Aquatic biosensor and equipment for monitoring water pollutants and water quality

Abstract

The invention provides an aquatic biosensor and equipment for monitoring water pollutants and water quality, living zooplankton is accommodated and exposed in a water sample to be detected through a biological sensing unit, and multi-modal data of the zooplankton is acquired through a multi-modal information acquisition unit; carrying out processing and feature extraction on the multi-modal data through a signal processing and feature extraction unit to obtain feature parameters; and performing data fusion analysis on the characteristic parameters by adopting an artificial intelligence model through a water quality evaluation and early warning decision unit to obtain a water quality comprehensive toxicity index, a pollutant type judgment result and an early warning grade. According to the technical scheme, the real-time data acquisition and analysis capability, the accurate early warning function and the pollution tracing capability can be realized, the system is suitable for various application scenes, and the continuous online water quality safety monitoring can be realized.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and biosensing technology, and more specifically, to aquatic biological sensors and devices for monitoring water pollutants and water quality. Background Technology

[0002] Currently, water quality monitoring mainly relies on traditional physicochemical monitoring technologies, which have inherent defects of "lag" and "one-sidedness". This technology uses specific instruments to quantitatively analyze known chemical substances (such as heavy metals, COD, ammonia nitrogen, etc.) in water bodies. Its limitations are: (1) Lag: The analysis process usually requires multiple steps such as on-site sampling and laboratory testing, which takes several hours to several days. It cannot respond to rapid changes in water quality in a timely manner and is difficult to provide early warning for sudden pollution events. (2) One-sidedness: It can only detect a limited number of pollutants in the target list and cannot identify the synergistic, antagonistic and other complex toxic effects between unknown compounds or chemical pollutants, which may underestimate the actual ecological risks of water bodies. The fact that the concentration of a chemical component meets the standard does not mean that the water quality is safe for organisms.

[0003] Secondly, while existing biomonitoring technologies can reflect the comprehensive biological effects of pollutants, they still have significant shortcomings in terms of sensitivity, information scope, and automation. Biomonitoring uses the response of living organisms to toxic stress to assess water quality, but its mainstream methods each have their own deficiencies: Fish toxicity testing: Although fish are used as indicator organisms, they have high ecological relevance, but their response speed is slow (usually several days), the experimental cost is high, the individual differences are large, and the ethical pressure is increasing, making them difficult to apply to high-frequency, online monitoring scenarios.

[0004] The luminescent bacteria method, based on the principle of inhibiting bacterial luminescence intensity, has the advantages of being rapid and simple. However, its indicator organisms are prokaryotes, which have vastly different physiological structures from higher aquatic organisms. The reliability of extrapolating the response results to real ecosystems is questionable, and the information provided is extremely limited (only luminescence intensity as an indicator), making it unable to identify the type of pollution.

[0005] Traditional zooplankton monitoring methods typically focus on zooplankton such as Daphnia magna, observing their mortality rates or behavioral abnormalities. However, this method heavily relies on manual microscopic observation, is highly subjective, inefficient, and cannot achieve quantification or standardization. Furthermore, it cannot capture and record rapid and subtle early behavioral changes, nor does it possess the capability for real-time online monitoring.

[0006] In summary, the current field of water quality monitoring urgently needs a technology that can simulate a real aquatic ecosystem, be sensitive to a variety of pollutants, provide abundant early warning information, and achieve automated real-time online monitoring. Summary of the Invention

[0007] In view of this, the present invention proposes an aquatic biological sensor and device for monitoring water pollutants and water quality, in order to solve the problems existing in the prior art.

[0008] To achieve the above objectives, the present invention proposes an aquatic biosensor for monitoring water pollutants and water quality, comprising: The system comprises a biosensing unit, a multimodal information acquisition unit, a signal processing and feature extraction unit, and a water quality assessment and early warning decision-making unit. The process involves using a biosensing unit to contain and expose live zooplankton in the water sample to be tested, a multimodal information acquisition unit to collect multimodal data of the zooplankton, a signal processing and feature extraction unit to process and extract features from the multimodal data to obtain feature parameters, and a water quality assessment and early warning decision-making unit to perform data fusion analysis on the feature parameters using an artificial intelligence model to obtain the comprehensive water toxicity index, pollutant type identification results, and early warning level.

[0009] Optionally, the biosensing unit includes a first microchamber and a second microchamber, with the first microchamber containing and exposing large zooplankton and the second microchamber containing and exposing small zooplankton; wherein the first microchamber and the second microchamber are connected to a common inlet and outlet via a microflow channel, and both the first microchamber and the second microchamber are provided with a temperature control module for temperature control and a light cycle control module for supplemental illumination.

[0010] Optionally, in the multimodal information acquisition unit, the multimodal data includes motion video data of large zooplankton and impedance data of small zooplankton; the motion video data is acquired through a microscopic optical system and an integrated vertical motion monitoring component, and the impedance scan data is acquired through a microelectrode array.

[0011] Optionally, in the signal processing and feature extraction unit, behavioral and impedance features are extracted from the multimodal data. Specifically, noise removal and target recognition and tracking are performed on the motion video data of large zooplankton. Based on the target recognition and tracking results, behavioral feature parameters are obtained, including average swimming speed, trajectory complexity, instantaneous velocity, acceleration, tumbling frequency, heart rate, and vertical motion rate. For small zooplankton, impedance features are preprocessed and feature extracted to obtain spectral feature parameters, including low-frequency impedance, high-frequency impedance, and characteristic frequency.

[0012] Optionally, in the water quality assessment and early warning decision-making unit, the artificial intelligence model adopts a fusion model based on machine learning algorithms. The impedance characteristic parameters of the small zooplankton and the behavioral characteristic parameters of the large zooplankton are used as input data for the artificial intelligence model. The output data are the comprehensive water quality toxicity index, pollutant type identification results, and early warning level. The pollutant type identification results include one or more pollutant types and corresponding classification probabilities. The pollutant types include heavy metals, organophosphorus pesticides, pyrethroid pesticides, and algal toxins.

[0013] Optionally, in the water quality assessment and early warning decision-making unit, the artificial intelligence model is a hybrid model of a convolutional neural network for processing impedance characteristic parameters of small zooplankton and a long short-term memory network for processing behavioral characteristic parameters of large zooplankton. The output data of the convolutional neural network and the long short-term memory network are weighted and fused through an attention mechanism, and the final result is output through a multi-branch task structure.

[0014] On the other hand, the present invention also provides a water quality monitoring device, including: an aquatic biological sensor as described above and a peripheral auxiliary module; the peripheral auxiliary module includes a water sample pretreatment module, a main control and communication module, a power supply module and a protective housing; The water sample pretreatment module filters, maintains constant temperature and gas balances the water sample before it enters the biosensor unit; the main control and communication module controls the operation of the aquatic biosensor and peripheral auxiliary modules, and uploads the output results of the water quality assessment and early warning decision unit; the power supply module provides power; and the protective shell provides physical protection for the aquatic biosensor.

