An experimental animal whole life cycle data acquisition and tracing system based on internet of things

Abstract

The application discloses an experimental animal whole life cycle data acquisition and tracing system based on Internet of Things linkage, which comprises a data acquisition module, an experimental design and resource management module, an ethics review and life cycle management module, a task execution module, a system security and permission management module and a data intelligent analysis module.The data acquisition module is used for acquiring animal identity and Internet of Things data; the experimental design and resource management module is used for experimental scheme design and resource scheduling; the ethics review and life cycle management module is used for animal ethics review and whole life cycle data management; the task execution module is used for task generation and execution; the system security and permission management module is used for security management of user permissions; and the data intelligent analysis module is used for intelligent analysis on animal behavior.The application proposes a multi-modal causal perception animal-device collaborative health monitoring algorithm for intelligent analysis on animal behavior, so that a more optimal scheme is provided for the experimental animal whole life cycle data acquisition and tracing system.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and multimodal time series analysis, specifically to an Internet of Things-linked system for collecting and tracing data on the entire life cycle of experimental animals. Background Technology

[0002] Computer vision technology is a technology that enables machines to automatically perceive, understand, and interpret their environment and objects through visual input of digital images / videos. Computer vision technology includes the construction of large-scale labeled datasets, the leap in computing power of GPU parallel computing hardware, and continuous innovation in optimization algorithms and network architectures such as ResNet and Transformer. It also involves data preprocessing and augmentation, model compression and acceleration to adapt to edge device deployment, and semi-supervised / self-supervised learning to reduce dependence on labeled data. Furthermore, 3D vision, video understanding, multimodal learning, generative visual models, and close integration with robotics and augmented reality together constitute a complete technology chain from perception and understanding to interaction and creation. The synergy of these technologies lays a solid foundation for animal behavior analysis in an IoT-linked experimental animal lifecycle data collection and traceability system.

[0003] Multimodal time series analysis is a technique that simultaneously collects, fuses, models, and jointly analyzes data from different sources that change over time to extract comprehensive information and support decision-making. This includes reusing basic models from other modalities to achieve efficient time series analysis, extending them to enhance performance, and conducting cross-modal interactive time series analysis. Specific methods include attention mechanism fusion, alignment, and transformation at the input, intermediate / output stages. These techniques have significantly improved the accuracy and interpretability of prediction and anomaly detection tasks. The continuous development of these techniques has laid a solid foundation for multimodal data fusion and analysis in an IoT-linked experimental animal lifecycle data collection and traceability system.

[0004] Furthermore, laboratory animals are a crucial foundation for life science research, and the completeness, accuracy, and traceability of their life-cycle data directly affect the reliability and reproducibility of research results. Currently, laboratory animal management largely relies on manual recording, which suffers from problems such as untimely data entry, susceptibility to errors, and difficulty in traceability. With the development of the Internet of Things and computer vision technologies, some automated laboratory animal management systems have emerged, but existing technologies still have the following shortcomings:

[0005] 1) The data collection dimensions are limited, mostly only environmental parameters can be collected, and it is impossible to achieve automated and continuous monitoring of animal behavior;

[0006] 2) The accuracy of detecting abnormal animal behavior is low. Traditional methods can only identify obvious behavioral abnormalities and cannot detect subtle health changes in the early stages.

[0007] 3) It cannot distinguish between abnormal animal health and behavioral changes caused by abnormal environmental equipment, resulting in a high misjudgment rate;

[0008] 4) Lack of a complete full lifecycle data management system, with ethical review, experiment execution, and data management being disconnected;

[0009] Therefore, there is an urgent need for an experimental animal management system that can achieve multimodal data fusion and acquisition, intelligent behavior analysis, and full life cycle traceability. Summary of the Invention

[0010] To address the aforementioned issues, this invention aims to provide an IoT-enabled system for collecting and tracing data throughout the entire life cycle of laboratory animals.

[0011] To achieve the above objectives, the present invention provides the following technical solution: an IoT-linked experimental animal lifecycle data collection and traceability system, comprising a data collection module, an experimental design and resource management module, an ethics review and lifecycle management module, a task execution module, a system security and access control module, and a data intelligent analysis module. The data collection module collects animal identification and IoT data. The experimental design and resource management module includes an experimental design unit and a facility resource management unit. The experimental design unit is used to develop experimental protocols online, and the facility resource management unit is used for the digital and visual management of physical resources. The ethics review and lifecycle management module includes an animal ethics review unit and a lifecycle data management unit. The animal ethics review unit is used for online review of animal ethics, and the lifecycle data management unit is used to establish a complete electronic archive of animal lifecycle data. The task execution module generates, pushes, and executes task lists. The system security and access control module manages user permissions securely. The data intelligent analysis module proposes a multimodal causal perception-based animal-device collaborative health monitoring algorithm for intelligent analysis of animal behavior. An IoT-linked experimental animal lifecycle data collection and traceability method includes the following steps:

