A method and system for time-limited feeding of animals

Abstract

The application provides a time-limited feeding method and system for animal experiments, which is applied to an intelligent door control system comprising a feeding cage, an electrically controlled door control channel, a multi-modal sensor array and an intelligent control unit. The method obtains a monitoring data stream of the state and behavior of animals through the multi-modal sensor array; realizes execution control of the time-limited feeding mode based on the monitoring data stream and preset parameters; outputs an individualized time-limited feeding intervention strategy based on execution data and the monitoring data stream; and realizes management and analysis of experimental data through a data management module. The application also provides a corresponding time-limited feeding system, which has dynamic adaptability, comprehensive monitoring dimensions, high flexibility and deep data integration capability, solves the problems of poor manual control accuracy, lack of intelligent adaptability and data deep mining in the prior art, and improves the accuracy, scientificity and animal welfare level of animal experiments.

Description

Technical Field

[0001] This invention relates to the field of biomedical experimental technology, and in particular to a method and system for timed feeding in animal experiments, used to precisely control and monitor the feeding time, behavior and physiological state of experimental animals, providing technical support for studying the interaction between the biological clock and metabolism and exploring intervention strategies for metabolic diseases. Background Technology

[0002] Time-Restricted Feeding (TRF) or Time-Restricted Eating (TRE), as an emerging dietary intervention model, has received widespread attention in the biomedical field in recent years. With changing modern lifestyles, irregular eating habits and lack of exercise have led to a year-on-year increase in the incidence of metabolic diseases such as obesity, diabetes, and cardiovascular disease, placing a huge burden on global health. Research has found that the biological clock is closely related to metabolic rhythms, and disruption of the biological clock directly leads to metabolic disorders. The core idea of ​​time-restricted feeding is to extend the nighttime fasting period by setting a fixed daily eating window (usually 4-12 hours) without strictly restricting total calorie intake, aiming to improve metabolic health by coordinating eating times with the biological clock rhythm.

[0003] The theoretical basis of time-restricted eating is that regular eating and fasting cycles help maintain normal circadian rhythms. Its underlying physiological mechanisms include: after prolonged fasting, the body shifts its energy source from glucose to fatty acids and ketone bodies, activating adaptive cellular responses; by adjusting the timing of eating to synchronize with the diurnal rhythm, it optimizes the function of metabolic organs (such as the liver and adipose tissue); and it regulates the structure and diversity of the gut microbiota, thereby affecting host metabolism. Common time-restricted eating patterns include 18:6, 16:8, and 14:10, among others, with the core focus being "when to eat" rather than "how much to eat."

[0004] Both animal studies and human research have shown that time-restricted eating has potential benefits in weight regulation, improving glycemic control (such as reducing fasting insulin, glycated hemoglobin, and diabetes conversion rate), and regulating lipid metabolism (such as reducing triglycerides and cholesterol). Therefore, time-restricted eating is considered an important research tool for studying the interaction between the biological clock and metabolism and exploring intervention strategies for metabolic diseases.

[0005] However, current methods for implementing timed feeding in animal models typically rely on researchers manually controlling the feeding and removal of feed, or using simple timers. These methods are not only time-consuming and labor-intensive, but also make it difficult to guarantee the accuracy of timing and operation.

[0006] The more advanced solutions in the current technology employ automated feed dispensing systems based on timed control. These systems control the opening and closing of feed troughs through a preset schedule, fulfilling the basic requirements of time-limited feeding experiments in animals. For example, while existing mouse energy metabolism monitoring systems on the market are powerful, integrating monitoring of energy metabolism, behavior, and physiological parameters, and offering options for automated feeding control, there is still room for improvement in areas specifically designed for flexible, intelligent, and adaptive time-limited feeding interventions.

[0007] The main technical shortcomings of current automated time-limited feeding implementation schemes include:

[0008] Lack of dynamic adaptability: Most systems are based on fixed schedules and lack the ability to dynamically adjust the feeding window according to the animal's real-time physiological state (such as weight changes, activity level, and actual intake) and environmental factors.

[0009] Insufficient monitoring dimensions: Simple timer switches cannot monitor in real time the details of an animal's actual feeding behavior (such as feeding frequency, amount of food consumed at one time, and microfeeding events) and its immediate impact on its physiological state.

[0010] Lack of flexibility and personalization: Fixed schedules are difficult to adapt to complex and ever-changing time-limited feeding experimental designs (such as switching between different modes and individualized adjustment needs), and are also difficult to cope with the individual differences that are common in research.

[0011] Limited data integration and analysis: Existing systems often lack the ability to efficiently integrate, manage and deeply analyze multimodal data such as feeding control, animal behavior and physiological state, which limits research efficiency and in-depth interpretation of results.

[0012] Furthermore, time-restricted feeding studies typically require high standards for animal husbandry environments, especially long-term experiments, necessitating assurance of animal welfare and data reliability. Current methods still fall short in terms of intelligent adaptation, in-depth data mining, and integration with complex physiological monitoring (such as continuous blood glucose and heart rate variability). Summary of the Invention

[0013] To address the aforementioned technical problems, this invention provides a method and system for timed feeding in animal experiments. This method and system feature intelligent self-adaptation, multi-dimensional monitoring, and in-depth data analysis, enabling precise control and comprehensive monitoring of timed feeding experiments in animals.

[0014] A first aspect of the present invention provides a method for timed feeding in animal experiments, applied to an intelligent gating system. The intelligent gating system includes a feeding box for experimental animals, an electrically controlled gating channel, a multimodal sensor array, and an intelligent control unit. The method includes:

[0015] The multimodal sensor array, which includes sensors for monitoring food intake, animal activity, physiological parameters, and environmental parameters, acquires a multimodal monitoring data stream of experimental animal status and behavior.

[0016] Based on the multimodal monitoring data stream and the preset time-limited feeding parameters, the execution control of the time-limited feeding mode is realized through the feeding time window control algorithm, dynamic adjustment algorithm and safety monitoring mechanism, and the execution data of the execution control is obtained.

[0017] Based on the execution data of the execution control and the multimodal monitoring data stream, a personalized time-limited feeding intervention strategy is output through an animal state assessment model, a diet pattern optimization algorithm, and an intervention strategy generator.

[0018] Based on the multimodal monitoring data stream, the execution data of the execution control, and the personalized time-limited feeding intervention strategy, experimental data management and analysis are realized through the data acquisition and storage module, the data processing and analysis module, and the visualization display interface.

[0019] Preferably, the acquisition of the multimodal monitoring data stream of experimental animal status and behavior includes:

[0020] The food intake monitoring sensor, which includes a weight sensor and a miniature camera, monitors food consumption in real time and filters out interference signals to obtain food intake data.

[0021] Animal movement patterns are identified and analyzed using the animal activity monitoring sensor containing an infrared activity detector or an accelerometer to obtain activity classification data;

[0022] Based on the physiological parameter monitoring sensor, the physiological parameters of the experimental animals are monitored to obtain physiological parameter data;

[0023] The environmental parameter monitoring sensor, which includes temperature, humidity, and air pressure sensors, collects and processes environmental parameters to obtain environmental change data.

[0024] Based on the food intake data, activity classification data, physiological parameter data, and environmental change data, time alignment and multi-source data fusion are performed to generate the multimodal monitoring data stream.

[0025] Preferably, the execution control for implementing the time-limited eating mode includes:

[0026] Based on the multimodal monitoring data stream and the preset time-limited feeding parameters, a feeding time window is determined through a time triggering mechanism, and a feeding time window opening and closing event table is generated in conjunction with status monitoring feedback.

[0027] Based on the feeding time window switch event table, the switching state of the electronically controlled gate channel is controlled by a state machine model;

[0028] Based on the multimodal monitoring data stream, an abnormal state detection algorithm constructed using a multi-level threshold model is used to identify potential health risks and abnormal behaviors. When a potential health risk or abnormal behavior is identified, an emergency response mechanism is triggered.

[0029] Based on the emergency response mechanism, combined with the feeding time window switch event table and the switch status of the electronic gate control channel, gate control execution instructions are generated and prioritized to realize the execution control of the time-limited feeding mode.

[0030] Preferably, the abnormal state detection algorithm constructed using a multi-level threshold model identifies potential health risks and abnormal behaviors. When a potential health risk or abnormal behavior is identified, an emergency response mechanism is triggered, including:

[0031] Based on the multimodal monitoring data stream, multi-scale processing is performed on the multimodal monitoring data stream to construct a multi-scale time representation. A cross-scale attention module is used to interact with features at different scales of the multi-scale time representation to form multi-time scale features.

[0032] Based on the aforementioned multi-timescale features, health status indicators and their uncertainties are calculated using a Bayesian neural network to identify anomaly types; wherein the anomaly types include high-confidence anomalies, high-uncertainty anomalies, and multimodal conflicts.

[0033] Based on the high-confidence anomalies, high-uncertainty anomalies, and multimodal conflicts, different response strategies are selected according to the anomaly type to trigger the emergency response mechanism.