[0015] Optionally, the aquatic biosensor is based on the multidimensional biological response signals of zooplankton and uses a dose-response relationship model to deduce the concentration of chemical pollutants in the water.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) The monitoring information dimensions are unprecedentedly rich, and the assessment is more comprehensive: Through the multimodal sensing strategy of "macrobehavior + microomics", and by utilizing the sensitivity differences of different groups of zooplankton, an unprecedented amount of biological effect information has been obtained, making the water quality assessment results closer to the real ecological effects and avoiding one-sidedness.

[0017] 2) The timeliness of early warning is greatly improved, enabling early alarms: Since it captures the most sensitive early stress signals of zooplankton (such as changes in characteristic biomarkers and subtle behavioral changes), rather than the final death endpoint, it can issue an alarm when the pollutant concentration is low, with the warning time being several hours to several days earlier than traditional methods.

[0018] 3) Possesses pollution type identification capabilities, supporting precise source tracing: The innovative AI fusion model can analyze complex biological response patterns and identify the types of major pollutants, providing key clues for tracing pollution sources and taking targeted measures, which is something that traditional biological monitoring technologies cannot achieve.

[0019] 4) High degree of automation and intelligence, suitable for long-term unattended operation: From water sample introduction, signal acquisition, data processing to early warning release, the whole process is automated, greatly reducing human intervention and enabling 24 / 7 uninterrupted online monitoring. It is especially suitable for long-term monitoring of key locations such as water sources and drainage outlets.

[0020] 5) Wide range of applications and strong versatility: The modular design of the equipment makes it easy to adapt to the monitoring needs of different water environments (such as rivers, lakes, reservoirs, and nearshore areas). By adjusting the types of zooplankton in the chamber and the parameters of the AI ​​model, the configuration can be optimized for the main risks in a specific area.

[0021] 6) Achieving the transformation from biological monitoring to chemical quantitative analysis: By establishing a dose-response relationship model, the concentration of chemical pollutants in water can be deduced based on the multidimensional biological response signals of zooplankton. This provides a scientific tool for screening priority pollutants for specific water areas and for setting water quality standards, filling the technological gap in the transformation from biological monitoring to chemical quantitative analysis. Attached Figure Description

[0022] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a diagram illustrating the overall system structure in an embodiment of the present invention. Figure 2 This is a schematic diagram of the microfluidic chip structure of the biosensor in an embodiment of the present invention; Figure 3 This is a diagram of an integrated monitoring device for three-dimensional behavioral characteristics of large zooplankton in an embodiment of the present invention; Figure 4 This is a diagram of an integrated monitoring device for impedance and vitality of small zooplankton in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the construction and workflow of a multi-dimensional data-driven artificial intelligence water quality assessment model in an embodiment of the present invention.

[0023] Figure 6 This is a flowchart illustrating the construction technology and equipment implementation of the multi-level aquatic organism monitoring system in this embodiment of the invention. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] Existing technologies, particularly traditional water quality monitoring techniques, are increasingly inadequate in meeting the demands for real-time, accurate, and multi-dimensional monitoring. This invention proposes an innovative water quality monitoring biosensor based on zooplankton and its supporting equipment. This technology organically integrates live biological sensing mechanisms, multimodal sensing technology, and advanced artificial intelligence algorithms to construct an innovative water quality monitoring technology system.

[0026] At the core technology architecture level, the device selects zooplankton from multiple ecological functional groups, such as cladocerans, rotifers, and copepods, as live sensing elements. These zooplankton are significantly sensitive to changes in water quality, and different groups exhibit different response characteristics to pollutants, laying a biological foundation for achieving multi-dimensional water quality monitoring.

[0027] Regarding monitoring strategies and algorithm design, this invention proposes a differentiated collaborative monitoring strategy that combines "behavioral analysis of large zooplankton with omics detection of small zooplankton." For both large and small zooplankton with a body length greater than 500 micrometers, computer vision and impedance monitoring are used for data acquisition. Based on this, a deep learning model is constructed, such as a water quality early warning algorithm model based on a CNN-LSTM hybrid neural network: the convolutional neural network (CNN) is used to automatically extract spatial features of multimodal data, and the long short-term memory network (LSTM) is responsible for processing time series information. Through the collaborative operation of the two, a quantitative assessment of the comprehensive toxicity of water quality is achieved, and it can accurately identify 12 common pollutant types, including heavy metal pollution, organic pollution, and nutrient excess, with an early warning accuracy rate of over 95%.

[0028] The device has significant advantages in terms of technology: it adopts a highly integrated design, integrating the biosensing unit and multimodal data acquisition module into a portable chassis; it has minute-level data acquisition and analysis capabilities; the monitoring indicators cover three major categories of parameters, including biological behavior, physiological and biochemical parameters, and environmental physicochemical parameters; its accurate early warning function and pollution source tracing capabilities effectively overcome the limitations of traditional physicochemical monitoring equipment.

[0029] This invention focuses on the interdisciplinary innovation of environmental monitoring and biosensing technologies, constructing an intelligent monitoring system and equipment with multi-technology integration characteristics. The system deeply integrates a live biosensor module, realizing the biotransformation of toxicity signals through the physiological responses of aquatic organisms to pollutants; an innovatively designed microfluidic chip structure enables automated sample preprocessing and precise manipulation, ensuring the timeliness and stability of monitoring; simultaneously, the onboard artificial intelligence data analysis module employs deep learning algorithms to perform real-time processing and feature extraction of multi-source heterogeneous data. This system and equipment can overcome the limitations of traditional monitoring technologies, achieving in-situ deployment, real-time dynamic monitoring, continuous online assessment of comprehensive water toxicity, and early intelligent warning of pollutants.

[0030] In view of the above, the present invention provides the following technical solution: a zooplankton biosensor for water quality monitoring, comprising: A biosensing unit is used to contain and expose live zooplankton in a water sample to be tested, serving as a sensitive element for sensing the toxic effects of water. A multimodal information acquisition unit, coupled to the biosensing unit, is used to synchronously and in real time acquire biological response signals of zooplankton exposed to the water sample under at least two different modes. The signal processing and feature extraction unit is communicatively connected to the multimodal information acquisition unit and is used to process the acquired multimodal biological response signals and extract quantitative feature parameters from them. The water quality assessment and early warning decision unit is communicatively connected to the signal processing and feature extraction unit. It is used to receive the quantitative feature parameters, perform data fusion analysis based on a preset artificial intelligence model, and output the comprehensive water toxicity index, pollutant type identification results, and early warning level.

[0031] Specifically, the biosensing unit includes: a first microchamber configured to contain large zooplankton; and a second microchamber configured to contain small zooplankton. The first and second micro-chambers are connected to a common inlet and outlet via a microflow channel to enable parallel or sequential flow of the water sample to be tested. Each micro-chamber is equipped with an independent temperature control module and a light cycle control module.

[0032] Specifically, the volume of the first microchamber is 50-200 μL, its internal structure is suitable for the swimming of cladocerans, and it is equipped with a first observation window; The second microchamber has a volume of 5-20 μL, its internal structure is suitable for the activity of rotifers, and it is equipped with microelectrodes.

[0033] Specifically, the multimodal information acquisition unit includes: The behavioral imaging acquisition module, targeting the first microchamber, includes at least one high-speed camera and a matching microscopic optical system for continuously capturing motion videos of large zooplankton; the impedance analysis acquisition module, targeting the second microchamber, uses microelectrode array technology to perform in-situ, non-destructive impedance scanning of single or multiple small zooplankton.

