Laboratory virus leakage monitoring system and method

Abstract

The invention discloses a laboratory virus leakage monitoring system and method, and the system comprises a data collection module, a virus monitoring module and an alarm module. The data acquisition module is used for acquiring real-time virus leakage influence factor data of a polluted area, a semi-polluted area and a clean area and inputting the data into the virus monitoring module; the virus monitoring module is used for preprocessing the real-time virus leakage factor data, integrating the preprocessed data to obtain a three-dimensional tensor, and inputting the three-dimensional tensor into an LSTM-GAN model for processing to obtain virus leakage abnormal data or normal data; and the alarm module is used for giving an alarm when obtaining the abnormal data of virus leakage. According to the invention, laboratory virus leakage monitoring can be realized.

Description

Technical Field

[0001] The present invention relates to the field of laboratory virus leakage monitoring, and in particular to a laboratory virus leakage monitoring system and method. Background Art

[0002] Amidst accelerating globalization and the frequent emergence of new infectious diseases, virology laboratories, as key sites for infectious disease research, diagnosis, and vaccine development, face increasingly severe biosafety challenges. In recent years, major infectious disease outbreaks have highlighted the central role of virology laboratories in the prevention and control system, while also exposing numerous shortcomings in existing laboratories' monitoring and early warning of virus leaks. According to the World Health Organization, approximately 65% ​​of high-level biosafety laboratories worldwide have experienced virus leaks of varying severity, and most of these incidents have failed to provide timely warnings due to outdated monitoring methods, leading to a sharp increase in the risk of virus spread.

[0003] Virology laboratories are complex, encompassing multiple zones, including contaminated, semi-contaminated, and clean areas. These zones are subject to frequent movement of personnel, equipment, and samples, and viruses can leak through a variety of pathways, including aerosols, surface contact, equipment failure, and human error. Traditional monitoring technologies struggle to simultaneously capture these multiple risk factors and are poorly adaptable to the dynamic laboratory environment. Consequently, they cannot meet the needs of high-biosafety laboratories for real-time, accurate monitoring and proactive early warning of virus leakage risks. Summary of the Invention

[0004] The purpose of the present invention is to provide a laboratory virus leakage monitoring system, aiming to solve the problem of laboratory virus leakage monitoring.

[0005] The present invention provides a laboratory virus leakage monitoring system, comprising:

[0006] Data acquisition module, virus monitoring module and alarm module;

[0007] The data acquisition module is used to collect real-time virus leakage impact factor data in contaminated areas, semi-contaminated areas and clean areas and input it into the virus monitoring module;

[0008] The virus monitoring module is used to preprocess the real-time virus leakage factor data, integrate the preprocessed data to obtain a three-dimensional tensor, and input the three-dimensional tensor into the LSTM-GAN model to obtain abnormal virus leakage data or normal data;

[0009] The alarm module is used to issue an alarm when abnormal data of virus leakage is obtained.

[0010] The present invention also provides a laboratory virus leakage monitoring method, comprising:

[0011] The data acquisition module collects real-time virus leakage impact factor data from contaminated areas, semi-contaminated areas, and clean areas and inputs it into the virus monitoring module;

[0012] The virus monitoring module pre-processes the real-time virus leakage factor data, integrates the pre-processed data to obtain a three-dimensional tensor, and inputs the three-dimensional tensor into the LSTM-GAN model to obtain virus leakage anomaly data;

[0013] When abnormal data of virus leakage is obtained, an alarm is issued through the alarm module.

[0014] By adopting the embodiment of the present invention, laboratory virus leakage monitoring and alarm can be realized.

[0015] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it is implemented in accordance with the contents of the specification, and in order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 Schematic diagram of a laboratory virus leakage monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] System Example

[0020] According to an embodiment of the present invention, a laboratory virus leakage monitoring system is provided. Figure 1 Schematic diagram of a laboratory virus leakage monitoring system according to an embodiment of the present invention. Figure 1 As shown, specifically including:

[0021] Data acquisition module, virus monitoring module and alarm module;

[0022] The data acquisition module is used to collect real-time virus leakage impact factor data in contaminated areas, semi-contaminated areas and clean areas and input it into the virus monitoring module;

[0023] The real-time virus leakage influencing factor data includes: virus sensor data, equipment monitoring data, video stream data, experimental environment data and experimental metadata.

