Intelligent linkage dimming method and system based on underwater biological recognition

Abstract

This invention discloses an intelligent linkage lighting method and system based on underwater biometrics. The method includes: acquiring multimodal sensing data and real-time environmental parameters within a target water body area; inputting the multimodal sensing data into a pre-constructed heterogeneous multimodal fusion recognition model to output the species, growth stage, and real-time physiological state parameters of underwater organisms currently passing through the target water body area; then matching an initial lighting scheme from a pre-constructed illumination knowledge base; dynamically correcting the initial lighting scheme based on real-time physiological state parameters and real-time environmental parameters to generate a current target lighting scheme; driving and controlling the corresponding light-emitting device to output illumination according to the current target lighting scheme; and dynamically adjusting the output illumination of the corresponding light-emitting device based on the actual illumination parameters within the target water body area until the acquired actual illumination parameters match the current target lighting scheme. This invention enables personalized lighting for underwater organisms.

Description

Technical Field

[0001] This invention relates to the field of underwater biological sensing and light control technology, and in particular to an intelligent linkage dimming method and system based on underwater biological recognition. Background Technology

[0002] In the fields of underwater aquaculture and scientific research, the lighting environment has a decisive impact on the behavioral rhythms, physiological metabolism, growth and development, and body coloration of organisms. Different species, and even different growth stages of the same species, exhibit significant differences in their requirements for spectral composition, light intensity, and photoperiod. For example, corals rely on specific wavelengths of light to maintain the photosynthesis of their symbiotic algae, while some fish require gentle light during their breeding season to reduce stress levels. Therefore, precise and personalized lighting for different underwater organisms is crucial for improving aquaculture efficiency, ensuring animal welfare, and obtaining high-quality scientific data.

[0003] Currently, underwater lighting systems generally employ uniform and fixed lighting schemes, deploying light sources with the same spectrum and intensity in the same aquaculture or observation water body, making it impossible to distinguish between biological species and their growth stages. While some systems have introduced timed switching or manual dimming functions, they still lack dynamic perception and intelligent decision-making capabilities based on individual biological characteristics. A few systems attempt to use a single underwater camera for biometric identification and adjust lighting accordingly based on the identification results. However, in typical complex environments such as turbid water, low light levels, and blurred target motion, the accuracy of single visual identification methods drops sharply, making it difficult to support reliable personalized lighting control. Therefore, there is an urgent need for an intelligent, interconnected control method capable of penetrating complex underwater environments, accurately sensing multi-dimensional biological information, and dynamically adjusting lighting based on identification results. Summary of the Invention

[0004] In order to overcome the shortcomings of the existing technology, one of the objectives of this invention is to provide an intelligent linkage lighting method based on underwater biological identification, which can solve the problem that the existing underwater biological lighting cannot provide precise lighting for underwater organisms, resulting in low efficiency of biological aquaculture or inability to obtain high-quality scientific research data.

[0005] The second objective of this invention is to provide an intelligent linkage lighting method based on underwater biological identification, which can solve the problem that existing underwater biological lighting methods cannot accurately illuminate underwater organisms, resulting in low efficiency in aquaculture or the inability to obtain high-quality scientific research data.

[0006] One of the objectives of this invention is achieved through the following technical solution: Intelligent linkage dimming methods based on underwater biometrics include: Data acquisition steps: Multi-spectral image acquisition units, active electric field imaging units, and acoustic signal acquisition units are set up at sensing nodes within the target water body area to acquire multi-modal sensing data in real time; and real-time environmental parameters are synchronously acquired through environmental sensors distributed within the target water body area; the multi-modal sensing data includes multi-spectral images, active electric field images, and acoustic signals; the real-time environmental parameters include water temperature, water flow velocity, turbidity, and dissolved oxygen; Reasoning steps: The multimodal sensing data is input into the pre-constructed heterogeneous multimodal fusion recognition model of the system. After feature extraction by the corresponding modal feature encoders, the feature data of the multimodal sensing data are fused based on the cross-modal cross-attention fusion mechanism. Then, combined with the multi-task decoder, the biological species, growth stage, and real-time physiological state parameters of the underwater organisms currently passing through the target water area are output synchronously. The real-time physiological state parameters are any one of stress index, health index, and metabolic activity. Matching and correction steps: Based on the species and growth stage of the underwater organisms currently passing through the target water body area, an initial lighting scheme is matched from the pre-built lighting knowledge base, and the initial lighting scheme is dynamically corrected based on the real-time physiological state parameters and real-time environmental parameters to generate the current target lighting scheme. Dynamic control steps: Drive and control the corresponding light-emitting device to output light to the corresponding target water body area according to the current target lighting scheme. At the same time, acquire the actual lighting parameters in the target water body area based on the distributed underwater lighting sensor, and dynamically adjust the output lighting of the corresponding light-emitting device according to the actual lighting parameters and the current target lighting scheme so that the acquired actual lighting parameters match the current target lighting scheme.