[0012] Step 1: Establish a unique identification for each laboratory animal and collect animal identification data and environmental IoT data in real time through IoT devices;

[0013] Step 2: Develop experimental plans online and digitally manage and schedule experimental facilities and resources;

[0014] Step 3: Complete the online animal ethics review and establish an electronic record of the animal's entire life cycle from birth to death;

[0015] Step 4: Automatically generate a task list based on the experimental plan and push it to the execution terminal, recording the task execution process;

[0016] Step 5: Employ a multimodal causal perception-based animal-equipment collaborative health monitoring algorithm to intelligently analyze animal behavior and equipment status and provide early warnings of anomalies.

[0017] Furthermore, the data acquisition module is used to collect animal identification and IoT data. It establishes an animal identification for each experimental animal through RFID chips and QR code ear tags, and deeply integrates with IoT devices through standard interfaces to collect IoT data in the animal room in real time.

[0018] Furthermore, the experimental design unit is used to develop experimental protocols online, defining research objectives, experimental groups, treatment factors, and observation indicators in detail, and assisting researchers in calculating sample size.

[0019] Furthermore, the facility resource management unit is used for the digital and visual management of physical resources. It digitally manages physical resources by recording supplier information, transportation conditions, receiving quarantine results, and managing the inventory of feed, bedding, and consumables, and displays the status of all cages in real time.

[0020] Furthermore, the animal ethics review unit is used for online review of animal ethics. Organizations can customize ethics review application form templates, review checklists, and submit, modify, review, vote on, and issue approval documents online according to their own requirements. All historical versions and review comments are automatically archived to form a complete ethics file.

[0021] Furthermore, the full life cycle data management unit is used to establish a complete electronic record of animal life cycle data, recording the animal's pedigree information, health information, and all operational information during the experimental process, creating a complete electronic record for each animal from birth to death.

[0022] Furthermore, the task execution module is used for generating, pushing, and executing task lists. It automatically generates daily task lists based on the experimental plan and pushes them to the mobile terminals of designated technicians. When executing tasks, staff scan animal / cage IDs for confirmation and record the operator, operation time, and results through system forms.

[0023] Furthermore, the system security and access control module is used to securely manage user permissions. Based on the identity of different users, it finely controls their permissions to menus, data fields, and operations, and records all users' login behavior and data change history.

[0024] Furthermore, the data intelligence analysis module proposes a multimodal causal perception-based animal-equipment collaborative health monitoring algorithm to intelligently analyze animal behavior and equipment malfunctions.

[0025] Furthermore, the multimodal causal perception-based animal-device collaborative health monitoring algorithm is as follows:

[0026] 1) Data preprocessing and feature extraction: Resampling, alignment and standardization are performed on video stream data and IoT sensor data with different timestamps. Spatiotemporal features are extracted for video stream data and sliding window extraction of statistical and frequency domain features is performed for sensor data.

[0027] 2) Animal Behavior Anomaly Detection: An autoencoder is trained to learn the behavior patterns of animals in their normal state. To more accurately capture local behavioral features related to health and enhance the representation learning of normal behavior patterns, a spatiotemporal perception contrastive reconstruction loss is proposed to replace the traditional mean squared error loss during the training phase of the autoencoder. At the same time, to solve the problem that the autoencoder pays the same attention to all time points and easily ignores brief and severe anomalies, a variational autoencoder based on the attention mechanism is proposed. During the inference phase, the trained model is used to calculate the anomaly score of real-time behavioral features. Then, since animals will exhibit social avoidance when they are sick, an additional dynamic social network is constructed for auxiliary analysis. The contact frequency between individuals is calculated by spatial distance, and the deviation of individual degree centrality is analyzed to discover abnormal individuals in the group.

[0028] 3) Equipment health monitoring and prediction: By learning the multidimensional parameter sequence of the equipment during normal operation through LSTM autoencoder, the long-term and short-term dependencies of the equipment status are captured. Then, based on the degradation trajectory, the remaining time before the equipment fails is predicted. When the equipment ages / fails, its operating trajectory will deviate from the normal manifold. Based on this, a nonlinear degradation model of the equipment health index is constructed, and the remaining time before the index first drops to the failure threshold is predicted.

[0029] 4) Causal relationships and intelligent decision-making:

[0030] ① If environmental monitoring data shows that all equipment is operating stably during the period when the animal abnormality occurs, and only a few animals in the same cage / room show abnormalities, then it is determined to be a health problem of the animal itself.

[0031] ② If environmental parameters fluctuate drastically during the period when the animal abnormality occurs, and multiple animals in the same cage / room exhibit similar abnormal behavior at the same time, then the equipment abnormality should be the primary cause.

