity sequence was subtracted hourly from the flow velocity response component caused by the hub's scheduling, thus obtaining the residual water temperature, residual water level, and residual flow velocity time series, respectively. At 20:00 on day 083, the original water temperature was 14.6℃, the scheduled water temperature disturbance was -1.74℃, and the residual water temperature was 16.34℃, thus reconstructing the natural thermal state of the river section at that moment.

[0045] Table 3. Separation results of residual sequences at station 3 (typical moments from day 083 to day 084)

[0046] The detection events of marked individuals such as F007 were sorted chronologically, and the spatial displacement direction and displacement rate between adjacent stations were calculated. The daily displacement rate distribution of each individual during non-migratory periods was statistically analyzed, and the upper quantile value was taken as the daily activity rate threshold (the threshold for the area near station 3 was 0.41 km / h). When the displacement direction continuously points upstream to the spawning functional area for more than 5 consecutive days and the displacement rate exceeds the threshold, it is marked as the migration initiation time point. A total of 14 historical initiation events were marked in year A, forming a historical initiation event set.

[0047] Table 4. Results of Identification of Migration Initiation Events in Typical Marked Individuals

[0048] Taking the start-up event of F007 on day 091 as an example, features within the 7-day window (days 084 to 090) prior to the start-up time are extracted. In step 501, the residual daily average water temperature sequence, residual daily water temperature change rate sequence, residual cumulative water level change, residual mean flow velocity, and dissolved oxygen mean are extracted based on the residual environmental time series. Each component is Z-score standardized and encoded as a residual environmental feature vector. In step 502, the frequency of scheduling disturbances, the maximum single disturbance amplitude, and the cumulative disturbance energy (obtained by summing the squares of the water level response waveform amplitudes at each time step according to the time step) within the same window are extracted based on the hub scheduling response signal time series. Each component is Z-score standardized and encoded as a scheduling disturbance statistical feature vector.

[0049] Table 5 Dual-channel feature vector of F007 startup event (window from day 084 to day 090)

[0050] The training set consists of 14 historical initiation events and non-initiation samples from the same period. The main channel decision tree group uses the residual environmental feature vector as input, the initiation label as the supervision signal, and the logarithmic loss function as the optimization objective. Iterative training is performed using a gradient boosting algorithm to fit the correlation between natural hydrological changes and fish schooling activity triggering, outputting the predicted initiation probability value of the main channel. The auxiliary channel decision tree group takes the statistical feature vector of scheduling disturbances as input, learns the moderating effect of scheduling disturbance intensity on behavior thresholds, and outputs the auxiliary channel adjustment coefficient. The weighted fusion unit is calculated according to the formula. ; The fusion weights are determined by minimizing the logarithmic loss function on the validation set, ultimately yielding... .

[0051] Table 6. Output results of the dual-channel model for typical samples in the validation set.

[0052] Sample 04 illustrates the moderating effect of the auxiliary channel: the scheduling disturbance is relatively strong at this moment ( However, the main channel determined that the natural hydrological conditions did not meet the activation threshold. The main channel dominates the final prediction with a high weight, and the final probability is only 0.36. The system did not issue an early warning, thus avoiding false alarms caused by scheduling disturbances.

[0053] On day 76 of year B, the system acquires real-time dispatch records for hub A and hub B, as well as raw environmental data for station 3. It performs dispatch response signal calculation and residual separation, extracts the daily feature vector, standardizes it according to the Z-score parameters used during training in year A, and inputs it into the dispatch perception fixed-point monitoring and early warning model. This data is collected over three consecutive days (days 74 to 76). The values ​​were 0.72, 0.78, and 0.83 respectively, all exceeding the preset threshold of 0.70. The system generated an early warning signal, indicating that the dominant natural environmental factor was a continuous rise in residual water temperature, and the intensity of the hub scheduling disturbance was moderate. Ten days after the warning, the system statistics showed that the actual cluster response ratio of 13 individuals, including F007, F012, and F019, was 0.79, compared to day 76. The prediction bias was 0.04, and the actual scheduling mode type for year B was recorded as "flood control with increased discharge," which was then used as a new sample label and fed back into the training dataset. The auxiliary channel decision tree group was incrementally trained on a quarterly basis, while the main channel parameters remained unchanged, allowing the model to adapt to the changes in the hub scheduling strategy for year B.