[0007] Furthermore, the heterogeneous multimodal fusion recognition model is obtained through the following steps: Training dataset construction steps: Acquire multispectral images, active electric field images, and acoustic signals of underwater organisms in the target water body, as well as corresponding labeled data of biological species, growth stages, and physiological state parameters, and construct the training dataset; wherein, the labeled data of physiological state parameters are obtained through behavioral analysis, blood biochemical detection, and respiratory metabolism measurement; Model construction steps: Construct a heterogeneous multimodal fusion recognition model; the heterogeneous multimodal fusion recognition model includes a multispectral encoder, an active electric field encoder, an acoustic encoder, a cross-modal cross-attention fusion network, and a multi-task decoder; wherein, the multispectral encoder is used to extract multispectral feature maps from the multispectral image; the active electric field encoder is used to extract conductivity distribution feature maps from the active electric field image; the acoustic encoder is used to extract acoustic feature vectors containing temporal information from the Mel spectrum of the acoustic signal; the cross-modal cross-attention fusion network is used to expand the multispectral feature map, conductivity distribution feature map, and acoustic feature vector into a labeled sequence, add learnable modal embedding position encoding, and then perform intermodal interaction through a multi-layer cross-attention mechanism to generate multimodal fusion features; the multi-task decoder includes a biological species classification head, a growth stage ordered regression head, and a physiological state parameter regression head; Model training steps: First, the multispectral encoder, active electric field encoder, and acoustic encoder are pre-trained using the datasets of the corresponding modalities. Then, after loading the pre-trained weights, the heterogeneous multimodal fusion recognition model is jointly fine-tuned end-to-end using the multi-task uncertainty weighted total loss until convergence. The multi-task uncertainty weighted total loss is composed of the weighted sum of species classification cross-entropy loss, growth stage ordered regression loss, and physiological state parameter mean square error loss.

[0008] Furthermore, the data acquisition step includes: when the active electric field imaging unit detects an electric field disturbance exceeding a preset threshold, it determines that an organism has entered the target water area and simultaneously wakes up the multispectral image acquisition unit and the acoustic signal acquisition unit to acquire data.

[0009] Furthermore, the data acquisition step specifically includes: installing electric field electrodes in the target water area and controlling the electric field electrodes to emit an alternating excitation electric field with a preset frequency and amplitude into the target water area; acquiring electric field disturbance signals caused by the difference in conductivity and dielectric constant of underwater organisms through a sensor array composed of multiple electrodes; and combining the electric field disturbance signals with the empty field reference signal when there are no organisms in the target water area using an electrical impedance tomography reconstruction algorithm to invert and generate an active electric field image reflecting the spatial distribution of the internal conductivity of underwater organisms. Multispectral images, i.e., characteristic bands reflecting the physiological state of underwater organisms, are acquired by a multispectral camera installed in the target water area; the characteristic bands include chlorophyll fluorescence bands and / or hemoglobin absorption bands. Acoustic signals are acquired by hydrophones installed in the target water body area.

[0010] Furthermore, the matching correction step specifically includes: inputting the biological species, growth stage, real-time physiological state parameters, and real-time environmental parameters output by the model as the state space into the constructed deep reinforcement learning agent; using the illumination parameters as the action space, the reinforcement learning agent calculates reward signals based on the behavioral feedback indicators of the underwater organisms to optimize the initial illumination scheme online in real time, thereby obtaining the current illumination scheme; the behavioral feedback indicators include the activity frequency, feeding amount, aggregation density, or stress performance of the underwater organisms; the illumination parameters include spectral composition, light intensity, and time variation curves; the reinforcement learning agent uses a proximal policy optimization algorithm to achieve optimization.

[0011] Furthermore, before the matching correction step, the method further includes: determining whether the real-time physiological state parameters exceed a preset safety threshold; if so, overwriting the current target lighting scheme with a preset safe and conservative lighting template, and simultaneously recording the state information when the safety mechanism is triggered and the original lighting action as a negative sample trajectory, storing it in the experience replay buffer of the deep reinforcement learning agent for training the policy network to learn the safe action boundary; if not, then performing the matching correction step.

[0012] Furthermore, the specific steps of driving and controlling the corresponding luminescent device to output illumination to the corresponding target water area according to the current target illumination scheme include: dynamically adjusting the output illumination of the corresponding luminescent device through a closed-loop compensation controller; the closed-loop compensation controller adopts a model predictive control algorithm, performs feedforward prediction compensation on the illumination output data of the luminescent device by combining a pre-established digital twin model of the optical characteristics of the water body, and performs real-time feedback compensation and coordinated adjustment based on the deviation between the actual illumination parameters and the target illumination parameters in the current target illumination scheme, so as to compensate for water body attenuation and ambient light changes, thereby making the actual illumination parameters match the target illumination parameters.

[0013] An intelligent linkage dimming system based on underwater biometrics, the intelligent linkage dimming system comprising: The data acquisition module is used to acquire multimodal sensing data in real time by setting up multispectral image acquisition units, active electric field imaging units, and acoustic signal acquisition units at sensing nodes within the target water body area; and to synchronously acquire real-time environmental parameters through environmental sensors distributed within the target water body area; the multimodal sensing data includes multispectral images, active electric field images, and acoustic signals; the real-time environmental parameters include water temperature, water flow velocity, turbidity, and dissolved oxygen; The inference module is used to input the multimodal sensing data into a pre-constructed heterogeneous multimodal fusion recognition model of the system. After feature extraction by the corresponding modal feature encoders, the feature data of the multimodal sensing data are fused based on a cross-modal cross-attention fusion mechanism. Then, combined with a multi-task decoder, the module synchronously outputs the species, growth stage, and real-time physiological state parameters of the underwater organisms currently passing through the target water area. The real-time physiological state parameters are any one of stress index, health index, and metabolic activity. The matching and correction module is used to match an initial lighting scheme from a pre-built lighting knowledge base based on the species and growth stage of the underwater organisms currently passing through the target water body area, and to dynamically correct the initial lighting scheme based on the real-time physiological state parameters and real-time environmental parameters to generate the current target lighting scheme. The dynamic control module is used to drive and control the corresponding light-emitting device to output light to the corresponding target water body area according to the current target lighting scheme. At the same time, it acquires the actual lighting parameters in the target water body area based on the distributed underwater lighting sensor, and dynamically adjusts the output lighting of the corresponding light-emitting device according to the actual lighting parameters and the current target lighting scheme, so that the acquired actual lighting parameters match the current target lighting scheme.