[0032] ③ If the cause cannot be clearly attributed, generate both a veterinary inspection work order and an equipment maintenance work order, and mark them as "requires joint investigation";

[0033] To address the misjudgment problem caused by univariate monitoring, a multivariate joint probability distribution model is proposed to improve the specificity of fault prediction. Decisions are made by calculating posterior probabilities. If the posterior probability of an animal abnormality exceeds a preset threshold, a veterinary inspection work order is generated. If the probability of equipment failure exceeds a preset threshold, an equipment maintenance work order is generated. This enables intelligent analysis of animal behavior and equipment failure.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] This invention proposes a multimodal causal perception-based animal-equipment collaborative health monitoring algorithm for intelligent analysis of animal behavior and equipment failure. The innovation of this invention lies in the fact that, firstly, during the training phase of the autoencoder, the algorithm proposes a spatiotemporal perception-based contrastive reconstruction loss to replace the traditional mean squared error loss, thereby more accurately capturing local behavioral features related to health and enhancing the representation learning of normal behavioral patterns. Secondly, it proposes an attention-based variational autoencoder to address the problem that the autoencoder pays the same attention to all time points and easily ignores brief and severe anomalies. Finally, it proposes a multivariate joint probability distribution model to solve the misjudgment problem caused by univariate monitoring, thereby realizing intelligent analysis of animal behavior and equipment failure.

[0036] Compared with traditional manual inspection methods: the early detection rate of animal health abnormalities increased from 32% to 94%, with an average of 2.3 days earlier detection; the false positive rate of behavioral abnormality detection decreased from 28% to 6%; the average repair time for equipment failures was shortened from 4.2 hours to 1.1 hours; the workload of manual labor was reduced by 65%; and the data integrity reached 100%. Attached Figure Description

[0037] The invention will be further illustrated with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the invention. For those skilled in the art, other drawings can be obtained based on the following drawings without any creative effort.

[0038] Figure 1 is an overall structural block diagram of the system of the present invention, showing the connection relationship and data flow between the data acquisition module, experimental design and resource management module, ethical review and life cycle management module, task execution module, system security and access control module and data intelligent analysis module.

[0039] Figure 2 is a diagram of the online recording interface for animal experimental data within the group of the present invention. Corresponding to the experimental design unit and the full life cycle data management unit of the present invention, it shows the real-time recording functions of experimental group management, basic animal information, weight changes, and drug administration information, as well as the automatic highlighting and warning function for abnormal data.

[0040] Figure 3 shows the multi-dimensional visualization interface of the experimental results of the present invention. Corresponding to the full life cycle data management unit and result display module of the present invention, it displays the trend line graph of tumor volume and weight changes, and supports comparative analysis and trend prediction of data from different experimental groups.

[0041] Figure 4 is a diagram of the individual experimental data detail query interface of the present invention. Corresponding to the full life cycle data management unit of the present invention, it displays detailed records of animal tumor volume, weight and drug dosage at different time points, and supports data export and batch processing.

[0042] Figure 5 is a diagram of the full-process management interface for the application for the use of experimental animals in this invention. Corresponding to the facility resource management unit and animal ethics review unit of this invention, it shows the functions of application status tracking, supplier information, animal strain, breeding area, and cost calculation, and supports adding new applications, batch deletion, and data export.

[0043] Figure 6 is a diagram of the online ethical review interface for experimental animals in this invention. Corresponding to the animal ethical review unit of this invention, it is designed in strict accordance with the 3R principle, and shows the functions of custom ethical review templates, harm-benefit analysis scoring, online submission and approval of application materials. All operations are fully recorded and automatically archived to form a complete ethical file. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] An IoT-linked experimental animal lifecycle data collection and traceability system includes a data collection module, an experimental design and resource management module, an ethics review and lifecycle management module, a task execution module, a system security and access control module, and a data intelligent analysis module. The data collection module collects animal identification information and IoT data. The experimental design and resource management module includes an experimental design unit and a facility resource management unit. The experimental design unit is used to develop experimental protocols online, while the facility resource management unit is used for the digital and visual management of physical resources. The ethics review and lifecycle management module includes an animal ethics review unit and a lifecycle data management unit. The animal ethics review unit is used for online review of animal ethics, while the lifecycle data management unit is used to establish a complete electronic archive of animal lifecycle data. The task execution module is used for generating, pushing, and executing task lists. The system security and access control module is used for secure management of user permissions. The data intelligent analysis module proposes a multimodal causal perception-based animal-device collaborative health monitoring algorithm for intelligent analysis of animal behavior.

