An on-line evaluation method and system for biological toxicity of groundwater composite pollution

Abstract

This application discloses an online assessment method and system for the biotoxicity of groundwater complex pollution, relating to the field of environmental monitoring technology. The system includes a multi-channel monitoring device, a data processing and calculation unit, and an early warning output unit. The method includes: acquiring time-series data of the movement trajectories of each fish in multiple independent monitoring channels; extracting multi-dimensional behavioral features from each monitoring channel; determining the baseline values ​​of the behavioral features of each monitoring channel at the current moment using a dynamic baseline model; comparing the multi-dimensional behavioral features with the baseline values ​​to calculate standardized feature offsets; fusing the standardized feature offsets of each monitoring channel to obtain a comprehensive feature vector, which is then input into a biotoxicity assessment model to output a comprehensive assessment result of the biotoxicity of the water body. This method intelligently learns the complex mapping relationship between complex pollution characteristics and biological responses, achieving more sensitive, accurate, and intelligent online monitoring and assessment of groundwater complex pollution.

Description

Technical Field

[0001] This application relates to the field of environmental monitoring technology, and in particular to an online method and system for assessing the biotoxicity of groundwater complex pollution. Background Technology

[0002] Groundwater in chemical industrial parks is often contaminated by a combination of heavy metals (such as chromium, arsenic, lead, cadmium, and copper) and organic matter (such as benzene compounds, phenols, halogenated hydrocarbons, and pesticides). The toxicity of these pollutants exhibits complex effects, including synergistic and antagonistic interactions, making it difficult for traditional physicochemical monitoring to directly reflect their comprehensive biological toxicity. Water bodies typically harbor a variety of organisms, and changes in water quality directly impact the behavior and activity of these organisms. Therefore, analyzing the dose-response relationship between changes in the behavior and activity of organisms in groundwater within chemical industrial parks and the pollutants can be an effective means of assessing the comprehensive toxicity of groundwater.

[0003] Most existing biotoxicity monitoring methods analyze water toxicity by detecting the luminescence intensity of luminescent bacteria or observing the growth rate of algae. These methods can provide simple early warning of biotoxicity in water, but they still have many shortcomings: (1) insufficient sensitivity, poor early warning capability for low concentrations and complex pollution; (2) poor specificity, making it difficult to distinguish the differentiated behavior patterns caused by different types of pollutants and unable to provide clues about the nature of pollution; (3) weak anti-interference capability, easily affected by non-toxic factors such as short-term changes in ambient light and slight vibrations of equipment, leading to false alarms; (4) single assessment dimension, only able to give a binary judgment of "toxic / non-toxic", unable to quantify and classify the intensity of toxicity and its development trend and predict trends. Summary of the Invention

[0004] This application addresses the aforementioned problems and technical needs by proposing an online assessment method and system for the biotoxicity of groundwater with complex pollution. The technical solution of this application is as follows:

[0005] An online method for assessing the biotoxicity of groundwater with combined pollution includes the following steps: Acquire the time-series data of the movement trajectory of each fish in multiple independent monitoring channels. The time-series data of the movement trajectory of each fish includes the location information of the current moment and multiple consecutive historical moments before it. Based on the time-series data of the movement trajectories of each fish, multidimensional behavioral features of each monitoring channel are extracted. These multidimensional behavioral features reflect the movement status and spatial distribution of the fish within the monitoring channels. Based on the historical behavioral characteristic data of each monitoring channel under pollution-free conditions, a dynamic baseline model for each monitoring channel is established, and the benchmark value of the behavioral characteristics of each monitoring channel at the current moment is determined using the dynamic baseline model. The multidimensional behavioral characteristics of each monitoring channel are compared with the corresponding behavioral characteristic baseline value, and the standardized feature offset is calculated. By integrating the standardized feature offsets of each monitoring channel, a comprehensive feature vector is obtained. The comprehensive feature vector contains the spatiotemporal context information of the behavioral characteristics of different monitoring channels at different times. The comprehensive feature vector is input into a pre-trained biotoxicity assessment model, which outputs a comprehensive assessment result of the biotoxicity of the water body. The comprehensive assessment result includes the toxicity level, pollution type and prediction of toxicity change trend.

[0006] The further technical solution is that the multidimensional behavioral characteristics include movement intensity characteristics, movement pattern characteristics, spatial distribution characteristics and social behavior characteristics. Movement intensity characteristics represent the movement intensity of each fish, movement pattern characteristics represent the movement pattern of each fish, spatial distribution characteristics represent the distribution preference of fish in the monitoring channel space, and social behavior characteristics represent the interaction relationship between fish in the monitoring channel. The characteristics of movement intensity include instantaneous velocity, average velocity, acceleration, movement distance, and movement activity index, which represents the proportion of fish whose average velocity exceeds a predetermined threshold per unit time. The characteristics of movement pattern include linear movement index, turning frequency, circular movement radius, and fractal dimension of movement path, which represents the proportion of fish moving in a straight line per unit time. The characteristics of spatial distribution include the proportion of dwell time in different height areas of the monitoring channel, vertical migration frequency, and preferred depth. The characteristics of social behavior include average distance between individuals, aggregation index, and parallel movement frequency, which represents the ratio of the spatial variance to the mean of the number of fish in the monitoring channel.

[0007] Its further technical solution is to establish a dynamic baseline model for any monitoring channel, including: Acquire environmental parameters and behavioral characteristic data of the monitoring channel within a predetermined historical period under pollution-free conditions; convert the timestamps within the historical period into periodic time features; Using time characteristics as independent variables and environmental parameters as covariates, a Gaussian process regression model was used to fit each behavioral characteristic to obtain the dynamic baseline model corresponding to the monitoring channel. The dynamic baseline model reflects the changes in the baseline values ​​of fish behavioral characteristics under normal diurnal rhythms and environmental disturbances.