[0034] Preferably, the output of the personalized time-restricted eating intervention strategy includes:

[0035] Based on the multimodal monitoring data stream, an animal health status assessment model is constructed using feature extraction algorithms and machine learning models.

[0036] Based on the animal health status assessment model and the execution data of the execution control, the results of the dietary pattern impact analysis are constructed through time series data analysis and correlation calculation.

[0037] Based on the analysis results of the impact of the dietary pattern and the preset research objectives, a personalized dietary pattern optimization scheme is generated through multi-objective optimization algorithm and dynamic sparse training.

[0038] Based on the aforementioned personalized dietary pattern optimization scheme, a personalized time-limited eating intervention strategy is output through a rule transformation engine and parameter mapping.

[0039] Preferably, the step of generating a personalized dietary pattern optimization scheme through a multi-objective optimization algorithm and dynamic sparse training includes:

[0040] Based on the results of the dietary pattern impact analysis, a joint feature representation including time-related features, physiological response features, and behavioral pattern features is extracted to form an embedding representation in a high-dimensional vector space;

[0041] Based on the embedded representation, an importance score matrix is ​​calculated through feature importance evaluation to identify key feature dimensions;

[0042] Due to limitations in computing resources, feature sparsification is performed on the key feature dimensions, and features whose importance is lower than a preset threshold are pruned to obtain pruned key feature dimensions. Then, the pruned key feature dimensions are allocated hierarchically to form a lightweight optimization model.

[0043] Based on the aforementioned lightweight optimization model and guided by the preset research objectives, a multi-objective optimization search is performed to generate a personalized dietary pattern optimization scheme.

[0044] Preferably, the implementation of experimental data management and analysis includes:

[0045] Based on the multimodal monitoring data stream, the execution data of the execution control, and the personalized time-limited feeding intervention strategy, the multi-channel data is traceably stored through a time-series database and a data synchronization mechanism to obtain the stored multi-channel data.

[0046] For the stored multi-channel data, a preprocessed dataset is generated through noise filtering, data standardization, and missing value handling.

[0047] The preprocessed dataset is processed using various data analysis methods to extract experimental results and pattern features; these methods include statistical analysis, machine learning, and time series analysis algorithms.

[0048] Based on the experimental results and pattern characteristics, experimental data management and analysis are achieved through interactive charts and multidimensional data visualization techniques.

[0049] Preferably, the method further includes:

[0050] Based on the execution results of the personalized time-limited eating intervention strategy and the multimodal monitoring data stream, a comprehensive effect score is calculated using an effect evaluation algorithm to generate an intervention effect evaluation result.

[0051] When the intervention effect evaluation result is lower than the preset effect threshold, the feedback adjustment mechanism is activated to conduct multi-angle investigation and generate adjustment suggestions.

[0052] Based on the aforementioned adjustment recommendations, the animal status assessment model and optimization strategy are updated to achieve iterative optimization of the intervention strategy.

[0053] Preferably, before acquiring the multimodal monitoring data stream of the experimental animal's state and behavior, the method further includes:

[0054] Based on experimental requirements and animal types, separate rearing spaces and dedicated passage areas were designed, and the basic structure of rearing boxes that meet SPF-level cleanliness standards was constructed.

[0055] Based on the basic structure of the feeding box, a sliding door gate driven by a micro servo motor and a position sensor are integrated to construct the electronically controlled gate channel;

[0056] Based on the basic structure of the feeding box and the electronically controlled gate channel, the food intake monitoring sensor, the animal activity status monitoring sensor, the physiological parameter monitoring sensor and the environmental parameter monitoring sensor are deployed to construct the multimodal sensor array;

[0057] Based on the multimodal sensor array and the electronically controlled gating channel, an embedded processor, a motor drive circuit, and a high-precision analog-to-digital converter module are integrated to construct an intelligent control unit.

[0058] A second aspect of the present invention provides an animal experimental timed feeding system, comprising:

[0059] Feeding enclosures are used to house laboratory animals;

[0060] An electrically controlled gate is installed on the feeding box to control the entry of laboratory animals into the feeding area;

[0061] A multimodal sensor array, including food intake monitoring sensors, animal activity status monitoring sensors, physiological parameter monitoring sensors, and environmental parameter monitoring sensors, is used to acquire multimodal monitoring data streams of laboratory animal status and behavior;

[0062] The intelligent control unit, connected to the electronically controlled gating channel and the multimodal sensor array, includes:

[0063] The data acquisition module is used to acquire multimodal monitoring data streams of experimental animal status and behavior through the multimodal sensor array containing food intake monitoring sensors, animal activity status monitoring sensors, physiological parameter monitoring sensors and environmental parameter monitoring sensors;

[0064] The feeding control module is used to control the execution of the time-limited feeding mode based on the multimodal monitoring data stream and preset time-limited feeding parameters, through feeding time window control algorithm, dynamic adjustment algorithm and safety monitoring mechanism, and to obtain execution data of the execution control.

[0065] The adaptive adjustment module is used to output a personalized time-limited feeding intervention strategy based on the execution data of the execution control and the multimodal monitoring data stream, through an animal state assessment model, a diet pattern optimization algorithm and an intervention strategy generator.

[0066] The data management module is used to manage and analyze experimental data based on the multimodal monitoring data stream, the execution data of the execution control, and the personalized time-limited feeding intervention strategy, through a data acquisition and storage module, a data processing and analysis module, and a visualization interface.

[0067] The present invention has the following beneficial effects:

[0068] 1. Through the hardware construction of the intelligent gate control system, a highly integrated transformation of the animal feeding box was achieved, ensuring that the system meets the SPF-level cleanliness standard, and the animal feeding behavior is precisely controlled through the electronic gate control channel.

[0069] 2. Through the multimodal sensing and monitoring system, comprehensive monitoring of animal food intake, activity status, physiological parameters and environmental parameters was achieved, and rich multidimensional data streams were obtained, providing a solid foundation for subsequent analysis and adjustment.

[0070] 3. Through the programmed control of the time-limited feeding mode, precise control of the feeding time window was achieved, and the safety and reliability of the experimental process were ensured through dynamic adjustment algorithms and safety monitoring mechanisms.

[0071] 4. Through adaptive adjustment algorithms, personalized intervention strategies based on individual animal differences and state changes were generated, which greatly improved the scientific rigor and accuracy of the experiment.

[0072] 5. Through the data management and experimental analysis platform, efficient management, processing, and visualization of experimental data were achieved, improving research efficiency and the ability to interpret results in depth. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1 This is a flowchart of the time-limited feeding method for animal experiments provided in the embodiments of the present invention;

[0075] Figure 2 This is a schematic diagram of the architecture of the intelligent gate control system provided in an embodiment of the present invention;

[0076] Figure 3 This is a schematic diagram of the multimodal sensing and monitoring system architecture provided in an embodiment of the present invention;

[0077] Figure 4 This is a schematic diagram of the structure of the time-limited feeding system for animal experiments provided in an embodiment of the present invention;

[0078] Figure 5 This is a schematic diagram of the animal experiment timed feeding device provided in an embodiment of the present invention. Detailed Implementation

[0079] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0080] Please see Figure 1 , Figure 1 This is a flowchart of an animal experiment time-limited feeding method provided in an embodiment of the present invention. For example... Figure 1 As shown, this embodiment provides a method for timed feeding in animal experiments, applied to an intelligent gating system. The intelligent gating system includes a feeding box for experimental animals, an electrically controlled gating channel, a multimodal sensor array, and an intelligent control unit. The method includes:

[0081] S100, by means of the multimodal sensor array containing a food intake monitoring sensor, an animal activity status monitoring sensor, a physiological parameter monitoring sensor and an environmental parameter monitoring sensor, a multimodal monitoring data stream of experimental animal status and behavior is acquired;

[0082] S200, based on the multimodal monitoring data stream and the preset time-limited feeding parameters, the execution control of the time-limited feeding mode is realized through the feeding time window control algorithm, dynamic adjustment algorithm and safety monitoring mechanism, and the execution data of the execution control is obtained;

[0083] Based on multimodal monitoring data streams and preset time-limited feeding parameters, the execution control of the time-limited feeding mode is achieved through feeding time window control algorithms, dynamic adjustment algorithms, and safety monitoring mechanisms, yielding execution control data. This step first provides configuration options for various preset modes through a user-friendly graphical interface, including the classic 16 / 8 mode (16 hours fasting, 8 hours feeding), 14 / 10 mode, and 5:2 intermittent fasting mode, and also supports researchers in customizing feeding windows according to specific research needs. The set parameters undergo scientific and safety verification to ensure that excessively long fasting periods do not pose a risk to experimental animals.

[0084] After configuration, a real-time clock-based time base is established to generate a precise window opening and closing event table. Unlike traditional systems, this system not only relies on a preset timetable but also monitors the animal's status in real time through a status monitoring feedback mechanism. For example, when it detects that an animal is eating and the feeding window is about to close, the closing operation is intelligently delayed to avoid stress reactions that may result from sudden interruption of feeding. The system uses a state machine model to manage the gating state, including multiple states such as standby, open request, opening in progress, open completed, and close request. A PID controller achieves millimeter-level precise gating positioning, while a current monitoring mechanism prevents the gating mechanism from jamming or overloading.

