An AI-based intelligent monitoring system for water quality in aquaculture

Abstract

The application discloses an AI-based intelligent water quality monitoring system for aquaculture, belonging to the field of machine learning, comprising: a scene modeling and parameter configuration module, a model training and deployment module, a real-time data acquisition and input construction module, a risk identification and sampling scheduling module, a synchronous sampling and quality fidelity module, a monitoring state confirmation and early warning output module, a data backhaul and closed-loop update module, and an abnormal behavior constraint adjustment module. The application achieves the technical effects of complete capture of short-term abnormal processes, accurate representation of multi-parameter synchronous changes, and accurate output of early warning and regulation basis through key risk window identification, rapid key parameter synchronous sampling, high-risk supplementary sampling fidelity, and state feedback closed-loop update, solving the problem of fixed low-frequency sampling, asynchronous sampling, and smooth processing easily covering short-term high-risk processes in the prior art.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to an AI-based intelligent monitoring system for aquaculture water quality. Background Technology

[0002] Since water quality parameters such as temperature, pH, dissolved oxygen, and ammonia nitrogen directly affect the growth and health of aquaculture organisms, current aquaculture production generally requires continuous monitoring of key water quality parameters. With the development of sensor, Internet of Things, and artificial intelligence technologies, AI-based water quality monitoring technology has begun to be applied to various users, including aquaculture enterprises, farm managers, and individual farmers, to achieve continuous perception, early warning of anomalies, and auxiliary management of aquaculture water bodies.

[0003] Existing water quality monitoring systems collect multi-parameter water quality data by deploying sensors such as dissolved oxygen, pH, ammonia nitrogen, turbidity, and water temperature in aquaculture water bodies. The collected data is then aggregated by a microcontroller and uploaded to the platform via a wireless network. The platform then performs data storage, visualization, and basic threshold monitoring. Simultaneously, it combines AI models to analyze time-series water quality data to predict, provide early warnings of anomalies, and assess risks for single key parameters or multiple water quality indicators. Based on the analysis results, it generates and executes coordinated controls such as oxygenation, water exchange, feeding adjustments, or temperature control at the cloud or edge, thereby achieving intelligent monitoring and regulation of aquaculture water quality.

[0004] For example, the Chinese invention patent with announcement number CN117195973B discloses a method and system for predicting aquaculture water quality parameters based on improved PSO, which includes: collecting water quality parameters at different locations and depths in the aquaculture pond and corresponding parameters; training an improved radial basis function (RBF) neural network using training set data; and optimizing the improved RBF neural network model parameters using an improved particle swarm optimization algorithm.

[0005] The above-mentioned technology has at least the following technical problems: In freshwater pond aquaculture, recirculating aquaculture systems (RAS), aquaponics or integrated ecological aquaculture, as well as marine cage and deep-sea aquaculture, aquaculture companies, farm managers, or individual farmers typically rely on continuous sampling data to assess water quality. However, key indicators such as dissolved oxygen and pH exhibit significant diurnal cyclical variations. For instance, dissolved oxygen in ponds often decreases continuously at night and approaches its lowest value around dawn, while gradually recovering during the day due to photosynthesis. pH also fluctuates intraday as carbon dioxide is consumed and accumulated. Therefore, when existing systems employ fixed low-frequency sampling, asynchronous sampling with different parameters, or sampling times that happen to avoid critical risk periods such as around dawn, after feeding, or before and after sudden weather changes, they can easily average and smooth out the originally short-duration but highly variable hypoxia or acid-base abnormality processes. This results in the data recorded on the platform only showing slow, fluctuating apparent changes, making it difficult to accurately reflect the formation process of water risks. For example, if a pond experiences a rapid drop in dissolved oxygen levels in the early morning, but the system only samples every 30 minutes or 1 hour, it often only records the two relatively normal values ​​before and after the initial drop, directly missing the dangerous trough in between. This distorts the basis for subsequent monitoring, early warning, and control, ultimately affecting the accurate judgment of the timing of risks and the intensity of intervention at the aquaculture site. Summary of the Invention

[0006] To address the technical problem in existing technologies where short-term water quality anomalies during critical risk periods are easily masked by fixed and asynchronous sampling, this invention provides an AI-based intelligent water quality monitoring system for aquaculture. The technical solution is as follows: An AI-based intelligent water quality monitoring system for aquaculture is provided, comprising: The scene modeling and parameter configuration module is used to create scene monitoring profiles and determine key parameter sets quickly.

[0007] The model training and deployment module is used to train the critical risk window identification model and deploy the critical risk window identification model on the cloud or edge execution terminal according to network stability, response time requirements, on-site computing power conditions and user execution permissions.

[0008] The real-time data acquisition and input construction module is used to receive real-time monitoring data and construct the model input sequence.

[0009] The risk identification and sampling scheduling module is used to output the stable state, the critical risk precursor state, the critical risk formation state, and to adjust the sampling frequency, perform rapid synchronous sampling of critical parameter groups, start critical parameter supplementation sampling, and maintain the current high-frequency sampling state, and drive the edge execution end to perform sampling level control.

[0010] The synchronous sampling and quality assurance module is used to perform synchronous sampling of rapid key parameter groups under enhanced sampling layer or locked sampling layer, and to perform quality control on the raw water quality data obtained in any sampling round.

[0011] The monitoring status confirmation and early warning output module is used to confirm the current monitoring status and divide the current monitoring status into stable monitoring status, suspicious change monitoring status and critical risk monitoring status, and generate monitoring results based on the current monitoring status.

[0012] The data feedback and closed-loop update module is used to feedback monitoring results and recorded data, and to perform model updates and redeployment when update trigger conditions are met.

[0013] The abnormal behavior constraint adjustment module is used to perform abnormal behavior constraint adjustment on the current monitoring process when the current monitoring point meets the preset abnormal conditions.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention provides an AI-based intelligent water quality monitoring system for aquaculture. By replacing fixed-frequency sampling with key risk window identification, it performs synchronous sampling of key parameters such as dissolved oxygen, pH, and water temperature. Under high-risk conditions, it performs supplementary sampling and anti-smoothing fidelity processing. Simultaneously, it feeds back the data quality status, synchronization completion status, and supplementary sampling status to the model to form a closed-loop update. This allows for more complete capture and characterization of short-term hypoxia processes before and after dawn, rapid fluctuations in local water quality after feeding, and water disturbances before and after sudden weather changes. This enables fidelity monitoring of short-term abnormal processes during key risk periods, accurate reflection of synchronous changes in multiple parameters, and accurate output of early warning timing and intervention basis. It effectively solves the problems in existing technologies where fixed low-frequency sampling, asynchronous sampling of different parameters, and subsequent smoothing processing easily mask short-term high-risk processes, leading to distorted characterization of risk formation processes, biased early warning judgments, and inaccurate on-site control basis.

[0015] 2. This invention uses a key risk window identification model to dynamically identify continuous time-series data, transforming the sampling strategy from fixed-period sampling to adaptive sampling based on risk status. This allows for timely increases in sampling frequency in the early stages of risk formation, preventing the omission of short-duration but highly variable anomalies due to excessively long sampling intervals. Simultaneously, by constructing a data feedback and model closed-loop update mechanism, key risk window data, low-confidence window data, and misidentified and missed identification records are used as update samples in model iteration. This enables the model to continuously optimize its key risk window identification capability during long-term operation, improving the system's ability to identify sudden water quality changes and enhancing operational stability, thereby further improving the long-term operational reliability and early warning accuracy of the monitoring system. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the business method of an AI-based intelligent water quality monitoring system for aquaculture, provided in this application embodiment; Figure 2 A schematic diagram of the structure of an AI-based intelligent water quality monitoring system for aquaculture provided in this application embodiment; Figure 3 The hierarchical sampling scheduling state transition diagram provided in the embodiments of this application; Figure 4 The model closed-loop update diagram provided in the embodiments of this application. Detailed Implementation

[0018] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present disclosure are shown in the drawings, it should be understood that embodiments of the present disclosure may be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure.