[0014] Furthermore, the heterogeneous multimodal fusion recognition model is obtained through the following steps: Training dataset construction steps: Acquire multispectral images, active electric field images, and acoustic signals of underwater organisms in the target water body, as well as corresponding labeled data of biological species, growth stages, and physiological state parameters, and construct the training dataset; wherein, the labeled data of physiological state parameters are obtained through behavioral analysis, blood biochemical detection, and respiratory metabolism measurement; Model construction steps: Construct a heterogeneous multimodal fusion recognition model; the heterogeneous multimodal fusion recognition model includes a multispectral encoder, an active electric field encoder, an acoustic encoder, a cross-modal cross-attention fusion network, and a multi-task decoder; wherein, the multispectral encoder is used to extract multispectral feature maps from the multispectral image; the active electric field encoder is used to extract conductivity distribution feature maps from the active electric field image; the acoustic encoder is used to extract acoustic feature vectors containing temporal information from the Mel spectrum of the acoustic signal; the cross-modal cross-attention fusion network is used to expand the multispectral feature map, conductivity distribution feature map, and acoustic feature vector into a labeled sequence, add learnable modal embedding position encoding, and then perform intermodal interaction through a multi-layer cross-attention mechanism to generate multimodal fusion features; the multi-task decoder includes a biological species classification head, a growth stage ordered regression head, and a physiological state parameter regression head; Model training steps: First, the multispectral encoder, active electric field encoder, and acoustic encoder are pre-trained using the datasets of the corresponding modalities. Then, after loading the pre-trained weights, the heterogeneous multimodal fusion recognition model is jointly fine-tuned end-to-end using the multi-task uncertainty weighted total loss until convergence. The multi-task uncertainty weighted total loss is composed of the weighted sum of species classification cross-entropy loss, growth stage ordered regression loss, and physiological state parameter mean square error loss.

[0015] Furthermore, the data acquisition module is also used to determine that an organism has entered the target water area when the active electric field imaging unit detects an electric field disturbance exceeding a preset threshold, and simultaneously wakes up the multispectral image acquisition unit and the acoustic signal acquisition unit to acquire data.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention introduces active electric field imaging, enabling the acquisition of images of the internal electrical properties of organisms in completely dark or extremely turbid waters. This fundamentally overcomes the limitations of single visual perception and significantly improves the robustness of biometric identification in complex environments. Simultaneously, it employs heterogeneous multimodal deep cross-attention fusion to achieve true interactive fusion of multispectral, electrical, and acoustic information, allowing for simultaneous identification of organism species, growth stages, and assessment of real-time physiological states. Furthermore, during illumination, it dynamically corrects the illumination by combining data on organism species, growth stages, real-time physiological states, and environmental sensors. It also dynamically adjusts the output illumination of the luminescent device based on actual illumination parameters, ensuring that the illumination of underwater organisms matches the target illumination. This guarantees personalized illumination and precise light tracking for underwater organisms, ensuring the uniformity of illumination in time and space. This technology can be applied to underwater aquaculture to improve farming efficiency and to scientific research to help obtain reliable research data. Attached Figure Description

[0017] Figure 1 The flowchart of the intelligent linkage dimming method based on underwater biometrics provided by the present invention is shown. Detailed Implementation

[0018] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments. Example 1

[0019] This invention provides an intelligent linkage dimming method based on underwater biometrics, such as... Figure 1 As shown, the intelligent linkage dimming method includes: Step S1: Multi-spectral image acquisition unit, active electric field imaging unit and acoustic signal acquisition unit are set up at the sensing nodes in the target water body area to acquire multi-modal sensing data in real time; and real-time environmental parameters are acquired synchronously by environmental sensors distributed in the target water body area; the modal sensing data includes multi-spectral images, active electric field images and acoustic signals; the real-time environmental parameters include water temperature, water flow velocity, turbidity and dissolved oxygen.

[0020] Specifically, this invention achieves underwater organism perception through active electric field imaging. Its core principle is to actively emit a weak, pre-set electric field into the water, and then measure the spatial distribution of this electric field after it is disturbed by surrounding objects using an electrode array. This allows for the inference of the object's position, shape, and even internal structure. Specifically, one or more pairs of transmitting electrodes can be used to inject a low-intensity, specific-frequency alternating current into the target water area, creating a quasi-static electric field. When an organism enters the target water area, its conductivity and dielectric constant differ from those of the water in the target area. This difference disturbs the aforementioned quasi-static electric field. For example, fish muscles and viscera have better conductivity than water, while fish swim bladders and fat have poor conductivity. This difference in electrical characteristics alters the distribution of electric field lines, increasing or decreasing the current density in some areas. Therefore, underwater organisms are identified by sensing this difference in electric field.

[0021] In other words, by deploying a sensor array consisting of multiple electrodes at the location where the organism passes, the electric field disturbance signal caused by the difference in conductivity and dielectric constant of the underwater organism is simultaneously measured. This signal is then combined with electrical impedance tomography (EIT) algorithms and a reference signal from a blank water area to invert and calculate a spatial image of the internal conductivity distribution of the underwater organism—also known as an active electric field image. Through the principle of active electric field imaging, information that is difficult to capture optically and acoustically, such as the internal structure, body shape, and real-time posture of the organism, can be obtained, making it particularly suitable for identifying biological objects in completely dark or extremely turbid water.