[0046] Preferably, the data acquisition module is used to collect animal identification and IoT data. It establishes an animal identification for each experimental animal through RFID chips and QR code ear tags, and deeply integrates with IoT devices through standard interfaces to collect IoT data such as temperature, humidity, pressure difference, and ammonia concentration in the animal room in real time.

[0047] Preferably, the experimental design unit is used to develop experimental protocols online, define research objectives, experimental groups, treatment factors, and observation indicators in detail, and assist researchers in calculating sample size.

[0048] Preferably, the facility resource management unit is used for the digital and visual management of physical resources. It performs digital management of physical resources by recording supplier information, transportation conditions, receiving quarantine results, and managing the inventory of feed, bedding, and consumables. It also displays the status of all cages in real time (idle, occupied, awaiting cleaning, under quarantine) and supports online reservation, review, and automated allocation.

[0049] Preferably, the animal ethics review unit is used for online review of animal ethics. It supports institutions to customize ethics review application form templates and review point lists according to their own requirements, and can configure multi-level and multi-role electronic approval processes with full traceability. It also supports online submission, modification, review, voting and approval of ethics application materials, and automatically archives all historical versions and review opinions to form a complete ethics file.

[0050] Preferably, the full life cycle data management unit is used to establish a complete electronic record of animal life cycle data, recording the animal's pedigree information: animal breed, genotype, parents, date of birth, weaning, toe amputation number, health information: daily health observation, vaccination, disease diagnosis and treatment, microbial test results, and all operational information during the experimental execution process: drug administration, surgery, imaging, death, to establish a complete electronic record for each animal from birth to death.

[0051] Preferably, the task execution module is used for the generation, push and execution of task lists. Based on the experimental plan, breeding cycle and compliance requirements, it automatically generates daily task lists for cage changing, weighing, medication administration and inspection, and pushes them to the mobile terminals of designated technicians. When executing tasks, staff scan animal / cage IDs for confirmation and record the operator, operation time and results through system forms.

[0052] Preferably, the system security and access control module is used to securely manage user permissions. Based on the identity of different users, it finely controls their permissions to menus, data fields and operations, records all users' login behavior and data change history, provides complete audit trail functions, and prevents data leakage and unauthorized operations.

[0053] Preferably, the data intelligent analysis module proposes a multimodal causal perception-based animal-equipment collaborative health monitoring algorithm to intelligently analyze the spontaneous behavior and social interaction of animals, identify health abnormalities / stress states at an early stage, and perform trend analysis on environmental and equipment data to achieve predictive maintenance.

[0054] Specifically, the multimodal causal perception-based animal-device collaborative health monitoring algorithm is as follows: First, video stream data and IoT sensor data with different timestamps are resampled, aligned, and standardized. ,in, For standardized sensor data, For the first Each sensor in time The original observations, For the first The mean of each sensor on the training set For the first The standard deviation of each sensor on the training set is used to extract spatiotemporal features from video stream data. ,in, The extracted spatiotemporal feature vector, The length of the time window for extracting video features, From time Time The video frame sequence, This is a 3D convolutional neural network function that uses a sliding window to extract statistical and frequency domain features from sensor data. ,in, The mean of the data within the window. The size of the sliding window. For time The original sensor readings, Let be the standard deviation of the data within the window. Then, an autoencoder is trained to learn the animal's behavioral patterns under normal conditions. When the animal is sick or stressed, its behavioral patterns deviate from normal, and the reconstruction error of the autoencoder increases significantly. To more accurately capture local behavioral features related to health and enhance the representation learning of normal behavioral patterns, a spatiotemporally aware contrastive reconstruction loss is proposed to replace the traditional mean squared error loss during the training phase of the autoencoder. Simultaneously, to address the problem that the autoencoder pays equal attention to all time points and easily overlooks brief, drastic anomalies, a variational autoencoder based on an attention mechanism is proposed. ,in, For the set of model parameters, The total number of training samples, For the sample index in the training dataset, For the first The loss function for each sample. The number of time steps for each sample For time-based attention weights, and ,in, The normalized exponential activation function is... For parameter vectors The transpose of , The hyperbolic tangent activation function is used. This is the weight matrix. For time The hidden state vector, For spatial attention mask matrix, For the first Training samples in time Input features, For decoder functions, For encoder functions, For the square operation of the L2 norm, To compare the weighting coefficients of the loss term, To compare the boundary coefficients in the loss, The latent feature vector extracted by the encoder, These are positive sample features from a historical database of normal behavior. The negative sample features are derived from a historical database of normal behavior. The inference phase involves calculating the anomaly score for the real-time behavioral features. ,in, For time Abnormal scores, For time The input feature vector, For sample indexes in the historical normal behavior database, To balance the weighting coefficients, The latent feature vector of the current sample. For positive sample feature vectors, The feature vector of the negative sample;