[0008] A further technical solution is to determine the baseline values ​​of behavioral feature i at the current time and multiple consecutive historical time points based on a dynamic baseline model, and to determine the mean of the baseline values ​​of behavioral feature i at the current time. and variance ; Calculate the standardized feature offset of behavioral feature i , These are the observed values ​​of behavioral feature i.

[0009] The further technical solution is that the biotoxicity assessment model is constructed based on a multi-task learning architecture. The biotoxicity assessment model includes convolutional layers, long short-term memory networks, and parallel output layers for toxicity level prediction, pollution type, and toxicity change trend. The convolutional layers are used to extract the basic features of the comprehensive feature vector. The long short-term memory network captures the temporal dependence of the comprehensive feature vector to obtain temporal features. The temporal features are then processed by the toxicity level prediction output layer to predict the toxicity intensity index. The temporal features are then processed by the pollution type output layer to predict the pollution type and its corresponding probability. Finally, the temporal features are then processed by the toxicity change trend output layer to predict the direction and magnitude of toxicity change within a predetermined future period.

[0010] The further technical solution is as follows: the toxicity level prediction output layer includes a first fully connected layer FC11, a ReLU activation function, a Dropout layer, a second fully connected layer FC12, a ReLU activation function, a third fully connected layer FC13, and a linear activation function, wherein the first fully connected layer FC11 contains 128 neurons, the second fully connected layer FC12 contains 64 neurons, and the third fully connected layer FC13 contains 1 neuron; the pollution type output layer includes a first fully connected layer FC21, a ReLU activation function, a BatchNorm layer, a second fully connected layer FC22, a ReLU activation function, a Dropout layer, a third fully connected layer FC23, and a Sigmoid activation function, all cascaded sequentially. The active function consists of three fully connected layers: the first fully connected layer FC21 contains 256 neurons, the second fully connected layer FC22 contains 128 neurons, and the third fully connected layer FC23 contains N neurons, where N is the total number of pollution types. The toxicity change trend output layer consists of a first fully connected layer FC31, a ReLU activation function, a second fully connected layer FC32, and a third fully connected layer FC33, a Softmax activation function, and a fourth fully connected layer FC34, a Sigmoid activation function, all cascaded in parallel. The first fully connected layer FC31 contains 64 neurons, the second fully connected layer FC32 contains 32 neurons, the third fully connected layer FC33 contains 3 neurons, and the fourth fully connected layer FC34 contains 1 neuron.

[0011] The further technical solution involves training a biotoxicity assessment model that includes: We acquired behavioral data of fish under different pollutant concentrations and types, and constructed a standard toxicity dataset. Each training sample in the standard toxicity dataset includes time-series data of fish movement trajectories and corresponding pollution type and toxicity level labels. For the motion trajectory time series data of each training sample in the standard toxicity dataset, a comprehensive feature vector is extracted; the comprehensive feature vector of each training sample is used as input, and the pollution type and toxicity level label of each training sample are used as output for model training, and the biotoxicity assessment model is obtained by minimizing the loss function. loss function , This represents the minimum mean square error between the predicted toxicity level and the toxicity level label output by the toxicity level prediction output layer. The binary cross-entropy loss function represents the difference between the predicted pollution type output by the pollution type output layer and the pollution type label. The multivariate cross-entropy loss function represents the direction of toxicity change output by the output layer and the actual direction of toxicity change. This represents the minimum mean square error between the toxicity change amplitude output by the toxicity change amplitude output layer and the actual toxicity change amplitude. , , , These are the weighting coefficients.

[0012] A further technical solution involves fusing the standardized feature offsets of each monitoring channel to obtain a comprehensive feature vector, including: The standardized feature offsets of each monitoring channel are constructed into a three-dimensional feature tensor according to time and monitoring channel dimensions. An attention mechanism is used to fuse the three-dimensional feature tensor to obtain a comprehensive feature vector. The attention mechanism includes a channel attention module and a temporal attention module. The channel attention module is used to learn and evaluate the importance weights of features of different monitoring channels, and the temporal attention module is used to capture features of key time periods of toxicity response.

[0013] Further technical solutions include: the online biotoxicity assessment method also includes: Based on the biotoxicity early warning results and their corresponding real-world conditions recorded by the system within a predetermined historical period, the actual false alarm rate within the predetermined historical period is determined; based on the error between the preset target false alarm rate and the actual false alarm rate, an early warning threshold is determined; when the toxicity level indicated by the comprehensive assessment result output by the biotoxicity assessment model exceeds the early warning threshold, an alarm is triggered and an early warning signal is generated.

[0014] An online assessment system for biotoxicity of groundwater complex pollution includes a multi-channel monitoring device, a data processing and calculation unit, and an early warning output unit; The multi-channel monitoring device includes multiple independent monitoring channels for accommodating test fish and water samples. Each monitoring channel is equipped with a video acquisition device and a water circulation device. The video acquisition device includes three cameras located above and on both sides of the monitoring channel and a backlight panel located behind the monitoring channel. The water circulation device includes a peristaltic pump and connecting pipes. The data processing and computing unit is connected to the video acquisition device. The data processing and computing unit receives and processes the data from the video acquisition device to extract the movement trajectory of the test fish, calculates multi-dimensional behavioral features and standardizes them based on the dynamic baseline model, fuses the standardized feature offsets and runs the biotoxicity assessment model to generate a comprehensive assessment result of the biotoxicity of the water body. The early warning output unit analyzes the comprehensive evaluation results generated by the data processing and calculation unit and issues an early warning.