[0022] Multispectral images are specifically acquired by a multispectral camera installed in the water to reflect the physiological state of underwater organisms, i.e., multispectral images; among which, the characteristic bands include chlorophyll fluorescence bands and / or hemoglobin absorption bands.

[0023] The acoustic signals are specifically acquired through hydrophones installed in the water.

[0024] More preferably, the present invention also achieves low power consumption by setting up a linked wake-up mechanism. Specifically, when the active electric field imaging unit detects an electric field disturbance exceeding a preset threshold, it determines that an organism has entered the target water area and simultaneously wakes up the multispectral image acquisition unit and the acoustic signal acquisition unit to collect data. By having the active electric field imaging unit simultaneously wake up the multispectral image acquisition unit and the acoustic signal acquisition unit based on the detection results, it is possible to avoid all acquisition units being in working state for a long time, thus preventing an increase in system power consumption.

[0025] Step S2: Input the multimodal sensing data into the pre-constructed heterogeneous multimodal fusion recognition model of the system, extract features through corresponding modal feature encoding, fuse the feature data of the multimodal sensing data based on the cross-modal cross-attention fusion mechanism, and then combine the multi-task decoder to synchronously output the biological species, growth stage and real-time physiological state parameters of the underwater organisms currently passing through the target water area.

[0026] Specifically, this invention uses feature extraction and data fusion of multiple modal sensing data to predict the species, growth stage, and real-time physiological state parameters of underwater organisms, enabling dynamic adjustment of illumination based on these parameters. The real-time physiological state parameters can be any one of stress index, health index, and metabolic activity.

[0027] Specifically, this invention employs a heterogeneous multimodal fusion recognition model to monitor underwater organisms. The heterogeneous multimodal fusion recognition model is trained by constructing a dataset through the collection and annotation of historical data. In practical use, the acquired multimodal sensing data is input into the constructed heterogeneous multimodal fusion recognition model. The multimodal sensing data undergoes feature extraction through corresponding modal feature encoding, and the feature data is fused based on a cross-modal cross-attention fusion mechanism. Combined with a multi-task decoder, the model synchronously outputs the species, growth stage, and real-time physiological state parameters of the underwater organisms, thereby enabling the monitoring of underwater organisms passing through the target water area.

[0028] More preferably, the training process of the heterogeneous multimodal fusion recognition model constructed in this invention specifically includes: First, a training dataset is acquired. This dataset includes multispectral images, active electric field images, and acoustic signals from the aquatic environment containing underwater organisms, along with corresponding annotations of biological species, growth stages, and physiological parameters. The physiological parameter annotations are obtained through behavioral analysis, blood biochemistry testing, or respiratory metabolism measurements. This data can be constructed by organizing historical measurement data of different underwater organisms and organisms at different stages within the target aquatic environment, or by simulating aquatic environments and utilizing historical measurement data of different organisms within corresponding aquatic bodies. For example, in controlled aquatic environments and wild breeding areas, multispectral images, active electric field images, and acoustic signals are simultaneously acquired and annotated by biological experts. The annotations include: biological species (e.g., grouper, large yellow croaker), growth stages (e.g., juvenile, adult, or specific age), and physiological parameters: near-true values ​​are obtained through behavioral analysis, blood sampling, and chlorophyll fluorescence measurements. In addition, the collected data can be preprocessed, such as by random cropping, rotation, noise injection (simulating turbid water), and spectral masking, to enhance the robustness of the data.

[0029] A heterogeneous multimodal fusion recognition model is constructed. This model includes a multispectral encoder, an active electric field encoder, an acoustic encoder, a cross-modal cross-attention fusion network, and a multi-task decoder. The multispectral encoder, active electric field encoder, and acoustic encoder are used for feature extraction from the multimodal sensing data.

[0030] A multispectral encoder is used to extract multispectral feature maps from multispectral images. Specifically, a multispectral image is input into the multispectral encoder to obtain a multispectral feature map; for example, the multispectral feature map includes chlorophyll fluorescence bands, hemoglobin absorption bands, etc. The above data is preprocessed by radiometric calibration, water body radiometric correction, and normalization, and then a lightweight CCN backbone network is used to encode the multispectral image into a spatial-spectral feature map, that is, a multispectral feature map.

[0031] An active electric field encoder is used to extract conductivity distribution feature maps from active electric field images. Specifically, the active electric field image, which reflects the distribution of electrical properties inside a biological body, is preprocessed, such as by differential reconstruction based on a spatial reference, adaptive thresholding for background removal, and normalization. Then, ResNet is used to encode the preprocessed active electric field image into a conductivity distribution feature map to capture internal structural information such as the contours of biological organs and the conductivity texture of tissues.

[0032] An acoustic encoder is used to extract acoustic feature vectors containing temporal information from the Mel spectrogram of an acoustic signal. Specifically, the underwater acoustic signal is first analyzed in time and frequency to obtain the Mel spectrogram. Then, it undergoes preprocessing techniques such as pre-emphasis, frame windowing, short-time Fourier transform, Mel filtering, and logarithmic calculation to form a fixed-size two-dimensional spectrum. A 2D-CNN is then used to extract local acoustic signature features, which are then fed along the time axis into a bidirectional LSTM to output the acoustic feature vector.

[0033] Dedicated feature encoders are constructed based on the sensing data of different modalities to realize feature extraction of sensing data of different modalities.

