A multi-level biological early warning fusion method and system based on luminous bacteria and fish

Abstract

The application provides a multi-level biological early warning fusion method and system based on luminous bacteria and fish, which comprises multi-source biological data synchronous acquisition; a multi-level biological early warning hardware platform is built, which comprises a luminous bacteria monitoring module, a behavior monitoring module of blue gourami, and a data synchronous acquisition unit; biological data preprocessing and abnormal feature extraction; a dynamic weight Bayesian SVM fusion model is constructed; graded early warning output and result feedback: based on a comprehensive toxicity index, four early warning threshold values are set, and graded early warning signals are output; timeliness and accuracy are taken into account: the advantages of rapid response of luminous bacteria in 5-15 min and stable verification of blue gourami in 10-60 min are fused, the single blue gourami early warning response time is shortened by 40%-60%, the single luminous bacteria early warning false alarm rate is reduced by 30%-50%, and extremely fast and accurate early warning is realized.

Description

Technical Field

[0001] This invention relates to the field of water quality biological early warning technology, and in particular to a multi-level biological early warning fusion method and system based on luminescent bacteria and fish. Background Technology

[0002] Biological early warning utilizes model organisms' specific responses to toxic substances to achieve real-time monitoring of water toxicity. Bioluminescent bacteria and medaka are core model organisms for water quality early warning. Bioluminescent bacteria (such as *Vibrio fischeri*) are sensitive to heavy metals and organic pollutants, with their luminescence intensity rapidly decaying after toxic exposure (response time 5-15 min), offering rapid early warning advantages. However, they suffer from drawbacks such as low species hierarchy, weak environmental resistance to interference, and susceptibility to water quality parameters (temperature, pH) affecting single indicators, resulting in a high false alarm rate. Medaka (a model vertebrate) exhibits stable responses to chronic toxicity and complex pollution, reflecting water toxicity through changes in behavioral characteristics (behavioral intensity, swimming speed, trajectory, aggregation degree), demonstrating high stability and reliability. However, their long response delay (10-60 min) cannot meet the rapid early warning requirements for sudden pollution events.

[0003] Existing multi-level biological early warning systems mostly adopt an "independent monitoring, simple threshold superposition" model. For example, the luminescent bacteria and medaka modules output alarm information separately, and the warning is triggered only when both alarm simultaneously. This model has three major defects: First, data is fragmented, as the response characteristics (time series, dimensionality, sensitivity) of the two types of organisms differ greatly, and independent analysis cannot uncover toxicity correlation information; second, the weights are fixed, and the warning weights of the two types of organisms are not dynamically adjusted according to the type and concentration of pollutants, resulting in poor adaptability for organic pollution relying on luminescent bacteria and heavy metals relying on medaka; third, the fusion accuracy is low, as simple threshold superposition cannot handle data noise and outliers, leading to high rates of missed reports (low-concentration compound pollution) and false alarms (single parameter fluctuations).

[0004] Existing algorithms, such as single SVM and Bayesian models, only model single biological data and cannot achieve deep fusion of multi-source biological data; traditional weighted fusion algorithms have static weights and cannot adapt to complex water quality scenarios. Therefore, there is an urgent need for a fusion algorithm that can simultaneously integrate the advantages of rapid response of bioluminescent bacteria and stable verification by killifish, dynamically adapt to pollutant characteristics, and accurately output comprehensive toxicity, so as to solve the problems of timeliness and accuracy contradiction and low data utilization of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a multi-level biological early warning fusion method and system based on luminescent bacteria and fish, which achieves deep fusion of two types of biological data, dynamically adapts to pollutant types, and improves the timeliness, accuracy, and reliability of water quality toxicity early warning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a multi-level biological early warning fusion method based on luminescent bacteria and fish, comprising the following steps: S1. Synchronous acquisition of multi-source biological data; A multi-level biological early warning hardware platform was built, including a luminescent bacteria monitoring module, a medaka behavior monitoring module, and a data synchronization and acquisition unit; S2. Biological data preprocessing and anomaly feature extraction; S3, Construction of Dynamic Weighted Bayesian SVM Fusion Model; S4. Tiered Early Warning Output and Result Feedback: Based on Comprehensive Toxicity Index Set four levels of early warning thresholds and output graded early warning signals.