[0024] Virus sensor data were collected using a photochemical immunosensor and a dual-mode particulate matter sensor;

[0025] Use biosafety cabinet differential pressure sensors and vibration sensors to collect equipment monitoring data;

[0026] Temperature and humidity sensors, thermal imaging sensors, pressure gradient sensors, airflow velocity sensors, and ultraviolet sensors are used to collect experimental environment data;

[0027] Collect experimental hazard level data, experimental personnel operation records and personnel entry and exit logs as experimental metadata.

[0028] In the embodiment of the present invention, the specific installation location and parameter settings of the data acquisition module are as follows:

[0029] 1.1 Three-dimensional monitoring network of polluted areas

[0030] The contaminated area is the core experimental area of ​​the virus laboratory and is also a high-risk area for the leakage of infectious viruses. Therefore, a three-dimensional monitoring network is set up in the contaminated area, mainly monitoring the air, surfaces, and equipment in the contaminated area.

[0031] Air medium monitoring: Install a photochemical immunosensor at the air outlet of the biosafety cabinet (H13-HEPA standard site) to capture the virus concentration in the aerosol in real time; deploy a dual-mode particulate matter sensor (0.3-10μm particle size resolution) at the end of the ventilation duct, and combine it with the CFD model to predict the diffusion path.

[0032] Surface medium monitoring: A dual-mode particulate matter sensor is used, a surface-enhanced photochemical sensor (contact type, sampling period 5 seconds / time) is embedded on the edge of the laboratory table, and a laser-induced breakdown spectroscopy (LIBS) scanner (non-contact type, full surface scanning 15 minutes / time) is installed on the centrifuge operation panel.

[0033] Equipment status monitoring: The internal differential pressure sensor of the biosafety cabinet is linked to the central negative pressure system; a three-axis MEMS vibration sensor is installed on the centrifuge door, and abnormal vibration is identified through the SVM model.

[0034] 1.2 Semi-polluted area linkage monitoring network

[0035] The semi-contaminated area is the sample pretreatment area. The semi-contaminated area linkage monitoring network deploys ultraviolet intensity comparison arrays, air flow velocity sensors, and surface temperature and humidity sensors at the transfer window, sample transfer channel access control system, and protective equipment area respectively.

[0036] 1.3 Clean Area Prevention Monitoring Network

[0037] The clean area is where people move around. The preventive monitoring network in the clean area deploys multi-stage filtering detection devices in the personnel / material channels. At the same time, a thermal imaging module is integrated into the access control system in the clean area, and redundant pressure gradient sensors are set up in the clean area.

[0038] 1.4 Experimental metadata acquisition

[0039] If there are experimenters doing experiments in the virus laboratory, the experiment hazard level, experiment operation records and personnel entry and exit logs are recorded in the system as experiment metadata.

[0040] The virus monitoring module is used to preprocess the real-time virus leakage factor data, integrate the preprocessed data to obtain a three-dimensional tensor, and input the three-dimensional tensor into the LSTM-GAN model to obtain abnormal virus leakage data or normal data;

[0041] The virus monitoring module is specifically used to preprocess real-time virus leakage factor data, integrate the preprocessed data into a three-dimensional tensor, and input the three-dimensional tensor into the LSTM-GAN model. The LSTM-GAN model adopts a three-level architecture of encoder-generator-discriminator. After receiving the three-dimensional tensor, the LSTM encoder extracts features and inputs the features into the generator of the GAN network. The generator obtains a normal baseline sequence based on historical normal virus leakage influencing factor data, inputs the three-dimensional tensor and the normal baseline sequence into the discriminator to obtain the normal and abnormal probabilities of the real-time virus leakage factor data. The generator generates the deviation of the normal baseline sequence. When the probability of abnormality is greater than a certain threshold and the deviation is less than a certain threshold, it is judged as a virus leak.

[0042] In this embodiment of the present invention, minimizing the loss function of the discriminator D improves the ability to distinguish between real data and generated data. The entire LSTM-GAN anomaly monitoring layer is continuously optimized through the adversarial training iterative process between the generator and the discriminator, ultimately making the generator fit the normal data distribution law and the discriminator accurately identify abnormal data.