[0019] It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure. In the description of the embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "this embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] like Figure 1 The diagram shown is a flowchart of a business method for an AI-based intelligent water quality monitoring system for aquaculture, provided in an embodiment of this application. Figure 2 The diagram shown is a structural schematic of an AI-based intelligent water quality monitoring system for aquaculture provided in an embodiment of this application, comprising: The scenario modeling and parameter configuration module is used to establish monitoring files for aquaculture scenarios, configure monitoring points and sensors, determine the set of monitoring parameters, and divide the monitoring parameters into fast and slow key parameter groups. First, a scenario monitoring file is established for the target aquaculture farm. The scenario monitoring file is established using a combination of fixed core fields and scenario-extended fields. Fixed core fields are fields registered for all types of aquaculture scenarios, preferably including aquaculture stage, monitoring point location, on-site execution equipment, daily operation plan, safe range of water quality parameters, sampling level parameters, communication method, and permission settings. Depending on the specific scenario, scenario-extended fields can be added accordingly. For example, in a freshwater pond scenario, light intensity, weather, feeding time, and aerator operation plan can be added; in a factory-style recirculating aquaculture system scenario, return water nodes, filtration unit status, circulation pump status, and process section identifiers can be added; and in a marine aquaculture scenario, salinity, water exchange conditions, and tidal information can be added.

[0022] After completing the scene documentation, unique identifiers are assigned to monitoring points, sensors, acquisition terminals, and execution terminals, and a monitoring point device mapping table is established. The mapping table should include the monitoring point identifier, parameter name, sensor type, acquisition terminal identifier, execution terminal identifier, communication address, primary sensor marker, and backup sensor marker. During mapping, the sensor measurement type is first checked to ensure it matches the required monitoring parameters, followed by checks on communication connectivity and time synchronization capabilities. Once these checks are passed, the information is written to the monitoring point device mapping table. If two sensors, a primary and a backup, are configured for the same parameter, their primary / backup relationship is recorded synchronously in the mapping table, and the primary sensor data is prioritized during real-time operation. If the primary sensor fails quality control or communication is interrupted, the system switches to the backup sensor according to the mapping table.

[0023] Next, a set of water quality parameters to be monitored is determined at each monitoring point. The preferred set of water quality parameters includes dissolved oxygen, pH, water temperature, ammonia nitrogen, and turbidity. Depending on the scenario, other parameters can be added, such as salinity and meteorological parameters (e.g., ambient temperature, relative humidity, rainfall, and weather conditions) for marine aquaculture scenarios, light parameters (e.g., light intensity, duration, and start and end times of sunshine) for freshwater pond scenarios, and conductivity and flow velocity for factory-scale recirculating aquaculture systems. Subsequently, the parameter set is grouped into fast-critical parameter groups and slow-critical parameter groups. Parameter grouping is performed according to pre-written decision rules in the scenario configuration. Preferably, four factors should be considered: the parameter's response speed to short-term risks, the parameter's real-time value to sampling scheduling, the time required for sensor output stabilization, and whether the parameter needs to be jointly judged with other parameters in the same decision round. If a parameter can reflect the approaching risk within the most recent 1-3 decision cycles and directly participate in decisions regarding enhanced sampling, locked sampling, synchronous sampling, or supplementary sampling, it is classified into the rapid critical parameter group. If a parameter is more suitable for characterizing risk consequences such as background load or pollution accumulation, it is classified into the slow critical parameter group. According to this rule, this embodiment classifies dissolved oxygen, pH, and water temperature into the rapid critical parameter group, and ammonia nitrogen and turbidity into the slow critical parameter group. Furthermore, a regular sampling cycle, an enhanced sampling cycle, and a locked sampling cycle are pre-set for the rapid critical parameter group, while the regular sampling cycle remains unchanged for the slow critical parameter group, or it is pre-set to perform sampling once every fixed number of cycles at the locked sampling level. The above sampling cycle and cycle parameters are all pre-set by the implementers based on the aquaculture scenario, equipment capabilities, and management requirements, and are written into the scenario monitoring file of the corresponding monitoring point. Once the same scenario is set, it is executed according to fixed rules during operation. For example, in a freshwater pond scenario, the standard sampling period for the rapid key parameter group can be set to 30 minutes, the enhanced sampling period to 10 minutes, and the locked sampling period to 2 minutes. In a factory-scale recirculating aquaculture system scenario, the standard sampling period for the rapid key parameter group can be set to 10 minutes, the enhanced sampling period to 3 minutes, and the locked sampling period to 1 minute. After configuring the monitoring points, each monitoring point is bound to its corresponding sensor, acquisition terminal, execution terminal, and site identifier to form a monitoring point device mapping table. For newly added parameters in different scenarios, technicians can group them according to the same rules; if a parameter cannot be grouped temporarily, it can be included in the slow key parameter group first, and then adjusted after sample accumulation. After grouping, the parameter group information is written into the monitoring point configuration file and serves as the basis for subsequent input sequence construction, synchronous sampling, and supplementary sampling execution.

[0024] The model training and deployment module is used to construct a training sample set based on historical continuous monitoring data, train a key risk window identification model, and determine the cloud or edge deployment method of the model according to network stability, response time requirements, on-site computing power conditions, and user execution permissions. During the AI ​​model cloud training phase, the system constructs a training sample set based on historical continuous monitoring data, execution logs, and risk records. The historical continuous monitoring data preferably includes parameter sampling values, sampling timestamps, sampling levels, data quality status, synchronous sampling completion status, supplementary sampling status, missing data markers, late data markers, unstable data markers, and scene information recorded chronologically at each monitoring point. The execution logs preferably include sampling frequency adjustment records, synchronous sampling records, supplementary sampling records, and early warning trigger records. The risk records preferably include status identification results, monitoring status confirmation results, records of key risk precursor status occurrence, key risk formation status confirmation records, risk resolution records, missed identification records, misidentification records, and final handling results. The samples are organized according to monitoring points, and each sample is arranged in continuous time sequence. Preferably, it includes multi-parameter water quality sequences from the most recent consecutive rounds, including dissolved oxygen, pH value, water temperature, current scene label, current time and location, current monitoring point identifier, current equipment status, missing data marker, synchronization completion marker, and re-sampling status. Historical monitoring data is preprocessed according to monitoring point and time sequence to form sample sequences that can be used for subsequent training. It should be noted that the "most recent rounds" mentioned above refers to the number of times the same monitoring point has completed a complete monitoring decision-making process, preferably 8-16 rounds recently. A complete monitoring decision-making process preferably includes sampling triggering for this round, data reception quality control, input sequence update, status identification, and sampling scheduling output.

[0025] The training sample set is divided into four segments based on event anchor points and consecutive round conditions: stable operation segment, critical risk precursor segment, critical risk formation segment, equipment disturbance segment, control intervention segment, and recovery and stabilization segment. Specifically, the stable operation segment describes the continuous change process of water quality parameters under normal operating conditions; the critical risk precursor segment describes the continuous offset, linked changes, and sampling level changes before risk formation; the critical risk formation segment describes the change trajectory and execution results of each parameter before and after risk confirmation; the equipment disturbance segment records parameter changes caused by on-site actions such as aeration, water exchange, circulation, or feeding; the control intervention segment records the response process of water quality changes after the execution of equipment actions; and the recovery and stabilization segment records the continuous process of each parameter returning to the normal range after the risk is resolved. This segmentation allows the training sample to simultaneously cover the normal process, the risk approach process, the risk formation process, the intervention process, and the recovery process.