[0034] Specifically, the behavioral image acquisition module further integrates a vertical motion monitoring component, which uses laser triangulation or confocal imaging technology to accurately measure the vertical positional changes of zooplankton, thereby quantifying their vertical migration behavior.

[0035] Specifically, the signal processing and feature extraction unit includes: The image processing subunit is configured to perform background subtraction, target recognition and tracking on the motion video, and calculate multiple behavioral feature parameters including but not limited to: average swimming speed, motion trajectory complexity, instantaneous velocity, acceleration, tumbling frequency, heart rate, and vertical motion rate. The spectral processing subunit is configured to preprocess impedance data (such as denoising and baseline correction) and extract low-frequency impedance, high-frequency impedance, characteristic frequencies, or perform principal component analysis to obtain principal component scores.

[0036] Specifically, the artificial intelligence model preset in the water quality assessment and early warning decision-making unit is a fusion model based on machine learning algorithms. This model is trained using multiple parameters extracted from the behavioral characteristics of macroplankton and the impedance characteristics of microplankton as input feature vectors; the model output includes: A continuous comprehensive water toxicity index; classification probabilities of one or more pollutant types (such as heavy metals, organophosphorus pesticides, pyrethroid pesticides, algal toxins); graded early warning signals (such as normal, attention, alert, alarm) triggered based on the toxicity index and classification probabilities.

[0037] Specifically, the artificial intelligence model is a hybrid model of convolutional neural networks and long short-term memory networks, wherein the convolutional neural network is used to process two-dimensional feature maps extracted from spectral data, and the long short-term memory network is used to process time series data composed of behavioral feature parameters.

[0038] The present invention also provides a water quality monitoring device, comprising the zooplankton biosensor as described above, and further comprising: The water sample pretreatment module is used to filter, keep the temperature constant, and balance the gas in the water sample before it enters the biosensing unit; the main control and communication module is used to coordinate and control the operation of the entire device and transmit the output results of the water quality assessment and early warning decision unit to the remote monitoring center via wired or wireless means; the power supply module provides power to all components of the device; and the protective shell provides physical protection and anti-biofouling function for the internal components.

[0039] Specifically, the device also integrates an automatic cleaning and maintenance module, which can periodically inject cleaning fluid and culture medium into the microchamber to maintain the activity of zooplankton and remove biofouling, enabling the device to operate unattended for a long time.

[0040] Specifically, the biosensor can deduce the concentration of chemical pollutants in water bodies based on the multidimensional biological response signals of zooplankton through a dose-response relationship model, thereby realizing the transformation from biological monitoring effects to quantitative analysis of chemical pollution.

[0041] The above technical solution will be described in detail with reference to the accompanying drawings: like Figure 1 As shown, the four-layer architecture design adopted in this invention includes a perception layer, a transmission layer, a processing layer, and an application layer.

[0042] The sensing layer consists of a biosensing unit and a multimodal information acquisition unit, responsible for real-time acquisition of planktonic behavior data and environmental parameters. The biosensing unit utilizes microfluidic chip technology to contain and expose live zooplankton in the water sample, enabling direct interaction between organisms and the environment. The multimodal information acquisition unit integrates multiple sensors, including video acquisition and impedance measurement, to simultaneously acquire the visual and electrical characteristics of plankton.

[0043] The transport layer is responsible for reliably and efficiently transmitting data from the perception layer to the processing layer. It employs a hybrid communication protocol, including short-range low-power protocols (LoRa, ZigBee) and wide-area networks (NB-IoT, 4G / 5G), ensuring real-time data transmission latency is less than 100ms. Simultaneously, edge computing nodes are deployed at the data source to achieve 90% localization of image recognition processing, reducing cloud load.

[0044] The processing layer adopts a "signal processing and feature extraction unit" architecture, which is responsible for real-time processing and feature extraction of the acquired data. This layer is built on a time series database (TSDB) and a streaming computing engine, and can store high-frequency sampled data (such as behavioral data every 10ms). It solves the clock synchronization problem of different devices through timestamp alignment technology.

[0045] The application layer includes water quality assessment and early warning decision-making units, built on a unified IoT cloud platform based on cloud computing technology. This layer integrates a multi-dimensional data-driven AI water quality assessment model, enabling functions such as calculating the comprehensive toxicity index of water, tracing pollution sources, and providing precise early warnings.

[0046] Specifically, the biosensing unit adopts a modular design, containing multiple independent biological culture chambers that can simultaneously accommodate different types of plankton (such as Daphnia magna and rotifers). Each chamber is equipped with an independent fluid control system that can precisely adjust environmental parameters such as water flow rate, temperature, and pH value. The chambers are made of transparent material, facilitating optical observation and imaging.

[0047] The multimodal information acquisition unit integrates a variety of sensor technologies, including: (1) a high-resolution camera system for capturing the behavioral trajectory and morphological characteristics of plankton; (2) an impedance measurement module for detecting the vitality and physiological state of small zooplankton; and (3) a water quality sensor array for real-time monitoring of environmental parameters such as dissolved oxygen, pH value, and temperature.

[0048] The signal processing and feature extraction unit employs FPGA hardware acceleration technology, coupled with the H.266 encoding protocol and low-latency transmission protocol, achieving video encoding latency as low as 30ms, decoding latency controlled within 50ms, and end-to-end latency consistently maintained at around 100ms. This unit enables real-time fusion processing of multimodal data, including functions such as behavior trajectory recognition, impedance parameter extraction, and environmental feature analysis.

[0049] The water quality assessment and early warning decision-making unit is built based on a deep learning model. This unit can calculate the comprehensive toxicity index of water quality in real time based on multimodal data, and predict the trend of water quality changes through machine learning algorithms, so as to achieve accurate early warning and pollution source tracing.

[0050] Standardized Interface and Data Bus Design: This invention employs a standardized interface design, supporting all mainstream fieldbus protocols, including Profibus, Profinet, Ethernet-IP, DeviceNet, CC-Link, and EtherCAT. Consistent bidirectional communication is achieved between the control system and the lowest sensor / actuator level via a standard unshielded three-core cable.

[0051] The data bus adopts CAN bus technology, supporting stable concurrent communication from multiple sensors (≥50). The CAN bus uses a typical linear topology, with all nodes connected in parallel to the two-wire bus, offering advantages such as bus topology, stability, and hard real-time performance. Simultaneously, the system also supports the IO-Link interface, a serial communication protocol similar to the I2C bus, enabling efficient communication between industrial automation controllers and sensors.

[0052] Efficient data transmission and collaborative processing mechanism: The system employs multi-sensor data fusion technology to achieve efficient data transmission and collaborative processing. Data fusion is performed at different information levels, including data-level fusion, feature-level fusion, and decision-level fusion.

[0053] At the data-level fusion level, the system directly concatenates raw data from different sensors to form a high-dimensional vector input to the deep learning model. At the feature-level fusion level, features are extracted from data from different sensors and then fused into feature vectors. At the decision-level fusion level, based on the data features obtained from feature-level fusion, discrimination, classification, and logical operations are performed to form the final decision result.