[0015] The beneficial technical effects of this application are: This application presents an online biotoxicity assessment system for groundwater complex pollution. By independently and synchronously monitoring the behavioral states of test fish in different control groups through multiple channels, it effectively avoids inconsistencies in monitoring results caused by individual differences, ensuring the stability and consistency of data acquisition and providing engineering support for long-term online monitoring. This system overcomes the shortcomings of traditional single-channel monitoring, such as low sensitivity, weak anti-interference, and limited assessment dimensions, providing crucial hardware support for the online intelligent assessment of biotoxicity in groundwater complex pollution. Furthermore, this multi-channel monitoring system can simultaneously collect behavioral data of test fish with different pollution concentrations, different pollutant types, or control groups during model construction, providing high-dimensional, multi-scenario sample support for training the biotoxicity assessment model, avoiding the limitations of single-channel data, and thus helping to improve the accuracy of toxicity detection.

[0016] This application discloses an online assessment method for the biotoxicity of groundwater compound pollution. By integrating multi-monitoring channels and multi-dimensional biological behavior information, and combining deep learning technology to intelligently learn the complex mapping relationship between compound pollution characteristics and biological responses, the method significantly improves the sensitivity, accuracy, and intelligence level of online monitoring of the biotoxicity of groundwater compound pollution.

[0017] By analyzing the impact of different pollutants in groundwater of chemical industrial parks on fish behavior, four high-dimensional features of the tested fish—movement intensity, patterns, spatial distribution, and social behavior—are extracted. This allows for the capture of subtle behavioral changes under low-concentration compound pollution, addressing the insufficient sensitivity of traditional single-parameter monitoring. It provides multi-dimensional evidence for pollution type identification and toxicity classification, forming the core foundation for achieving highly sensitive and specific online intelligent assessment. The dynamic baseline model, based on Gaussian process regression, adapts to normal behavioral fluctuations caused by diurnal cycles, temperature, and other environmental factors. Standardized feature offsets enhance the specific response to toxic stress, avoiding missed detections due to non-toxic interference. Attention-based spatiotemporal feature fusion automatically weights key information from different channels and time points, improving the ability to identify the synergistic effects of compound pollution. Considering the specific impact of different pollutants on fish behavior, a multi-task deep learning model is designed to establish an intelligent mapping from behavioral characteristics to toxicity levels and pollution types. This model can output toxicity levels, pollution type tendencies, and trend predictions in parallel, overcoming the limitations of traditional binary "toxic / non-toxic" judgments and providing quantitative evidence for pollution source tracing. Attached Figure Description

[0018] Figure 1 This is a structural diagram of an online biotoxicity assessment system in an example.

[0019] Figure 2 This is a flowchart of the online biotoxicity assessment method.

[0020] Figure 3 This is a diagram of the architecture of a biotoxicity assessment model in practice.

[0021] Attached reference numerals: 1. Multi-channel monitoring device, 2. Industrial control computer, 3. Camera, 4. Peristaltic pump, 5. Backlight panel, 6. Connecting pipeline. Detailed Implementation

[0022] The specific embodiments of this application will be further described below with reference to the accompanying drawings.

[0023] This application discloses an online assessment system for the biotoxicity of groundwater with complex pollution. Please refer to [link / reference]. Figure 1 The structural diagram shown illustrates that the system includes a multi-channel monitoring device, a data processing and computing unit, and an early warning output unit. The multi-channel monitoring device 1 includes multiple independent monitoring channels that accommodate the test fish and the water samples to be tested. Figure 1 The diagram shows eight monitoring channels, each equipped with a video acquisition device and a water circulation device. The video acquisition device is used to capture high-definition images of fish behavior within each monitoring channel. The video acquisition device includes three cameras 3 located above and on both sides of the monitoring channel, and a backlight panel 5 located behind the monitoring channel. Figure 1As shown, the black backplate in the middle of the multi-channel monitoring device 1 is the backlight panel 5, and four monitoring channels are arranged in front of and behind the backlight panel 5. It should be noted that... Figure 1 The structural layout is for illustrative purposes only. In practice, the monitoring channels can be arranged horizontally or vertically. The backlight panel can also be placed on the rear or other sides of the monitoring channel, but this needs to be coordinated with the camera's placement, ensuring the camera can clearly capture the fish within the monitoring channel. The monitoring channel itself must be constructed of transparent material to facilitate video recording of the fish swimming inside. A backlight panel with a light source is placed on one side of the monitoring channel to provide illumination and improve the camera's image quality. The water circulation device includes a peristaltic pump 4 and connecting pipes 6. The connecting pipes connect the container holding the groundwater sample to the multi-channel monitoring device, and the peristaltic pump enables the circulation of inlet and outlet water.

[0024] The data processing and computing unit is communicatively connected to the video acquisition device. It receives and processes data from the video acquisition device to extract the movement trajectory of the test fish, calculates multidimensional behavioral features, standardizes them based on a dynamic baseline model, fuses the standardized feature offsets, and runs a biotoxicity assessment model to generate a comprehensive assessment result of aquatic biotoxicity. The early warning output unit analyzes the comprehensive assessment result generated by the data processing and computing unit and issues an early warning. Both the data processing and computing unit and the early warning output unit are deployed in the industrial control computer 2. The data processing and computing unit is used for data processing and control of the online biotoxicity assessment method, while the early warning output unit is used for data storage and early warning information dissemination.

[0025] Based on the aforementioned online biotoxicity assessment system, this application discloses an online biotoxicity assessment method for groundwater complex pollution. Please refer to [reference needed]. Figure 2 The flowchart shown illustrates the specific steps of this method as follows: Step 1: Obtain the time-series data of the movement trajectory of each fish in multiple independent monitoring channels. The time-series data of the movement trajectory of each fish includes the location information of the current moment and multiple consecutive historical moments before.