[0043] In an embodiment of the present invention, a historical database is established, and the historical database includes: virus sensor data, equipment monitoring data, video stream data, experimental environment data and experimental metadata.

[0044] In the embodiment of the present invention, virus sensor data, equipment monitoring data, video stream data, experimental environment data and experimental metadata multi-dimensional time series data are used for multi-source data integration (time window T = 60 seconds, step size 10 seconds), and after normalization processing (logarithmic transformation, Z-score normalization), a three-dimensional tensor X∈R is formed. T×N×F (N is the number of sensors, F is the feature dimension, and T is the time window).

[0045] The alarm module is used to issue an alarm when abnormal data of virus leakage is obtained.

[0046] The system also includes: a virus leakage path prediction module, which is used to obtain the leakage point by combining the point position of the sensor corresponding to the abnormal data when a virus leakage is determined, and inversely deduce the virus leakage path according to the leakage point.

[0047] The virus leakage path prediction module is specifically used to:

[0048] A real-time parameterized gas diffusion model is established, which is used to simulate the real-time viral aerosol diffusion path. The diffusion coefficient of the gas diffusion model is calibrated in real time using a filtering algorithm according to the leakage point, and the parameters in the real-time parameterized gas diffusion model are dynamically corrected based on airflow disturbance events.

[0049] Airflow disturbance events include: opening and closing of access control and starting and stopping of disinfection equipment.

[0050] The spatiotemporal features of laboratory dynamic grid nodes are extracted based on the spatiotemporal graph convolutional network to predict the trend of virus concentration changes;

[0051] Through the spatiotemporal graph convolutional network, the spatiotemporal evolution pattern of virus concentration is learned from historical monitoring data, such as local diffusion anomalies caused by human movement;

[0052] Based on the cascade attention mechanism, a diffusion path probability distribution map is generated according to the predicted virus concentration trend and the real-time parameterized gas diffusion model;

[0053] Cascaded attention mechanism: As the fusion center of the two, it dynamically determines the contribution ratio of the gas diffusion model and the spatiotemporal graph convolutional network in different spatial regions and at different time points through hierarchical attention weight distribution, solving the heterogeneous complementarity problem between physical models and data-driven models.

[0054] The essence of the cascaded attention mechanism is to allow the model to dynamically switch perspectives between physical laws and data features through hierarchical focusing: spatial attention solves the problem of where the diffusion concentration is high, and temporal attention solves the problem of the time of virus spread.

[0055] The impact area levels are divided according to the probability distribution map, and a dynamic heat map of the virus leakage path is generated based on the impact area levels.

[0056] In an embodiment of the present invention, the system also includes: an early warning and visualization module, which is used to calculate the leakage risk value based on the viral aerosol diffusion path, output the risk level based on the risk value, issue an early warning based on the risk level, and visualize the viral aerosol diffusion path and risk level.

[0057] In the embodiment of the present invention, the warning risk levels are as follows:

[0058]

[0059] If the risk level reaches level 4-5, a hierarchical linkage mechanism will be adopted (for example, if an abnormality is triggered in the contaminated area, measures will also be taken in the semi-contaminated area and clean area).

[0060] In an embodiment of the present invention, the visualization terminal uses an AR visualization terminal to realize a holographic risk sandbox display (including virus spread simulation, real-time data overlay, and emergency navigation) through AR glasses, supports gesture / voice interactive disposal guidance, and realizes multi-terminal data synchronization for collaborative decision-making.

[0061] Among them, the virus diffusion simulation in the holographic risk sandbox renders a dynamic three-dimensional particle flow in a real laboratory scene to intuitively display the virus diffusion path in the next 10 minutes, with the core path highlighted in red and the secondary path presented in an orange gradient; at the same time, the sensor readings and corresponding risk level icons are displayed in real time in the form of a floating layer on the surface of the equipment and area; in addition, the emergency navigation function generates a green safe evacuation path that avoids the contaminated area in real time based on the impact range deduction results.

[0062] The interactive disposal guidance supports calling the disposal plan library through gesture operations, and can drag and drop the control panel on the virtual interface to adjust equipment parameters (such as UV intensity and ventilation direction). It also supports voice commands to link emergency plans. For example, saying "Start level 3 disinfection" can automatically trigger the preset process.