[0054] The system also employs advanced timing synchronization technology. It achieves cross-device time alignment by combining PTP (Precision Time Protocol) or NTP with hardware pulse synchronization signals (such as GPIO triggering). A "time alignment buffer" structure is used to maintain a circular queue with timestamp indexes for each data stream, ensuring accurate synchronization of multimodal data.

[0055] The collaborative processing mechanism is based on a distributed architecture design, achieving hierarchical data processing through the coordinated work of edge computing and cloud computing. Edge computing nodes are responsible for the initial processing and feature extraction of real-time data, while the cloud platform is responsible for training complex models and deep analysis. This architecture ensures both real-time data processing and full utilization of cloud computing resources.

[0056] like Figure 2 A schematic diagram of the structure of the biosensor microfluidic chip is provided: The biosensor microfluidic chip of the present invention adopts an innovative design of large and small chambers. By integrating monitoring channels of different sizes on a single chip, it can meet the monitoring needs of various types of planktonic organisms (including planktonic algae, protozoa, rotifers, cladocerans, etc.).

[0057] The chip design is based on the size differences of plankton, and includes three different chamber sizes: an algae detection channel, a protozoan and rotifer detection channel, and a cladoceran and copepod detection channel. This design fully considers the morphological characteristics and movement properties of different plankton, ensuring that each organism can be monitored in the most suitable environment.

[0058] The chamber design employs a non-enclosed structure, with the cover plate and substrate secured by mortise and tenon joints, hot melt adhesive bonding, or spring clips, forming a convex-concave mortise and tenon structure. This detachable design solves the problems of narrow channels, complex structures, easy clogging, and difficulty in cleaning inherent in traditional microfluidic chips. When the chip requires cleaning, simply separate the substrate and cover plate, rinse with distilled water and 75% ethanol, and it can be reused repeatedly.

[0059] Design details of the small chamber (for algae detection): The small chamber is mainly used for small plankton and algae detection. Its design parameters are precisely calculated: depth 0.05 - 0.3 mm, width 0.5 - 3 mm. This size design can accommodate single-cell algae and small colonial algae, while restricting their movement in the vertical direction, facilitating optical detection and image analysis.

[0060] The small chamber adopts a "ji" - shaped symmetric structure, consisting of two connecting segments and a detection segment. The detection segment is located on the upper surface of the substrate, facilitating the acquisition of impedance data. The cross - section of the chamber is square, with a depth of 20 μm\5 mm and a width of 50 μm\10 mm. The design with a width greater than the depth is conducive to observing the movement of algae in the horizontal direction.

[0061] The material selection for the small chamber includes transparent materials such as glass, quartz, polydimethylsiloxane (PDMS), polymethyl methacrylate (PMMA), polycarbonate (PC), etc. Among them, PDMS material has good biocompatibility, optical transparency, and easy - processing characteristics, and is the most commonly used material for microfluidic chips. Glass and quartz materials have better chemical stability and optical properties, and are suitable for high - precision detection.

[0062] Design details of the large chamber (for large zooplankton detection): The large chamber is mainly used for large zooplankton detection, including cladoceran and copepod detection channels. Its design parameters are: depth 0.8 - 3 mm, width 5 - 10 mm. This larger space design can accommodate larger zooplankton such as Daphnia magna and Cyclops, ensuring that they can move freely in the chamber.

[0063] The large chamber adopts a linear design and is arranged in parallel on the chip. The depth and width of the chamber are optimized, which not only ensures the activity space of large zooplankton but also facilitates optical imaging and behavior analysis. The bottom of the chamber is flat and smooth, reducing interference with the movement of animals.

[0064] The large chamber is also equipped with a special fluid control system, which can precisely adjust the water flow velocity and direction. By setting micro - pumps and valves at the inlet and outlet of the chamber, the circulation and replacement of the culture solution can be realized to maintain a suitable living environment. At the same time, the fluid control system can also generate specific water flow stimuli to test the behavioral responses of zooplankton.

[0065] Fluid control and sample injection system: The fluid control system of the microfluidic chip adopts advanced micro - pump and micro - valve technologies to achieve precise flow control and direction adjustment. The system integrates multiple micro - pumps, which can independently control the fluid flow of each chamber, and the flow velocity adjustment range is 0.1 - 100 μL / min.

[0066] The sample introduction system is designed with automation in mind, including a sample reservoir, injection pump, filter, and waste collector. The sample is delivered to the chip inlet via the injection pump, passes through the filter to remove large particles, and then enters the detection chamber. The injection process is driven by negative pressure to ensure stable sample delivery and precise control.

[0067] The chip also features a unique bubble removal system. Since dissolved gases in water easily form bubbles in the microchannels, affecting detection results, the system uses a combination of ultrasonic vibration and temperature control to remove these bubbles. By placing an ultrasonic vibrator at the inlet, dissolved gases are converted into tiny bubbles, which are then discharged through a special exhaust channel.

[0068] The fluid control software features a graphical user interface, allowing users to set parameters such as flow rate, flow direction, and replacement cycle via a touchscreen. The system also has an automatic calibration function, which automatically adjusts control parameters based on changes in ambient temperature and pressure to ensure the accuracy and stability of fluid control.

[0069] like Figure 3 The integrated monitoring device for three-dimensional behavioral characteristics of large zooplankton shown is designed specifically for large zooplankton such as Daphnia magna. It features a portable, box-like structure assembled from black acrylic material. The device measures 300×200×200mm and weighs approximately 5kg, making it easy to carry and use in the field.

[0070] The equipment mainly consists of four core components: the enclosure, the camera unit, the lighting unit, and the zooplankton containment device. The interior of the enclosure has a matte black finish to reduce the impact of light reflection on the shooting quality. Acrylic tracks are connected to the bottom outer surfaces of the left and right side panels of the enclosure, facilitating the installation and adjustment of the movable panels. A movable panel is located under the front panel of the enclosure, allowing for adjustment of the observation angle as needed.

[0071] The camera unit is located on the upper surface of the enclosure's top panel and features a non-fixed connection design, allowing for adjustment of its position and angle as needed. The lighting unit is located on the lower surface of the top panel and is magnetically attached, allowing for easy adjustment of its position to ensure optimal lighting conditions. The zooplankton enclosure is made of transparent material and located at the bottom of the enclosure, also featuring a non-fixed connection design for easy replacement and cleaning.

[0072] The equipment employs an innovative dual-camera tracking system, capable of simultaneously capturing the subtle movements of zooplankton from multiple angles with high precision and reliability. The dual-camera system includes two high-resolution industrial cameras with a resolution of 1920×1080 pixels and a frame rate of 60fps, clearly capturing the rapid movements of zooplankton.

[0073] The 3D reconstruction algorithm is based on the principle of stereo vision and achieves 3D coordinate reconstruction by calculating the disparity between two cameras. The system first calibrates the cameras to obtain their intrinsic and extrinsic parameters, and then calculates the position of zooplankton in 3D space using feature point matching and triangulation algorithms. The 3D positioning accuracy can reach the sub-millimeter level, accurately recording the spatial movement trajectory of zooplankton.