[0055] Next, since animals exhibit social avoidance when sick, a dynamic social network is constructed. Anomalies are detected by analyzing changes in the connection relationships between individuals within the network, and the frequency of contact between individuals is calculated based on spatial distance. ,in, for Time Individual With individuals Contact frequency between them It is an exponential function. for Time Individual With individuals Spatial distance between them To control the scale parameter of the contact frequency decay rate, and to calculate the deviation of the degree centrality of an individual. ,in, for Time Individual Degree centrality, for Average degree centrality of all individuals over time for The standard deviation of time-degree centrality is then used to learn the multidimensional parameter sequence during normal device operation via an LSTM autoencoder. During device aging / failure, the device's operating trajectory deviates from the normal manifold. The calculation of the long- and short-term dependencies of the device state is then performed. ,in, For LSTM autoencoders in time The hidden state, For the computation function of the Long Short-Term Memory network unit, For LSTM autoencoders in time Input features, For time The hidden state, for The parameter set of the unit, the current health state of the system is determined by the ability to reconstruct the input sequence from the hidden state, and the device health index is... ,in, For time The equipment health index, For LSTM autoencoders in time The reconstructed output, We take the historical average values ​​of the parameters observed by the equipment under normal conditions, and then predict the remaining time before equipment failure based on the degradation trajectory. Assuming that the equipment's health indicators degrade non-linearly over time, the degradation model for the health indicators is as follows: ,in, For time The equipment health indicators, Initial health indicators for the equipment. The degradation rate coefficient, To observe the noise, the time point at which the health indicator first drops to the failure threshold is... ,in, For the remaining service life, The preset failure threshold, Using the current time as an example, a causal relationship is established between abnormal animal behavior and abnormal environmental equipment. If environmental monitoring data shows stable equipment operation during the period in which the animal abnormality occurs, the problem is attributed to the animal's own health. Conversely, if environmental parameters fluctuate drastically, the problem is primarily attributed to equipment malfunction. To address the misjudgment problem caused by univariate monitoring, a multivariate joint probability distribution model is proposed to improve the specificity of fault prediction. The joint probability distribution is as follows: , ,in, Given environmental and behavioral data, this represents the posterior probability of an animal's abnormality. For the proportional sign, This represents the likelihood probability of environmental data under abnormal animal conditions. The likelihood probability of animal behavior data under abnormal conditions. Let be the prior probability of animal abnormality. Given the probability of equipment failure, after calculating the posterior probability, if If the threshold is exceeded, a veterinary inspection work order will be generated. If the threshold is exceeded, an equipment maintenance work order is generated to achieve intelligent analysis of animal behavior and equipment malfunctions.

[0056] To enable those skilled in the art to fully implement the technical solution of this invention based on the contents of this specification, this invention fully discloses the model structure, training process, input / output data processing, parameter adjustment, and deployment scheme of the multimodal causal perception-based animal-device collaborative health monitoring algorithm, as detailed below:

[0057] The training dataset used in this algorithm comes from the real operating environment of multiple collaborative experimental animal centers / facilities, covering multimodal data of commonly used experimental animals such as mice, rats, and rabbits under normal feeding, experimental operation, and abnormal health conditions. Specifically, it includes: more than 1,500 hours of high-definition monitoring video data and synchronously collected IoT sensor time-series data, with a total of more than 300,000 labeled samples. Among them, the video data was labeled by professional experimental animal veterinarians and researchers with more than 5 years of experience using manual and semi-automatic labeling tools. Bounding boxes / key points / behavioral categories were labeled for normal autonomous behavior and abnormal behavior, and multiple rounds of expert cross-validation were carried out to ensure that the labeling consistency was higher than 96%. The normal / abnormal labels of the sensor data were jointly labeled by combining the equipment maintenance records, veterinary inspection logs, and animal clinical signs. The dataset was divided into training set, validation set, and test set in a ratio of 7:1.5:1.5, and various data augmentation techniques such as random pruning, brightness / contrast / hue perturbation, time series noise injection, and Gaussian blur were used to expand the sample size to significantly improve the robustness and generalization ability of the model.

[0058] During model training, this invention employs a hybrid training strategy combining end-to-end supervised learning and self-supervised learning. First, a 3D convolutional neural network is used to pre-train the video spatiotemporal feature extraction module and the sensor statistical / frequency domain feature extraction module. Then, an attention-based variational autoencoder and an LSTM autoencoder are jointly optimized. Hyperparameter tuning utilizes Bayesian optimization combined with 5-fold cross-validation to systematically search and optimize key parameters, including the weight coefficients in the spatiotemporal perception contrast reconstruction loss. (0.1~1.0), boundary coefficient (0.5~2.0), Time Attention Weight Learning rate, sliding window size (50~200 sampling points), contact frequency attenuation scale parameters (10~50 pixels), number of LSTM hidden layers and number of units, through early stopping mechanism (patience=10) and multi-objective monitoring such as reconstruction error on the validation set, anomaly detection F1-score and causal attribution accuracy, effectively prevent overfitting and finally obtain the optimal model parameters;

[0059] The input of this algorithm is a multimodal temporal feature vector after resampling, alignment and standardization. The output includes animal behavior abnormality score, equipment health index, predicted value of remaining service life and causal attribution result under multivariate joint probability distribution. The relationship between input and output is explicitly established through attention mechanism, contrast loss and joint probability distribution model to ensure that the model has good interpretability and causal reasoning ability.