[0063] Multi-terminal collaboration can support data synchronization between the command center's large screen, mobile APP and AR glasses, thereby ensuring collaborative decision-making among emergency personnel, managers and external experts.

[0064] In the embodiment of the present invention, the real-time monitoring of virus leakage in the contaminated area of ​​a BSL-3 virus laboratory is illustrated as an example;

[0065] 1. Sensor deployment:

[0066] Air medium monitoring: Install a photochemical immunosensor (detection accuracy: 0.1 PFU / m 3) to capture the virus concentration in the aerosol in real time; deploy a dual-mode particulate matter sensor (0.3-10μm particle size resolution) at the end of the ventilation duct, and combine it with the CFD model to predict the diffusion path.

[0067] Surface medium monitoring: A surface-enhanced photochemical sensor (contact type, sampling period 5 seconds / time) is embedded in the edge of the experimental table, and a laser-induced breakdown spectroscopy (LIBS) scanner (non-contact type, full surface scanning 15 minutes / time) is installed on the centrifuge operation panel.

[0068] Equipment status monitoring: The internal differential pressure sensor of the biosafety cabinet (range ±50Pa, accuracy ±0.5Pa) is linked to the central negative pressure system; a three-axis MEMS vibration sensor (bandwidth 10-500Hz) is installed on the centrifuge door, and abnormal vibration is identified through the SVM model (energy entropy threshold >2.5J / Hz).

[0069] 2. Data collection and transmission:

[0070] The sensor data is transmitted to the edge computing gateway via the LoRa wireless protocol, encapsulated into a standardized JSON format (including spatiotemporal tags: three-dimensional coordinates + millisecond timestamps) after CRC verification, and pushed to the intelligent monitoring module through the Kafka message queue.

[0071] 3. Anomaly Detection Process

[0072] Data standardization: Virus sensor data is logarithmically transformed and normalized, and vibration data is linearly normalized to the [0,1] range to suppress exponential fluctuations in concentration data. Experimental environment data is Z-score standardized to better highlight outliers. Equipment monitoring data is linearly normalized. In the experimental metadata, the hazard level (1-5 levels unique hot encoding), personnel entry and exit logs (personnel density within the time window), and operation records (such as centrifuge start and stop status) are converted into discrete feature vectors and spliced ​​with the continuous sensor data to unify the dimensions.

[0073] Extract relevant data from the historical database and perform standardization operations such as cleaning and normalization to keep consistent with the standardized processing method of integrated data.

[0074] LSTM-GAN model execution:

[0075] The LSTM encoder extracts time series features from multi-source data with a step size of 10 seconds within a 60-second time window; the generator G synthesizes a normal benchmark risk sequence based on historical normal data; the discriminator D calculates the KL divergence and Euclidean distance between real-time data and the benchmark, with weights α = 0.6 and β = 0.4, and determines an anomaly when the comprehensive deviation is greater than 1.2; the CFD model combines real-time airflow velocity and pressure gradient, calibrates the diffusion coefficient through Kalman filtering, and inverts the coordinates of the leakage source with an error of ≤ 0.3m.

[0076] LSTM-GAN anomaly detection model architecture

[0077] (1) LSTM encoder temporal feature extraction

[0078] The three-dimensional tensor of real-time monitoring data is used as the input of LSTM, and LSTM is allowed to learn the temporal features in the data. The input data X is encoded through the bidirectional LSTM network and the encoding features are extracted: H∈R T×D , where T represents the number of time steps; D is the dimension of the LSTM hidden state vector; each row of the output feature H corresponds to the hidden state vector of one time step, that is, the tth row of H is H t , H t The expression is H t =LSTM(X t ,H t-1 ), X t represents the tth row of X and the t-1th row of H t-1 .

[0079] Generator G (normal risk benchmark sequence generation)

[0080] The non-leaked historical data of the laboratory is selected as the training benchmark for the generator G. The generator G learns important information such as the distribution pattern and feature correlation of normal data from the historical data. Combining this information with the hidden state in the long-term dependency features output by the LSTM layer enables the generator to generate samples that are closer to the distribution of real normal data. The generator G takes the LSTM encoded feature H as input and generates a benchmark sequence G(H)∈R that is close to the distribution of real normal data. T×N×F , as the risk baseline under normal conditions.