[0074] The device also integrates depth camera technology, enabling direct acquisition of depth information from zooplankton. Utilizing structured light or Time-of-Flight (ToF) principles, the depth camera outputs real-time depth images, providing an additional dimension of information for 3D behavior analysis. The addition of depth information significantly improves the accuracy of behavior recognition, especially when zooplankton occlude each other.

[0075] The video acquisition system uses a high-speed image acquisition card, supporting simultaneous acquisition from dual cameras at a capture frequency of up to 100fps. The system is equipped with dedicated image acquisition software capable of real-time display and recording of video data. The software offers multiple shooting modes, including continuous shooting, timed shooting, and triggered shooting, to meet various experimental needs.

[0076] The image processing system is based on deep learning algorithms. It employs k-means clustering for automatic background subtraction, uses machine learning methods (random forest and support vector machine) for zooplankton classification, and uses the Simple Online Real-Time Tracking (SORT) algorithm to track the location of each zooplankton. This algorithm demonstrates excellent performance in terms of recognition accuracy, recall, F1 score, and ID switching, achieving a recognition accuracy of 79.64% and a recall of 80.63%.

[0077] The image processing workflow includes four steps: image preprocessing, object detection, feature extraction, and trajectory tracking. Image preprocessing includes operations such as grayscale conversion, filtering, and binarization to improve image quality. Object detection employs a deep learning model to accurately identify zooplankton and locate their positions. Feature extraction algorithms extract morphological, kinematic, and behavioral features of zooplankton. The trajectory tracking algorithm uses a Kalman filter to predict the zooplankton's trajectory and performs data association using the Hungarian algorithm.

[0078] Behavioral feature extraction algorithms can automatically analyze various behavioral characteristics of zooplankton, including movement speed, direction of movement, activity range, and swimming patterns. The algorithm first separates zooplankton using background subtraction techniques, and then extracts individuals using morphological operations and connected component analysis.

[0079] For behavioral analysis of Daphnia magna, the system can identify various behavioral patterns, including normal swimming, jumping, rotating, and stillness. By analyzing the curvature, speed changes, and spatial distribution of the swimming trajectory, the behavioral state of Daphnia magna can be determined. For example, during normal swimming, the trajectory is relatively smooth and the speed is stable; when stimulated, it will exhibit rapid jumping behavior; when sick or dead, it will remain still.

[0080] The system can also analyze the group behavior characteristics of zooplankton, including group cohesion, inter-individual spacing, and synchronicity. Through multi-target tracking technology, the system can simultaneously track multiple individuals and analyze their interactions. These group behavior characteristics are of great significance for assessing water pollution and ecotoxicity.

[0081] like Figure 4 The image shows an integrated monitoring device for impedance and vitality of small zooplankton. The integrated impedance and vitality monitoring device for small zooplankton is based on microfluidic bioimpedance sensing technology, a label-free detection technique based on electrical principles. The basic principle of impedance measurement is to identify and analyze cell state by utilizing the differences in the electrical properties of cells at different frequencies.

[0082] When an electric current passes through a solution containing zooplankton, it encounters varying impedances. The resistance of the cell membrane to electric current depends on the membrane's integrity, permeability, and capacitance. Normal cells have intact cell membranes and exhibit high impedance; while damaged or dead cells, due to the loss of membrane integrity, show a significant decrease in impedance.

[0083] The device employs microelectrode array technology, integrating multiple microelectrodes onto a microfluidic chip. The electrode materials are inert, such as gold, platinum, or carbon, ensuring no chemical reactions occur during the electrochemical process. The electrode spacing is designed to be 10-100 μm, accommodating the detection of zooplankton of varying sizes.

[0084] Impedance measurements were performed using AC excitation with a frequency range of 100 Hz to 10 MHz. At different frequencies, the cell membrane exhibits different electrical properties: at low frequencies, it primarily reflects the resistive properties of the cell membrane, while at high frequencies, it primarily reflects the capacitive properties. By analyzing the change in impedance with frequency, the electrical fingerprint of the cell can be obtained, which can be used to identify cell types and assess cell viability.

[0085] For monitoring small zooplankton such as rotifers, the device employs a special microfluidic chip design. The rotifer detection channel is designed with a depth of 0.1-0.8 mm and a width of 2-8 mm, capable of accommodating rotifers with a body length of 100-500 μm.

[0086] The chip employs a three-layer structure design, comprising an upper cover plate, a middle microchannel layer, and a lower electrode layer. The microchannels are made of polydimethylsiloxane (PDMS) material and fabricated using soft photolithography. The electrode layer is made of silicon or glass substrates and fabricated using photolithography and metal deposition processes. The three-layer structure is sealed using plasma bonding technology to ensure fluid-tight performance.

[0087] Rotifer monitoring employs flow cytometry, where samples are driven through a detection zone by a micro-pump. Within the detection zone, rotifers pass through the electrode gaps one by one, generating characteristic impedance change signals. The system can record the impedance change curve of each rotifer in real time and determine its size, morphology, and viability based on the curve characteristics.

[0088] The electrode design employs a microelectrode array structure, including excitation and measurement electrodes. The excitation electrode provides alternating current, while the measurement electrode detects voltage changes. The electrode array uses an interdigitated electrode structure, which generates a uniform electric field distribution, improving detection accuracy.

[0089] The signal acquisition circuit employs a high-precision impedance analyzer, capable of measuring both the real and imaginary parts of impedance. The circuit includes modules such as a signal generator, transimpedance amplifier, filter, and A / D converter. The signal generator produces sinusoidal excitation signals of different frequencies with a frequency accuracy of 0.1%; the transimpedance amplifier converts current signals into voltage signals, with an adjustable gain range of 10-1000 times; the filter removes noise and interference signals; and the A / D converter uses a 16-bit high-precision chip with a sampling rate of 1MHz.

[0090] The circuit also integrates automatic gain control (AGC), which automatically adjusts the amplifier gain based on signal strength to ensure the signal is always within the optimal measurement range. Additionally, the circuit features overvoltage and overcurrent protection to prevent electrode damage and sample contamination.

[0091] The signal processing algorithm employs digital lock-in amplifier (DPSA) technology, which can extract weak impedance signals from noise. By comparing the phase with the excitation signal, the algorithm separates the real and imaginary parts of the impedance, improving measurement accuracy and anti-interference capability.

[0092] The viability assessment algorithm is based on impedance spectroscopy analysis, which evaluates the viability status of rotifers by analyzing changes in their electrical properties. The algorithm first extracts characteristic parameters from the impedance spectrum, including low-frequency impedance, high-frequency impedance, and characteristic frequency.

[0093] Normal and dead rotifers exhibit significant differences in impedance spectra. Normal rotifers, with their intact cell membranes, show high impedance at low frequencies and decreasing impedance at high frequencies; dead rotifers, due to cell membrane damage, show less impedance variation with frequency.

[0094] like Figure 5 The diagram shows the construction and workflow of a multi-dimensional data-driven artificial intelligence water quality assessment model. The multidimensional data-driven AI water quality assessment model employs a deep learning architecture, integrating convolutional neural networks (CNN), recurrent neural networks (RNN), and hybrid architectures. The overall model architecture comprises five main parts: a data input layer, a feature extraction layer, a fusion layer, a decision layer, and an output layer.