[0085] The safety monitoring mechanism is a core component of the S200 procedure. It employs a multi-level threshold model to continuously analyze the monitoring data stream, categorizing detected anomalies into three levels: warning, moderate risk, and high risk. When a potential health risk is detected, the corresponding emergency response mechanism is activated based on the severity of the risk. For example, if an animal has barely consumed food for 24 hours and is experiencing rapid weight loss, the feed passage may be temporarily opened, even within the originally scheduled fasting period. All emergency events are communicated to researchers through multiple channels, including mobile app push notifications, SMS, or email, ensuring timely handling of emergencies. The entire execution control process uses a priority ranking mechanism, ensuring that safety-related instructions always have the highest execution priority, followed by dynamically adjusted instructions, and finally, basic time window instructions.

[0086] S300, based on the execution data of the execution control and the multimodal monitoring data stream, outputs a personalized time-limited feeding intervention strategy through an animal state assessment model, a diet pattern optimization algorithm and an intervention strategy generator;

[0087] Based on execution control data and multimodal monitoring data streams, personalized time-limited feeding intervention strategies are output through animal status assessment models, dietary pattern optimization algorithms, and intervention strategy generators. This step first analyzes key health indicators in the monitoring data stream using an animal health status assessment model, including weight change trends, activity frequency and intensity, and feeding behavior patterns. Machine learning models are then used to quantitatively assess the animal's current health status, providing foundational data support for subsequent decision-making.

[0088] Subsequently, the dietary pattern impact analysis algorithm, through time-series data analysis and correlation calculations, delves into the effects of currently applied time-restricted feeding patterns on animal health. This algorithm not only focuses on simple direct associations but also attempts to separate the influence of time-restricted feeding interventions from other environmental factors through causal inference. Based on these analytical results, personalized dietary pattern optimization schemes are generated using multi-objective optimization algorithms and dynamic sparse training techniques. This process considers both research objectives and individual animal characteristics, seeking the optimal balance between experimental scientific value and animal welfare.

[0089] The intervention strategy generator, the final step in step S300, transforms abstract optimization schemes into concrete, executable control instructions. The strategy generator employs a knowledge graph-based rule transformation engine, incorporating expert knowledge and experience from the field of time-limited eating research. For example, when the optimization algorithm suggests "increasing the morning eating window to improve metabolic rate," the rule transformation engine parses this abstract goal into a series of specific instructions, such as "opening the morning eating window 30 minutes earlier." The parameter mapping module further quantifies these rules into precise system parameters, comprehensively considering hardware limitations, safety boundaries, and operational precision. Finally, the execution plan generation algorithm organizes these parameters into a time-series execution plan, generating a sequence of instructions that the control system can directly understand, ensuring the accurate execution of personalized intervention strategies.

[0090] S400, based on the multimodal monitoring data stream, the execution data of the execution control, and the personalized time-limited feeding intervention strategy, realizes experimental data management and analysis through a data acquisition and storage module, a data processing and analysis module, and a visualization display interface.

[0091] Based on multimodal monitoring data streams, execution data from control measures, and personalized time-limited feeding intervention strategies, experimental data management and analysis are achieved through data acquisition and storage modules, data processing and analysis modules, and visualization interfaces. This step first ensures the complete preservation of raw, high-resolution data from all sensors using a time-series database and establishing a multi-channel data synchronization mechanism to guarantee data traceability and integrity. The data preprocessing module performs noise filtering, data standardization, and missing value handling on the stored raw data, generating a reliable preprocessed dataset for subsequent analysis.

[0092] It offers a rich library of multi-dimensional analysis algorithms, including statistical analysis tools, machine learning models, time series analysis methods, behavioral pattern analysis techniques, and energy metabolism calculation formulas. Researchers can leverage these tools to conduct in-depth data mining, supporting user-defined data grouping and normalization. Analysis results are presented through an intuitive visualization system, offering interactive charts, heatmaps, trajectory graphs, and other visualization methods, supporting comprehensive display of multi-dimensional data and customized report templates.

[0093] The data management module also integrates research report and conclusion generator functions. Through conclusion derivation algorithms and natural language generation technology, it assists researchers in summarizing experimental findings and forming preliminary conclusions. It provides standardized data output formats, is compatible with commonly used analysis software, and facilitates more professional data analysis for researchers. Furthermore, the system integrates an alarm function based on user-defined thresholds, which promptly notifies researchers when key indicators exceed preset ranges. Through this comprehensive data management and analysis platform, researchers can fully grasp the experimental process and results, improve research efficiency and data mining depth, and provide strong data support for interpreting the scientific mechanisms of time-restricted feeding.

[0094] In particular, it supports a continuous improvement mechanism for intervention strategies. An effectiveness evaluation algorithm calculates the overall effectiveness score of intervention measures. When the effectiveness evaluation result is lower than expected, a feedback adjustment mechanism is activated to analyze the reasons and generate targeted adjustment suggestions. An incremental learning strategy is adopted, continuously updating the health status assessment model and optimization strategies as experimental data accumulates, making the system increasingly adaptable to specific experimental subjects. This adaptive closed-loop design is a key advantage that distinguishes this system from traditional fixed-parameter systems, enabling it to gradually optimize personalized intervention plans for different animal groups, significantly improving the scientific value of the experiment and the level of animal welfare.

[0095] In one specific embodiment, prior to acquiring the multimodal monitoring data stream of experimental animal status and behavior, the method further includes:

[0096] Based on experimental requirements and animal types, separate rearing spaces and dedicated passage areas were designed, and the basic structure of rearing boxes that meet SPF-level cleanliness standards was constructed.

[0097] Based on the basic structure of the feeding box, a sliding door gate driven by a micro servo motor and a position sensor are integrated to construct the electronically controlled gate channel;

[0098] Based on the basic structure of the feeding box and the electronically controlled gate channel, the food intake monitoring sensor, the animal activity status monitoring sensor, the physiological parameter monitoring sensor and the environmental parameter monitoring sensor are deployed to construct the multimodal sensor array;

[0099] Based on the multimodal sensor array and the electronically controlled gating channel, an embedded processor, a motor drive circuit, and a high-precision analog-to-digital converter module are integrated to construct an intelligent control unit.

[0100] In practical implementation, the design first utilizes standard laboratory animal housing boxes, selecting materials resistant to autoclaving and routine cleaning processes to meet SPF or higher animal facility cleanliness standards. By adding partitions and dedicated passageways, a segmented housing space is constructed, creating a controllable animal living area and feed supply area. The design must consider good ventilation and gas exchange, drawing inspiration from advanced metabolic cage designs to ensure a stable cage environment and reduce CO2 and ammonia accumulation. Cage lids and other components can be made of transparent materials to facilitate observation and maintain normal circadian rhythms in the animals. The design should maximize animal living space and integrate environmental enrichment components to improve animal welfare.

[0101] Please see Figure 2 , Figure 2 This is a schematic diagram of the architecture of the intelligent gate control system provided in an embodiment of the present invention. Figure 2 As shown, a sliding gate or rotating gate mechanism driven by a miniature servo motor or stepper motor is designed based on a partitioned feeding space. It includes anti-jamming protection devices and precise position sensors. The motor and control circuitry must be appropriately waterproofed and insulated to withstand potentially humid environments and cleaning needs. The gate control can be designed for food or liquids, and even interchangeable.

[0102] Integrated high-precision weight sensors are used to monitor food and water consumption and animal weight. Integrated infrared activity detectors or accelerometers are used to monitor animal activity status and location. Integrated miniature cameras are used to assist in monitoring feeding behavior, activity patterns, and abnormal conditions. Integrated temperature and humidity sensors, light sensors, and sound sensors are used to monitor environmental parameters. Other sensors can be expanded for integration, such as H2 and CH4 sensors for microbiome research, or 13CO2 sensors for stable isotope tracing. Sensor deployment should consider ease of cleaning and maintenance.

[0103] Based on an embedded processor platform (such as ARM Cortex series or ESP32). Integrates motor drive circuitry, a high-precision analog-to-digital converter (ADC), and data storage modules (such as SD card or Flash). The control unit should be designed for low power consumption and consider electrical insulation requirements. Employs a modular design for easy future functional expansion and maintenance. Designs low-power power management circuitry to support AC or battery-powered solutions. Provides wired (e.g., Ethernet) / wireless (e.g., Wi-Fi, Bluetooth) communication interfaces, as well as a USB interface for local configuration and data download.

[0104] In one specific embodiment, acquiring the multimodal monitoring data stream of experimental animal status and behavior includes:

[0105] The food intake monitoring sensor, which includes a weight sensor and a miniature camera, monitors food consumption in real time and filters out interference signals to obtain food intake data.

[0106] Animal movement patterns are identified and analyzed using the animal activity monitoring sensor containing an infrared activity detector or an accelerometer to obtain activity classification data;

[0107] Based on the physiological parameter monitoring sensor, the physiological parameters of the experimental animals are monitored to obtain physiological parameter data;

[0108] The environmental parameter monitoring sensor, which includes temperature, humidity, and air pressure sensors, collects and processes environmental parameters to obtain environmental change data.