[0034] A cross-modal cross-attention fusion network is used to expand multispectral feature maps, conductivity distribution feature maps, and acoustic feature vectors into labeled sequences, add learnable modal embedding position codes, and then generate multimodal fusion features through multi-layer cross-attention mechanisms to facilitate intermodal interaction. This invention achieves the fusion of perceptual data between multiple modalities through a cross-modal cross-attention fusion network. For example, in the multispectral branch, the features of the perceptual data from the multispectral branch are used as queries to fuse the key and value of the active electric field image and acoustic signal, thereby allowing spectral features to perceive electric field and sound information. Specifically, for example, 2-4 cross-attention layers are stacked, with parallel computation in three directions within each layer: multispectral feature map fusion with conductivity distribution feature map and acoustic features; conductivity distribution feature map fusion with multispectral feature map and acoustic features; acoustic features fusion with multispectral feature map and conductivity distribution feature map, etc. Through this interactive fusion, sufficient information flow between modalities can be achieved, ultimately ensuring that the output of each modality incorporates information from other modal data.

[0035] Furthermore, to enable cross-attention calculation of feature data from multiple modalities, this invention also requires aligning and encoding the multimodal feature data. Specifically, the feature maps of the multispectral image and the active electric field image are flattened into sequences, with each spatial location representing a token, and learnable modality type embeddings and 2D sinusoidal positional encodings are added. The acoustic feature vector of the acoustic signal is copied and expanded into a sequence with the same number of tokens as the feature maps of the multispectral image and the active electric field image, and the same modality embeddings are added. Thus, the three resulting sequences are all of the same length. Then, based on linear projection, the tokens of each feature are uniformly mapped to the same dimension, resulting in three aligned sequences, which are then fed into the cross-modal cross-attention fusion network for fusion.

[0036] The multi-task decoder includes a species classification head, a growth stage ordered regression head, and a physiological state parameter regression head. Starting from the fused feature vector, three independent task heads are set to achieve inference with different outputs. Specifically, the species classification head outputs the organism type, including the organism type label and confidence score; the growth stage ordered regression head outputs the organism's growth stage; and the physiological state parameter regression head outputs the organism's physiological state parameters, which can include any one of the following: stress index, health index, and metabolic activity estimate.

[0037] Model training: First, the multispectral encoder, active electric field encoder and acoustic encoder are pre-trained using the datasets of the corresponding modalities. Then, after loading the pre-trained weights, the heterogeneous multimodal fusion recognition model is jointly fine-tuned end-to-end using the multi-task uncertainty weighted total loss until convergence.

[0038] In addition, during model training, an uncertainty-weighted loss function is constructed. This loss function is derived by weighted summation of species classification cross-entropy loss, growth stage ordered regression loss, and physiological state parameter mean square error loss. The weight coefficients of each loss can be adaptively adjusted during training.

[0039] The multispectral encoder, active electric field encoder, and acoustic encoder were pre-trained using datasets for their respective modalities. After pre-training, the entire fusion network and multi-task decoder were trained end-to-end using a multi-task loss algorithm. During training, the AdamW optimizer was employed for optimization. Furthermore, during system deployment, the regression head for real-time physiological state parameters can be self-supervised and corrected based on feedback rewards, such as changes in biological behavior, to enhance the accuracy of the evaluation.

[0040] Step S3: Based on the species and growth stage of the underwater organisms in the target water area, an initial lighting scheme is obtained by matching from the pre-built lighting knowledge base. The initial lighting scheme is then dynamically modified based on real-time physiological state parameters and real-time environmental parameters to generate the current target lighting scheme.

[0041] The lighting knowledge base is a system that pre-determines lighting schemes for different types of organisms at different growth stages based on experience or historical data.

[0042] When the species and growth stage of the underwater organism are deduced, the initial lighting scheme is first obtained by matching the lighting knowledge base. Then, the initial lighting scheme is dynamically modified based on real-time physiological state parameters and real-time environmental parameters to obtain the current target lighting scheme.

[0043] When dynamically correcting the initial lighting scheme, this invention also incorporates a deep reinforcement learning agent. This involves inputting the organism species, growth stage, real-time physiological parameters, and environmental sensor data into the reinforcement learning agent as the state space. Based on the initial lighting scheme, the agent uses the light spectrum, light intensity, and time-varying curves as the action space, and calculates reward signals based on the underwater organism's behavioral feedback indicators. This continuously optimizes the lighting strategy online, allowing the target lighting scheme to change automatically with the organism's long-term state, thereby achieving dynamic optimization of the lighting scheme.

[0044] Among them, the behavioral feedback indicators of underwater organisms can be any one of the following: activity frequency, food intake, aggregation density, or stress performance of underwater organisms.

[0045] Furthermore, the reinforcement learning agent employs a proximal policy optimization algorithm, aiming to optimize underwater organisms to be "healthy, low-stress, and fast-growing," while simultaneously conserving energy. A reward function is then constructed based on these objectives. The reward function's design metrics include a health index, stress penalty, activity reward, and energy consumption penalty. The health index can be directly mapped from the real-time physiological state parameters output by the model to the reward. The stress penalty can be directly mapped from the stress index in the real-time physiological state parameters; the higher the stress index, the greater the penalty. The activity reward can be obtained by detecting objective behavioral feedback indicators of the organism, such as food intake, feeding frequency, or normal activity frequency. The energy consumption penalty refers to the actual power consumption of the LEDs in the full-spectrum LED array. The reward function is constructed based on these metrics, for example: R = w1·(1-stress index) + w2·health index + w3·normalized food intake value - w4·energy consumption penalty term; R is the total reward score, and w1, w2, w3, and w4 are the weighting coefficients. The reward function here can be constructed according to the actual situation, and the above indicators are not limited to those given in this embodiment.