[0095] The data input layer supports various data types, including image data (from video surveillance), time-series data (from impedance data), and text data (from environmental reports). The input data is first standardized to convert data of different dimensions into a unified numerical range, facilitating model learning.

[0096] The feature extraction layer employs a multimodal feature extraction network, including CNN for visual feature extraction, LSTM for temporal feature modeling, and Transformer for global feature learning. CNN networks can automatically extract spatial features from images, such as the morphology and movement trajectories of zooplankton; LSTM networks can handle long-term dependencies in temporal data, such as the temporal variation trends of water quality parameters; and Transformer networks can model complex relationships between different features.

[0097] The fusion layer employs an attention mechanism to achieve adaptive fusion of multimodal features. This attention mechanism dynamically adjusts the weights of different modal features, highlighting important information and suppressing noise interference. The fused feature vector contains comprehensive information from multiple data sources, providing rich feature representations for subsequent water quality assessment.

[0098] The decision-making layer performs water quality assessment and early warning decisions based on the fused feature vectors. The decision-making layer employs a multilayer perceptron (MLP) structure, mapping high-dimensional features to the assessment results through nonlinear transformations. The model can output assessment results across multiple dimensions, including the comprehensive water toxicity index, pollution level, and early warning information.

[0099] Data preprocessing is a crucial step in model building, including data cleaning, feature selection, and data augmentation. Data cleaning primarily addresses missing values, outliers, and noisy data. For missing values, interpolation or predictive models are used to impute them; for outliers, statistical methods or machine learning algorithms are used for identification and processing; and for noisy data, filtering algorithms are used for smoothing.

[0100] Feature engineering comprises three aspects: feature extraction, feature selection, and feature transformation. Feature extraction extracts meaningful features from raw data, such as extracting zooplankton behavior from videos or water quality parameter variation features from sensor data. Feature selection uses statistical methods or machine learning algorithms to select the features that contribute most to water quality assessment, reducing feature dimensionality and improving model efficiency.

[0101] Feature transformation converts raw features into a form more suitable for model learning. For example, it can convert categorical features into numerical features, or nonlinear relationships into linear relationships. Commonly used feature transformation methods include normalization, standardization, and principal component analysis (PCA).

[0102] Data augmentation techniques are used to expand training datasets and improve the generalization ability of models. For image data, methods such as rotation, flipping, scaling, and adding noise are used to generate new samples; for time-series data, methods such as sliding windows, interpolation, and synthesis are used to increase the amount of data. Data augmentation techniques are particularly suitable for handling imbalanced sample problems, ensuring a balanced distribution of data across categories.

[0103] The model integrates multiple machine learning algorithms, including random forest, support vector machine, and neural network. The random forest algorithm is used for feature importance evaluation and preliminary classification; the support vector machine is used for nonlinear classification and regression; and the neural network is used for complex pattern recognition and prediction.

[0104] The choice of deep learning algorithms is based on task characteristics and data properties. CNN networks are suitable for image processing tasks and can automatically extract spatial features from images; RNN networks, especially LSTM and GRU, are suitable for time-series data processing and can model long-term dependencies; Transformer networks are suitable for processing data with complex relationships and can achieve global modeling.

[0105] Model optimization employs various techniques, including hyperparameter tuning, regularization, and early stopping. Hyperparameter tuning utilizes grid search or Bayesian optimization methods to find the optimal combination of model parameters; regularization techniques include L1 and L2 regularization to prevent overfitting; early stopping stops training when validation set performance no longer improves, thus avoiding overtraining.

[0106] The model also employs ensemble learning, which weights and fuses the predictions from multiple models. Ensemble methods can improve the stability and generalization ability of a model, while reducing the bias and variance of a single model. Commonly used ensemble methods include voting, averaging, and stacking.

[0107] The model training process includes steps such as data preparation, model initialization, training iterations, and validation and evaluation. First, the dataset is divided into training, validation, and test sets, typically in a 7:1:2 ratio. The training set is used for model parameter learning, the validation set for hyperparameter tuning and model selection, and the test set for final performance evaluation.

[0108] Model initialization employs either pre-trained weights or random initialization methods. For CNN networks, weights pre-trained on large-scale image datasets are typically used for initialization; for other networks, random initialization methods are employed. The choice of initialization parameters has a significant impact on the model's convergence speed and final performance.

[0109] The training process employs stochastic gradient descent (SGD) and its variants, such as Adagrad, Adadelta, and Adam. The learning rate is dynamically adjusted, typically starting at 0.01-0.1 and adjusting during training based on validation set performance. Batch size is usually set between 32 and 256, selected according to hardware resources and data size.

[0110] The model's workflow includes data acquisition, real-time processing, assessment and early warning, and result output. In real-time monitoring mode, the system continuously collects multimodal data, performs real-time analysis using a trained model, and outputs water quality assessment results and early warning information. In batch processing mode, the system processes historical data to generate water quality change trend reports and pollution source analysis results.

[0111] The model also possesses online learning capabilities, continuously updating its parameters based on new data to improve its adaptability. Online learning employs an incremental learning algorithm, learning new knowledge without affecting existing knowledge. This mechanism is particularly suitable for dynamic changes in water quality, ensuring the model consistently maintains optimal predictive performance.

[0112] like Figure 6 A flowchart illustrating the construction technology and equipment implementation of an aquatic organism monitoring system is provided. The multi-level aquatic organism monitoring system employs a zooplankton monitoring layer. This layer focuses on monitoring zooplankton such as protozoa, rotifers, cladocerans, and copepods. Utilizing the microfluidic chip technology of this invention, it is possible to simultaneously monitor the species and abundance changes of multiple zooplankton species. As a crucial component of aquatic ecosystems, changes in the community structure of zooplankton can sensitively reflect water quality conditions.

[0113] The hardware system adopts a distributed architecture design, comprising three main parts: field monitoring units, a data transmission network, and a central control system. The field monitoring units are deployed at monitoring points and integrate various sensors and monitoring equipment, enabling automatic data collection and processing. The data transmission network uses a combination of wired and wireless methods to ensure reliable data transmission. The central control system is responsible for the centralized management, analysis, and display of data.

[0114] The design of the field monitoring unit fully considers the characteristics of different aquatic environments. In river environments, the monitoring unit adopts a buoy design, which can automatically adjust with changes in water level; in lake environments, the monitoring unit adopts a fixed design and is installed on a pier or platform; in marine environments, the monitoring unit adopts a submersible design, which can carry out long-term fixed-point monitoring.

[0115] The hardware system also includes a power management module, employing a combination of solar panels and wind turbines to ensure long-term stable operation without external power. The system is equipped with a large-capacity battery, capable of maintaining normal operation for more than 7 days during continuous rainy weather.

[0116] The software system adopts a modular design, including a data acquisition module, a data processing module, a data analysis module, an early warning and decision-making module, and a user interface module. The data acquisition module is responsible for communicating with various sensors and monitoring equipment to collect monitoring data in real time; the data processing module preprocesses and performs quality control on the collected data; the data analysis module uses various algorithms to analyze and mine the data; the early warning and decision-making module generates early warning information based on the analysis results; and the user interface module provides a user-friendly human-computer interaction interface.