[0060] To meet the low-latency requirements of real-time health monitoring in laboratory animal facilities, this algorithm supports edge deployment. Through lightweight techniques such as model pruning, channel pruning, INT8 quantization, and knowledge distillation, the core inference model size is compressed to 1 / 4 to 1 / 8 of the original model, while performance loss is controlled within 5%. On embedded platforms of NVIDIA Jetson Nano / Orin series, Huawei Ascend Lite, and other domestic edge AI accelerator cards, the end-to-end inference latency of single-channel video + multi-sensor data can be stably controlled within 120 to 250 milliseconds, fully meeting the requirements of real-time anomaly early warning in the Internet of Things. At the same time, this system adopts a cloud-edge collaborative architecture: edge devices are responsible for real-time feature extraction and preliminary anomaly detection, while local servers / cloud are responsible for complex multi-animal social network analysis and causal decision-making, achieving the optimal balance between computational efficiency and accuracy.

[0061] The detailed description of the above-mentioned training dataset construction method, model training process, hyperparameter tuning method, input-output correlation and edge deployment scheme ensures that those skilled in the art can reproduce and implement the technical solution of this invention without creative effort, and solves the problems of low accuracy in detecting abnormal behavior in experimental animals, difficulty in causal attribution and insufficient real-time performance in the prior art.

[0062] The IoT-linked experimental animal lifecycle data acquisition and traceability system, using a mouse numbered "M01" in the experimental animal facility as an example, demonstrates a complete implementation. Assuming that at 14:00 on May 20, 2026, the system collects the following multimodal data: video stream from a high-definition camera, data from temperature and humidity sensors in the independent ventilation cage system, and differential pressure sensor data. The specific implementation is as follows:

[0063] 1) Data Synchronization and Feature Extraction: First, data with different timestamps are aligned and standardized. Then, for the data with different timestamps... Raw readings from each sensor Based on the statistics of the training set Standardization, extraction Time before A sequence of video frames of a certain length is used to extract spatiotemporal feature vectors through a 3D convolutional neural network. Then, the sensor data stream is processed using a length of... A sliding window is used to calculate the mean within the window. and standard deviation This is used to characterize the current steady-state level of the environment;

[0064] 2) Animal behavior anomaly detection: The extracted video spatiotemporal feature vectors The input is fed into an attention-based variational autoencoder for inference, and the model computation... outlier score at time The model captured a reduction in the mouse's fine movements and prolonged stillness at the current moment, resulting in a significant increase in reconstruction error. Simultaneously, the contrast loss term also captured deviations from the normal pattern, ultimately leading to the calculated... Then, further analysis was conducted on the social interactions among animals, and the frequency of contact between individuals was calculated based on spatial distance. Then calculate the individual Degree of centrality deviation Assuming the average contact rate of the group is Standard deviation The target mouse M01, due to its poor condition, experienced a significant decrease in contact frequency, resulting in a decrease in its degree centrality deviation. If the preset threshold is 1.5, and this value is significantly greater than the set abnormal threshold of 1.5, it indicates that M01 exhibits typical social avoidance behavior.

[0065] 3) Equipment Health Monitoring and Prediction: The operating status of the wind turbine is continuously monitored through an LSTM self-encoder, and the equipment is monitored over time. The health status of a device is determined by its ability to reconstruct the input sequence from its hidden states; the device health index is calculated. Assuming the reconstruction error calculated at the current time is Total fluctuation Then the equipment health index This value is less than the ideal health value of 1.0, indicating that the device is beginning to show slight degradation, but has not yet exceeded the safety threshold. Assuming the device health index... Nonlinear degradation over time Based on the current degradation rate, the system predicts that there are approximately 17 days remaining until the failure threshold of 0.6 is reached, and the current operating status of the equipment is determined to be stable.