[0007] Furthermore, in S1, The luminescent bacteria monitoring module collects real-time time-series data on luminescence intensity. Sampling frequency: 1 time / minute; The medaka behavior monitoring module collects video data through a high-definition camera and extracts swimming speed through image processing. Trajectory complexity Group aggregation Three-dimensional behavioral feature time-series data, with a sampling rate of 1 time / minute; The data synchronization acquisition unit aligns the two types of data using timestamps to form a synchronized dataset. .

[0008] Furthermore, S2 includes: Preprocessing of bioluminescent bacteria data: Time series data of luminescence intensity Wavelet threshold denoising is used to remove noise caused by temperature and pH fluctuations; Calculate the light emission suppression rate , The luminescence intensity is for the blank control group; Extract the mean inhibition rate using a sliding window. Coefficient of variation This constitutes the abnormal feature vector of luminescent bacteria. .

[0009] Furthermore, S2 also includes: Preprocessing of medaka fish behavioral data: swimming speed Trajectory complexity Group aggregation Normalize them separately; Principal component analysis (PCA) was used for dimensionality reduction to extract the first principal component. As a comprehensive behavioral characteristic; Identify behavioral anomalies using a One-Class SVM model and output the anomaly probability. This constitutes the abnormal feature vector of the medaka fish. .

[0010] Furthermore, S3 specifically refers to: The structure of the S301 dynamic weighted Bayesian SVM fusion model includes an input layer, a feature fusion layer, and a decision output layer. The input layer receives the abnormal feature vector of luminescent bacteria. Abnormal feature vectors of medaka ; S302. Dynamic Weight Calculation: Based on prior knowledge of pollutant types and real-time data characteristics, a dynamic weight allocation formula is constructed: ; ; in, Weights for luminescent bacteria; Weighting of medaka; This is the adjustment coefficient; This represents the maximum emission suppression rate of the current window. The warning threshold for luminescent bacteria; the weight of luminescent bacteria at low concentrations of organic pollution. Increased weight of medaka when heavy metals or multiple pollutants are present. Increase; S303, Feature Fusion and Decision Making; Fusion of Feature Vectors ; Will Input the trained dynamic weighted Bayesian SVM fusion model and output preliminary toxicity classification results; The comprehensive toxicity index is calculated by applying Bayesian probability correction and combining prior probabilities from historical warning data. ; in, This is a correction factor; The toxicity score output by the dynamically weighted Bayesian SVM fusion model; For Bayesian posterior probability; The value ranges from 0 to 100, with higher values ​​indicating greater toxicity.

[0011] Furthermore, S4 specifically includes: Level I: The water quality is normal and is being monitored as usual. Level II: Low-toxicity pollution; encrypted monitoring initiated. Level III: Moderate toxic pollution triggers an early warning and sends out pollution alarm information; Level IV: In cases of severe toxic pollution, an emergency response is activated, linking the water quality tracing module and simultaneously feeding the early warning results back to the data acquisition module. The sampling frequency and monitoring parameters are then dynamically adjusted to form a closed-loop early warning system.

[0012] Furthermore, it includes at least one processor, and a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor to implement the multi-level biological early warning fusion method based on luminescent bacteria and medaka.