[0109] Based on the food intake data, activity classification data, physiological parameter data, and environmental change data, time alignment and multi-source data fusion are performed to generate the multimodal monitoring data stream.

[0110] Please see Figure 3 , Figure 3 This is a schematic diagram of the multimodal sensing and monitoring system architecture provided in an embodiment of the present invention. Figure 3 As shown, in the specific implementation process, it first relies on a high-precision weight sensor and a miniature camera. High-precision ADC data acquisition and digital signal filtering are used to achieve a high sampling rate, capturing details of feeding behavior, including micro-feeding events. Combined with image recognition algorithms, feeding posture, duration, etc., are analyzed. A spill-proof food container is considered.

[0111] Based on an XYZ triaxial infrared motion detector or accelerometer, and employing a motion pattern recognition algorithm (distinguishing between coarse and fine motion, calculating activity intensity, tracking position, and classifying behavior), the system should be able to differentiate between real-time activity and static obstacles detected by the sensor.

[0112] It can be integrated with non-invasive physiological parameter acquisition devices or is compatible with implantable telemetry devices. Monitoring is performed through signal processing algorithms and abnormal state detection models.

[0113] Based on sensors for temperature, humidity, light intensity, and air pressure, environmental parameters are collected and processed, and a circadian rhythm recognition algorithm is used for monitoring. External environmental factors that may affect the experiment are recorded.

[0114] Based on the aforementioned monitoring data, a multimodal monitoring data stream comprehensively reflecting animal status is generated through time alignment algorithms, multi-source data fusion techniques, and feature extraction methods. In multimodal data fusion processing, the system faces a core technical challenge: how to extract consistent feature representations from sensor data with different dimensions and distribution characteristics. To address this challenge, this system employs a dimensionless convergence technique based on an approximate Gaussian mixture diffusion model.

[0115] In the diffusion model, the time-dependent noise scheduling function β(t) adopts a cosine scheduling strategy:

[0116]

[0117] Where β min =10 -4 ,β max =0.02, where T is the total number of diffusion steps, set to 1000. This scheduling method maintains low noise in the early stages of diffusion, which helps preserve the detailed features of the data, while increasing the noise intensity in the later stages to accelerate convergence to the standard distribution.

[0118] The number of components M in the Gaussian mixture model is adaptively determined based on data complexity, initially set to 5, and dynamically adjusted during training using the Bayesian Information Criterion (BIC), eventually converging to between 3 and 8. The mixture weights πj are iteratively optimized using the Expectation-Maximization (EM) algorithm, with the iteration stopping when the change in the log-likelihood function is less than 10. -6 Or it can reach the maximum number of iterations, 100.

[0119] Adaptive scaling factor λ in dimensionless convergence mechanism i (t) is initially set to 1.0, and the Wasserstein distance objective function is optimized using gradient descent:

[0120]

[0121] The target distribution q(t) was chosen as a standard normal distribution N(0, I). The optimization process used the Adam optimizer with a learning rate of 0.01 and a total of 500 iterations. To improve computational efficiency, the Wasserstein distance was approximated using the Sinkhorn algorithm, with a regularization coefficient ε set to 0.01.

[0122] The raw data collected by each sensor is first considered as a distribution in multiple high-dimensional spaces, which typically exhibit complex non-Gaussian properties. Traditional methods often solve the problem of dimensional inconsistencies through simple standardization, but this approach cannot handle the inherent distributional differences in the data.

[0123] This technique introduces a diffusion model as the theoretical foundation for data representation. The diffusion model maps any complex distribution to a standard space by defining a continuous transformation process from a complex distribution to a simple distribution (typically a standard Gaussian distribution). Specifically, for data xi from sensor i, the system defines the diffusion process:

[0124]

[0125] Where β(t) is a predefined time-dependent noise scheduling function, and dw is a standard Wiener process. This stochastic differential equation describes how the data is gradually "contaminated" by noise until it approaches a Gaussian distribution.

[0126] The innovation of this system lies in representing the intermediate process distribution using an approximate Gaussian Mixture Model (GMM) and introducing a dimensionless convergence mechanism. Specifically, the system discretizes the diffusion process into K time steps, where the data distribution is approximated by a Gaussian Mixture Model at each time step t:

[0127]

[0128] Where M is the number of mixed components, π j It is a mixed weight.

[0129] The dimensionless convergence mechanism introduces an adaptive scaling factor λ. i (t) is used to handle the dimensionless issues of data from different sensors:

[0130]

[0131] λ i The value of (t) is automatically determined by minimizing the distributional differences of data from different sensors in the potential space. The Wasserstein distance is used as a measure of these distributional differences.

[0132]

[0133] Where q(t) is a predefined target distribution.

[0134] The advantages of this method are: (1) no need to know the specific distribution of the data in advance; (2) it can handle complex distributions such as multi-peak and skewed distributions; and (3) it naturally achieves data noise reduction and anomaly detection through the diffusion process.

[0135] In animal monitoring experiments, this technology enables the system to extract unified feature representations from data with different characteristics, such as food consumption, activity intensity, and body temperature changes, significantly improving the accuracy and robustness of subsequent analysis. For example, when food consumption data exhibits periodic fluctuations while activity data shows sudden changes, traditional methods struggle to effectively integrate them, while this technology can capture the intrinsic correlations between these different patterns.

[0136] In one specific embodiment, the execution control of the time-limited eating mode includes:

[0137] Based on the multimodal monitoring data stream and the preset time-limited feeding parameters, a feeding time window is determined through a time triggering mechanism, and a feeding time window opening and closing event table is generated in conjunction with status monitoring feedback.

[0138] Based on the feeding time window switch event table, the switching state of the electronically controlled gate channel is controlled by a state machine model;

[0139] Based on the multimodal monitoring data stream, an abnormal state detection algorithm constructed using a multi-level threshold model is used to identify potential health risks and abnormal behaviors. When a potential health risk or abnormal behavior is identified, an emergency response mechanism is triggered.

[0140] Based on the emergency response mechanism, combined with the feeding time window switch event table and the switch status of the electronic gate control channel, gate control execution instructions are generated and prioritized to realize the execution control of the time-limited feeding mode.

[0141] Specifically, the execution control of the time-limited feeding mode is a key component of the system's core function, translating the experimental parameters set by researchers into precise physical control actions. First, the system determines the feeding time window through a time-triggered mechanism. This mechanism, based on a high-precision real-time clock chip, provides millisecond-level time accuracy, ensuring that the opening and closing of the feeding window strictly adheres to the preset time. The system supports various complex time strategies, including fixed time windows (e.g., allowing feeding from 8:00 AM to 4:00 PM daily), segmented time windows (e.g., allowing feeding from 9:00 AM to 11:00 AM and 5:00 PM to 7:00 PM respectively), and periodically varying windows (e.g., using different feeding times on weekdays and weekends). The time-triggered mechanism considers not only absolute time but also the biological rhythms of the experimental animals, allowing for the setting of relative time windows based on the light cycle to better simulate foraging behavior in natural environments.

[0142] Building upon time-triggered mechanisms, a state monitoring feedback mechanism is integrated to dynamically adjust gating behavior. For example, when an animal is detected to be actively feeding and the time window is about to close, the system will appropriately extend the window time according to preset rules to avoid stress reactions that may result from sudden interruption of feeding behavior. Similarly, if the animal is detected to be asleep at the beginning of the window opening, the system may postpone the actual opening time of the feeding window, waiting for the animal to wake up naturally. This feedback adjustment based on behavioral state makes feeding window control more in line with the animal's natural behavioral patterns, reducing the interference of experimental intervention on the animal's normal physiological rhythms. By combining time-triggered and state feedback, a detailed feeding time window opening and closing event table is generated. This event table not only contains basic time information but also integrates key parameters such as trigger conditions, execution actions, and priorities. Gating operations are performed based on this event table to achieve precise time-limited feeding control. The event table supports multiple time modes (such as absolute time and relative time) and conditional expressions, and can be dynamically adjusted according to the animal's state, combining strict time management with flexible adaptability, providing a technical guarantee for the precise execution of experimental protocols.

[0143] Based on the generated feeding time window switching event table, a comprehensive state machine model is used to control the switching state of the electrically controlled gate channel. This state machine model decomposes the gate operation into multiple refined states, including standby state, opening preparation state, opening execution state, opening completion state, closing preparation state, closing execution state, and closing completion state. Each state has clearly defined entry and exit conditions and fault-tolerant handling mechanisms. For example, before initiating the opening execution, the system checks the motor drive circuit status and gate position sensor feedback to confirm that the system is in a controllable state; during the opening process, it monitors the motor current and gate position changes in real time to ensure smooth completion of the action; if abnormal resistance or position deviation is detected, a preset abnormal handling procedure is triggered, such as pause, retry, or alarm. The application of the state machine model makes the gate execution process more stable and reliable, capable of handling various expected and unexpected situations.