[0117] The algorithm design employs a multi-scale analysis method to analyze the monitoring data from both temporal and spatial dimensions. In the temporal dimension, time series analysis is used to identify periodic and trend changes in the data; in the spatial dimension, spatial interpolation and kriging methods are used to plot the spatial distribution of water quality parameters.

[0118] The software system also integrates various artificial intelligence algorithms, including cluster analysis, classification and recognition, and prediction models. Cluster analysis is used to identify similar monitoring points and similar water quality conditions; classification and recognition are used to determine water quality levels and pollution types; and prediction models are used to predict water quality change trends and pollution source locations.

[0119] The software boasts powerful data visualization capabilities, displaying monitoring results in various formats such as charts, maps, and animations. Users can view monitoring data from any time and location through an interactive interface, performing data comparison analysis and trend prediction. The system also supports data export, enabling the export of monitoring data and analysis results into various file formats.

[0120] The equipment integration adopts a standardized and modular design concept, decomposing the complex monitoring system into multiple functional modules, each with a clearly defined function and interface specification. Modules are connected through standard interfaces to ensure the system's scalability and maintainability.

[0121] The integrated solution also includes auxiliary equipment such as sampling pumps, flow meters, and controllers. Sampling pumps are used to collect water and sediment samples, with a flow rate range of 10-1000 mL / min; flow meters are used to measure water flow with an accuracy of ±1%; and controllers are used for automated system control and support multiple communication protocols.

[0122] The equipment is integrated using a rack-mount design, housing all devices within a standard rack for easy installation and maintenance. The rack is waterproof, dustproof, and shockproof, capable of withstanding various harsh environmental conditions. The rack's internal layout is rational, ensuring excellent heat dissipation and guaranteeing long-term stable operation of the equipment.

[0123] The system is also equipped with a comprehensive operation and maintenance management system, including functions such as equipment status monitoring, fault diagnosis, and remote maintenance. Through IoT technology, maintenance personnel can monitor the equipment's operating status in real time and promptly identify and handle faults. The system also features an automatic calibration function, periodically calibrating sensors to ensure the accuracy of monitoring data.

[0124] The following descriptions illustrate different embodiments of the above technical solutions: Example 1: Establishment of an acute water toxicity assessment model based on the behavioral characteristics of large Daphnia magna (taking heavy metals as an example) Step 1: Standard toxicity exposure and data collection.

[0125] Before conducting the experiment, the experimental equipment was thoroughly cleaned and disinfected. The first microchamber of the biosensor unit was repeatedly rinsed with ultrapure water to avoid residual impurities interfering with the experimental results. Potassium dichromate was used as the standard heavy metal poison. Chemical reagent preparation specifications were strictly followed. The reagent was accurately weighed using a high-precision electronic balance (accuracy up to 0.0001g) to prepare five gradient concentration solutions of 0.1mg / L, 0.5mg / L, 1.0mg / L, 2.0mg / L, and 5.0mg / L. Each concentration solution was prepared in triplicate to ensure concentration accuracy.

[0126] The *Daphnia magna* used in the experiment were healthy individuals cultured in the laboratory for a long period. Individuals with a body length of 2-3 mm and strong mobility were selected. Ten *Daphnia magna* were placed in the first microchamber of the biosensor unit, each exposed to solutions of different concentrations. Simultaneously, a high-definition high-speed camera (4K resolution, 240fps) from the multimodal information acquisition unit was used. To ensure video recording quality, a uniform and shadow-free ring light source was arranged around the microchamber. The swimming behavior of the *Daphnia magna* was continuously recorded over 30 minutes. Five replicates were set up for each concentration group to ensure a sufficient number of valid individual data for each concentration group.

[0127] Step 2: Behavior trajectory tracking and feature extraction.

[0128] The signal processing and feature extraction unit performs automated analysis of the acquired video. First, it uses a YOLOv5 deep learning algorithm optimized through transfer learning. During training, it fine-tunes the algorithm using a dataset containing 10,000 images of large daphnia under different poses and lighting conditions to improve recognition accuracy and enable real-time identification of individual large daphnia in each frame. Next, it employs an improved version of the DeepSORT multi-object tracking algorithm, which introduces a metric combining Mahalanobis distance and cosine distance to more accurately correlate across frames and generate continuous motion trajectories for each individual.

[0129] Based on trajectory data, several key behavioral parameters were quantitatively calculated using specially developed behavioral analysis software. Average swimming speed was calculated by measuring the individual's displacement per unit time; acceleration was determined based on the rate of change of velocity over time; the curvature of the movement path was quantified using the integral of curvature method; and the vertical distribution height was obtained by analyzing the statistical distribution of the individual's position in the vertical space of the micro-chamber. Additionally, auxiliary behavioral parameters such as the frequency of changes in movement direction and the proportion of pauses were extracted to provide richer data dimensions for subsequent model training.

[0130] Step 3: Regression model training and validation.

[0131] The extracted behavioral feature data was used as the input variable (X), and the corresponding known heavy metal concentration was used as the output label (Y). To select the most suitable machine learning algorithm, CNN-LSTM, Support Vector Machine Regression (SVR), Random Forest Regression (RFR), and Gradient Boosting Regression (GBR) were used for model training. During training, grid search combined with cross-validation (5-fold cross-validation) was used to optimize the key parameters of each algorithm, such as the kernel function type and penalty factor C of SVR, and the number of trees and maximum depth of RFR.

[0132] The model learns the complex mapping relationship between behavioral characteristics and toxicity concentration. After training, the model's prediction accuracy is validated using a reserved test set (20% of the total data), in addition to the coefficient of determination (R²). 2 In addition to using the mean squared error (MSE) as the primary evaluation metric, other metrics such as mean squared error (MSE) and mean absolute error (MAE) are introduced to evaluate performance from different perspectives. Furthermore, scatter plots of predicted and actual values, as well as histograms of error distribution, are plotted to visually demonstrate the model's predictive effectiveness.

[0133] Step 4: Model integration and online application.

[0134] The optimal-performing model (assuming a CNN-LSTM model) is integrated into the water quality assessment and early warning decision-making unit. In actual monitoring, the system acquires behavioral data of *Daphnia magna* at a frequency of 10 times per second through a real-time data acquisition module. This data is then input into the model, outputting a comprehensive toxicity prediction value expressed as potassium dichromate equivalent concentration. To ensure the reliability of the early warning, the system sets three warning thresholds: a yellow warning is triggered when the predicted value exceeds 0.1 mg / L, indicating possible mild pollution; an orange warning is triggered when it exceeds 0.5 mg / L, indicating worsening pollution; and a red warning is triggered when it exceeds 1.0 mg / L, indicating severe pollution. When the corresponding threshold is reached, the system automatically sends heavy metal pollution warnings to relevant departments and personnel via SMS, email, and other means, achieving rapid and quantitative water quality assessment.

[0135] Example 2: Integrated water quality early warning decision algorithm that combines macro- and micro zooplankton response information: Step 1: Multi-source data preprocessing and feature-level fusion.