[0026] After constructing the sample set, a key risk window identification model is trained using common machine learning models, preferably a GRU (Gated Recurrent Unit) model. However, technicians can choose a suitable machine learning model for training based on actual conditions; this embodiment does not impose any constraints on this. During training, the model takes a multi-parameter continuous sequence and corresponding state information from the input window as input, and the current round's state and its scheduling label as output. The output preferably includes whether the current state is stable, whether it is in a key risk precursor state, whether it is in a key risk formation state, whether the sampling frequency needs adjustment, whether rapid key parameter group synchronization sampling needs to be performed, whether key parameter supplementation needs to be initiated, and whether the current high-frequency sampling state needs to be maintained. During training, in order to ensure that the output can directly serve the subsequent adjustment of the sampling frequency, each input window needs to be accompanied by a labeling result before training. The labeling result preferably includes two parts: the first part is the monitoring state category corresponding to the current window, such as stable state, critical risk leading state, or critical risk formation state; the second part is the scheduling result corresponding to the current window, such as whether the sampling frequency needs to be adjusted, whether synchronous sampling needs to be performed, whether supplementary sampling needs to be started, and whether the locked sampling level needs to be maintained.

[0027] After the key risk window identification model is trained, it enters the deployment determination phase. Deployment determination can be performed according to preset rules. Preferably, if the site requires automatic state identification and sampling scheduling within a sampling round, and the on-site execution terminal performs pre-deployment testing on the target on-site terminal, and the terminal can output the scheduling result of the current round within a preset inference time limit after a complete input window is formed, and can continuously cache data from at least two complete input windows, then the terminal is deemed to meet the on-site deployment conditions. In this case, the model is deployed on the edge execution terminal, which is responsible for real-time inference and sampling scheduling. If the on-site network is stable, the user retains manual confirmation permissions, and the above on-site deployment conditions are not met, the model is deployed in the cloud, where inference is completed and sampling strategies are distributed to the user or execution terminal. If both on-site real-time performance and unified update management are required, an edge inference plus cloud update deployment method is adopted. Network stability can be evaluated based on communication latency, packet loss rate, continuous interruption duration, and number of disconnections; on-site computing power conditions can be evaluated based on the execution terminal processor idle rate, available memory, local caching capacity, and single-round inference duration. The thresholds for the aforementioned evaluation indicators are preset by technical personnel and written into the deployment configuration; this embodiment does not impose constraints on the specific values. Regardless of the deployment method, the real-time execution sequence should remain consistent: sampling first, then quality control, then construction of the input sequence, followed by state identification, and finally, distribution of the sampling scheduling results.

[0028] The real-time data acquisition and input construction module receives real-time monitoring data uploaded from each monitoring point, performs data quality control, and constructs the model input sequence for the current monitoring point by combining current time information, scene information, equipment operating status, data quality status, sampling level change status, synchronous sampling completion status, and supplementary sampling execution status. After the system enters the running state, it continuously receives real-time monitoring data uploaded from each monitoring point and maintains the input sequence for each monitoring point. The input sequence preferably includes fast and slow key parameter group data from the most recent 12 rounds, current time information, current scene information, current equipment operating status, current data quality status, sampling level change status from the most recent rounds, synchronous sampling completion status from the most recent rounds, and supplementary sampling execution status from the most recent rounds. Each round of data in the input sequence must be accompanied by a timestamp and round number for easy subsequent alignment and tracking.

[0029] When the system is first started or the model is first deployed, if the full input window length has not yet been collected, a partial input window is constructed based on the number of valid rounds collected. A start-up placeholder, a no-historical-data marker, and an incomplete status marker are added to the missing positions at the beginning. The minimum number of start-up rounds is preset by technical personnel and written into the model configuration. Before reaching the minimum number of start-up rounds, the system allows output of regular sample-and-hold, synchronous sampling suggestions, and supplementary sampling suggestions. After reaching the minimum number of start-up rounds, it can participate in augmented sampling judgment. Once the full input window length is reached, all scheduling results are output according to normal rules. This arrangement allows the system to run in the initial stage while avoiding erroneous jumps due to insufficient historical windows. The full input window length refers to an input sequence whose length is equal to the preset input window length.

[0030] The risk identification and sampling scheduling module is used to feed the model input sequence into the critical risk window identification model, identify whether the current monitoring point is in a stable state, a critical risk precursor state, or a critical risk formation state, and output whether to adjust the sampling frequency, whether to perform rapid critical parameter group synchronous sampling, whether to start critical parameter supplementation sampling, and whether to continue the current high-frequency sampling state. Based on the model output results, it performs hierarchical sampling scheduling, including regular sampling, enhanced sampling, or locked sampling, such as... Figure 3The diagram shown is a hierarchical sampling scheduling state transition diagram provided in an embodiment of this application. After updating the input sequence in each round, the system sends the input window of the current monitoring point into the critical risk window identification model to obtain the state identification result and scheduling judgment result of the current round. Preferably, the core states are divided into three categories: stable state, critical risk precursor state, and critical risk formation state. A stable state indicates that the current parameter change falls within the safe range of the parameter value registered in the scene monitoring file, and no enhanced sampling, locked sampling, synchronous sampling, or supplementary sampling has been triggered in the last three consecutive rounds, and no risk approaching characteristics have formed. A critical risk leading state indicates that a shift trend before risk formation has appeared in the current continuous sequence, that is, at least two rounds in the last three consecutive rounds have changed in the same direction, and the difference between the current value and the most recent high-confidence reference value of the parameter reaches the trend judgment threshold corresponding to the parameter, at least two parameters in the rapid critical parameter group simultaneously meet their respective trend judgment conditions in the same synchronous sampling round, or meet the pre-registered parameter combination change rules, or the current round has missing data, late data, data that has not passed stability confirmation, data that has not been synchronized, or data that has been judged as invalid by quality control, as well as situations where the current parameter value and the most recent high-confidence reference value of the parameter work together, requiring an increase in monitoring density. A critical risk formation state indicates that the current monitoring point has entered a stage where continuous sampling according to the locked sampling cycle is required, and the number of locked sampling level holding rounds is not lower than the preset minimum holding rounds of locked sampling, requiring continuous monitoring according to the locked sampling cycle and outputting risk results. In addition to the core state, the model also outputs scheduling flags, including whether to adjust the sampling frequency, whether to perform synchronous sampling, whether to start supplementary sampling, and whether to maintain the current high-frequency sampling state.