[0054] The entire data flow process demonstrates a clear logical chain: Step 1 obtains the original environmental time series and scheduling records, which are then transformed into scheduling response signals for each station by the hydrodynamic model in Step 2. Step 3 then performs time-by-time difference calculations to obtain the residual sequence, eliminating the false precursor features caused by scheduling. Step 4 provides supervision labels for the feature extraction in Step 5 by using the set of historical initiation events marked in the acoustically detected events. The dual-channel feature vector output in Step 5 is directly used as the input for model training in Step 6. Step 7 reuses the standardized parameters from Step 5 for real-time feature extraction, ensuring consistency between training and inference. Step 8's incremental update is driven by the prediction bias value from Step 7, forming a closed-loop adaptive update path for the model.

[0055] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for passive tracking and fixed-point monitoring of rare fish species in protected areas, characterized in that, include: Acquire upstream water conservancy hub scheduling records, environmental time series data from each fixed monitoring station, and acoustic detection event data of labeled individuals; Based on the scheduling record data, the hub scheduling response signal of each station is calculated using a one-dimensional hydrodynamic propagation delay model. The environmental time series and the response signal are then subjected to time-by-time difference calculation to obtain the residual environmental time series. Based on the displacement direction and displacement rate of the marked individuals in the detected event data, historical cluster migration initiation events are marked; Based on the residual environment time series, a residual environment feature vector is extracted; based on the response signal, a scheduling disturbance statistical feature vector is extracted. A dual-channel gradient boosting decision tree ensemble model is trained using the residual environmental feature vector as the main channel input and the scheduling disturbance statistical feature vector as the auxiliary channel input, and an early warning model is output. Real-time data is input into the early warning model after residual separation and feature extraction. When the probability value of cluster activity activation exceeds the preset threshold for several consecutive days, an early warning signal is generated.

2. The method for passive tracking and fixed-point monitoring of rare fish species in protected areas according to claim 1, characterized in that, The inputs to the one-dimensional hydrodynamic propagation delay model are the time and magnitude of the change in the discharge flow of the hub, the cross-sectional parameters of the river channel, and the distance between each station and the hub. The outputs are the arrival time, peak amplitude, attenuation time, and corresponding parameters of the water level response waveform at each station and the water temperature disturbance waveform.

3. The method for passive tracking and fixed-point monitoring of rare fish species in protected areas according to claim 2, characterized in that, When multiple water conservancy projects release water in a coordinated manner, the one-dimensional hydrodynamic propagation delay model calculates the response waveforms of each project independently and then superimposes them at the same station. When the response waveforms of multiple projects overlap in time, the time-by-time linear superposition value of each waveform is taken as the synthetic scheduling response signal of that station.

4. The method for passive tracking and fixed-point monitoring of rare fish species in protected areas according to claim 1, characterized in that, The marked historical cluster migration start event includes: For each marked individual, the acoustic detection event data are sorted by time, and the spatial displacement direction and displacement rate between adjacent fixed-point monitoring stations are calculated. When the displacement direction continuously points to the functional area for no less than a preset number of days and the displacement rate exceeds the daily activity rate threshold, mark the corresponding cluster migration start time point. The daily activity rate threshold is the upper quantile of the statistical distribution of displacement rates of each marked individual during non-migration periods.

5. The method for passive tracking and fixed-point monitoring of rare fish species in protected areas according to claim 1, characterized in that, The extraction of the residual environment feature vector includes: For each event in the historical cluster migration initiation event, extract the residual daily average water temperature sequence, residual daily water temperature change rate sequence, residual cumulative water level change, residual flow velocity mean, and dissolved oxygen mean within a set number of days before the initiation time point. After Z-score standardization processing, each component is encoded into the residual environmental feature vector.