[0081] The generator network uses a fully connected layer + transposed convolutional layer, with the output layer activation function being Tanh, to ensure that the output range is consistent with the normalized real-world data. The first fully connected module in the generator network includes the LeakyReLU activation function and three stacked fully connected layers with 256, 512, and 1024 neurons, respectively. The transposed convolutional module in the generator network uses the LeakyReLU activation function between convolutional layers and a batch normalization layer to improve training stability.

[0082] The adversarial training goal of the generator G is to minimize the distribution difference between the generated data and the real normal data, continuously learn the distribution characteristics of the real normal data, and generate a benchmark sequence that is closer to the real normal state. In this way, it is more difficult for the discriminator to distinguish between the generated benchmark sequence and the real normal data, which helps the generator capture the pattern of the normal data, and then enables the discriminator to more accurately identify abnormal data that deviates from the normal pattern. Loss function for:

[0083]

[0084] Where D represents the discriminator; G(H) represents the benchmark sequence that generates a distribution close to the real normal data; P data is the real normal data distribution; Represents the mathematical expectation, which is the distribution of all normal data from the real normal data P data The encoding features H sampled in the averaging process are considered, considering the possible results of all input encoding features H.

[0085] Discriminator D (abnormal discrimination and adversarial game)

[0086] The discriminator D simultaneously receives real-time data X and generates a benchmark sequence G(H) that is close to the real normal data distribution. After extracting features through a multi-layer convolutional neural network (CNN) + fully connected layer, it outputs a binary classification probability (normal / abnormal) and a deviation score.

[0087] The second fully connected layer in the discriminator network includes a ReLU activation function and three stacked fully connected layers with 512, 256, and 128 neurons, respectively. The multi-layer CNN uses a ReLU activation function and three stacked layers with increasing numbers of convolutions and layers, respectively, 64, 128, and 256. The output layer contains two neurons, one of which uses a sigmoid activation function to output the binary classification probability, and the other uses a linear activation function to output the deviation score.

[0088] The adversarial goal of the discriminator D is to maximize the ability to distinguish between real data and generated data. By minimizing the loss function, it continuously learns to distinguish between real normal data and generated benchmark sequences, thereby being able to more accurately identify abnormal data that deviates from the normal pattern. for:

[0089]

[0090] Where D(X) represents the output of the judgement D, They represent the distribution of real data P dataThe mathematical expectation of the real sample X and the encoded feature H sampled in is obtained, that is, considering all possible real samples and generated samples.

[0091] The deviation calculation combines KL divergence (distribution difference) and Euclidean distance (numerical difference), and its formula is:

[0092] Deviation=α·KL(X‖G(H))+β·Euclidean(X,G(H))

[0093] Among them, KL(X‖G(H)) means using two X‖G(H) as divergence, and β is the weight coefficient, which is used to adjust the relative importance of KL divergence and Euclidean distance in calculating anomaly monitoring values.

[0094] In this embodiment of the present invention, the LSTM-GAN model is trained as follows:

[0095] This example details the training process of an anomaly monitoring model based on LSTM-GAN in a real-time dynamic monitoring and early warning system for infectious diseases in a virus laboratory. This model uses a generative adversarial network architecture to achieve dynamic feature learning and anomaly detection of laboratory environmental data.

[0096] 1. Data preparation

[0097] The training dataset is derived from multi-dimensional monitoring data collected during the laboratory's historical normal operation, covering all aspects of information such as air virus concentration, surface contact residues, equipment operating status, and personnel operation trajectories. The following operations are performed during the data preprocessing phase:

[0098] (1) Data cleaning: Remove outliers caused by sensor failures and use sliding window median filtering to process noisy data.

[0099] (2) Timing alignment: Align the timestamps of monitoring data with different sampling frequencies (such as video stream 25fps and air sensor 1Hz) to build a unified timing sequence.

[0100] (3) Feature standardization: Perform Z-score standardization on each dimension of data to eliminate dimensional differences.

[0101] (4) Data enhancement: Virtual normal samples are generated by adding Gaussian noise (μ = 0, σ = 0.05) to expand the dataset size.

[0102] Finally, a training set consisting of 120,000 time windows is formed, each window is 10 minutes long and contains multi-channel time series data with 300 time steps.