[0046] In addition, before step S3, the following steps are included: determining whether the real-time physiological state parameters exceed the preset safety threshold. If so, the current target lighting scheme is overwritten with the preset safe and conservative lighting template. At the same time, the state information when the safety mechanism is triggered and the original lighting action are recorded as negative sample trajectories and stored in the experience replay buffer of the deep reinforcement learning agent for training the policy network to learn the safety action boundary. If not, step S3 is executed, that is, the initial lighting scheme is obtained by matching the biological species and growth stage from the lighting knowledge base.

[0047] Step S4: Drive and control the light output of the light-emitting device according to the current target lighting scheme, and obtain the actual lighting parameters in the target water area based on the distributed underwater lighting sensors installed in the target underwater area. Then, dynamically modulate the light output of the light-emitting device according to the actual lighting parameters and the current target lighting scheme until the obtained actual lighting parameters match the current target lighting scheme.

[0048] Specifically, firstly, a full-spectrum LED array matching the target water area is driven to output light to the water area where the underwater organism is currently located, based on the current target lighting scheme. Simultaneously, distributed lighting sensors acquire the actual lighting parameters in the water. Since objects within the water may cause deviations between the actual lighting parameters and those in the current target lighting scheme, distributed lighting sensors are also needed to acquire the actual lighting parameters in real time for dynamic adjustment.

[0049] The deviation between the actual illumination parameters and the illumination parameters in the current target illumination scheme is calculated, and the output illumination of the full-spectrum programmable LED array is dynamically adjusted according to the deviation, so that the actual illumination parameters of the underwater illumination match the illumination parameters in the current target illumination scheme.

[0050] In this invention, a closed-loop compensation controller is employed to dynamically adjust the output illumination of the full-spectrum programmable LED array based on deviations. The closed-loop compensation controller uses a model predictive control algorithm. It performs feedforward predictive compensation of the full-spectrum programmable LED array output by combining a pre-established digital twin model of the water's optical characteristics, and coordinates adjustment based on feedback compensation of real-time deviations to compensate for water attenuation and ambient light changes, thereby matching the actual illumination parameters with the illumination parameters in the current target illumination scheme. Example 2

[0051] Based on Embodiment 1, the present invention also provides an intelligent linkage dimming system based on underwater biometrics, comprising: The data acquisition module is used to acquire multimodal sensing data in real time by setting up multispectral image acquisition units, active electric field imaging units, and acoustic signal acquisition units at sensing nodes within the target water body area; and to synchronously acquire real-time environmental parameters through environmental sensors distributed within the target water body area; the multimodal sensing data includes multispectral images, active electric field images, and acoustic signals; the real-time environmental parameters include water temperature, water flow velocity, turbidity, and dissolved oxygen; The inference module is used to input the multimodal sensing data into a pre-constructed heterogeneous multimodal fusion recognition model of the system. After feature extraction by the corresponding modal feature encoders, the feature data of the multimodal sensing data are fused based on a cross-modal cross-attention fusion mechanism. Then, combined with a multi-task decoder, the module synchronously outputs the species, growth stage, and real-time physiological state parameters of the underwater organisms currently passing through the target water area. The real-time physiological state parameters are any one of stress index, health index, and metabolic activity. The matching and correction module is used to match an initial lighting scheme from a pre-built lighting knowledge base based on the species and growth stage of the underwater organisms currently passing through the target water body area, and to dynamically correct the initial lighting scheme based on the real-time physiological state parameters and real-time environmental parameters to generate the current target lighting scheme. The dynamic control module is used to drive and control the corresponding light-emitting device to output light to the corresponding target water body area according to the current target lighting scheme. At the same time, it acquires the actual lighting parameters in the target water body area based on the distributed underwater lighting sensor, and dynamically adjusts the output lighting of the corresponding light-emitting device according to the actual lighting parameters and the current target lighting scheme, so that the acquired actual lighting parameters match the current target lighting scheme.

[0052] Furthermore, the heterogeneous multimodal fusion recognition model is obtained through the following steps: Training dataset construction steps: Acquire multispectral images, active electric field images, and acoustic signals of underwater organisms in the target water body, as well as corresponding labeled data of biological species, growth stages, and physiological state parameters, and construct the training dataset; wherein, the labeled data of physiological state parameters are obtained through behavioral analysis, blood biochemical detection, and respiratory metabolism measurement; Model construction steps: Construct a heterogeneous multimodal fusion recognition model; the heterogeneous multimodal fusion recognition model includes a multispectral encoder, an active electric field encoder, an acoustic encoder, a cross-modal cross-attention fusion network, and a multi-task decoder; wherein, the multispectral encoder is used to extract multispectral feature maps from the multispectral image; the active electric field encoder is used to extract conductivity distribution feature maps from the active electric field image; the acoustic encoder is used to extract acoustic feature vectors containing temporal information from the Mel spectrum of the acoustic signal; the cross-modal cross-attention fusion network is used to expand the multispectral feature map, conductivity distribution feature map, and acoustic feature vector into a labeled sequence, add learnable modal embedding position encoding, and then perform intermodal interaction through a multi-layer cross-attention mechanism to generate multimodal fusion features; the multi-task decoder includes a biological species classification head, a growth stage ordered regression head, and a physiological state parameter regression head; Model training steps: First, the multispectral encoder, active electric field encoder, and acoustic encoder are pre-trained using the datasets of the corresponding modalities. Then, after loading the pre-trained weights, the heterogeneous multimodal fusion recognition model is jointly fine-tuned end-to-end using the multi-task uncertainty weighted total loss until convergence. The multi-task uncertainty weighted total loss is composed of the weighted sum of species classification cross-entropy loss, growth stage ordered regression loss, and physiological state parameter mean square error loss.