[0031] To ensure that status results can be directly incorporated into the execution process, the system reads the model output, including the current sampling level, the current round's status result, the cumulative number of consecutive leading rounds, the minimum holding rounds for locked sampling, the synchronization completion status, and the supplementary sampling completion status, and generates the scheduling instruction for this round. The cumulative number of consecutive leading rounds (the actual number of rounds the current monitoring point has continuously output key risk leading states), the minimum holding rounds for locked sampling (the minimum number of rounds required to maintain the status before downgrading after entering the locked sampling level), and the number of downgrading confirmation rounds (the number of rounds required to continuously meet downgrading conditions before downgrading from the locked sampling level to the enhanced sampling level, or from the enhanced sampling level to the regular sampling level) are all pre-configured by technical personnel in the monitoring point configuration. Based on the model output, the system drives the edge control unit to execute sampling level control for the current monitoring point. The preferred sampling levels include a regular sampling layer, an enhanced sampling layer, and a locked sampling layer. The regular sampling layer corresponds to the daily monitoring cycle; the enhanced sampling layer corresponds to the encrypted monitoring cycle of the leading phase; and the locked sampling layer corresponds to the high-frequency monitoring cycle of the risk formation phase. The specific execution rules are as follows: When the model output is in a stable state and the scheduling flag indicates no adjustment of the sampling frequency, the system maintains the current sampling level. If the current round is in the regular sampling level, it continues to execute according to the regular sampling cycle; if the current round is in the enhanced sampling level or the locked sampling level, it continues to check whether the degradation conditions are met. When the model output is a critical risk precursor state and the scheduling flag indicates yes to the adjustment of the sampling frequency, the system first determines the current sampling level. If it is currently in the regular sampling level, it immediately switches to the enhanced sampling level and calls the preset enhanced sampling cycle for that monitoring point; if it is already in the enhanced sampling level, it accumulates and records the precursor duration rounds; if it is already in the locked sampling level, it maintains the locked sampling cycle and continues execution. When the model output is a critical risk formation state, the system prioritizes checking the current sampling level. If the current sampling layer is enhanced, it switches to the locked sampling layer and invokes the preset locked sampling cycle for that monitoring point. If it is already in the locked sampling layer, it remains in the locked sampling state, and the sampling cycle is incremented by one. If it is temporarily in the regular sampling layer due to the initial operation phase or abnormal scenario, it is first promoted to the enhanced sampling layer, and in the next round, it is determined whether to switch to the locked sampling layer. The specific switching rules need to be pre-configured by technical personnel according to the scenario scheduling requirements. When the model continuously outputs the key risk precursor status for several consecutive rounds, and the cumulative number of precursor rounds reaches the number of continuous precursor confirmation rounds, it is allowed to switch from the enhanced sampling layer to the locked sampling layer even if the key risk formation status has not been output in the current round, so as to increase the monitoring density in advance. The number of continuous precursor confirmation rounds is also manually set by technical personnel, and this embodiment does not limit it. When the model output indicates "Keep Current Sampling Level," the system reads the number of rounds the current monitoring point has been continuously in the locked sampling level. If the minimum number of rounds for locking sampling has not yet been reached, the locked sampling level will continue to be maintained, and no downgrading will be performed. If the minimum number of rounds has been reached, the system will determine whether downgrading is allowed based on subsequent continuous confirmation results. Downgrading requires the following conditions to be met simultaneously: the preset number of downgrading confirmation rounds has been reached; no critical risk precursor status or critical risk formation status has been output within these consecutive rounds; rapid critical parameter group synchronous sampling has been completed; critical supplementary sampling has been completed; and there is no critical data in the unprocessed low-confidence window data within the current window. Only after these conditions are met is it allowed to downgrade from the locked sampling level to the enhanced sampling level, or from the enhanced sampling level to the regular sampling level. This number of downgrading confirmation rounds is pre-configured by technical personnel.

[0032] For example, the rapid critical parameter group configuration for a freshwater pond monitoring station is as follows: a regular sampling cycle of 30 minutes, an enhanced sampling cycle of 10 minutes, a locked sampling cycle of 2 minutes, two rounds of continuous confirmation of the leading sample, five rounds of minimum lock sampling duration, and three rounds of degrade confirmation. When the leading critical risk status is output continuously in rounds 1 and 2, the enhanced sampling cycle is switched to the third round; when the leading critical risk status is output in round 4, the locked sampling cycle is switched to the fifth round; the locked sampling is maintained continuously from rounds 5 to 9; if no leading critical risk status or leading critical risk status is output continuously from rounds 10 to 12, and synchronous sampling, supplementary sampling, and quality control are all completed, the enhanced sampling cycle is reverted to the third round; if the degrade conditions are met continuously again from rounds 13 to 15, the regular sampling cycle is reverted to the sixth round.

[0033] For synchronous sampling and supplementary sampling, the system directly triggers them based on the corresponding markers in the model output. When the scheduling results require synchronous sampling, sampling of the fast key parameter group should be initiated within the same decision round; when the scheduling results require supplementary sampling, supplementary sampling tasks should be initiated for key parameters that are missing, late, unstable, not synchronized, or fail quality control.

[0034] The synchronous sampling and quality assurance module is used to perform synchronous sampling and sampling stability verification on rapid key parameter groups. It performs range rationality checks, abrupt jump checks, stability checks, continuity checks, and sensor health checks on the sampled data. Under high-risk monitoring conditions, it performs supplementary sampling, supplementary synchronous sampling, and anti-smoothing quality assurance processing. Under the enhanced sampling layer or locked sampling layer, the system performs synchronous sampling on rapid key parameter groups and records the sampling start time, sampling end time, and sampling status for each parameter. The purpose of recording time and status is threefold: first, to determine whether the rapid key parameter group has completed sampling within the same decision-making round, supporting parameter linkage analysis; second, to identify late data, data awaiting confirmation, and asynchronous data; and third, to provide a basis for subsequent quality control, supplementary sampling execution, and sensor health assessment.

[0035] For sensors with response time, the system performs confirmation according to the stability determination rule corresponding to the sensor after triggering sampling. The stability determination rule preferably includes a minimum waiting time, a stable observation time, an allowable fluctuation bandwidth, and a maximum confirmation time. The minimum waiting time refers to the shortest time required from the moment the sensor receives the sampling trigger command to the moment it is allowed to begin reading and determining whether the sensor's output value has entered a stable state. The stable observation time refers to the observation period after the minimum waiting time ends, during which the system continuously reads the sensor's output value and determines whether it is stable. The allowable fluctuation bandwidth refers to the maximum allowable fluctuation range among multiple consecutive output values ​​within the stable observation time, which can be represented by the difference between the maximum and minimum output values ​​within that time period. The maximum confirmation time refers to the preset confirmation deadline allowed from the moment the sampling trigger command is issued until the system must complete the stability confirmation of this sampling or record the sampling result as unstable data; this is pre-configured by technical personnel. If the sensor enters a stable observation period after the minimum waiting time, and the variation range of the continuous readings within the stable observation period consistently falls within the upper limit of the allowable difference between the maximum and minimum values ​​of multiple consecutive sampling results within the stable observation period, then the current sampled value is recorded as a valid value. If the stability judgment is not met within the maximum confirmation period, it is recorded as a value to be confirmed and written into the pending confirmation status flag for use in the next round of supplementary sampling or quality control. For example, if the difference between the maximum and minimum values ​​of the dissolved oxygen sensor does not exceed 0.2 mg / L; the pH sensor does not exceed 0.05; and the water temperature sensor does not exceed 0.3℃ within the stable observation period, it is determined that the stable condition has been reached. The above values ​​can be manually set by technicians, and this embodiment does not limit them.

[0036] Next, the system performs quality control on each round of sampling results. Quality control includes, in sequence, range rationality checks, adjacent sample value jump checks, stability checks, continuity checks, and sensor health checks. Range rationality checks include sensor physical range checks and scene reliability range checks; adjacent sample value jump checks compare the current value with the most recent data from the same monitoring point and for the same parameter that simultaneously passed physical range checks, scene reliability range checks, stability confirmation checks, time sequence checks, and sensor health checks; stability checks are performed based on stability determination results, which include stable pass results and unstable results. Sample values ​​that have not yet completed stability confirmation before the confirmation deadline can be recorded as values ​​awaiting confirmation; continuity checks determine whether there are any critical missing, late, out-of-order, or incomplete synchronization issues in the current window; sensor health checks combine sensor self-test status, communication status, calibration status, response timeout conditions, and historical anomaly records to determine whether the current sensor is in a normal, degraded, or faulty state. The pass conditions, warning conditions, and failure conditions for each check can be preset by technicians according to parameter type and equipment performance; this embodiment does not limit their specific values.