[0144] A safety monitoring mechanism is an indispensable safeguard in the implementation and control of time-limited feeding. This mechanism, based on multimodal monitoring data streams, employs an anomaly detection algorithm constructed using a multi-level threshold model to continuously monitor the animal's health status and behavior. The algorithm sets three levels of safety thresholds: warning, moderate risk, and high risk. For example, a warning may be triggered if an animal's food intake is less than 30% of the expected intake within 24 hours; a moderate risk level may be escalated if insufficient food intake persists for 48 consecutive hours and weight loss exceeds 5%; and a high risk level may be determined if a significant decrease in activity or abnormal behavioral patterns are further observed. Different risk levels correspond to different response strategies: a warning level primarily records and notifies researchers; moderate risk may lead to automatic adjustment of feeding window parameters; and a high risk level will immediately trigger emergency measures, such as temporarily opening the food access channel and issuing emergency notifications through various channels (mobile app push notifications, SMS, email, etc.).

[0145] When the safety monitoring mechanism identifies a potential health risk or abnormal behavior, it triggers an emergency response mechanism. This mechanism first assesses the urgency of the risk and the type of intervention required, then, based on the current feeding window opening / closing event table and the status of the electrically controlled gating channels, generates appropriate response instructions. For example, for a detected risk of severe dehydration, the system might temporarily open the drinking channel without changing the status of the food channel; for a detected risk of severe energy deficiency, the system might temporarily open the food channel for a limited time within the originally planned fasting period. All emergency response operations are recorded in detail, including triggering conditions, execution time, intervention type, and duration, facilitating subsequent analysis and adjustment of experimental protocols by researchers.

[0146] After comprehensively considering the time event table, safety monitoring results, and emergency response requirements, a dedicated priority processing mechanism generates the final gating execution instruction sequence. This mechanism ensures that safety-related instructions always have the highest execution priority, followed by emergency response instructions, then dynamic adjustment instructions, and finally basic time window instructions. This hierarchical priority design ensures that animal health and welfare are always guaranteed to the highest degree while meeting basic research needs. The final gating execution instructions are processed by the motor control signal conversion module and converted into specific motor drive signals, including precise position control parameters, speed curves, and torque limits, ensuring smooth, accurate, and reliable gating actions. The entire execution process adopts a closed-loop control design, using position sensors and current monitoring for real-time feedback to ensure accurate execution of each control instruction and timely adjustment or termination of operation in case of unexpected situations.

[0147] It is worth emphasizing that a comprehensive fault-tolerant design was adopted in the process of implementing the time-limited feeding mode control. This includes hardware redundancy (such as backup power supply and dual sensors), software fault tolerance (such as anomaly detection and safe mode switching), and communication backup (multi-channel alarm mechanism). Even in the event of partial hardware failure or software anomaly, basic safety monitoring and control functions can still be maintained, ensuring that experimental animals are not put in danger due to system failure. At the same time, all actions and abnormal events are recorded in the system log in real time, providing complete event tracking information for fault diagnosis and experimental data analysis.

[0148] Through the complex and sophisticated control mechanism described above, highly precise execution control of the timed feeding mode is achieved. Compared with traditional simple timer switches, this system can not only strictly control the feeding time window according to the preset plan, but also make intelligent adjustments based on the animal's real-time status and provide timely protection in case of potential risks. This significantly improves the accuracy, reliability, and safety of animal timed feeding experiments, providing researchers with an ideal experimental platform.

[0149] The programmed control flow for time-limited eating provided in this embodiment of the invention first offers a user graphical interface. It supports multiple preset modes (16 / 8, 14 / 10, 5:2, Eat-Stop-Eat, etc.) and custom mode configurations. Access control rules based on time windows or intake amounts can be set. It includes parameter verification and safety prompt mechanisms.

[0150] Based on configured time-limited feeding parameters and monitoring data streams, this algorithm achieves precise time window control under different feeding modes through a time-triggered mechanism, state monitoring feedback, and a precise gating control algorithm. The algorithm first establishes an absolute time base based on a real-time clock (RTC) and generates a precise time event table according to the configured parameters. The time-triggered mechanism triggers gating state transitions at the time points specified in the event table, such as opening the feed channel at the start of the feeding window and closing it at the end. The algorithm uses a state machine model to manage gating states, including multiple states such as standby, open request, opening in progress, open completed, and close request. Notably, this algorithm not only relies on time triggering but also integrates monitoring data streams for state monitoring feedback. For example, when it detects that an animal is feeding and the window is about to close, the system intelligently delays the closing operation to avoid stress caused by sudden interruption. The algorithm also achieves precise gating control, using a PID controller to achieve millimeter-level precision in the door opening and closing position, and using current monitoring to prevent the gating mechanism from jamming or overloading.

[0151] Based on monitoring data streams and execution status, this system dynamically fine-tunes the feeding time window through an animal status assessment algorithm, an adaptive rule engine, and an adjustment decision model. This algorithm addresses the problem that static timetables cannot adapt to individual animal differences and status changes. First, the algorithm analyzes the monitoring data stream in real time through the animal status assessment module, calculating key health indicators including weight change rate, activity level, and feeding patterns. The adaptive rule engine generates adjustment suggestions based on these indicators and preset safety thresholds; for example, it suggests expanding the feeding window when an animal's weight loss exceeds the safety threshold. The adjustment decision model prioritizes and resolves conflicts among the adjustment suggestions, considering the balance between experimental objectives and animal welfare, ultimately forming a specific time window adjustment strategy. This dynamic adjustment capability allows the system to better adapt to individual animal differences while maintaining experimental objectives, improving the scientific rigor and humaneness of the experiment.

[0152] In one specific embodiment, the abnormal state detection algorithm constructed using a multi-level threshold model identifies potential health risks and abnormal behaviors. When a potential health risk or abnormal behavior is identified, an emergency response mechanism is triggered, including:

[0153] Based on the multimodal monitoring data stream, multi-scale processing is performed on the multimodal monitoring data stream to construct a multi-scale time representation. A cross-scale attention module is used to interact with features at different scales of the multi-scale time representation to form multi-time scale features.

[0154] Specifically, the multi-scale time representation is constructed using a multi-level sliding window method. The original data is collected at a frequency of 10Hz, and then minute-level and hour-level data representations are constructed through mean pooling. Specifically, for each sensor data stream X, a time pyramid structure is constructed: X1 (original 10Hz sampling), X2 (1-minute mean), and X3 (1-hour mean). Features of each scale data are extracted through a 3-layer convolutional neural network with kernel sizes of [3, 5, 7] and the number of channels of [32, 64, 128]. The activation function used is LeakyReLU.

[0155] The cross-scale attention module employs a multi-head attention mechanism with 4 heads and an embedding dimension of 128. For each feature representation at scale i, a query matrix Q_i is generated through linear projection, while features at other scales j generate corresponding key-value matrices K_j and V_j. Attention calculation uses scaled dot product attention.

[0156]

[0157] Where d is the dimension of the attention module (set to 128), and the temperature coefficient. Used for scaling to avoid the gradient vanishing problem.

[0158] Based on the aforementioned multi-timescale features, health status indicators and their uncertainties are calculated using a Bayesian neural network to identify anomaly types; wherein the anomaly types include high-confidence anomalies, high-uncertainty anomalies, and multimodal conflicts.

[0159] Specifically, the Bayesian neural network model is trained based on historical experimental data, which includes records of the behavior and physiological indicators of healthy animals under standard experimental conditions, as well as a library of labeled abnormal state samples. The model architecture consists of four fully connected layers with the number of nodes [256, 128, 64, 32]. The activation function of the intermediate layers is ReLU, and the output layer uses the Sigmoid function to map to the abnormal probability space.

[0160] The training process employed a stochastic gradient descent optimizer with a learning rate of 0.001, a batch size of 64, and 200 training epochs. To achieve Bayesian inference capabilities, Monte Carlo Dropout was introduced after each fully connected layer with a dropout rate of 0.2. Five-fold cross-validation was used to ensure the model's generalization ability during training, and F1 score and AUC were used as evaluation metrics.

[0161] During the inference phase, 20 forward propagations are performed (using different dropout masks) to obtain the distribution characteristics of health status indicators. The mean of the indicators is then calculated. With variance σ², the system can simultaneously obtain the health assessment results and their confidence levels.

[0162]

[0163] Where T is the number of forward propagations (20), f(x, w) i ) represents the result of a single inference, w i Let be the model parameters for the i-th random sampling.

[0164] After training, the model is optimized into a lightweight model through model quantization and parameter pruning techniques, and then converted to TensorFlow Lite format for deployment on embedded processors. The final model size is controlled within 5MB, and the single inference time does not exceed 100ms, meeting real-time monitoring requirements.

[0165] Based on the high-confidence anomalies, high-uncertainty anomalies, and multimodal conflicts, different response strategies are selected according to the anomaly type to trigger the emergency response mechanism.