6. The method for passive tracking and fixed-point monitoring of rare fish species in protected areas according to claim 5, characterized in that, The extraction of the statistical feature vector of the scheduling disturbance includes: Based on the hub scheduling response signal, the frequency of scheduling disturbances, the maximum single disturbance amplitude, and the cumulative disturbance energy within the same window are extracted. Each component is then Z-score standardized and encoded into the scheduling disturbance statistical feature vector. The cumulative disturbance energy is the total amount obtained by summing the squares of the amplitudes of the water level response waveforms at each moment within a set number of days window according to the time step.

7. The method for passive tracking and fixed-point monitoring of rare fish species in protected areas according to claim 1, characterized in that, The dual-channel gradient boosting decision tree ensemble model includes a main channel decision tree group, an auxiliary channel decision tree group, and a weighted fusion unit. The main channel decision tree group takes the residual environment feature vector as input, the start-up labels of historical start-up events as supervision signals, and the log loss function as the optimization objective to perform gradient boosting training, and outputs the main channel start-up probability prediction value. The auxiliary channel decision tree group takes the scheduling perturbation statistical feature vector as input, uses gradient boosting training to learn the adjustment effect of scheduling perturbation intensity on behavioral response threshold, and outputs auxiliary channel adjustment coefficient; The weighted fusion unit performs a weighted calculation on the predicted value of the main channel start probability and the adjustment coefficient of the auxiliary channel, and outputs the final predicted value of the cluster activity start probability. The main channel decision tree group and the auxiliary channel decision tree group are trained and updated independently, and the outputs of the two groups are only combined in the weighted fusion unit.

8. The method for passive tracking and fixed-point monitoring of rare fish species in protected areas according to claim 7, characterized in that, The weighted fusion unit multiplies the predicted value of the main channel start probability by the main channel fusion weight, multiplies the auxiliary channel adjustment coefficient by the auxiliary channel fusion weight, and sums the two products to obtain the final predicted value of the cluster activity start probability. The sum of the main channel fusion weight and the auxiliary channel fusion weight is one. The fusion weight is determined by minimizing the logarithmic loss function between the predicted value of the cluster activity initiation probability and the actual initiation label on the validation set.

9. A method for passive tracking and fixed-point monitoring of rare fish species in protected areas according to claim 1, characterized in that, It also includes incremental update steps: In the period following the early warning, the acoustic detection event data of the marked individuals are tracked, the actual cluster response ratio is statistically analyzed, and the prediction deviation between the cluster activity initiation probability value and the actual cluster response ratio is calculated. The predicted deviation value and the actual hub scheduling mode type for the corresponding time period are fed back as new samples into the training dataset; Incremental training is performed on the main channel decision tree group and the auxiliary channel decision tree group at a set period. The incremental updates of the two channels are independent of each other. When there are insufficient new samples in a certain channel, the parameters of that channel remain unchanged.

10. A passive tracking and fixed-point monitoring system for rare fish species in protected areas, used to execute the passive tracking and fixed-point monitoring method for rare fish species in protected areas as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire upstream water conservancy hub scheduling records, environmental time series data of each fixed monitoring station, and acoustic detection event data of labeled individuals; The scheduling response calculation module is used to calculate the hub scheduling response signal of each station based on the scheduling record data using a one-dimensional hydrodynamic propagation delay model. The residual separation module is used to perform time-by-time difference calculation between the environmental time series and the response signal to obtain the residual environmental time series. The event tagging module is used to tag historical cluster migration start events based on the displacement direction and displacement rate of the tagged individuals in the detected event data; The feature extraction module is used to extract residual environment feature vectors based on the residual environment time series and to extract scheduling disturbance statistical feature vectors based on the response signal. The model training module is used to train a dual-channel gradient boosting decision tree ensemble model with the residual environmental feature vector as the main channel input and the scheduling disturbance statistical feature vector as the auxiliary channel input, and output an early warning model. The early warning output module is used to input real-time data into the early warning model after residual separation and feature extraction. When the probability value of cluster activity starts exceeds a preset threshold for several consecutive days, an early warning signal is generated.

Images