[0037] After summarizing the quality control results, the system categorizes the current data into three types according to preset merging rules: high-confidence data, low-confidence data, or invalid data, and writes the corresponding quality status back to the input sequence. Data that simultaneously meets the following conditions is categorized as high-confidence data: passing physical range and scenario confidence range checks, having completed stability confirmation and obtained a stable pass result, having no critical missing or delayed data, having no incomplete supplementary data collection, and the sensor currently being in normal condition. Data that exceeds the sensor's physical range, has not reached stability after the confirmation deadline and cannot be retained as a pending confirmation value, has a faulty sensor and no alternative data source, or still cannot meet any of the replacement conditions after supplementary data collection, is categorized as invalid data. The remaining data that can be retained but requires further confirmation is categorized as low-confidence data. The written-back content preferably includes the current round parameter value, quality check results, primary / backup sensor source markers, synchronization completion markers, and pending confirmation markers. When the current monitoring point is in an enhanced sampling layer or a locked sampling layer, the system enters a high-risk monitoring state and prioritizes critical parameter supplementary data collection and fidelity processing. For missing, incompletely synchronized, pending-confirmation, and quality control failed items in the rapid critical parameter group, the system initiates supplementary sampling based on the scheduling results of this round. The specific supplementary sampling operation is as follows: When a critical parameter is missing, late, has not completed stability confirmation, has not entered the synchronization time window, or fails quality control in the current round, the system generates a supplementary sampling task for that parameter. The supplementary sampling task prioritizes resampling the parameter's primary sensor. If the primary sensor still does not return usable data within the supplementary sampling time limit, then the backup sensor is called for supplementary sampling. During supplementary sampling, the sampling start time, end time, stability confirmation result, and quality control result are continuously recorded. When the supplementary sampling result meets the substitution conditions, the supplementary value replaces the original missing or pending-confirmation value of the current round; when the supplementary sampling result does not meet the substitution conditions, the original status flag of the current round is retained, and this round is recorded as a low-confidence round.

[0038] After supplementary data collection, if the supplementary data passes quality control and meets the replacement conditions (i.e., simultaneously passes physical range checks, scenario confidence range checks, stability confirmation checks, time sequence checks, and sensor health checks), then the supplementary value replaces the original value to be confirmed or the missing bit. If the data still does not meet the stability criteria after supplementary collection, or remains invalid after quality control, then the original time sequence and missing, pending, or low-confidence states are retained. "Still unstable" means that after supplementary collection, the data still does not meet the stability criteria within the maximum confirmation time of the corresponding sensor, or although a reading is obtained, the quality control result still does not meet the usability conditions. "Still incomplete" means that in the set of key parameters required for joint analysis in the current round, there is still at least one missing, late, asynchronous, pending, or invalid item. For incomplete data segments within high-risk windows, the system marks them as low-confidence window data and continues monitoring at the current sampling level to ensure that the original information of the risk segments is preserved.

[0039] The monitoring status confirmation and early warning output module is used to confirm the current water quality monitoring status and output monitoring results and early warning results. After completing sampling scheduling, synchronous sampling, quality control, and high-risk fidelity processing, the system confirms the current monitoring status by combining the current round status identification results, the current sampling level, the current quality category, the synchronous completion status, and the supplementary sampling status. Preferably, the monitoring status is divided into stable monitoring status, suspicious change monitoring status, and critical risk monitoring status. Technical personnel can also add or delete statuses according to the actual situation. This embodiment only provides the above-mentioned statuses as examples and does not mean that it is limited to the above three monitoring statuses. When the model outputs a stable status, the current key parameter group data is complete, the quality control is based on high confidence results, and there are no incomplete supplementary sampling items, the system confirms the current round as a stable monitoring status. If, in the most recent three consecutive rounds, at least two parameters in the rapid key parameter group reach the trend judgment threshold, or if one rapid key parameter reaches the trend judgment threshold and the current round has incomplete synchronization or supplementary sampling, low-confidence window data, or items awaiting confirmation, but has not yet met the conditions for the formation of a critical risk, then the current round's status identification result is a critical risk formation status. Alternatively, if at least one parameter in the rapid key parameter group of the current round exceeds the upper or lower limit of the parameter safety range and is confirmed by synchronization sampling or supplementary sampling, then it is confirmed as a suspicious change monitoring status. If the current round has already output a critical risk formation status, or has entered the locked sampling layer and the corresponding risk conditions persist, then it is confirmed as a critical risk monitoring status. The monitoring status confirmation result, along with the sampling level, quality status, and execution status, is written into the monitoring record.

[0040] The system generates the monitoring results for the current round based on the monitoring status. The monitoring results preferably include the parameter values ​​of the current monitoring point, the current sampling level, the current monitoring status, whether the system is currently in the critical risk precursor or formation stage, the current data confidence level, and whether an early warning is needed. If the current monitoring status is stable, a normal monitoring result is output; if the current monitoring status is suspicious change, a prompt to continue monitoring is output, and it may indicate the presence of low-confidence data, values ​​awaiting confirmation, or precursor signs; if the current monitoring status is critical risk, a risk monitoring result is generated, and an early warning judgment is triggered. The conditions for early warning judgment are preset by technical personnel in the scenario configuration. In this embodiment, any of the following situations can be used as early warning trigger conditions: the current round outputs a critical risk formation status and the corresponding critical data is of high confidence; the precursor status has not been lifted after reaching the required number of consecutive precursor confirmation rounds; any parameter in the rapid critical parameter group exceeds the upper or lower limit of the parameter safety range registered in the scenario file, and is confirmed by synchronous sampling or supplementary sampling; the system is currently in a locked sampling layer and the critical risk status has not been lifted for several consecutive rounds. When the conditions are met, the system outputs early warning information to the user terminal. If the system is configured for automatic execution mode, it will synchronously output linkage instructions or linkage suggestions to the linkage module of the execution device.

[0041] The data feedback and closed-loop update module is used to upload monitoring results, key risk window data, low-confidence window data, early warning records, linkage records, and misjudgment records to the cloud platform, and to perform candidate model training, validation, replacement, and redeployment to form a closed-loop model update, such as... Figure 4 The diagram shown is a model closed-loop update diagram provided in an embodiment of this application, and its specific implementation is as follows: The system uploads monitoring results, key risk window data, low confidence window data, early warning records, linkage records, missed identification records, and misidentification records to the cloud platform or centralized management terminal, and performs closed-loop updates according to preset rules. Closed-loop updates operate on a combination of periodic checks and conditional triggers; the update process only begins when the update trigger conditions are met. The preferred update trigger conditions include the following categories: First, missed identification triggers updates when a confirmed key risk window is not promptly identified by the current model, or when the model fails to output enhanced sampling or locked sampling strategies before the risk forms, and the cumulative number of such missed identification events reaches a threshold set by technical personnel. Prompt identification refers to outputting the key risk precursor state at least two rounds before the key risk confirmation round, or outputting the key risk formation state at the latest in the current round of key risk confirmation. Second, misidentification triggers updates when the model frequently identifies stable states as precursor or formation states, leading to repeated frequency increases, synchronous sampling, or supplementary sampling on-site, and the cumulative number of such misidentification events reaches a preset threshold. Third, updates are triggered by changes in the scenario, i.e., changes in the breeding stage, season, process node, monitoring equipment, monitoring point deployment, or management strategy, reaching the scenario change conditions defined by the technicians. Fourth, updates are triggered by sample accumulation, i.e., the high-confidence critical risk window samples, critical risk leading samples, recovery and stabilization samples, and stable samples accumulated by the system reach the pre-set sample thresholds. All of the above thresholds are pre-configured by the technicians, and this embodiment does not limit their specific values.