[0013] The beneficial effects of this invention are: it balances timeliness and accuracy: it combines the advantages of rapid response of luminescent bacteria (5-15 min) and stable verification of medaka fish (10-60 min), which shortens the early warning response time by 40%-60% compared with single medaka fish and reduces the false alarm rate of early warning by 30%-50% compared with single luminescent bacteria, thus achieving extremely rapid and accurate early warning; Dynamically adapts to pollutant types: Through a dynamic weight allocation mechanism, the weights of the two types of biological data are automatically adjusted for organic pollution, heavy metal pollution, and compound pollution, solving the problem of poor adaptability of traditional fixed weights and improving the accuracy of compound pollution identification by more than 25%. Deep fusion of multi-source data: By employing wavelet denoising, PCA dimensionality reduction, and One-Class SVM anomaly detection preprocessing, combined with a Bayesian-SVM fusion model, the toxicity correlation information of the two types of biological data is fully explored, improving data utilization by 40% and reducing the underreporting rate of low-concentration pollution by 60%. Accurate and controllable tiered early warning: Based on the comprehensive toxicity index, the four-level tiered early warning system adapts to the response needs of different pollution levels, avoiding over-warning or under-warning. It is applicable to water quality early warning in multiple scenarios such as water sources, river basins, and industrial park drainage outlets, and is highly practical. Attached Figure Description

[0014] Figure 1 This is a flowchart of a multi-level biological early warning fusion method based on luminescent bacteria and fish; Figure 2 This is a schematic diagram of the structure of the dynamic weighted Bayesian SVM fusion model. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] Please see Figure 1 A multi-level biological early warning fusion method based on luminescent bacteria and fish includes the following steps: S1. Synchronous acquisition of multi-source biological data; A multi-level biological early warning hardware platform was built, including a luminescent bacteria monitoring module, a medaka behavior monitoring module, and a data synchronization and acquisition unit; S2. Biological data preprocessing and anomaly feature extraction; S3, Construction of Dynamic Weighted Bayesian SVM Fusion Model; S4. Tiered Early Warning Output and Result Feedback: Based on Comprehensive Toxicity Index Set four levels of early warning thresholds and output graded early warning signals.

[0017] In S1, The luminescent bacteria monitoring module collects real-time time-series data on luminescence intensity. Sampling frequency: 1 time / minute; The killifish behavior monitoring module collects video data through a high-definition camera and extracts swimming speed through image processing. Trajectory complexity Group aggregation Three-dimensional behavioral feature time-series data, with a sampling rate of 1 time / minute; The data synchronization acquisition unit aligns the two types of data using timestamps to form a synchronized dataset. .

[0018] S2 includes: Preprocessing of bioluminescent bacteria data: Time series data of luminescence intensity Wavelet threshold denoising is used to remove noise caused by temperature and pH fluctuations; Calculate the light emission suppression rate , The luminescence intensity is for the blank control group; Extract the mean inhibition rate using a sliding window. Coefficient of variation This constitutes the abnormal feature vector of luminescent bacteria. The window length is 5 minutes.

[0019] S2 further includes: Preprocessing of medaka fish behavioral data: swimming speed Trajectory complexity Group aggregation Normalize them separately and map them to interval; Principal component analysis (PCA) was used for dimensionality reduction to extract the first principal component. As a comprehensive behavioral characteristic; Identify behavioral anomalies using a One-Class SVM model and output the anomaly probability. This constitutes the abnormal feature vector of the medaka fish. .

[0020] Please see Figure 2 Specifically, S3 is: The structure of the S301 dynamic weighted Bayesian SVM fusion model includes an input layer, a feature fusion layer, and a decision output layer. The input layer receives the abnormal feature vector of luminescent bacteria. Abnormal feature vectors of medaka ; The feature fusion layer calculates the fusion weights of two features through a dynamic weight allocation unit; The decision output layer uses a Bayesian SVM hybrid model to output a comprehensive toxicity index. ; S302. Dynamic Weight Calculation: Based on prior knowledge of pollutant types and real-time data characteristics, a dynamic weight allocation formula is constructed: ; ; in, Weights for luminescent bacteria; Weighting of medaka; This is the adjustment coefficient, and its value is [value missing]. ; This represents the maximum emission suppression rate of the current window. The warning threshold for luminescent bacteria is set at 75%; the weight of luminescent bacteria at low concentrations of organic pollution. Increased, i.e., 0.7~0.9; weight of medaka fish in cases of heavy metal or complex pollution. Increased, 0.6~0.8; S303, Feature Fusion and Decision Making; Fusion of Feature Vectors ; Will Input the trained dynamic weighted Bayesian SVM fusion model and output preliminary toxicity classification results; The comprehensive toxicity index is calculated by applying Bayesian probability correction and combining prior probabilities from historical warning data. ; in, Here, is the correction factor, , , ; The toxicity score (0~100) output by the dynamic weighted Bayesian SVM fusion model. is the Bayesian posterior probability (0~1); The value ranges from 0 to 100, with higher values ​​indicating greater toxicity.