[0066] 4) Causal Association and Intelligent Decision-Making: Establishing a causal relationship between abnormal animal behavior and abnormal environmental equipment by first calculating the posterior probability of the animal's own abnormality. Assume the prior probability of animal abnormality Under abnormal animal conditions, the likelihood of observing current stable environmental data was determined. The likelihood of observing current high-abnormal-score behavioral data The posterior probability of the animal's own abnormality is then... At the same time, the probability of equipment failure is calculated based on the equipment health index. Assuming the system has a preset animal anomaly alarm threshold of 0.02 and a device malfunction alarm threshold of 0.15, the final anomaly assessment is as follows: Although the probability of device malfunction is... The alarm threshold has been reached, but the objective environmental monitoring data did not show any abnormal fluctuations. Meanwhile, the animal's abnormal behavior score is extremely high. And accompanied by significant social deviation and the posterior probability of the animal's own abnormalities. When the animal abnormality alarm threshold is reached, the system determines that it is a health problem of the animal itself. It then generates a high-level veterinary inspection work order and pushes it to the mobile terminal of the relevant personnel. The work order includes the animal ID of experimental mouse M01, the time of abnormality, the quantitative score of abnormal behavior, the deviation of social network, and the associated environmental data snapshot, and prompts "high probability of abnormal animal health, it is recommended to immediately review clinical signs". After receiving the work order, the on-duty veterinarian can promptly conduct a targeted clinical examination on M01.

[0067] An IoT-linked system for collecting and tracing data on the entire lifecycle of laboratory animals is proposed. This system integrates IoT sensors, video surveillance, and automated equipment to achieve continuous, automated, and unified management of multi-dimensional data on the physiology, behavior, and environment of laboratory animals from birth to death. This ensures data integrity and traceability, thereby improving the reliability and reproducibility of research data. The system integrates modules for data acquisition, experimental design and resource management, ethical review and lifecycle management, task execution, system security and access control, and intelligent data analysis. It proposes a multimodal causal perception-based animal-equipment collaborative health monitoring algorithm for intelligent analysis of animal behavior and equipment malfunctions. The innovation of this invention lies in the fact that the multimodal causal perception-based animal-equipment collaborative health monitoring algorithm first proposes a spatiotemporal perception-based approach during the training phase of the autoencoder. This invention replaces traditional mean squared error loss with reconstruction loss to more accurately capture health-related local behavioral features and enhance the representation learning of normal behavioral patterns. Simultaneously, it proposes an attention-based variational autoencoder to address the problem of autoencoders focusing on the same amount of attention across all time points, easily overlooking brief but severe anomalies. Finally, it proposes a multivariate joint probability distribution model to solve the misjudgment problem caused by univariate monitoring. This enables intelligent analysis of animal behavior and equipment malfunctions, effectively improving the performance of an IoT-linked experimental animal lifecycle data acquisition and traceability system. It provides more intelligent and accurate technical support for such a system. Furthermore, this invention relates to computer vision and multimodal temporal analysis techniques, offering an efficient and convenient method and system for experimental animal lifecycle data acquisition and traceability, contributing significant application value to lifecycle data acquisition and management.

[0068] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An IoT-linked system for collecting and tracing data on the entire life cycle of laboratory animals, characterized in that, The system comprises a data acquisition module, an experimental design and resource management module, an ethics review and lifecycle management module, a task execution module, a system security and access control module, and a data intelligence analysis module. The data acquisition module collects animal identification data and IoT data. The experimental design and resource management module includes an experimental design unit and a facility resource management unit. The experimental design unit is used to develop experimental protocols online, while the facility resource management unit is used for the digital and visual management of physical resources. The ethics review and lifecycle management module includes an animal ethics review unit and a full lifecycle data management unit. The animal ethics review unit is used for online review of animal ethics, while the full lifecycle data management unit is used to establish complete electronic archives of animal lifecycle data. The task execution module generates, pushes, and executes task lists. The system security and access control module manages user permissions securely. The data intelligence analysis module incorporates a multimodal causal perception-based animal-equipment collaborative health monitoring algorithm for intelligent analysis and anomaly warning of animal behavior and equipment malfunctions.

2. The IoT-linked experimental animal life-cycle data acquisition and traceability system according to claim 1, characterized in that, The data acquisition module is used to collect animal identification and IoT data. It establishes an animal identification for each experimental animal through RFID chips and QR code ear tags, and deeply integrates with IoT devices through standard interfaces to collect IoT data in the animal room in real time.

3. The IoT-linked experimental animal life-cycle data acquisition and traceability system according to claim 1, characterized in that, The experimental design unit is used to develop experimental protocols online, define research objectives, experimental groups, treatment factors, and observation indicators in detail, and assist researchers in calculating sample size.

4. The IoT-linked experimental animal life-cycle data acquisition and traceability system according to claim 1, characterized in that, The facility resource management unit is used for the digital and visual management of physical resources. It digitally manages physical resources by recording supplier information, transportation conditions, receiving quarantine results, and managing the inventory of feed, bedding, and consumables, and displays the status of all cages in real time.