[0042] After an update is triggered, the system filters updated samples from archived data. During the screening process, the system retains high-confidence critical risk window samples, critical risk precursor samples, recovery and stabilization samples, missed identification samples, and misidentified samples, while removing samples that do not meet the update conditions. Preferred removal conditions include: parameter values ​​exceeding the sensor's physical range, out-of-order timestamps that cannot be recovered, conflicting and unconfirmed sources of critical labels, missing critical execution logs, and critical parameters continuously missing within the training window exceeding the upper limit set by technical personnel. After screening, the new samples are merged with the original basic training samples to form an updated training set, preserving historical effective experience and incorporating new scenario features. The merging process is as follows: First, the new samples and the original basic training samples are categorized into stable operation samples, critical risk precursor samples, critical risk formation samples, equipment disturbance samples, control intervention samples, and recovery and stabilization samples, respectively. Then, within each category, the new samples are appended to the original basic training samples. If the number of new samples in a certain category exceeds the number of original basic training samples in that category multiplied by a preset ratio, then sampling is performed based on time distribution and monitoring point distribution. Finally, the training set, validation set, and test set are divided according to category to form an updated training set. The preset ratio is set manually by technical personnel and is not limited in this embodiment.

[0043] Subsequently, the candidate update model is retrained using the updated training set. The training method is the same as that for the critical window risk identification model; the input remains the multi-parameter sequence and state sequence within a continuous window, and the output remains the core state and scheduling label. After training, the candidate update model is compared with the current model on an independent validation sample set. The preferred comparison indicators include four items. The first item is the ability to identify critical risk windows in advance, i.e., the proportion of windows in which the model outputs a leading state or forming state before the actual confirmation of a critical risk. The second item is the ability to control false alarms in stable states, i.e., the proportion of incorrect output of leading states or forming states within a confirmed stable window. The third item is the stability of sampling frequency adjustment, i.e., the number of times or the proportion of unnecessary level switching occurs within a unit monitoring cycle. The fourth item is the effectiveness of synchronous sampling and supplementary sampling triggering, i.e., the proportion of triggered synchronous sampling or supplementary sampling that ultimately improves data integrity, stability, or the credibility of risk judgment. The calculation methods, qualification thresholds, and replacement thresholds of the above comparison indicators are pre-written into the model update configuration by technical personnel; this embodiment does not limit their specific values.

[0044] After comparing the candidate updated model with the current model, the system performs a model replacement decision. If the candidate updated model meets the following conditions, it is allowed to replace the current model: Its false alarm control capability, sampling frequency adjustment stability, and the effectiveness of synchronous sampling and supplementary sampling triggering are no lower than the current model; its ability to identify key risk windows in advance reaches the improvement threshold set by technical personnel; or its adaptability after changes in specific scenarios meets the preset deployment conditions. If the candidate updated model improves the early identification capability but simultaneously leads to an increase in false alarms, exceeding the limit for the number of level switching attempts, or a decrease in the effectiveness of supplementary sampling triggering, the current model is retained, samples are accumulated, and the system awaits the next round of updates. After the model replacement is approved, the system redeploys the updated model according to the same deployment method as described above. When using edge deployment, the verified updated model is deployed to the edge execution end, and the execution end completes the replacement within the maintenance window. When using cloud deployment, the current model is replaced in the cloud, and scheduling results continue to be distributed externally. When using edge inference plus cloud update, only the edge inference model version is replaced, and the cloud archiving configuration and version records are updated synchronously.

[0045] Example 2, based on Example 1, addresses the issue that the critical risk window identification model may encounter behavior types not covered during training, or may still fail to provide stable and consistent identification results for the current monitoring status even when the current monitoring data is complete and meets the usage requirements. Therefore, an abnormal behavior constraint adjustment module is added to perform abnormal behavior constraint adjustment on the current monitoring process when the current monitoring point meets preset abnormal conditions. The specific implementation is as follows: When the system reaches the risk identification and sampling scheduling module in Example 1, if the current monitoring point meets any of the following conditions, it is determined to enter the first-level abnormal behavior constraint state. The specific conditions include: First, within three consecutive sampling rounds, the model alternately outputs a stable state and a critical risk leading state for the same monitoring point, and the original data within these three sampling rounds can be used as valid data after quality control; Second, within three consecutive sampling rounds, the model outputs twice requiring adjustment of the sampling frequency, and once without requiring adjustment of the sampling frequency, resulting in inconsistent scheduling results; Third, dissolved oxygen, pH value, or water temperature in the rapid critical parameter group have shifted towards the risk direction for two consecutive sampling rounds, but the model still has not output the critical risk formation state. Upon entering the Level 1 Abnormal Behavior Constraint state, the system immediately performs the following adjustments: If the current monitoring point is in the regular sampling layer, it switches to the enhanced sampling layer; if the current monitoring point is already in the enhanced sampling layer, it remains in the enhanced sampling layer and is not allowed to directly drop back to the regular sampling layer within the current abnormal behavior window; synchronous sampling is performed for three consecutive sampling rounds on the fast key parameter group; a supplementary sampling is performed on the key parameters that are missing, late, unstable, or not synchronized in the current round; the consecutive confirmation rounds of the current monitoring point are increased by two sampling rounds based on the original preset value; and the monitoring window is marked as the Level 1 Abnormal Behavior window. Through the above processing, when the model determines that the system is not yet stable, it prioritizes increasing the data acquisition density and data completeness, avoiding the omission of abnormal processes due to premature frequency reduction. It should be noted that the above three consecutive sampling rounds are not fixed to three. This embodiment is only for illustrative purposes, and the number of consecutive rounds can be freely set by technical personnel according to the actual situation. This embodiment does not impose any constraints on this.

[0046] After the execution of the Level 1 abnormal behavior constraint state, if the current monitoring point continues to meet any of the following conditions, it will be upgraded to the Level 2 abnormal behavior constraint state: First, after three consecutive enhanced sampling rounds and three consecutive synchronous sampling rounds, the model is still switching back and forth between the critical risk leading state and the critical risk formation state; Second, the rapid key parameter group has shown abnormal changes for three consecutive sampling rounds, that is, at least one rapid key parameter has reached the corresponding trend judgment threshold, or at least two rapid key parameters have reached their respective trend judgment thresholds simultaneously in the same synchronous sampling round, and synchronous sampling has been completed in each round, but the model still has not output stable and consistent sampling scheduling results; Third, two supplementary samplings have been performed within the current abnormal behavior window, but there are still gaps in key parameters that affect the confirmation of the state. Upon entering the Level 2 Abnormal Behavior Constraint State, the system immediately performs the following adjustments: Switching the current monitoring point to the locked sampling layer; Maintaining the locked sampling layer for at least five consecutive sampling rounds without premature exit; Performing five consecutive sampling rounds of synchronous sampling for the rapid key parameter group; For each unstable or incompletely synchronized key parameter, adding one more sampling opportunity on top of the original sampling; Suspending the normal smoothing and value compensation processing within the current abnormal behavior window, retaining only the original time sequence, missing markers, pending confirmation markers, and synchronization completion status; and marking this monitoring window as a Level 2 Abnormal Behavior window.