[0026] Given that fish, as vertebrates, have physiological metabolism and organ systems (such as the nervous and endocrine systems) more similar to higher organisms (including humans), their toxic responses to pollutants (such as neurotoxicity and reproductive toxicity) better reflect the potential risks to ecosystems and human health. Compared to lower organisms such as bacteria and algae, the toxic responses of fish more closely reflect the real impacts of complex ecosystems. Moreover, fish movement and social behavior have rich observable dimensions, which can be converted into objective data through video tracking technology. Compared to single-indicator organisms (such as the luminescence intensity of bioluminescent bacteria), fish behavior provides more comprehensive toxicity information. In addition, fish can survive for extended periods in monitoring channels, supporting dynamic toxicity assessments (such as behavioral changes during the incubation period, outbreak period, and recovery period). Therefore, this application uses fish as the test organism for biotoxicity testing. The specific fish species used can be any standard model organism, such as killifish or zebrafish.

[0027] High-definition cameras deployed on multiple independent monitoring channels synchronously collect real-time video stream data of fish within each channel. Several test fish are placed in each monitoring channel. Background modeling and subtraction are performed on each frame of the video stream data to identify and track the centroid position of each fish, obtaining the temporal data of each fish's motion trajectory. While existing techniques can be used to identify the centroid position of fish from image data, this application first employs a kNN background segmentation algorithm for background modeling. Specifically, real-time video stream data of fish in each channel is collected for an initial run of N minutes, input into the kNN model, and the background model is calculated. Then, each frame of the subsequent video stream is subtracted from the background model to obtain the position of the fish in the foreground. Finally, the position of each fish in each channel is input into a Kalman filter for tracking.

[0028] Step 2: Based on the time-series data of the movement trajectories of each fish, extract multidimensional behavioral features of each monitoring channel. These multidimensional behavioral features reflect the movement status and spatial distribution of the fish within the monitoring channel.

[0029] Analysis of the impact of different pollutants in groundwater of industrial parks on fish behavior reveals that: heavy metal pollution primarily leads to reduced activity intensity (e.g., lethargy, slowed swimming); organic pollution often induces abnormal movement patterns (e.g., alternating excitation / inhibition, non-directional swimming); and combined pollution may simultaneously affect spatial distribution (e.g., reduced vertical migration) and social behavior (e.g., decreased aggregation). Therefore, by extracting multidimensional behavioral features through classification, the mapping relationship between different pollutants and feature subsets can be clarified, providing a basis for subsequent pollution type identification.

[0030] In one embodiment, multidimensional behavioral characteristics include movement intensity characteristics, movement pattern characteristics, spatial distribution characteristics, and social behavior characteristics. Movement intensity characteristics represent the movement intensity of each fish, directly reflecting changes in the overall activity level of the fish and serving as a basic response indicator to toxic exposure. Movement pattern characteristics represent the movement patterns of each fish, reflecting abnormalities in trajectory regularity and capturing specific behaviors such as non-directional swimming caused by neurotoxicity or physiological damage. Spatial distribution characteristics represent the distribution preferences of fish within the monitoring channel, reflecting changes in fish's habitat selection; heavy metal or organic pollution may cause them to avoid specific water layers. Social behavior characteristics represent the interaction relationships between fish within the monitoring channel, reflecting the disruption of group coordination when multiple fish are in the same compartment; organic pollution often triggers group behavioral disorders involving alternating excitation and inhibition.

[0031] The characteristics of movement intensity include instantaneous velocity, average velocity, acceleration, movement distance, and movement activity index. The movement activity index represents the proportion of fish whose average velocity exceeds a predetermined threshold per unit time. The specific value of the predetermined threshold for average velocity can be customized based on experimental experience. The characteristics of movement pattern include linear motion index, turning frequency, circular motion radius, and fractal dimension of the movement path. The linear motion index represents the proportion of fish moving in a straight line per unit time. The characteristics of spatial distribution include the proportion of fish staying in different height areas of the monitoring channel, vertical migration frequency, and preferred depth. The characteristics of social behavior include average distance between individuals, aggregation index, and parallel movement frequency. The aggregation index represents the ratio of the spatial variance to the mean of the number of fish in the monitoring channel.

[0032] All of the above features can be calculated based on fish movement trajectory data. Among them, the fractal dimension of the movement path is a mathematical index that quantitatively describes the complexity of the movement trajectory, reflecting the tortuosity and spatial filling characteristics of the path. The higher the value, the more complex and irregular the path, and the more abnormal the fish behavior. Box counting or correlation integral methods are usually used to segment the movement trajectory at different scales, and the dimension value is obtained by fitting the logarithmic relationship between scale change and coverage.

[0033] The clustering index is quantified by calculating the ratio of the spatial variance to the mean of the number of fish in the monitoring channel. Specifically, the monitoring channel is divided into multiple continuous units (such as virtual grids), the number of fish in each unit is counted, and the variance of the fish count within each unit is determined. and average The aggregation index of the monitoring channel was then calculated. VMR≈1 indicates that the spatial distribution of fish is random (Poisson distribution), with no tendency to aggregate or be uniform; VMR>1 indicates that fish are aggregated. The larger the value, the higher the degree of aggregation; VMR<1 indicates that fish are uniformly or dispersedly distributed.

[0034] Step 3: Based on the historical behavioral characteristic data of each monitoring channel under pollution-free conditions, establish a dynamic baseline model for each monitoring channel, and use the dynamic baseline model to determine the benchmark value of the behavioral characteristics of each monitoring channel at the current moment.

[0035] To determine the normal behavioral characteristics of fish in unpolluted water, a dynamic baseline model of their behavior under non-toxic pollution conditions was established for each monitoring channel, serving as a benchmark. Specifically, establishing a dynamic baseline model for any monitoring channel includes: (1) Obtain the environmental parameters of the monitoring channel in a predetermined historical period under pollution-free conditions and the corresponding behavioral characteristic data of the monitoring channel, and obtain the environmental parameter sequence and behavioral characteristic sequence that change over time. The environmental parameters include pH, dissolved oxygen, conductivity, turbidity and temperature.