[0053] Furthermore, the data acquisition module is also used to determine that an organism has entered the target water area when the active electric field imaging unit detects an electric field disturbance exceeding a preset threshold, and simultaneously wakes up the multispectral image acquisition unit and the acoustic signal acquisition unit to acquire data.

[0054] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. An intelligent linkage dimming method based on underwater biological recognition, characterized in that, include: Data acquisition steps: Multispectral image acquisition units, active electric field imaging units, and acoustic signal acquisition units are set up at sensing nodes within the target water body area to acquire multimodal sensing data in real time; The system also synchronously acquires real-time environmental parameters through environmental sensors distributed within the target water body area; the multimodal sensing data includes multispectral images, active electric field images, and acoustic signals; the real-time environmental parameters include water temperature, water flow velocity, turbidity, and dissolved oxygen. Reasoning steps: The multimodal sensing data is input into the heterogeneous multimodal fusion recognition model pre-constructed by the system. After feature extraction by the corresponding modal feature encoder, the feature data of the multimodal sensing data is fused based on the cross-modal cross-attention fusion mechanism. Then, combined with the multi-task decoder, the biological species, growth stage and real-time physiological state parameters of the underwater organisms that are currently passing through the target water area are output synchronously. The real-time physiological state parameter is any one of the stress index, health index, and metabolic activity. Matching and correction steps: Based on the species and growth stage of the underwater organisms currently passing through the target water body area, an initial lighting scheme is matched from the pre-built lighting knowledge base, and the initial lighting scheme is dynamically corrected based on the real-time physiological state parameters and real-time environmental parameters to generate the current target lighting scheme. Dynamic control steps: Drive and control the corresponding light-emitting device to output light to the corresponding target water body area according to the current target lighting scheme. At the same time, acquire the actual lighting parameters in the target water body area based on the distributed underwater lighting sensor, and dynamically adjust the output lighting of the corresponding light-emitting device according to the actual lighting parameters and the current target lighting scheme so that the acquired actual lighting parameters match the current target lighting scheme. 2.The intelligent linkage dimming method based on underwater biological recognition according to claim 1, wherein, The heterogeneous multimodal fusion recognition model is obtained through the following steps: Training dataset construction steps: Acquire multispectral images, active electric field images, and acoustic signals of underwater organisms in the target water body, as well as corresponding labeled data of biological species, growth stages, and physiological state parameters, and construct the training dataset; wherein, the labeled data of physiological state parameters are obtained through behavioral analysis, blood biochemical detection, and respiratory metabolism measurement; Model construction steps: Construct a heterogeneous multimodal fusion recognition model; the heterogeneous multimodal fusion recognition model includes a multispectral encoder, an active electric field encoder, an acoustic encoder, a cross-modal cross-attention fusion network, and a multi-task decoder; wherein, the multispectral encoder is used to extract multispectral feature maps from the multispectral image; the active electric field encoder is used to extract conductivity distribution feature maps from the active electric field image; the acoustic encoder is used to extract acoustic feature vectors containing temporal information from the Mel spectrum of the acoustic signal; the cross-modal cross-attention fusion network is used to expand the multispectral feature map, conductivity distribution feature map, and acoustic feature vector into a labeled sequence, add learnable modal embedding position encoding, and then perform intermodal interaction through a multi-layer cross-attention mechanism to generate multimodal fusion features; the multi-task decoder includes a biological species classification head, a growth stage ordered regression head, and a physiological state parameter regression head; Model training steps: First, the multispectral encoder, active electric field encoder, and acoustic encoder are pre-trained using the datasets of the corresponding modalities. Then, after loading the pre-trained weights, the heterogeneous multimodal fusion recognition model is jointly fine-tuned end-to-end using the multi-task uncertainty weighted total loss until convergence. The multi-task uncertainty weighted total loss is composed of the weighted sum of species classification cross-entropy loss, growth stage ordered regression loss, and physiological state parameter mean square error loss. 3.The intelligent linkage dimming method based on underwater biometric identification of claim 1, wherein, The data acquisition steps include: when the active electric field imaging unit detects an electric field disturbance exceeding a preset threshold, it determines that an organism has entered the target water area and simultaneously wakes up the multispectral image acquisition unit and the acoustic signal acquisition unit to acquire data. 4.The intelligent linkage dimming method based on underwater biometric identification of claim 1, wherein, The data acquisition steps specifically include: installing electric field electrodes in the target water area and controlling the electric field electrodes to emit an alternating excitation electric field with a preset frequency and amplitude into the target water area; collecting electric field disturbance signals caused by the difference in conductivity and dielectric constant of underwater organisms through a sensor array composed of multiple electrodes; and combining the electric field disturbance signals with the empty field reference signal when there are no organisms in the target water area using an electrical impedance tomography reconstruction algorithm to generate an active electric field image reflecting the spatial distribution of the internal conductivity of underwater organisms. Multispectral images, i.e., characteristic bands reflecting the physiological state of underwater organisms, are acquired by a multispectral camera installed in the target water area; the characteristic bands include chlorophyll fluorescence bands and / or hemoglobin absorption bands. Acoustic signals are acquired by hydrophones installed in the target water body area. 5.The intelligent linkage dimming method based on underwater biometric identification of claim 1, wherein, The matching and correction step specifically includes: inputting the biological species, growth stage, real-time physiological state parameters, and real-time environmental parameters output by the model as the state space into the constructed deep reinforcement learning agent; using the illumination parameters as the action space, the reinforcement learning agent calculates reward signals based on the behavioral feedback indicators of the underwater organisms to optimize the initial illumination scheme online in real time, thereby obtaining the current illumination scheme; the behavioral feedback indicators include the activity frequency, feeding amount, aggregation density, or stress performance of the underwater organisms; the illumination parameters include spectral composition, light intensity, and time variation curves; the reinforcement learning agent uses a proximal policy optimization algorithm to achieve optimization. 6.The intelligent linkage dimming method based on underwater biometric identification according to claim 5, wherein, Before the matching correction step, the method further includes: determining whether the real-time physiological state parameters exceed a preset safety threshold. If so, the current target lighting scheme is overwritten with a preset safe and conservative lighting template, and the state information when the safety mechanism is triggered and the original lighting action are recorded as negative sample trajectories and stored in the experience replay buffer of the deep reinforcement learning agent for training the policy network to learn the safe action boundary. If not, the matching correction step is executed. 7.The intelligent linkage dimming method based on underwater biometric identification of claim 1, wherein, The specific implementation of driving and controlling the corresponding luminescent device to output illumination to the corresponding target water area according to the current target illumination scheme includes: dynamically adjusting the output illumination of the corresponding luminescent device through a closed-loop compensation controller; the closed-loop compensation controller adopts a model predictive control algorithm, performs feedforward prediction compensation on the illumination output data of the luminescent device by combining a pre-established digital twin model of the optical characteristics of the water body, and performs real-time feedback compensation and coordinated adjustment based on the deviation between the actual illumination parameters and the target illumination parameters in the current target illumination scheme, so as to compensate for water body attenuation and changes in ambient light, thereby matching the actual illumination parameters with the target illumination parameters.