[0166] In its implementation, a safety monitoring mechanism is crucial for ensuring animal health and experimental reliability. This mechanism continuously analyzes the monitoring data stream, identifying potential health risks and abnormal behaviors through anomaly detection algorithms. The algorithm employs a multi-level threshold model, categorizing anomalies into three levels: warning, moderate risk, and high risk. For example, if an animal's food intake is close to zero within 24 hours, or its weight decreases rapidly, the system triggers an alarm for the corresponding level. An emergency response mechanism automatically initiates corresponding interventions based on the type and severity of the anomaly, such as temporarily opening the feed channel when severe starvation risk is detected, or adjusting environmental parameters when abnormal behavior is observed. The alarm triggering system notifies researchers through multiple channels, including mobile application push notifications, SMS, and email, ensuring timely handling of emergencies. The entire safety monitoring mechanism employs a redundant design, maintaining basic monitoring functionality even if some sensors fail.

[0167] Safety monitoring mechanisms are a crucial element in ensuring the health of subjects in animal experiments, requiring the rapid and accurate identification of potential health risks and system malfunctions. However, traditional monitoring methods face two main challenges: first, it is difficult to simultaneously capture anomalies at different time scales (from second-level fluctuations to daily cycle changes); second, it is difficult to distinguish between sensor malfunctions and genuine physiological abnormalities.

[0168] To address these issues, this system employs a multi-scale temporal fusion network technology based on an attention mechanism, specifically designed for fault diagnosis of uncertain patterns in multimodal data.

[0169] First, the system organizes the data streams from each sensor into a multi-scale time representation. For data stream X, the system constructs a time pyramid structure:

[0170] X1: Data at the original sampling frequency (e.g., 10Hz);

[0171] X2: Downsampled data at a medium time scale (e.g., 1-minute average);

[0172] X3: Long-term scale data (such as hourly trends).

[0173] Data at each scale is processed by specific encoders E1, E2, and E3 to generate corresponding feature representations.

[0174] Multi-scale temporal fusion is achieved through an attention mechanism. A cross-scale attention module is introduced to enable features from different time scales to interact with each other.

[0175]

[0176] Q i , V_j are the query matrix originating from scale i and the key-value matrix originating from scale j, respectively. This enables the system to adaptively focus on anomalous patterns at different time scales, for example, to correlate long-term trend changes with short-term sudden events.

[0177] Fault diagnosis based on uncertain modes is the core innovation of this technology. The system not only predicts abnormal states but also estimates the uncertainty of the predictions, which is achieved through a Bayesian neural network:

[0178] p(y|x, θ)=∫p(y|x, w)·p(w|θ, D)dw

[0179] Where y is the health status indicator, x is the multimodal input, w is the model parameter, θ is the distribution hyperparameter, and D is the historical data.

[0180] The Monte Carlo Dropout technique is used to approximate the posterior distribution of the parameters, and the predicted distribution is obtained through multiple forward propagations instead of a single-point estimate:

[0181]

[0182] This method enables the system to distinguish several key cases:

[0183] High-confidence abnormality: Physiological indicators are clearly abnormal, requiring immediate intervention;

[0184] High uncertainty anomaly: This may be due to sensor malfunction or an unknown pattern, requiring further verification;

[0185] Multimodal conflict: Data from different sensors indicate contradictory states, which may indicate a partial fault.

[0186] In practical applications, when the system detects an anomaly, it will adopt different response strategies based on the level of uncertainty:

[0187] For high-confidence anomalies, safety interventions are triggered directly (such as opening feed passages);

[0188] For anomalies with high uncertainty, increase the sampling frequency and initiate cross-validation procedures.

[0189] For modal conflicts, isolate the suspected sensors and continue monitoring based on redundant data.

[0190] This attention-based multi-scale temporal fusion network significantly improves the system's reliability, enabling it to accurately distinguish between real physiological abnormalities, environmental disturbances, and equipment malfunctions in complex biological experimental environments, thereby ensuring the health and safety of laboratory animals and the quality of data. For example, the system can accurately distinguish between normal, brief fasting behaviors in animals and feed supply interruptions caused by equipment malfunctions, avoiding unnecessary interventions or underreporting of risk situations.

[0191] Based on time window control, dynamic adjustment, and safety monitoring results, this module generates and executes precise gating commands through an instruction sequence generation algorithm, priority processing, and motor control signal conversion. This module translates research parameters into specific physical control actions, ensuring accurate implementation of experimental strategies and animal safety. Based on the basic time window plan, dynamic adjustment results, and safety monitoring output, the instruction sequence generation algorithm produces a complete gating action plan. This algorithm employs an event-driven design pattern, integrating control requirements from different sources onto a unified timeline. The priority processing module ensures that safety-related instructions always have the highest execution priority, followed by dynamically adjusted instructions, and finally, basic time window instructions. The motor control signal conversion module translates high-level instructions into specific motor drive signals, considering acceleration / deceleration curves, position feedback, and torque limitations to ensure the smoothness and reliability of the gating action. During execution, the system continuously monitors the gating execution status and applies closed-loop control to ensure precise completion of the action. If unexpected resistance or positional deviation is encountered, the control parameters are automatically adjusted or an anomaly handling process is triggered.

[0192] These sub-steps together form a complete closed-loop control system, realizing end-to-end management from setting research parameters to physical execution, state monitoring, and dynamic adjustment. This system not only accurately executes various complex time-limited feeding patterns but also possesses intelligent adaptation and safety assurance capabilities, significantly improving the accuracy, safety, and scientific value of animal experiments.

[0193] In one specific embodiment, the output of the personalized time-restricted eating intervention strategy includes:

[0194] Based on the multimodal monitoring data stream, an animal health status assessment model is constructed using feature extraction algorithms and machine learning models.

[0195] Based on the animal health status assessment model and the execution data of the execution control, the results of the dietary pattern impact analysis are constructed through time series data analysis and correlation calculation.

[0196] Based on the analysis results of the impact of the dietary pattern and the preset research objectives, a personalized dietary pattern optimization scheme is generated through multi-objective optimization algorithm and dynamic sparse training.

[0197] Based on the aforementioned personalized dietary pattern optimization scheme, a personalized time-limited eating intervention strategy is output through a rule transformation engine and parameter mapping.

[0198] In the specific implementation process, firstly, based on the multimodal monitoring data stream, an assessment model that can quantitatively assess the health status of animals is constructed through feature extraction algorithms, health indicator calculations, and machine learning models.

[0199] Based on health status assessment results and time-restricted eating implementation data, an algorithm for analyzing the impact of dietary patterns on health is constructed through time-series data analysis, correlation calculation, and causal inference methods.

[0200] In one specific embodiment, the step of generating a personalized dietary pattern optimization scheme through a multi-objective optimization algorithm and dynamic sparse training includes:

[0201] Based on the results of the dietary pattern impact analysis, a joint feature representation including time-related features, physiological response features, and behavioral pattern features is extracted to form an embedding representation in a high-dimensional vector space;

[0202] Based on the embedded representation, an importance score matrix is ​​calculated through feature importance evaluation to identify key feature dimensions;

[0203] Due to limitations in computing resources, feature sparsification is performed on the key feature dimensions, and features whose importance is lower than a preset threshold are pruned to obtain pruned key feature dimensions. Then, the pruned key feature dimensions are allocated hierarchically to form a lightweight optimization model.

[0204] Based on the aforementioned lightweight optimization model and guided by the preset research objectives, a multi-objective optimization search is performed to generate a personalized dietary pattern optimization scheme.

[0205] In the design of personalized diet pattern optimization algorithms, the system needs to process a large amount of multidimensional time-series data to extract individual animal characteristics and formulate the optimal intervention strategy accordingly. Traditional methods face two main challenges: first, embedded environments with limited computing resources cannot support complex models; second, the characteristics of different individuals vary significantly, making it difficult for static models to adapt.

[0206] The dynamic sparse training process first calculates the feature importance score matrix from the joint feature representation. Feature importance evaluation employs a gradient-based method.

[0207]

[0208] Where L is the composite objective function, which is a weighted sum of the health status prediction loss (mean squared error) and the behavior pattern retention loss (KL divergence):

[0209]

[0210] The weighting coefficients α = 0.7 and β = 0.3 reflect the balance between the main objectives of health status prediction and behavioral stability.

[0211] The pruning threshold τ in the feature sparsification process is initially set to the lower quartile of the importance score distribution and dynamically adjusted during experiments. When the system resource utilization rate exceeds 80%, τ is increased to the lower tertiary; when the utilization rate is below 40%, τ is decreased to the lower quintile, achieving a dynamic balance between computational resources and model accuracy.

[0212] The computational hierarchy divides features into three levels: features with importance scores in the top 30% are processed locally on edge devices in real time; features in the 30%-70% range are processed in batches at a low frequency (every 10 minutes); and the remaining features are processed once per hour or uploaded to the server for in-depth analysis.

[0213] The multi-objective optimization search employed an improved version of the NSGA-II algorithm, with optimization objectives including: (1) weight maintenance index; (2) activity balance score; and (3) food utilization efficiency. Constraints included minimum daily energy intake requirements and maximum fasting duration limits. The population size was set to 100, the number of generations to 50, the crossover probability to 0.8, and the mutation probability to 0.1. The optimal solution was finally selected using the Pareto frontier and the decision-maker's preference function.