[0103] 2. Model initialization

[0104] The LSTM-GAN model is constructed using a three-level architecture of encoder-generator-discriminator:

[0105] (1) LSTM encoder:

[0106] Input dimensions: (batch_size, sequence_length, feature_dim) = (128, 300, 16);

[0107] Hidden layer: bidirectional LSTM layer with 256 hidden units and the forget gate bias initialized to 1.0;

[0108] Output: Time series feature tensor H∈R 128×256 ;

[0109] (2) Generator G:

[0110] Fully connected layer: 3-layer fully connected network with 256, 512, and 1024 neurons respectively, using LeakyReLU activation (α = 0.2);

[0111] Transposed convolution layer: 3 layers of transposed convolution, kernel size 4×4, stride 2, number of channels 512-256-128, each layer is followed by batch normalization;

[0112] Output layer: Tanh activation function, generating a benchmark sequence with the same distribution as the real data.

[0113] G(H)∈R^(128×300×16);

[0114] (3) Discriminator D:

[0115] Feature extraction layer: 4-layer convolutional neural network, convolution kernel number 64-128-256-512, kernel size 3×3, stride 2.

[0116] Fully connected layer: 3-layer fully connected network with 512-256-128 neurons and ReLU activation.

[0117] Output layer: dual-branch structure.

[0118] Classification branch: Sigmoid activation, output binary classification probability D prob ∈R 128 ;

[0119] Regression branch: linear activation, output deviation score Deviation∈R 128 .

[0120] 3. Training process

[0121] The generator and discriminator are trained using an alternating iterative optimization strategy:

[0122] (1) Generator optimization:

[0123] Fixed the discriminator parameters and input the encoded feature H to generate the benchmark sequence G(H);

[0124] Loss function:

[0125] Optimizer: Adam (lr = 0.0002, β1 = 0.5, β2 = 0.999), batch size 128, Learning Rate (lr), represents the learning rate;

[0126] Update strategy: Adjust the generator parameters through gradient ascent to maximize the probability of discriminator misjudgment.

[0127] (2) Discriminator optimization:

[0128] Fixed generator parameters, receiving real data X and generating sequence G(H);

[0129] Loss function:

[0130] Optimizer: SGD (lr=0.001, momentum=0.9), batch size 256;

[0131] Update strategy: Adjust the discriminator parameters through gradient descent to improve the ability to distinguish between real and generated data.

[0132] (3) Dynamic balance adjustment:

[0133] Execute 1 generator iteration after every 5 discriminator iterations;

[0134] A gradient penalty term (λ=10) is introduced to prevent mode collapse;

[0135] Dynamically adjust the learning rate: when the discriminator accuracy is >85%, the generator learning rate decays by 10%;

[0136] 4. Training termination conditions

[0137] Training is terminated when any of the following conditions are met:

[0138] (1) The AUC value of the Deviation of the validation set for 10 consecutive epochs increased by <0.001;

[0139] (2) Reaching the preset maximum number of iterations (500 epochs);

[0140] (3) The discriminator’s misclassification rate for the generated sequence is less than 5% and maintained for 10 epochs;

[0141] The final trained model achieved the following results on the test set:

[0142] Anomaly detection accuracy: 97.2%;

[0143] Pearson correlation coefficient between deviation score and actual leakage: 0.89;

[0144] False positive rate: 0.7%.

[0145] 5. Model Evaluation

[0146] Cross-validation strategy is used to evaluate model performance:

[0147] (1) Time domain evaluation: Sliding window validation, window overlap rate 50%, calculation of 10-fold cross-validation average indicators;

[0148] (2) Frequency domain evaluation: Fourier transform is performed on the time series data to verify the model's anomaly detection capability in the 0.01-0.1 Hz frequency band (corresponding to the human operating frequency);

[0149] (3) Adversarial evaluation: injecting synthetic abnormal data (amplitude 0.5σ-3σ) to test the detection sensitivity of the model;

[0150] The evaluation results show that the trained LSTM-GAN model can effectively capture the dynamic characteristics of laboratory environmental data and achieve accurate prediction of virus leakage risks.