[0136] The system simultaneously acquires behavioral feature vectors (such as velocity Vd, trajectory complexity Td, and vertical migration index Zd) from *Daphnia macrocarpa* and Raman spectral feature vectors (such as characteristic peak intensity Rs and principal component scores Pc) from copepods, supplemented by corresponding auxiliary behavioral feature vectors (such as swimming velocity Vf and population distribution Zf) obtained from impedance monitoring. First, the two types of heterogeneous data are standardized to eliminate the influence of dimensions. Then, a feature concatenation method is used to merge the processed feature vectors into a unified, high-dimensional fused feature vector, which can be represented as Fused_Vector=[Vd,Td,Zd,Rs,Pc,Vf,Zf...]. This step integrates response information from different biological levels into the same information space.

[0137] Step 2: Construction and training of a multi-task model based on deep learning.

[0138] A deep learning model with a multi-task learning framework is constructed. This model takes a fused feature vector as input, uses a CNN module and an LSTM as initial feature extraction methods, a convolutional neural network to process the two-dimensional feature map extracted from spectral data, and a long short-term memory network to process time-series data composed of behavioral feature parameters. These are weighted and fused through an attention mechanism, with the fused feature vector serving as input. The core of the model is a shared deep neural network (such as a multi-layer fully connected network) used to learn the deep correlation between the responses of two types of zooplankton. After this shared layer, the model branches into two specific output branches: Branch 1 (Regression Task): Responsible for outputting a continuous Comprehensive Toxicity Index (CTI), which comprehensively reflects the overall intensity of toxicity.

[0139] Branch 2 (Classification Task): Responsible for outputting the probability distribution of pollutant types (e.g., [Heavy metals: 0.7, Organophosphorus pesticides: 0.25, Others: 0.05]).

[0140] The model is trained using a historical dataset containing known toxicity types and concentrations. By jointly optimizing the loss functions for regression and classification tasks, the model can accurately complete both tasks simultaneously.

[0141] Step 3: Dynamic weight adaptation and decision rule formulation.

[0142] In model application, a dynamic weighting mechanism is introduced. The system dynamically adjusts the contribution weights of Daphnia macrocarpa and copepod features in the final decision based on the quality of real-time data (such as signal-to-noise ratio) and the historical response sensitivity of each group to different pollutants. For example, when the system suspects pesticide pollution, the weight of copepod spectral features, which are more sensitive to pesticides, is automatically increased. Based on CTI values ​​and pollutant classification probabilities, graded early warning rules are formulated. Normal (CTI<0.3): No warning.

[0143] Focus on (0.3≤CTI<0.6): Record data and highlight trends.

[0144] Alert (0.6≤CTI<0.8, and probability of a certain type of pollutant >50%): Issues a primary alert, indicating suspected pollutants.

[0145] Alert (CTI≥0.8 and pollutant probability>70%): Issue the highest level alert, identify the type of pollution, and initiate emergency response procedures.

[0146] Step 4: System integration and closed-loop feedback.

[0147] The fusion decision-making algorithm is integrated into the core software of the water quality assessment and early warning decision-making unit. The system runs the algorithm in real time and can incrementally learn from the model using continuously accumulated new data, achieving continuous performance optimization and improving early warning accuracy, forming an intelligent closed loop with adaptive capabilities.

[0148] Experimental verification shows that the accuracy of the integrated toxicity assessment using the fusion model reaches 94%, which is 7-15 percentage points higher than the method using Daphnia magna or copepods alone.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An aquatic biosensor for monitoring water pollutants and water quality, characterized in that, include: Biosensing unit, multimodal information acquisition unit, signal processing and feature extraction unit, and water quality assessment and early warning decision-making unit; The process involves using a biosensing unit to contain and expose live zooplankton in the water sample to be tested, a multimodal information acquisition unit to collect multimodal data of the zooplankton, a signal processing and feature extraction unit to process and extract features from the multimodal data to obtain feature parameters, and a water quality assessment and early warning decision-making unit to perform data fusion analysis on the feature parameters using an artificial intelligence model to obtain the comprehensive water toxicity index, pollutant type identification results, and early warning level.

2. The aquatic biosensor according to claim 1, characterized in that, The biosensing unit includes a first microchamber and a second microchamber. The first microchamber contains and exposes large zooplankton, and the second microchamber contains and exposes small zooplankton. The first and second microchambers are connected to a common inlet and outlet via a microflow channel. Both the first and second microchambers are equipped with a temperature control module for temperature control and a light cycle control module for supplemental illumination.

3. The aquatic biosensor according to claim 1, characterized in that, In the multimodal information acquisition unit, the multimodal data includes motion video data of large zooplankton and impedance data of small zooplankton; the motion video data is acquired through a microscopic optical system and an integrated vertical motion monitoring component, and the impedance scan data is acquired through a microelectrode array.

4. The aquatic biosensor according to claim 2, characterized in that, In the signal processing and feature extraction unit, behavioral and impedance features are extracted from the multimodal data. Noise removal and target recognition and tracking are performed on the motion video data of large zooplankton. Based on the target recognition and tracking results, behavioral feature parameters are obtained, including average swimming speed, motion trajectory complexity, instantaneous velocity, acceleration, tumbling frequency, heart rate, and vertical motion rate. The impedance features of small zooplankton are preprocessed and feature extracted to obtain spectral feature parameters, including low-frequency impedance, high-frequency impedance, and characteristic frequency.

5. The aquatic biosensor according to claim 4, characterized in that, In the water quality assessment and early warning decision-making unit, the artificial intelligence model adopts a fusion model based on machine learning algorithms. The impedance characteristic parameters of small zooplankton and the behavioral characteristic parameters of large zooplankton are used as input data for the artificial intelligence model. The output data are the comprehensive water quality toxicity index, pollutant type identification results, and early warning level. The pollutant type identification results include one or more pollutant types and corresponding classification probabilities. The pollutant types include heavy metals, organophosphorus pesticides, pyrethroid pesticides, and algal toxins.

6. The aquatic biosensor according to claim 1, characterized in that, In the water quality assessment and early warning decision-making unit, the artificial intelligence model is a hybrid model of a convolutional neural network for processing impedance characteristic parameters of small zooplankton and a long short-term memory network for processing behavioral characteristic parameters of large zooplankton. The output data of the convolutional neural network and the long short-term memory network are weighted and fused through an attention mechanism, and the final result is output through a multi-branch task structure.

7. A water quality monitoring device, characterized in that, include: The aquatic biological sensor and peripheral auxiliary module as described in any one of claims 1-6 above; The peripheral auxiliary modules include a water sample pretreatment module, a main control and communication module, a power supply module, and a protective housing; The water sample pretreatment module filters, maintains constant temperature and gas balances the water sample before it enters the biosensor unit; the main control and communication module controls the operation of the aquatic biosensor and peripheral auxiliary modules, and uploads the output results of the water quality assessment and early warning decision unit; the power supply module provides power; and the protective shell provides physical protection for the aquatic biosensor.

8. The water quality monitoring equipment according to claim 7, characterized in that, The aquatic biosensor is based on the multidimensional biological response signals of zooplankton and uses a dose-response relationship model to deduce the concentration of chemical pollutants in the water.

Images