[0214] To address these challenges, this system employs an embedded joint feature extraction technique based on dynamic sparse training. The core idea of ​​this technique is to significantly reduce the computational burden while maintaining expressive power by dynamically adjusting the model complexity and feature space dimensionality.

[0215] First, a joint feature representation is constructed for each individual animal. This representation integrates information from multiple dimensions:

[0216] Time-related characteristics: capturing the time patterns of animal feeding and activity.

[0217] Physiological response characteristics: including changes in weight, metabolic indicators, and other responses to dietary interventions.

[0218] Behavioral patterns reflect an animal's habits and preferences.

[0219] These features are encoded as an embedding representation E∈R^d in a high-dimensional vector space, where d is the initial dimension of the feature space.

[0220] Dynamic sparse training is achieved through two key mechanisms: feature importance assessment and dynamic structure adjustment.

[0221] Feature importance assessment quantifies the contribution of each feature dimension by introducing an importance score matrix S:

[0222]

[0223] Where L is the objective function (e.g., health status prediction error). The importance score reflects the degree of influence of a specific feature dimension on model performance.

[0224] Dynamic structure adjustment adaptively adjusts the model structure according to the computing environment and task requirements:

[0225] Feature sparsification: Pruning features with importance below a threshold τ:

[0226]

[0227] in

[0228] Computational hierarchy allocation: Dynamically allocate computing resources to different processing levels based on feature importance.

[0229] High importance feature: Processed locally on the device in real time;

[0230] Medium importance characteristic: Batch processing at low frequency;

[0231] Low importance feature: can be selectively uploaded to the cloud for periodic analysis.

[0232] Knowledge distillation: By distilling knowledge from complex models into lightweight models, the number of parameters is reduced while maintaining performance.

[0233] min θ KL(p T (y|x)‖p S (y|x;θ))

[0234] Where p T and p S These are the output probability distributions of the teacher model and the student model, respectively.

[0235] This dynamic sparse training method enables the system to run efficiently on embedded devices while maintaining the model's adaptability to individual differences. In practical applications, the system can dynamically adjust the model complexity for different individuals and at different times, such as retaining more feature dimensions in the early stages of experimentation to capture baseline behavior, while reducing dimensions in the stable phase to improve efficiency.

[0236] This method allows the system to extract the unique characteristics of each individual animal from limited observational data, and generate more precise personalized dietary intervention strategies, significantly improving experimental results and scientific research value.

[0237] The dynamic sparse training process first calculates the feature importance score matrix from the joint feature representation. Feature importance evaluation employs a gradient-based method.

[0238]

[0239] Where L is the composite objective function, which is a weighted sum of the health status prediction loss (mean squared error) and the behavior pattern retention loss (KL divergence):

[0240]

[0241] The weighting coefficients α = 0.7 and β = 0.3 reflect the balance between the main objectives of health status prediction and behavioral stability.

[0242] The pruning threshold τ in the feature sparsification process is initially set to the lower quartile of the importance score distribution and dynamically adjusted during experiments. When the system resource utilization rate exceeds 80%, τ is increased to the lower tertiary; when the utilization rate is below 40%, τ is decreased to the lower quintile, achieving a dynamic balance between computational resources and model accuracy.

[0243] The computational hierarchy divides features into three levels: features with importance scores in the top 30% are processed locally on edge devices in real time; features in the 30%-70% range are processed in batches at a low frequency (every 10 minutes); and the remaining features are processed once per hour or uploaded to the server for in-depth analysis.

[0244] The multi-objective optimization search employed an improved version of the NSGA-II algorithm, with optimization objectives including: (1) weight maintenance index; (2) activity balance score; and (3) food utilization efficiency. Constraints included minimum daily energy intake requirements and maximum fasting duration limits. The population size was set to 100, the number of generations to 50, the crossover probability to 0.8, and the mutation probability to 0.1. The optimal solution was finally selected using the Pareto frontier and the decision-maker's preference function.

[0245] Based on the optimization results, an intervention strategy generator capable of outputting specific gating adjustment parameters was developed using a rule transformation engine, parameter mapping, and execution plan generation algorithm. The intervention strategy generator aims to transform the generated optimization schemes into concrete, executable control instructions. This process first uses a rule transformation engine to parse abstract optimization objectives into specific operational rules. The rule transformation engine employs a knowledge graph-based reasoning model, incorporating expert knowledge and experimental experience from the field of time-limited feeding research. For example, when the optimization algorithm suggests the abstract objective of "increasing the duration of the morning feeding window to improve metabolic rate," the rule transformation engine transforms it into a series of specific rules, such as "advancing the start time of the morning feeding window by 30 minutes" and "adjusting the feed delivery rate to 1.2 times the standard value." The parameter mapping module quantifies these rules into precise system parameters, taking into account hardware limitations, safety boundaries, and operational precision. For example, the gating opening time is mapped not only to a specific clock time but also to the acceleration characteristics and positional accuracy of the gating mechanism. The execution plan generation algorithm, the final step in the intervention strategy generator, organizes the quantified parameters into a time-series execution plan and generates a sequence of instructions that can be directly understood by the control system. This algorithm employs an event-driven architecture, enabling it to handle complex conditional branches and temporal dependencies, ensuring that interventions are executed in the correct order and at the right time. The output of the entire intervention strategy generator is a complete execution plan document, containing specific time points, gating parameters, monitoring thresholds, and anomaly handling strategies, which can be directly executed by the control system.

[0246] In one specific embodiment, the method further includes:

[0247] Based on the execution results of the personalized time-limited eating intervention strategy and the multimodal monitoring data stream, a comprehensive effect score is calculated using an effect evaluation algorithm to generate an intervention effect evaluation result.

[0248] When the intervention effect evaluation result is lower than the preset effect threshold, the feedback adjustment mechanism is activated to conduct multi-angle investigation and generate adjustment suggestions.

[0249] Based on the aforementioned adjustment recommendations, the animal status assessment model and optimization strategy are updated to achieve iterative optimization of the intervention strategy.

[0250] Intervention effect evaluation and iterative optimization enable the system's self-learning capability, allowing intervention strategies to be continuously improved based on actual results. The effect evaluation algorithm quantifies the effectiveness of the intervention by comparing changes in animal status before and after the intervention. This algorithm employs a multi-indicator comprehensive scoring system, comparing data from various dimensions such as weight change, food intake, activity level, and physiological indicators with the expected goals to calculate a comprehensive effect score. To handle random fluctuations and individual differences during the experiment, the algorithm introduces Bayesian statistical methods to quantify uncertainty, distinguishing between deterministic changes and random fluctuations in the intervention effect. The feedback adjustment mechanism generates targeted adjustment suggestions based on the evaluation results. When the intervention effect is detected to be less than expected, the mechanism analyzes possible causes, such as insufficient dosage, inappropriate timing, or individual sensitivity differences, and proposes corresponding adjustment directions. Notably, the feedback mechanism not only focuses on short-term effects but also tracks long-term trends, preventing the system from falling into local optimization traps. The iterative optimization method applies the adjustment suggestions to the health status assessment model, impact analysis algorithm, and optimization strategy, forming a complete optimization closed loop. This method employs an incremental learning strategy, continuously updating the health status assessment model, impact analysis algorithm, and optimization strategy as experimental data accumulates, thus increasing the system's adaptability to specific experimental subjects. This adaptive learning process enables the system to gradually develop personalized intervention strategies tailored to the characteristics of specific animal groups, even starting with the same initial parameters.

[0251] The coordinated operation of these steps endows the system with true intelligent adaptive capabilities. Unlike traditional pre-programmed control systems, this system can continuously learn and improve during the experiment, adapting to individual differences and state changes in different experimental subjects. For example, when the system detects that a group of animals is not well-adapted to the standard 16 / 8 time-restricted feeding regimen, it will make multiple rounds of iterative adjustments, potentially ultimately forming a more personalized regimen suitable for that group of animals, such as 14 / 10 or segmented feeding. This capability greatly improves the scientific value of the experiment and the level of animal welfare, while reducing the workload of researchers. Through the closed-loop design of intervention strategy generator and iterative optimization of effect evaluation, the system achieves a complete intelligent transformation from data to knowledge, and from monitoring to intervention, marking a significant breakthrough in the precision and intelligence of animal experimental technology.

[0252] In one specific embodiment, the implementation of experimental data management and analysis includes:

[0253] Based on the multimodal monitoring data stream, the execution data of the execution control, and the personalized time-limited feeding intervention strategy, the multi-channel data is traceably stored through a time-series database and a data synchronization mechanism to obtain the stored multi-channel data.

[0254] For the stored multi-channel data, a preprocessed dataset is generated through noise filtering, data standardization, and missing value handling.

[0255] The preprocessed dataset is processed using various data analysis methods to extract experimental results and pattern features; these methods include statistical analysis, machine learning, and time series analysis algorithms.

[0256] Based on the experimental results and pattern characteristics, experimental data management and analysis are achieved through interactive charts and multidimensional data visualization techniques.

[0257] In the actual implementation process, the experimental data acquisition and storage system development ensures the traceable storage of all raw sensor data. A reliable database is used for storage. Synchronous acquisition and display of multi-channel data is supported.