[0036] (2) Convert the timestamps within the historical period into periodic time features.

[0037] Because fish behavior (such as activity level and movement patterns) exhibits diurnal cyclical patterns (e.g., active during the day and dormant at night), and the original timestamps are linearly increasing values, they cannot directly reflect this cyclical nature. Therefore, it is necessary to "periodize" the time features. Taking the daily cycle as an example, the timestamp is converted into sine / cosine components of the daily cycle to capture the diurnal pattern. Specifically, the "number of hours in a day" (e.g., t=0~23) is separated from the timestamp; then the number of hours is mapped to the range of [0, 2π) radians; finally, the sine / cosine feature corresponding to this radian value is generated as the time feature.

[0038] (3) Using time characteristics as independent variables and environmental parameters as covariates, a Gaussian process regression (GPR) model is used to fit each behavioral characteristic to obtain the dynamic baseline model corresponding to the monitoring channel. The dynamic baseline model reflects the baseline value changes of fish behavioral characteristics under normal diurnal rhythm and environmental disturbance. The GPR used here is a non-parametric Bayesian method that uses kernel functions to characterize the spatiotemporal correlation and uncertainty of the data. Its commonly used kernel functions are a combination of periodic kernels (such as ExpSineSquared) and radial basis function (RBF) kernels to simultaneously describe periodic and smooth changes.

[0039] This application transforms timestamps into periodic features and introduces environmental covariates to enable the dynamic baseline model to accurately capture the diurnal rhythms and environmental dependence of medaka behavior, thereby more accurately distinguishing between "normal behavioral fluctuations" and "toxic stress responses." The trained GPR model can predict the mean and variance of behavioral characteristic baseline values ​​for any given time point and corresponding environmental conditions, taking its confidence interval (e.g., 95%) as the boundary of the normal range. This range dynamically changes with input time and environmental conditions, forming a "dynamic baseline."

[0040] Compared to traditional static baselines that use fixed thresholds, dynamic baselines, modeled using GPR, are time-dependent and can explicitly model periodic trends, allowing the baseline to adjust automatically daily. They are also environmentally adaptable, incorporating environmental covariates to enable the baseline to respond to normal changes caused by external factors such as pH, dissolved oxygen, conductivity, turbidity, and temperature. Furthermore, they can quantify uncertainty, as the prediction variance provided by GPR varies with data sparsity and noise levels, allowing the normal range width to be dynamically expanded and reduced false alarms.

[0041] Furthermore, the multidimensional behavioral characteristics of each monitoring channel are compared with the corresponding behavioral characteristic baseline values ​​to calculate the standardized feature offset.

[0042] For any behavioral feature i, the baseline values ​​of behavioral feature i at the current time and multiple consecutive historical time points are determined based on the dynamic baseline model, and the mean of the baseline values ​​of behavioral feature i at the current time is also determined. and variance ; Calculate the standardized feature offset of behavioral feature i , These are the observed values ​​of behavioral feature i.

[0043] Step 4: Integrate the standardized feature offsets of each monitoring channel to obtain a comprehensive feature vector. The comprehensive feature vector contains the spatiotemporal context information of the behavioral characteristics of different monitoring channels at different times.

[0044] Specifically, the process of fusing the standardized feature offsets of each monitoring channel to obtain a comprehensive feature vector includes: constructing a three-dimensional feature tensor (channel × time × feature) based on the time and monitoring channel dimensions of each monitoring channel; fusing the three-dimensional feature tensor using an attention mechanism to obtain the comprehensive feature vector; the attention mechanism includes a channel attention module and a temporal attention module. Since different channels may correspond to different pollution concentrations or response stages, the channel attention module is used to learn and evaluate the importance weights of features from different monitoring channels; since the toxic response has different key stages in the time dimension, including the incubation period, outbreak period, and recovery period, the temporal attention module is used to capture the features of key time periods of the toxic response.

[0045] Step 5: Input the comprehensive feature vector into the pre-trained biotoxicity assessment model and output the comprehensive assessment results of the biotoxicity of the water body. The comprehensive assessment results include the toxicity level, pollution type and toxicity change trend prediction.

[0046] Because different pollutants have specific effects on fish behavior, the degree and type of pollution can be determined by analyzing fish behavioral characteristics. For example, the greater the deviation between observed fish behavioral characteristics and baseline values, the more the observed values ​​exceed the normal range and the higher the degree of abnormality. Grading thresholds (such as mild, moderate, and severe pollution) can be set to correspond to the magnitude of the deviation. Organic pollution causes a sharp drop in dissolved oxygen, primarily through neurotoxicity and endocrine disruption. Fish often exhibit alternating excitation / inhibition or non-directional abnormal swimming. Alternating excitation / inhibition behavior mainly corresponds to dynamic fluctuations in a subset of movement intensity characteristics, specifically: Instantaneous velocity and acceleration: Drastic fluctuations occur in the short term, such as alternating sudden high-speed swimming (excitation phase) and stagnation or slow movement (inhibition phase). Motion activity index: The proportion of speeds exceeding the threshold per unit time shows a periodic change of "high-low-high," reflecting unstable switching of activity levels. Non-directional abnormal swimming behavior mainly corresponds to disorder in a subset of movement pattern characteristics, specifically: A significant decrease in the linear motion index: Fish no longer maintain a stable linear movement path, and the proportion of linear movement per unit time decreases. Increased turning frequency: Swimming direction changes frequently, the number of turns increases significantly, and the path exhibits irregular twists and turns. Increased fractal dimension: The complexity of the movement path increases, changing from a regular trajectory to a fractal characteristic similar to a random walk. Heavy metal pollution leads to increased metal ion concentration and conductivity; oxidative stress, ion balance disruption, and organ damage result in a general slowing and weakness in fish movement, manifested as a general slowing of movement, i.e., a significant decrease in average speed and acceleration, a shortened total distance traveled, and a persistently low level of the kinetic activity index (a decrease in the proportion of speeds exceeding the threshold). The movement pattern is relatively "regular but sluggish," the linear movement index may be normal or slightly decreased, but the overall movement path lacks vitality, the turning frequency decreases, and the path tends to be simplified (such as prolonged stillness or slow straight-line movement).