[0021] Specifically, S4 is: Level I (Safe, Green): The water quality is normal and is being monitored as usual. Grade II (mild abnormality, yellow): Low toxicity pollution; encrypt monitoring initiated, sampling frequency 1 time / 30 seconds; Level III (Moderately Abnormal, Orange): Moderate toxic pollution triggers an early warning and sends out pollution alarm information; Level IV (Severe abnormality, red): In cases of severe toxic pollution, an emergency response is activated, linking the water quality tracing module and simultaneously feeding the early warning results back to the data acquisition module. The sampling frequency and monitoring parameters are then dynamically adjusted to form a closed-loop early warning system.

[0022] It includes at least one processor and a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor to implement the multi-level biological early warning fusion method based on luminescent bacteria and medaka.

[0023] Example 1: Early warning of sudden organic pollution in water source: Data collection: A hardware platform was built, and a bioluminescent bacteria monitoring module (Vibrio fischeri) collected time-series data on luminescence intensity. The medaka module collects swimming speed. Trajectory complexity Group aggregation Timestamp synchronization, sampling frequency 1 time / minute; Data preprocessing: Data on bioluminescent bacteria: Calculation of luminescence suppression rate after wavelet denoising Sliding window, 5 minutes, extract , Abnormal feature vectors of luminescent bacteria .

[0024] Medaka data: After normalization, PCA will retain the first principal component. One-Class SVM Model Identifies Anomaly Probability Abnormal feature vectors of medaka ; Fusion model calculation: Dynamic weights: , , .

[0025] Fusion characteristics: ; Overall toxicity index: score Bayesian posterior probability , .

[0026] Early warning output: Output a Level II (yellow) mild anomaly warning and initiate encrypted monitoring.

[0027] Example 2: Early warning of heavy metal composite pollution at industrial park drainage outlets: Data collection: Simultaneous data collection of luminescent bacteria and medaka showed a rapid decrease in luminescence intensity, a slowing of medaka swimming speed, and an increase in aggregation.

[0028] Data preprocessing: Data on bioluminescent bacteria: , Abnormal feature vectors of luminescent bacteria ;

[0029] Data on medaka fish: , Abnormal feature vectors of medaka .

[0030] Fusion model calculation: Dynamic weights: , , .

[0031] Fusion characteristics: ; Overall Toxicity Index: SVM Score Yes posterior probability , ; Early warning output: Output a Level III (orange) moderate anomaly warning, trigger an alarm and start encrypted monitoring.

[0032] Video capture of medaka fish was performed, and V / C / G three-dimensional behavioral temporal sequences were extracted. Normalization and PCA dimensionality reduction were then used to obtain the first principal component. Input a pre-trained One-Class SVM, calculate the distance of a sample from the normal distribution, and output the anomaly probability. Combined into a feature vector of medaka fish The data is fed into a dynamic weighted Bayesian algorithm, SVM, and luminescent bacteria characteristics to calculate the comprehensive toxicity index. .

[0033] How does the One-Class SVM model work? One-Class SVM (Single-Class Support Vector Machine) was used to process the three-dimensional behavioral features of medaka fish, specifically their swimming speed. Trajectory complexity Group aggregation .

[0034] Objective: To model and identify abnormal toxic behaviors that deviate from the normal pattern using only behavioral samples of killifish under normal water conditions, and output the probability of such abnormalities. Conventional binary SVMs require labeled samples of both normal and abnormal classes. However, in water quality monitoring, abnormal pollution samples are scarce and it is difficult to fully cover all types of toxins. One-Class SVMs only require normal samples, making them a perfect fit for this scenario.