[0047] After the execution of the Level 2 abnormal behavior constraint state, if the current monitoring point continues to meet any of the following conditions, it will be upgraded to the Level 3 abnormal behavior constraint state: First, after locking sampling, synchronous sampling and supplementary sampling, the model still cannot form a stable and consistent state judgment within five consecutive sampling rounds; Second, after manual review or post-test confirmation, the current abnormal behavior window belongs to a new type of behavior that has not appeared in the training samples; Third, the same type of abnormal behavior window appears more than three times in the same scenario and the same breeding stage. Upon entering the Level 3 abnormal behavior constraint state, the system performs the following adjustments: Maintaining the current monitoring point in the locked sampling layer until the current abnormal behavior window ends; freezing the automatic frequency reduction action, disallowing the system from automatically reducing back to the enhanced sampling layer or the regular sampling layer; setting the synchronous sampling task of the rapid key parameter group as the highest priority task for the current monitoring point; continuing to perform supplementary sampling for incomplete and unstable sampling parameters until the current abnormal behavior window ends or is manually terminated; writing all original monitoring data, data quality status, synchronization completion status, supplementary sampling status, monitoring status confirmation results, and final handling results within the abnormal behavior window into the candidate update sample pool; if the system is in automatic linkage mode, adding a manual confirmation step for control strategies involving direct actions of the executing equipment, but allowing the system to continue automatically executing monitoring strategies that only involve sampling frequency adjustment, synchronous sampling, and supplementary sampling.

[0048] Once the abnormal behavior constraint state has been activated, the system also sets exit conditions. Exiting the abnormal behavior constraint state is only permitted if the following conditions are met simultaneously within three consecutive sampling rounds: First, the model's assessment of the current monitoring point's output state remains consistent, no longer switching between stable state, critical risk precursor state, and critical risk formation state; Second, there are no incomplete supplementary sampling items at the current monitoring point; Third, the rapid critical parameter group has completed synchronous sampling; Fourth, after state confirmation, the current monitoring point has returned to a stable monitoring state, or has entered a suspicious change monitoring state with consistent change direction. The suspicious change monitoring state is consistent with the state definition in the monitoring state confirmation and early warning output module of Example 1, meaning that the current monitoring point has not yet entered the critical risk monitoring state, but has experienced at least one of the following: continuous offset, incomplete synchronous sampling, incomplete supplementary sampling, or low-confidence window data, requiring continued enhanced monitoring and awaiting subsequent confirmation. After exiting the Level 1 abnormal behavior constraint state, the current monitoring point resumes the normal stratified sampling rules in Example 1; after exiting the Level 2 or Level 3 abnormal behavior constraint state, the current abnormal behavior window will still be retained in the candidate update sample pool for subsequent sample screening and model updates.

[0049] In this embodiment, the abnormal behavior constraint adjustment step is used in conjunction with the data feedback and closed-loop update module in Embodiment 1. Specifically, the system does not immediately update the model due to a single abnormal behavior. Instead, it first increases the sampling frequency, strengthens synchronous sampling, increases the opportunity for supplementary sampling, pauses smoothing processing, and extends the high-frequency monitoring time through the abnormal behavior constraint adjustment step. After the system accumulates a window of abnormal behavior samples that meet the update conditions, it then enters the model update process in Embodiment 1. The reason for this setting is that if the model is modified immediately when a behavior not covered in the training phase occurs for the first time, the model update result is likely to be unstable due to insufficient sample quantity, unclear labels, or incomplete recording of the abnormal process. However, through the abnormal behavior constraint adjustment step in this embodiment, the original evolution process of this type of behavior, the relationship of key parameter changes, the supplementary sampling results, and the state confirmation results can be completely recorded first, and then used as the basis for subsequent model updates, thereby improving the reliability and effectiveness of the model closed-loop update.

[0050] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the above functions can be divided into different functional modules to complete all or part of the functions described above.

[0051] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0052] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units, located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0053] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0054] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the solution, or all or part of the technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0055] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based intelligent water quality monitoring system for aquaculture, characterized in that, include: The module includes: scene modeling and parameter configuration, model training and deployment, real-time data acquisition and input construction, risk identification and sampling scheduling, synchronous sampling and quality assurance, monitoring status confirmation and early warning output, and data feedback and closed-loop update. Among them, the scene modeling and parameter configuration module is used to establish scene monitoring files and determine the rapid key parameter group; The model training and deployment module is used to train the critical risk window identification model and deploy the critical risk window identification model on the cloud or edge execution terminal according to network stability, response time requirements, on-site computing power conditions and user execution permissions. The real-time data acquisition and input construction module is used to receive real-time monitoring data and construct the model input sequence; The risk identification and sampling scheduling module is used to output stable state, critical risk precursor state, critical risk formation state, adjust sampling frequency, perform rapid key parameter group synchronous sampling, start key parameter supplementation sampling and maintain the current high-frequency sampling state, and drive the edge execution end to perform sampling level control. The synchronous sampling and quality assurance module is used to perform synchronous sampling of rapid key parameter groups under the enhanced sampling layer or the locked sampling layer, and to perform quality control on the raw water quality data obtained in any sampling round. The monitoring status confirmation and early warning output module is used to confirm the monitoring status, divide the current monitoring status into stable monitoring status, suspicious change monitoring status and critical risk monitoring status, and generate monitoring results; The data feedback and closed-loop update module is used to feedback monitoring results and recorded data, and to perform model updates and redeployment when update trigger conditions are met.

2. The AI-based intelligent water quality monitoring system for aquaculture as described in claim 1, characterized in that: The scene modeling and parameter configuration module includes: Establish a monitoring file for aquaculture scenarios, configure monitoring points and sensors, determine the set of monitoring parameters, and divide the set of monitoring parameters into a fast key parameter group and a slow key parameter group. The model training and deployment module constructs a training sample set based on historical continuous monitoring data, execution logs and risk records during the training phase. The training sample set is organized according to monitoring points. Each training sample is arranged in continuous time sequence and includes a multi-parameter water quality sequence of several consecutive historical rounds, current scene label, current time and location, current monitoring point identifier, current equipment status, missing marker, synchronization completion marker and supplementary sampling status. The training sample set is divided into the following segments according to the operation process: stable operation segment, critical risk leading segment, critical risk formation segment, equipment disturbance segment, control intervention segment, and recovery and stabilization segment. The critical risk window identification model takes a multi-parameter continuous sequence and corresponding state information in the input window as input, and the state of the current round and its scheduling label as output.

3. The AI-based intelligent water quality monitoring system for aquaculture as described in claim 2, characterized in that: The model input sequence constructed by the real-time data acquisition and input construction module includes at least the fast key parameter group data, slow key parameter group data, current time information, current scene information, current device operating status, current data quality status, historical sampling level change status, historical synchronous sampling completion status, and historical supplementary sampling execution status within the historical window length. When the system is first started or the critical risk window identification model is just launched, before the complete window length has been collected, the real-time data acquisition and input construction module constructs a partial input window according to the valid rounds that have been collected, and writes a start placeholder mark, a no historical data mark, and an incomplete status mark.