[0047] Therefore, this application utilizes deep learning methods to construct a biotoxicity assessment model. Specifically, the biotoxicity assessment model is built based on a multi-task learning architecture, as shown in the following figure. Figure 3As shown, the system includes convolutional layers, a long short-term memory network, and parallel output layers for toxicity level prediction, pollution type, and toxicity change trend. The toxicity level prediction output layer performs a regression task, the pollution type output layer performs a multi-label classification task, and the toxicity change trend output layer performs a time series prediction task. The convolutional layers are used to extract the basic features of the comprehensive feature vector. The long short-term memory network captures the temporal dependence of the comprehensive feature vector to obtain temporal features. The temporal features are then processed by the toxicity level prediction output layer to predict the toxicity intensity index (corresponding to a value of 0 to 100). The temporal features are then processed by the pollution type output layer to predict the pollution type and its corresponding probability, for example, the output vector [heavy metal tendency score, organic pollution tendency score, compound pollution tendency score]. The temporal features are then processed by the toxicity change trend output layer to predict the direction (increasing, stable, decreasing) and magnitude of toxicity change within a predetermined time period (such as the next hour).

[0048] The network structures for each task branch are as follows: the toxicity level prediction output layer consists of a first fully connected layer FC11, a ReLU activation function, a Dropout layer, a second fully connected layer FC12, a ReLU activation function, a third fully connected layer FC13, and a linear activation function, all cascaded sequentially. The first fully connected layer FC11 learns the nonlinear mapping from spatiotemporal features to toxicity intensity, capturing higher-order toxicity representations. The Dropout layer prevents overfitting and improves generalization ability. The second fully connected layer FC12 further refines the features, optimizing the representation for the regression task. The third fully connected layer FC13 is the output layer, directly outputting the toxicity intensity index. The first fully connected layer FC11 contains 128 neurons, the second fully connected layer FC12 contains 64 neurons, and the third fully connected layer FC13 contains 1 neuron. The pollution type output layer consists of a first fully connected layer FC21, a ReLU activation function, a BatchNorm layer, a second fully connected layer FC22, a ReLU activation function, a Dropout layer, a third fully connected layer FC23, and a Sigmoid activation function, all cascaded sequentially. The first fully connected layer FC21 is a relatively wide network that learns complex type-discriminating features. The BatchNorm layer is used to stabilize training and accelerate convergence. The second fully connected layer FC22 extracts type-specific feature combinations. The Dropout layer has a high dropout rate to prevent erroneous associations between types. The third fully connected layer FC23 is the output layer, with each neuron corresponding to a pollution type and outputting an independent probability. The first fully connected layer FC21 contains 256 neurons, the second fully connected layer FC22 contains 128 neurons, and the third fully connected layer FC23 contains N neurons, where N is the total number of pollution types. The toxicity change trend output layer consists of a first fully connected layer FC31, a ReLU activation function, a second fully connected layer FC32, and a parallel third fully connected layer FC33, a Softmax activation function, and a fourth fully connected layer FC34, a Sigmoid activation function. The first fully connected layer FC31 learns the latent pattern of trend changes; the second fully connected layer FC32 separates the intensity and direction features of the trend; the third and fourth fully connected layers FC33 and FC34 are dual-head output layers. The third fully connected layer FC33 is the trend direction head, outputting the trend direction category, and the fourth fully connected layer FC34 is the trend intensity head, outputting the magnitude of change. The first fully connected layer FC31 contains 64 neurons, the second fully connected layer FC32 contains 32 neurons, the third fully connected layer FC33 contains 3 neurons, and the fourth fully connected layer FC34 contains 1 neuron.

[0049] Based on the established model architecture, the trained biotoxicity assessment model includes: Under controlled laboratory conditions, a water environment monitoring channel containing different concentrations and types (single and compound) of heavy metals and organic pollutants was constructed to obtain fish behavior data under different pollutant concentrations and types. A standard toxicity dataset was also constructed, in which each training sample included fish movement trajectory time-series data and corresponding pollution type and toxicity level labels. For the motion trajectory time series data of each training sample in the standard toxicity dataset, perform the processing steps 1 to 4 to extract the comprehensive feature vector; use the comprehensive feature vector of each training sample as input, and use the pollution type and toxicity level label of each training sample as output to train the model, and obtain the biotoxicity assessment model by minimizing the loss function. loss function , This represents the minimum mean square error between the predicted toxicity level and the toxicity level label output by the toxicity level prediction output layer. The binary cross-entropy loss function represents the difference between the predicted pollution type output by the pollution type output layer and the pollution type label. The multivariate cross-entropy loss function represents the direction of toxicity change output by the output layer and the actual direction of toxicity change. This represents the minimum mean square error between the toxicity change amplitude output by the toxicity change amplitude output layer and the actual toxicity change amplitude. , , , These are weighting coefficients, which can be set according to the importance of the task. Since toxicity level prediction is the primary task, it has the highest weight. The value with the largest value is used for pollution type prediction, which is an important auxiliary task, and its weight is second only to the largest value. Value less than The task of predicting the direction and magnitude of toxicity changes is of the lowest importance and therefore has a lower weight. Based on this, this application sets... =0.4, =0.3, =0.2, =0.1.

[0050] Furthermore, based on the toxicity level, pollution type tendency and trend prediction output by the model, combined with the preset and dynamically adjustable early warning threshold, the final monitoring results and early warning signals can be generated.