[0035] Complete construction process of one-class SVM model: Primitive temporal characteristics of medaka: swimming speed Trajectory complexity Group aggregation ; Feature normalization: all features are mapped to the [0,1] interval to eliminate dimensional differences; PCA dimensionality reduction: Three-dimensional behavioral features are compressed into a single-dimensional comprehensive behavior first principal component. This reduces data redundancy and model computation. Input vector for one-class SVM model: One-dimensional feature sequence All samples were standard samples collected under conditions of non-toxicity and normal water quality.

[0036] Only the comprehensive behavioral characteristics of medaka fish in long-term stable, unpolluted water sources are collected as positive samples (normal samples), without the need for any polluted or abnormal samples to participate in the training. Medaka fish in normal water quality exhibit stable behavior and form a fixed distribution; when the water is toxic, their swimming speed and aggregation patterns change drastically, and their behavioral characteristics deviate from the historical normal distribution, thus being identified as abnormal.

[0037] How does the dynamic weighted Bayesian SVM fusion model work? Overall model definition: The Dynamic Weighted Bayesian SVM fusion model is a hierarchical fusion decision model for dual-source heterogeneous biological time-series features, with a three-layer core structure: Dynamic weight allocation layer: Real-time adaptive allocation of weights for two feature fusion classes: luminescent bacteria and medaka. Medaka weight This solves the problem of poor adaptation to pollution types in traditional fixed weights; SVM Discriminant Layer: Performs nonlinear toxicity discrimination on the weighted fused comprehensive features, and outputs a toxicity score in the range of 0~100; Bayesian probabilistic correction layer: Introduces prior historical water quality probabilities to correct the deterministic scoring of SVM, outputting a comprehensive toxicity index with uncertainty. It is used for four-level graded early warning.

[0038] Complete construction process of dynamic weighted Bayesian SVM fusion model: The model is divided into four independent building modules: dynamic weight allocation module, SVM classification and scoring module, Bayesian probability calculation module, and so on. Fusion output module; Dynamic weight allocation module: This has already been explained.

[0039] SVM sub-model construction (toxicity scoring module); Input weighted fusion features Output a continuous toxicity score from 0 to 100. Different levels of toxicity are distinguished.

[0040] Dataset creation: A large number of labeled samples were collected: the labels were divided into four categories (Ⅰ safe, Ⅱ mild, Ⅲ moderate, Ⅳ severe), and each sample corresponded to a set. Calculate the corresponding fusion feature vector in advance As input for SVM training.

[0041] Model training: Multi-class SVM was used for training, with the original output being class labels; then regression mapping was performed to convert the classification distance into a continuous score toxicity rating in the range of 0 to 100. : The closer the distance from the hyperplane is to the highly toxic category, the closer the score is to 100. It leans towards the normal category, with a score close to 0.

[0042] Bayesian sub-model construction (probability correction module)

[0043] Calculate the posterior probability based on historical monitoring data. It characterizes the reliability of the current toxicity assessment and is used to smooth out transient noise.

[0044] Statistical prior probability : Statistics on long-term historical monitoring data: A1: Water safety incident; A2: Mild pollution; A3: Moderate pollution; A4: Severe pollution ; Obtain the basic prior probabilities of various types of pollution occurring.

[0045] Constructing likelihood probability ; Statistics on fusion characteristics under each type of pollution Assuming a multidimensional Gaussian distribution, calculate the mean and covariance matrix corresponding to each label class as the likelihood function.

[0046] Construction of the Bayesian posterior formula: ; The overall confidence probability is obtained by integration. : Combine the posterior probabilities of moderate and severe contamination to form the overall contamination confidence level: ; The value ranges from 0 to 1. The larger the value, the higher the credibility of the presence of toxic pollution in the current water body.

[0047] The prior probabilities and Gaussian distribution mean / covariance of various types of pollution are permanently stored and directly substituted into real-time data during inference. calculate .

[0048] Overall toxicity index Construction of integrated output unit (decision-level integration).

[0049] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be defined by the appended claims.