[0151] 4. Virus Path Prediction:

[0152] Establishing a real-time parameterized gas diffusion model, wherein the gas diffusion model adopts a Gaussian plume model, and the real-time parameterized gas diffusion model is used to simulate the real-time viral aerosol diffusion path;

[0153] The diffusion coefficient of the gas diffusion model is calibrated in real time using a Kalman filter algorithm according to the leakage point. The calibration process specifically includes: establishing a state equation and an observation equation, and performing state prediction and updating based on the two equations to obtain an optimal gas diffusion path;

[0154] Dynamically modify parameters in a real-time parameterized gas diffusion model based on airflow disturbance events, such as the opening and closing of access control panels and the start and stop of disinfection equipment.

[0155] When airflow disturbances are detected, external input items are added to the state prediction and update process. The external input items are the wind speed mutation and the disturbance impact matrix. The disturbance impact matrix is ​​obtained by simulating a real-time parameterized gas diffusion model.

[0156] The spatiotemporal features of laboratory dynamic grid nodes are extracted based on the spatiotemporal graph convolutional network to predict the trend of virus concentration changes, including:

[0157] Build a dynamic grid, where the grid nodes contain multi-dimensional data such as virus concentration, temperature, humidity, etc. at each node at time t;

[0158] The spatiotemporal features are obtained through the spatiotemporal convolutional network, and the spatiotemporal features are fused through jump connections to output the predicted trend of virus concentration changes;

[0159] Generate a diffusion path probability distribution map based on the predicted virus concentration trend and real-time parameterized gas diffusion model, including:

[0160] The predicted virus concentration trend and the real-time parameterized gas diffusion model are fused using Bayesian probability, and the diffusion path probability distribution map is obtained through Monte Carlo sampling and partitioning.

[0161] The impact area is divided into different levels based on the probability distribution map, and a dynamic heat map of the virus leakage path is generated based on the impact area level. Specifically, it includes:

[0162] A dynamic heat map of the virus leakage path is obtained based on color mapping and real-time rendering of the affected area level.

[0163] 5. Early Warning Implementation:

[0164] By quantifying risk levels, triggering multi-level response strategies, and relying on AR technology to achieve holographic visualization and multi-terminal collaborative decision-making, laboratory safety is comprehensively guaranteed, including:

[0165] When the risk level is level 3, the ventilation system of the contaminated area will be automatically shut down and local ultraviolet disinfection will be started (irradiation dose ≥ 1000μW / cm 2 ).

[0166] The AR terminal renders a three-dimensional particle flow (core path highlighted in red) and pushes the emergency navigation path to the personnel terminal.

[0167] The beneficial effects are as follows:

[0168] 1. Aiming at the limitations of existing single-point decentralized monitoring, the present invention uses a data acquisition module to collect real-time virus leakage impact factor data from contaminated areas, semi-contaminated areas, and clean areas, achieving multi-data fusion;

[0169] 2. This paper constructs a dynamic anomaly monitoring model based on LSTM-GAN to monitor anomalies in real-time monitoring data, integrates historical data with real-time data, performs time series feature analysis and generates risk sequences, and uses neural networks to improve the accuracy of anomaly data identification and reduce false positive and false negative rates.

[0170] 3. The present invention constructs a real-time parameterized gas diffusion model to predict the spread of the virus and establishes a diffusion risk level to facilitate the prejudgment of risks.

[0171] 4. The present invention uses AR to visualize the virus leakage path, and the AR terminal realizes holographic risk display and interaction.

[0172] Method Example 1

[0173] According to an embodiment of the present invention, a method for monitoring laboratory virus leakage is provided, which specifically includes:

[0174] The data acquisition module collects real-time virus leakage impact factor data from contaminated areas, semi-contaminated areas, and clean areas and inputs it into the virus monitoring module;

[0175] The virus monitoring module pre-processes the real-time virus leakage factor data, integrates the pre-processed data to obtain a three-dimensional tensor, and inputs the three-dimensional tensor into the LSTM-GAN model to obtain virus leakage anomaly data;

[0176] When abnormal data of virus leakage is obtained, an alarm is issued through the alarm module.

[0177] The embodiment of the present invention is a system embodiment corresponding to the above-mentioned method embodiment. The specific operations of each module can be understood by referring to the description of the method embodiment, which will not be repeated here.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. These modifications or replacements of the technical solutions of the embodiments of the present invention do not cause the essence of the corresponding technical solutions to deviate from the scope of this solution.