[0051] In one embodiment, the online biotoxicity assessment method further includes: determining the actual false alarm rate within a predetermined historical period based on the biotoxicity warning results and their corresponding actual situations recorded by the system. The predetermined historical period is customized according to actual application conditions, such as being adjusted daily. That is, the warning situation is monitored for one day, and the warning threshold for the day is set according to the warning results of the previous day. The actual situation corresponding to the warning results can be determined through subsequent manual confirmation or verification results from other reliable sensors. The warning threshold is determined based on the error between the preset target false alarm rate and the actual false alarm rate. When the toxicity level indicated by the comprehensive assessment result output by the biotoxicity assessment model exceeds the warning threshold, an alarm is triggered and a warning signal is generated.

[0052] The specific method for dynamically adjusting and determining the warning threshold based on preset control rules is as follows: when the total number of warnings Total_Alert = 0, the warning threshold T_new = T_old is determined; otherwise, the candidate warning threshold T_c = T_old + Kp * (R_actual - R_desired) is determined. When the candidate warning threshold T_min≤T_c≤T_max, the warning threshold T_new = T_c is determined. When T_c < T_min, the warning threshold T_new = T_min is determined. When T_c > T_max, the warning threshold T_new = T_max is determined. Wherein, the actual false alarm rate R_actual = FP / Total_Alerts, FP is the number of false alarms, R_desired is the target false alarm rate, Kp is the proportional gain, T_old is the historical warning threshold, the historical warning threshold at the initial moment can be determined according to the normal range calculated by the dynamic baseline model, T_max is the upper limit of the warning threshold, and T_min is the lower limit of the warning threshold.

[0053] Error = R_actual - R_desired represents the error between the target false alarm rate and the actual false alarm rate. The specific value of the target false alarm rate R_desired can be set empirically, for example, to 5%. Delta = Kp * (R_actual - R_desired) represents the adjustment amount. If Error > 0 (too many false alarms), Delta is positive, T_new increases, the threshold becomes stricter, and warnings are reduced. If Error < 0 (too few false alarms), Delta is negative, T_new decreases, the threshold becomes more lenient, and warnings are increased (to be wary of missed detections). Furthermore, boundary limits must be imposed on T_new to prevent over-adjustment. The upper and lower limits of the warning threshold can be set empirically; this application sets T_max = 80 and T_min = 50.

[0054] The above descriptions are merely preferred embodiments of this application, and this application is not limited to the above embodiments. It is understood that other improvements and variations that can be directly derived or conceived by those skilled in the art without departing from the spirit and concept of this application should be considered to be included within the protection scope of this application.

Claims

1. A method for online assessment of the biotoxicity of groundwater with combined pollution, characterized in that, The online biotoxicity assessment method includes: Acquire the time-series data of the movement trajectory of each fish in multiple independent monitoring channels. The time-series data of the movement trajectory of each fish includes the location information of the current moment and multiple consecutive historical moments before it. Based on the time-series data of the movement trajectories of each fish, multidimensional behavioral features are extracted from each monitoring channel. These multidimensional behavioral features reflect the movement state and spatial distribution of the fish within the monitoring channel. Based on the historical behavioral characteristic data of each monitoring channel under pollution-free conditions, a dynamic baseline model for each monitoring channel is established, and the benchmark value of the behavioral characteristics of each monitoring channel at the current moment is determined using the dynamic baseline model; the multidimensional behavioral characteristics of each monitoring channel are compared with the corresponding behavioral characteristic baseline value, and the standardized feature offset is calculated. By fusing the standardized feature offsets of each monitoring channel, a comprehensive feature vector is obtained. The comprehensive feature vector contains the spatiotemporal context information of the behavioral features of different monitoring channels at different times. The comprehensive feature vector is input into a pre-trained biotoxicity assessment model, which outputs a comprehensive assessment result of the biotoxicity of the water body. The comprehensive assessment result includes the toxicity level, pollution type, and prediction of toxicity change trend.

2. The online biotoxicity assessment method according to claim 1, characterized in that, The multidimensional behavioral characteristics include movement intensity characteristics, movement pattern characteristics, spatial distribution characteristics, and social behavior characteristics. Movement intensity characteristics represent the movement intensity of each fish, movement pattern characteristics represent the movement pattern of each fish, spatial distribution characteristics represent the distribution preference of fish in the monitoring channel space, and social behavior characteristics represent the interaction relationship between fish in the monitoring channel. The movement intensity characteristics include instantaneous velocity, average velocity, acceleration, movement distance, and movement activity index, whereby the movement activity index represents the proportion of fish whose average velocity exceeds a predetermined threshold per unit time. The movement pattern characteristics include linear motion index, turning frequency, circular motion radius, and fractal dimension of the movement path, whereby the linear motion index represents the proportion of fish moving in a straight line per unit time. The spatial distribution characteristics include the proportion of dwell time in different height areas of the monitoring channel, vertical migration frequency, and preferred depth. The social behavior characteristics include average distance between individuals, aggregation index, and parallel movement frequency, whereby the aggregation index represents the ratio of the spatial variance to the mean of the number of fish in the monitoring channel.

3. The online biotoxicity assessment method according to claim 1, characterized in that, Establishing a dynamic baseline model for any monitoring channel includes: The environmental parameters of the monitoring channel during a predetermined historical period under pollution-free conditions, as well as the behavioral characteristic data of the monitoring channel, are obtained; the timestamps during the historical period are converted into periodic time features. Using the time characteristics as independent variables and environmental parameters as covariates, a Gaussian process regression model is used to fit each behavioral characteristic to obtain the dynamic baseline model corresponding to the monitoring channel. The dynamic baseline model reflects the changes in the baseline values ​​of fish behavioral characteristics under normal diurnal rhythms and environmental disturbances.