4. The AI-based intelligent water quality monitoring system for aquaculture as described in claim 1, characterized in that: The risk identification and sampling scheduling module performs the following control based on the output results of the key risk window identification model: When the output of the critical risk window identification model indicates a stable state and the sampling frequency is not adjusted, the current sampling level is maintained. When the model output indicates a critical risk leader state and the sampling frequency is adjusted, if the current sampling level is a regular sampling level, then switch to an enhanced sampling level; if the current sampling level is an enhanced sampling level, then accumulate consecutive leader rounds; if the current sampling level is a locked sampling level, then maintain the locked sampling level. When the model output indicates the formation status of a key risk, if the current sampling level is the enhanced sampling level, then switch to the locked sampling level; if the current sampling level is the locked sampling level, then remain at the locked sampling level. When the model output indicates a critical risk leader status for several consecutive rounds and the cumulative number of consecutive leader rounds reaches the number of leader continuous confirmation rounds, the enhanced sampling layer can be switched to the locked sampling layer. When the model output indicates that the current high-frequency sampling state should be maintained, if the current monitoring point is in the locked sampling layer and the number of continuous sampling rounds since the current monitoring point switched to the locked sampling layer has not reached the preset minimum number of locked sampling rounds, then downgrading is prohibited. Only when multiple consecutive sampling rounds at the current monitoring point reach the preset number of downgrade confirmation rounds, no critical risk precursor status or critical risk formation status is output within the consecutive rounds, the rapid critical parameter group synchronous sampling has been completed, the critical parameter supplementation has been completed, and there is no unprocessed low-confidence critical data in the current window, it is allowed to descend from the locked sampling layer to the enhanced sampling layer or from the enhanced sampling layer to the regular sampling layer.

5. The AI-based intelligent water quality monitoring system for aquaculture as described in claim 4, characterized in that: The risk identification and sampling scheduling module also includes: When the output result of the critical risk window identification model indicates that rapid critical parameter group synchronous sampling should be performed, sampling of the rapid critical parameter group should be initiated within the same decision round. When the output result of the critical risk window identification model indicates that critical parameter supplementation should be initiated, a supplementation task should be initiated for critical parameters that are missing, late, unstable, not synchronized, or fail quality control. The synchronous sampling and quality fidelity module is used to perform synchronous sampling on the fast key parameter group under the enhanced sampling layer or the locked sampling layer, and record the sampling start time, sampling end time and sampling status of each parameter; For sensors with response time, the synchronous sampling and quality fidelity module performs sampling stability confirmation according to the corresponding stability determination rules. The stability determination rules include at least the minimum waiting time, the stable observation time, the allowable fluctuation bandwidth, and the maximum confirmation time. When the stability determination rule is met, the current sampled value is recorded as a valid value; when the stability determination rule is not met, the current sampled value is recorded as a value to be confirmed and written to the pending confirmation status flag.

6. The AI-based intelligent water quality monitoring system for aquaculture as described in claim 5, characterized in that: The synchronous sampling and quality fidelity module also includes: Quality control is performed on the raw water quality data obtained from any sampling round. The quality control includes at least range reasonableness check, adjacent sampling value jump check, stability check, continuity check, and sensor health check. After summarizing the quality control results, the current data is categorized into high-confidence data, low-confidence data, or invalid data. The current round parameter values, quality inspection results, primary and backup sensor source markers, synchronization completion markers, and pending confirmation markers are written back to the model input sequence. The synchronous sampling and quality assurance module enters a high-risk monitoring state when the current monitoring point is in the enhanced sampling layer or the locked sampling layer, and performs supplementary sampling of key parameters for missing items, incomplete synchronization items, pending confirmation items, and quality control failure items in the rapid key parameter group; When supplementary data is available, the supplementary data replaces the original value to be confirmed or the missing bit. If the data is still unavailable after supplementary data collection, the original time sequence and missing status, pending confirmation status and low confidence status are retained, and incomplete data segments within the high-risk window are marked as low confidence window data.

7. The AI-based intelligent water quality monitoring system for aquaculture as described in claim 1, characterized in that: The monitoring status confirmation and early warning output module includes: After completing sampling level control, rapid key parameter group synchronous sampling, quality control and key parameter supplementation, the current monitoring status is confirmed by combining the current round status identification results, current sampling level, current quality category, synchronous completion status and supplementation status, and the current monitoring status is divided into stable monitoring status, suspicious change monitoring status and key risk monitoring status. The monitoring status confirmation and early warning output module is also used to generate monitoring results based on the current monitoring status. The monitoring results include at least the parameter values ​​of the current monitoring point, the current sampling level, the current monitoring status, whether it is currently in the critical risk leading stage or formation stage, the current data confidence status, and whether an early warning is needed.

8. The AI-based intelligent water quality monitoring system for aquaculture as described in claim 1, characterized in that: The data backhaul and closed-loop update module includes: The model closed-loop update is triggered when the cumulative number of missed identification events reaches a preset threshold, the cumulative number of misidentified events reaches a preset threshold, the scene changes meet the preset scene change conditions, or the sample accumulation reaches a preset sample threshold. The closed-loop update of the model includes screening high-confidence key risk window samples, key risk leading samples, recovery and stabilization samples, missed identification samples and misidentified samples from the archived data, and removing samples that do not meet the update conditions. The newly added samples are merged with the original basic training samples to form an updated training set. The candidate updated model is trained using the updated training set. The candidate updated model is compared with the current model on an independent validation sample set. When the candidate updated model meets the preset replacement conditions, the candidate updated model is replaced with the current model and redeployed according to the original deployment method.

9. An AI-based intelligent water quality monitoring system for aquaculture, characterized in that, Also includes: An abnormal behavior constraint adjustment module is located between the risk identification and sampling scheduling module and the synchronous sampling and quality assurance module. It is used to perform abnormal behavior constraint adjustment on the current monitoring process when the current monitoring point meets the preset abnormal conditions. When it is determined that the first-level abnormal behavior constraint state has been entered, the current monitoring point is switched to the enhanced sampling layer or kept in the enhanced sampling layer. The rapid key parameter group is sampled synchronously for multiple consecutive sampling rounds. The key parameters that are missing, late, unstable or not synchronized in the current round are supplemented and the continuous confirmation rounds of the current monitoring point are adjusted. When the upgrade conditions are met continuously under the first-level abnormal behavior constraint state, the current monitoring point is moved to the second-level abnormal behavior constraint state and switched to the locked sampling layer. The rapid key parameter group is sampled synchronously for multiple consecutive sampling rounds. For key parameters that are unstable or whose synchronization is incomplete, additional sampling is added, and the smoothing compensation processing within the current abnormal behavior window is suspended. When the upgrade conditions are met continuously under the Level 2 abnormal behavior constraint state, the current monitoring point is put into the Level 3 abnormal behavior constraint state and the automatic frequency reduction action is frozen. The synchronous sampling task of the rapid key parameter group is set as the highest priority task, and the supplementary sampling continues to be performed. All original monitoring data, data quality status, synchronization completion status, supplementary sampling status, monitoring status confirmation results and final handling results in the current abnormal behavior window are written into the candidate update sample pool.

10. The AI-based intelligent water quality monitoring system for aquaculture as described in claim 9, characterized in that: The abnormal behavior constraint adjustment module further includes: The abnormal behavior constraint state will exit when the following conditions are met simultaneously within three consecutive sampling rounds: the model's state judgment of the current monitoring point is consistent; there are no incomplete supplementary sampling items at the current monitoring point; the rapid key parameter group has completed synchronous sampling; and the current monitoring point has returned to a stable monitoring state or entered a suspicious change monitoring state after state confirmation. After exiting the Level 1 abnormal behavior constraint state, the current monitoring point resumes normal stratified sampling rules; After exiting the Level 2 or Level 3 abnormal behavior constraint state, the current abnormal behavior window remains in the candidate update sample pool and participates in subsequent update sample selection and model closed-loop update.

Images