5. The IoT-linked experimental animal life-cycle data acquisition and traceability system according to claim 1, characterized in that, The animal ethics review unit is used for online review of animal ethics. Organizations can customize ethics review application form templates and review checklists according to their own requirements, as well as submit, modify, review, vote and issue approval documents online for ethics application materials, and automatically archive all historical versions and review comments to form a complete ethics file.

6. The IoT-linked experimental animal life-cycle data acquisition and traceability system according to claim 1, characterized in that, The full life cycle data management unit is used to establish a complete electronic record of animal life cycle data, recording the animal's pedigree information, health information, and all operational information during the experimental process, creating a complete electronic record for each animal from birth to death.

7. The IoT-linked experimental animal life-cycle data acquisition and traceability system according to claim 1, characterized in that, The task execution module is used for generating, pushing, and executing task lists. It automatically generates daily task lists based on the experimental plan and pushes them to the mobile terminals of designated technicians. When executing tasks, staff scan animal / cage IDs for confirmation and record the operator, operation time, and results through system forms.

8. The IoT-linked experimental animal life-cycle data acquisition and traceability system according to claim 1, characterized in that, The system security and access control module is used to securely manage user permissions. It finely controls the permissions of different users to menus, data fields and operations based on their identities, and records the login behavior and data change history of all users.

9. The IoT-linked experimental animal life-cycle data acquisition and traceability system according to claim 1, characterized in that, The multimodal causal perception-based animal-device collaborative health monitoring algorithm is as follows: 1) Data preprocessing and feature extraction: Resampling, alignment and standardization are performed on video stream data and IoT sensor data with different timestamps. Spatiotemporal features are extracted for video stream data and sliding window extraction of statistical and frequency domain features is performed for sensor data. 2) Animal Behavior Anomaly Detection: An autoencoder is trained to learn the behavior patterns of animals in their normal state. To more accurately capture local behavioral features related to health and enhance the representation learning of normal behavior patterns, a spatiotemporal perception contrastive reconstruction loss is proposed to replace the traditional mean squared error loss during the training phase of the autoencoder. At the same time, to solve the problem that the autoencoder pays the same attention to all time points and easily ignores brief and severe anomalies, a variational autoencoder based on the attention mechanism is proposed. During the inference phase, the trained model is used to calculate the anomaly score of real-time behavioral features. Then, since animals will exhibit social avoidance when they are sick, an additional dynamic social network is constructed for auxiliary analysis. The contact frequency between individuals is calculated by spatial distance, and the deviation of individual degree centrality is analyzed to discover abnormal individuals in the group. 3) Equipment health monitoring and prediction: By learning the multidimensional parameter sequence of the equipment during normal operation through LSTM autoencoder, the long-term and short-term dependencies of the equipment status are captured. Then, based on the degradation trajectory, the remaining time before the equipment fails is predicted. When the equipment ages / fails, its operating trajectory will deviate from the normal manifold. Based on this, a nonlinear degradation model of the equipment health index is constructed, and the remaining time before the index first drops to the failure threshold is predicted. 4) Causal relationships and intelligent decision-making: ① If environmental monitoring data shows that all equipment is operating stably during the period when the animal abnormality occurs, and only a few animals in the same cage / room show abnormalities, then it is determined to be a health problem of the animal itself. ② If environmental parameters fluctuate drastically during the period when the animal abnormality occurs, and multiple animals in the same cage / room exhibit similar abnormal behavior at the same time, then the equipment abnormality should be the primary cause. ③ If the cause cannot be clearly attributed, generate both a veterinary inspection work order and an equipment maintenance work order, and mark them as "requires joint investigation"; To address the misjudgment problem caused by univariate monitoring, a multivariate joint probability distribution model is proposed to improve the specificity of fault prediction. Decisions are made by calculating posterior probabilities. If the posterior probability of an animal abnormality exceeds a preset threshold, a veterinary inspection work order is generated. If the probability of equipment failure exceeds a preset threshold, an equipment maintenance work order is generated. This enables intelligent analysis of animal behavior and equipment failure.

10. A method for collecting and tracing data on the entire life cycle of laboratory animals using Internet of Things (IoT) linkage, characterized in that: The system implementation according to any one of claims 1 to 9 includes the following steps: Step 1: Establish a unique identification for each laboratory animal and collect animal identification data and environmental IoT data in real time through IoT devices; Step 2: Develop experimental plans online and digitally manage and schedule experimental facilities and resources; Step 3: Complete the online animal ethics review and establish an electronic record of the animal's entire life cycle from birth to death; Step 4: Automatically generate a task list based on the experimental plan and push it to the execution terminal, recording the task execution process; Step 5: Employ a multimodal causal perception-based animal-equipment collaborative health monitoring algorithm to intelligently analyze animal behavior and equipment status and provide early warnings of anomalies.

Images