4. The online biotoxicity assessment method according to claim 1, characterized in that, For any behavioral feature i, the baseline values ​​of behavioral feature i at the current time and multiple consecutive historical time points are determined based on the dynamic baseline model, and the mean of the baseline values ​​of behavioral feature i at the current time is also determined. and variance ; Calculate the standardized feature offset of behavioral feature i , These are the observed values ​​of behavioral feature i.

5. The online biotoxicity assessment method according to claim 1, characterized in that, The biotoxicity assessment model is constructed based on a multi-task learning architecture. The biotoxicity assessment model includes convolutional layers, long short-term memory networks, and parallel output layers for toxicity level prediction, pollution type, and toxicity change trend. The convolutional layers are used to extract the basic features of the comprehensive feature vector. The long short-term memory network captures the temporal dependence of the comprehensive feature vector to obtain temporal features. The temporal features are then processed by the toxicity level prediction output layer to predict the toxicity intensity index. The temporal features are then processed by the pollution type output layer to predict the pollution type and its corresponding probability. Finally, the temporal features are processed by the toxicity change trend output layer to predict the direction and magnitude of toxicity change within a predetermined future time period.

6. The online biotoxicity assessment method according to claim 5, characterized in that, The toxicity level prediction output layer comprises a first fully connected layer FC11, a ReLU activation function, a Dropout layer, a second fully connected layer FC12, a ReLU activation function, a third fully connected layer FC13, and a linear activation function, wherein the first fully connected layer FC11 contains 128 neurons, the second fully connected layer FC12 contains 64 neurons, and the third fully connected layer FC13 contains 1 neuron; the pollution type output layer comprises a first fully connected layer FC21, a ReLU activation function, a BatchNorm layer, a second fully connected layer FC22, a ReLU activation function, a Dropout layer, a third fully connected layer FC23, and a Sigmoid activation function, all cascaded together. The first fully connected layer FC21 contains 256 neurons, the second fully connected layer FC22 contains 128 neurons, and the third fully connected layer FC23 contains N neurons, where N is the total number of pollution types. The toxicity change trend output layer includes a first fully connected layer FC31, a ReLU activation function, a second fully connected layer FC32, and a third fully connected layer FC33, a Softmax activation function, and a fourth fully connected layer FC34, a Sigmoid activation function, all cascaded in sequence. The first fully connected layer FC31 contains 64 neurons, the second fully connected layer FC32 contains 32 neurons, the third fully connected layer FC33 contains 3 neurons, and the fourth fully connected layer FC34 contains 1 neuron.

7. The online biotoxicity assessment method according to claim 6, characterized in that, The biotoxicity assessment model obtained through training includes: We acquired behavioral data of fish under different pollutant concentrations and types, and constructed a standard toxicity dataset. Each training sample in the standard toxicity dataset includes time-series data of fish movement trajectories and corresponding pollution type and toxicity level labels. For the motion trajectory time series data of each training sample in the standard toxicity dataset, a comprehensive feature vector is extracted; the comprehensive feature vector of each training sample is used as input, and the pollution type and toxicity level label of each training sample are used as output for model training, and the biotoxicity assessment model is obtained by minimizing the loss function. The loss function , This represents the minimum mean square error between the predicted toxicity level and the toxicity level label output by the toxicity level prediction output layer. The binary cross-entropy loss function represents the difference between the predicted pollution type output by the pollution type output layer and the pollution type label. The multivariate cross-entropy loss function represents the direction of toxicity change output by the output layer and the actual direction of toxicity change. This represents the minimum mean square error between the toxicity change amplitude output by the toxicity change amplitude output layer and the actual toxicity change amplitude. , , , These are the weighting coefficients.

8. The online biotoxicity assessment method according to claim 1, characterized in that, The integrated feature vector obtained by fusing the standardized feature offsets of each monitoring channel includes: The standardized feature offsets of each monitoring channel are constructed into a three-dimensional feature tensor according to time and monitoring channel dimensions; the three-dimensional feature tensor is fused using an attention mechanism to obtain a comprehensive feature vector; the attention mechanism includes a channel attention module and a temporal attention module. The channel attention module is used to learn and evaluate the importance weights of features of different monitoring channels, and the temporal attention module is used to capture features of key time periods of toxicity response.

9. The online biotoxicity assessment method according to claim 1, characterized in that, The online biotoxicity assessment method also includes: Based on the biotoxicity early warning results and their corresponding real-world conditions recorded by the system within a predetermined historical period, the actual false alarm rate within the predetermined historical period is determined; based on the error between the preset target false alarm rate and the actual false alarm rate, an early warning threshold is determined; when the toxicity level indicated by the comprehensive assessment result output by the biotoxicity assessment model exceeds the early warning threshold, an alarm is triggered and an early warning signal is generated.

10. An online assessment system for the biotoxicity of groundwater with combined pollution, characterized in that, The system for implementing the method as described in any one of claims 1-9 includes a multi-channel monitoring device, a data processing and computing unit, and an early warning output unit; The multi-channel monitoring device includes multiple independent monitoring channels for accommodating test fish and water samples. Each monitoring channel is equipped with a video acquisition device and a water circulation device. The video acquisition device includes three cameras located above and on both sides of the monitoring channel and a backlight panel located behind the monitoring channel. The water circulation device includes a peristaltic pump and connecting pipes. The data processing and computing unit is communicatively connected to the video acquisition device. The data processing and computing unit receives and processes the data from the video acquisition device to extract the movement trajectory of the test fish, calculates multidimensional behavioral features and standardizes them based on a dynamic baseline model, fuses the standardized feature offsets and runs a biotoxicity assessment model to generate a comprehensive assessment result of aquatic biotoxicity. The early warning output unit analyzes and issues an early warning based on the comprehensive evaluation results generated by the data processing and calculation unit.

Images