[0050] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A multi-level biological early warning fusion method based on luminescent bacteria and fish, characterized in that, Includes the following steps: S1. Synchronous acquisition of multi-source biological data; A multi-level biological early warning hardware platform was built, including a luminescent bacteria monitoring module, a medaka behavior monitoring module, and a data synchronization and acquisition unit; S2. Biological data preprocessing and anomaly feature extraction; S3, Construction of Dynamic Weighted Bayesian SVM Fusion Model; Specifically, S3 is: The structure of the S301 dynamic weighted Bayesian SVM fusion model includes an input layer, a feature fusion layer, and a decision output layer. The input layer receives the abnormal feature vector of luminescent bacteria. Abnormal feature vectors of medaka ; S302. Dynamic Weight Calculation: Based on prior knowledge of pollutant types and real-time data characteristics, a dynamic weight allocation formula is constructed: ； ； in, Weights for luminescent bacteria; Weighting of medaka; This is the adjustment coefficient; This represents the maximum emission suppression rate of the current window. The warning threshold for luminescent bacteria; the weight of luminescent bacteria at low concentrations of organic pollution. Increased weight of medaka when heavy metals or multiple pollutants are present. Increase; S303, Feature Fusion and Decision Making; Fusion of Feature Vectors ； Will Input the trained dynamic weighted Bayesian SVM fusion model and output preliminary toxicity classification results; The comprehensive toxicity index is calculated by applying Bayesian probability correction and combining prior probabilities from historical warning data. ； in, This is a correction factor; The toxicity score output by the dynamically weighted Bayesian SVM fusion model; For Bayesian posterior probability; The value ranges from 0 to 100, with higher values ​​indicating stronger toxicity. S4. Tiered Early Warning Output and Result Feedback: Based on Comprehensive Toxicity Index Set four levels of early warning thresholds and output graded early warning signals.

2. The multi-level biological early warning fusion method based on luminescent bacteria and fish according to claim 1, characterized in that: In S1, The luminescent bacteria monitoring module collects real-time time-series data on luminescence intensity. Sampling frequency: 1 time / minute; The medaka behavior monitoring module collects video data through a high-definition camera and extracts swimming speed through image processing. Trajectory complexity Group aggregation Three-dimensional behavioral feature time-series data, with a sampling rate of 1 time / minute; The data synchronization acquisition unit aligns the two types of data using timestamps to form a synchronized dataset. .

3. The multi-level biological early warning fusion method based on luminescent bacteria and fish according to claim 2, characterized in that, S2 includes: Preprocessing of bioluminescent bacteria data: Time series data of luminescence intensity Wavelet threshold denoising is used to remove noise caused by temperature and pH fluctuations; Calculate the light emission suppression rate , The luminescence intensity is for the blank control group; Extract the mean inhibition rate using a sliding window. Coefficient of variation This constitutes the abnormal feature vector of luminescent bacteria. .

4. The multi-level biological early warning fusion method based on luminescent bacteria and fish according to claim 3, characterized in that, S2 further includes: Preprocessing of medaka fish behavioral data: swimming speed Trajectory complexity Group aggregation Normalize them separately; Principal component analysis (PCA) was used for dimensionality reduction to extract the first principal component. As a comprehensive behavioral characteristic; Identify behavioral anomalies using a One-Class SVM model and output the anomaly probability. This constitutes the abnormal feature vector of the medaka fish. .

5. The multi-level biological early warning fusion method based on luminescent bacteria and fish according to claim 4, characterized in that, Specifically, S4 is: Level I: The water quality is normal and is being monitored as usual. Level II: Low-toxicity pollution; encrypted monitoring initiated. Level III: Moderate toxic pollution triggers an early warning and sends out pollution alarm information; Level IV: In cases of severe toxic pollution, an emergency response is activated, linking the water quality tracing module and simultaneously feeding the early warning results back to the data acquisition module. The sampling frequency and monitoring parameters are then dynamically adjusted to form a closed-loop early warning system.

6. A multi-level biological early warning fusion system based on luminescent bacteria and fish, characterized in that, It includes at least one processor and a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor to implement the multi-level biological early warning fusion method based on luminescent bacteria and fish as described in any one of claims 1 to 5.

Images