Claims

1. A laboratory virus leak monitoring system, characterized in that: include, Data acquisition module, virus monitoring module and alarm module; The data acquisition module is used to collect real-time virus leakage impact factor data in contaminated areas, semi-contaminated areas and clean areas and input it into the virus monitoring module; The virus monitoring module is used to preprocess the real-time virus leakage factor data, integrate the preprocessed data to obtain a three-dimensional tensor, and input the three-dimensional tensor into the LSTM-GAN model to obtain abnormal virus leakage data or normal data; The alarm module is used to issue an alarm when abnormal data of virus leakage is obtained.

2. The system according to claim 1, wherein: The real-time virus leakage influencing factor data includes: virus sensor data, equipment monitoring data, video stream data, experimental environment data and experimental metadata.

3. The system according to claim 1, wherein: The virus monitoring module is specifically used to preprocess real-time virus leakage factor data, integrate the preprocessed data into a three-dimensional tensor, and input the three-dimensional tensor into the LSTM-GAN model. The LSTM-GAN model adopts a three-level architecture of encoder-generator-discriminator. After receiving the three-dimensional tensor, the LSTM encoder extracts features and inputs the features into the generator of the GAN network. The generator obtains a normal baseline sequence based on historical normal virus leakage influencing factor data, inputs the three-dimensional tensor and the normal baseline sequence into the discriminator to obtain the normal and abnormal probabilities of the real-time virus leakage factor data. The generator generates the deviation of the normal baseline sequence. When the probability of abnormality is greater than a certain threshold and the deviation is less than a certain threshold, it is judged as a virus leak.

4. The system according to claim 2, wherein: Virus sensor data were collected using a photochemical immunosensor and a dual-mode particulate matter sensor; Use biosafety cabinet differential pressure sensors and vibration sensors to collect equipment monitoring data; Temperature and humidity sensors, pressure gradient sensors, thermal imaging sensors, airflow velocity sensors, and ultraviolet sensors are used to collect experimental environment data; Collect experimental hazard level data, experimental personnel operation records and personnel entry and exit logs as experimental metadata.

5. The system according to claim 1, wherein: The system also includes: a virus leakage path prediction module, which is used to obtain the leakage point according to the sensor point corresponding to the abnormal data when it is determined that there is a virus leakage, and invert the virus aerosol diffusion path according to the leakage point.

6. The system according to claim 5, characterized in that The virus leakage path prediction module is specifically used to: Establishing a real-time parameterized gas diffusion model for simulating the real-time viral aerosol diffusion path, using a filtering algorithm to calibrate the diffusion coefficient of the gas diffusion model in real time based on the leak point, and dynamically correcting the parameters in the real-time parameterized gas diffusion model based on airflow disturbance events; According to the spatiotemporal graph convolutional network, the spatiotemporal characteristics of the laboratory dynamic grid nodes are extracted, and the virus concentration change trend is predicted based on the spatiotemporal characteristics. The diffusion path probability distribution map is generated based on the predicted virus concentration change trend and the real-time parameterized gas diffusion model. The affected area level is divided according to the probability distribution map, and a dynamic heat map of the virus leakage path is generated according to the affected area level.

7. The system according to claim 6, characterized in that The virus leakage path prediction module is specifically used to: calibrate the diffusion coefficient in real time according to the leakage point using the Kalman filter algorithm.

8. The system according to claim 6, wherein: The airflow disturbance events include: opening and closing of access control and starting and stopping of disinfection equipment.

9. The system according to claim 8, characterized in that The system also includes: an early warning and visualization module, which is used to calculate the leakage risk value according to the viral aerosol diffusion path, output the risk level according to the risk value, issue an early warning according to the risk level, and visualize the viral aerosol diffusion path and risk level.

10. A laboratory virus leakage monitoring method, characterized in that: include, The data acquisition module collects real-time virus leakage influencing factor data of contaminated areas, semi-contaminated areas and clean areas and inputs it into the virus monitoring module; The virus monitoring module pre-processes the real-time virus leakage factor data, integrates the pre-processed data to obtain a three-dimensional tensor, and inputs the three-dimensional tensor into the LSTM-GAN model to obtain virus leakage anomaly data; When abnormal data of virus leakage is obtained, an alarm is issued through the alarm module.

Images