8. The intelligent linkage dimming system based on underwater biological recognition, characterized in that, The intelligent linkage dimming system includes: The data acquisition module is used to acquire multimodal sensing data in real time by setting up multispectral image acquisition units, active electric field imaging units, and acoustic signal acquisition units at sensing nodes within the target water body area; and to synchronously acquire real-time environmental parameters through environmental sensors distributed within the target water body area; the multimodal sensing data includes multispectral images, active electric field images, and acoustic signals; the real-time environmental parameters include water temperature, water flow velocity, turbidity, and dissolved oxygen; The inference module is used to input the multimodal sensing data into a pre-constructed heterogeneous multimodal fusion recognition model of the system. After feature extraction by the corresponding modal feature encoders, the feature data of the multimodal sensing data are fused based on a cross-modal cross-attention fusion mechanism. Then, combined with a multi-task decoder, the module synchronously outputs the species, growth stage, and real-time physiological state parameters of the underwater organisms currently passing through the target water area. The real-time physiological state parameters are any one of stress index, health index, and metabolic activity. The matching and correction module is used to match an initial lighting scheme from a pre-built lighting knowledge base based on the species and growth stage of the underwater organisms currently passing through the target water body area, and to dynamically correct the initial lighting scheme based on the real-time physiological state parameters and real-time environmental parameters to generate the current target lighting scheme. The dynamic control module is used to drive and control the corresponding light-emitting device to output light to the corresponding target water body area according to the current target lighting scheme. At the same time, it acquires the actual lighting parameters in the target water body area based on the distributed underwater lighting sensor, and dynamically adjusts the output lighting of the corresponding light-emitting device according to the actual lighting parameters and the current target lighting scheme, so that the acquired actual lighting parameters match the current target lighting scheme.

9. The intelligent linkage dimming system based on underwater biometrics according to claim 8, characterized in that, The heterogeneous multimodal fusion recognition model is obtained through the following steps: Training dataset construction steps: Acquire multispectral images, active electric field images, and acoustic signals of underwater organisms in the target water body, as well as corresponding labeled data of biological species, growth stages, and physiological state parameters, and construct the training dataset; wherein, the labeled data of physiological state parameters are obtained through behavioral analysis, blood biochemical detection, and respiratory metabolism measurement; Model construction steps: Construct a heterogeneous multimodal fusion recognition model; the heterogeneous multimodal fusion recognition model includes a multispectral encoder, an active electric field encoder, an acoustic encoder, a cross-modal cross-attention fusion network, and a multi-task decoder; wherein, the multispectral encoder is used to extract multispectral feature maps from the multispectral image; the active electric field encoder is used to extract conductivity distribution feature maps from the active electric field image; the acoustic encoder is used to extract acoustic feature vectors containing temporal information from the Mel spectrum of the acoustic signal; the cross-modal cross-attention fusion network is used to expand the multispectral feature map, conductivity distribution feature map, and acoustic feature vector into a labeled sequence, add learnable modal embedding position encoding, and then perform intermodal interaction through a multi-layer cross-attention mechanism to generate multimodal fusion features; the multi-task decoder includes a biological species classification head, a growth stage ordered regression head, and a physiological state parameter regression head; Model training steps: First, the multispectral encoder, active electric field encoder, and acoustic encoder are pre-trained using the datasets of the corresponding modalities. Then, after loading the pre-trained weights, the heterogeneous multimodal fusion recognition model is jointly fine-tuned end-to-end using the multi-task uncertainty weighted total loss until convergence. The multi-task uncertainty weighted total loss is composed of the weighted sum of species classification cross-entropy loss, growth stage ordered regression loss, and physiological state parameter mean square error loss.

10. The intelligent linkage dimming system based on underwater biometrics according to claim 8, characterized in that, The data acquisition module is also used to determine that an organism has entered the target water area when the active electric field imaging unit detects an electric field disturbance exceeding a preset threshold, and simultaneously wakes up the multispectral image acquisition unit and the acoustic signal acquisition unit to acquire data.

Images