[0258] The data preprocessing and feature extraction module includes noise filtering, data standardization / calibration, missing value handling, and feature extraction.

[0259] The multi-dimensional analysis algorithm library includes statistical analysis, machine learning, time series analysis, behavioral pattern analysis, and energy metabolism calculation. It supports user-defined data grouping and normalization.

[0260] The visualization system provides interactive charts. It supports multi-dimensional data visualization and customized report templates.

[0261] The research report and conclusion generator is developed based on analysis results and visualizations. It utilizes conclusion derivation algorithms, natural language generation technology, and scientific report formatting tools to assist researchers in forming research conclusions. It provides standardized data output formats, potentially compatible with common analysis software or formats. An integrated alarm function allows for remote notifications based on user-defined thresholds.

[0262] The experimental data management adopts a hierarchical storage architecture, combining a time-series database (InfluxDB) and a relational database (PostgreSQL). Raw sensor data is collected at a frequency of 10Hz, filtered in real time, and stored in the time-series database at a frequency of 1Hz, and retained for 30 days; minute-level data after downsampling is retained for 180 days; hourly summary data and experimental event data are stored in the relational database for a long time.

[0263] The data preprocessing process includes: (1) median filtering to remove outliers with a window size of 5; (2) Kalman filtering to smooth the time series data with process noise covariance Q = 0.01 and measurement noise covariance R = 0.1; (3) noise removal based on wavelet transform using Daubechies wavelet with a decomposition level of 4; and (4) missing value imputation using a combination of forward filling and linear interpolation.

[0264] Multidimensional data analysis employs a variety of algorithms: statistical analysis includes descriptive statistics, hypothesis testing (t-test, ANOVA), and nonparametric statistics; time series analysis includes autoregressive models, fast Fourier transform, and wavelet analysis; machine learning methods include random forest (100 trees, depth limit 12), support vector machine (RBF kernel, C=1.0), and deep learning models.

[0265] The visualization technology, based on D3.js and Plotly, enables interactive charts and supports various visualization formats, including time series plots, heatmaps, correlation matrices, behavioral state transition diagrams, and feeding pattern radar charts. It supports multi-dimensional data filtering, zooming, and drill-down, as well as customizable comparative views, facilitating in-depth analysis and hypothesis testing for researchers.

[0266] Furthermore, the experimental data management adopts a hierarchical storage architecture, combining a time-series database (InfluxDB) and a relational database (PostgreSQL). Raw sensor data is collected at a frequency of 10Hz, filtered in real time, and stored in the time-series database at a frequency of 1Hz, and retained for 30 days; minute-level data after downsampling is retained for 180 days; hourly summary data and experimental event data are stored in the relational database long-term.

[0267] The data preprocessing process includes: (1) median filtering to remove outliers with a window size of 5; (2) Kalman filtering to smooth the time series data with process noise covariance Q = 0.01 and measurement noise covariance R = 0.1; (3) noise removal based on wavelet transform using Daubechies wavelet with a decomposition level of 4; and (4) missing value imputation using a combination of forward filling and linear interpolation.

[0268] Multidimensional data analysis employs a variety of algorithms: statistical analysis includes descriptive statistics, hypothesis testing (t-test, ANOVA), and nonparametric statistics; time series analysis includes autoregressive models, fast Fourier transform, and wavelet analysis; machine learning methods include random forest (100 trees, depth limit 12), support vector machine (RBF kernel, C=1.0), and deep learning models.

[0269] The visualization technology, based on D3.js and Plotly, enables interactive charts and supports various visualization formats, including time series plots, heatmaps, correlation matrices, behavioral state transition diagrams, and feeding pattern radar charts. It supports multi-dimensional data filtering, zooming, and drill-down, as well as customizable comparative views, facilitating in-depth analysis and hypothesis testing for researchers.

[0270] Please see Figure 2 and Figure 4 , Figure 2 This is a schematic diagram of the architecture of the intelligent gate control system provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the animal experiment time-limited feeding system provided in an embodiment of the present invention. Figure 2 As shown in the diagram, the architecture of the intelligent access control system includes:

[0271] Feeding box 10 is used to house laboratory animals;

[0272] An electrically controlled gate channel 20 is installed on the feeding box and is used to control the laboratory animals to enter the food area;

[0273] The multimodal sensor array 30 includes a food intake monitoring sensor, an animal activity status monitoring sensor, a physiological parameter monitoring sensor, and an environmental parameter monitoring sensor, used to acquire multimodal monitoring data streams of experimental animal status and behavior;

[0274] The intelligent control unit 40 is connected to the electronically controlled gating channel and the multimodal sensor array, such as... Figure 4 As shown, it includes:

[0275] Data acquisition module 41 is used to acquire multimodal monitoring data streams of experimental animal status and behavior through the multimodal sensor array containing food intake monitoring sensors, animal activity status monitoring sensors, physiological parameter monitoring sensors and environmental parameter monitoring sensors;

[0276] The feeding control module 42 is used to control the execution of the time-limited feeding mode based on the multimodal monitoring data stream and preset time-limited feeding parameters, through a feeding time window control algorithm, a dynamic adjustment algorithm and a safety monitoring mechanism, and to obtain the execution data of the execution control.

[0277] The adaptive adjustment module 43 is used to output a personalized time-limited feeding intervention strategy based on the execution data of the execution control and the multimodal monitoring data stream, through an animal state assessment model, a diet pattern optimization algorithm and an intervention strategy generator.

[0278] The data management module 44 is used to manage and analyze experimental data based on the multimodal monitoring data stream, the execution data of the execution control, and the personalized time-limited feeding intervention strategy, through a data acquisition and storage module, a data processing and analysis module, and a visualization display interface.

[0279] In addition, such as Figure 5As shown in the figure, this embodiment of the invention also provides an animal experimental timed feeding device, including a cage cover 51 and a support frame 52. A food container 53 and a rotating wheel device 54 are disposed at the bottom of the cage cover, and a weight sensor 55 and a temperature, humidity, and pressure sensor 56 are disposed at the top of the cage cover. The rotating wheel device includes functions for voluntary, workload-based, and motor skill-based running, including an enable / disable function. Furthermore, the system's food intake monitoring sensors include a weight sensor and a miniature camera for real-time monitoring of food consumption and filtering out interference signals to obtain food intake data; animal activity status monitoring sensors include an infrared activity detector or an accelerometer to identify and analyze animal movement patterns to obtain activity classification data; physiological parameter monitoring sensors are used to monitor the physiological parameters of experimental animals to obtain physiological parameter data; and environmental parameter monitoring sensors include temperature, humidity, and pressure sensors to collect and process environmental parameters to obtain environmental change data.

[0280] The embodiments described in this invention are merely preferred implementations and are not intended to limit the scope of protection of this invention. All variations, modifications, or equivalent substitutions made within the scope of protection claimed by this invention are included within the scope of protection of this invention.

Claims

1. A method for timed feeding in animal experiments, characterized in that, The method is applied to an intelligent gating system, which includes a feeding box for laboratory animals, an electrically controlled gating channel, a multimodal sensor array, and an intelligent control unit. The multimodal sensor array, which includes sensors for monitoring food intake, animal activity, physiological parameters, and environmental parameters, acquires a multimodal monitoring data stream of experimental animal status and behavior. Based on the multimodal monitoring data stream and preset time-limited feeding parameters, the execution control of the time-limited feeding mode is achieved through a feeding time window control algorithm, a dynamic adjustment algorithm, and a safety monitoring mechanism, resulting in execution control data. The execution control of the time-limited feeding mode includes: determining the feeding time window through a time-triggered mechanism based on the multimodal monitoring data stream and the preset time-limited feeding parameters, and generating a feeding time window switch event table based on state monitoring feedback; controlling the switching state of the electrically controlled gate channel through a state machine model based on the feeding time window switch event table; identifying potential health risks and abnormal behaviors using an abnormal state detection algorithm based on the multimodal monitoring data stream, and triggering an emergency response mechanism when a potential health risk or abnormal behavior is identified; and based on the emergency response mechanism, combined with the feeding time window... The event table of the gate switch and the on / off state of the electronically controlled gate channel are used to generate and prioritize gate execution instructions to achieve execution control of the time-limited eating mode. The step of using an abnormal state detection algorithm to identify potential health risks and abnormal behaviors includes: performing multi-scale processing on the multi-modal monitoring data stream to construct a multi-scale time representation; using a cross-scale attention module to interact with features at different scales of the multi-scale time representation to form multi-time-scale features; calculating health status indicators and their uncertainties using a Bayesian neural network based on the multi-time-scale features to identify abnormal types; wherein the abnormal types include high-confidence abnormalities, high-uncertainty abnormalities, and multi-modal conflicts; and selecting different response strategies based on the abnormal types. Based on the execution data of the execution control and the multimodal monitoring data stream, a personalized time-limited feeding intervention strategy is output through an animal state assessment model, a dietary pattern optimization algorithm, and an intervention strategy generator. Specifically, this includes: constructing an animal state assessment model based on the multimodal monitoring data stream using feature extraction algorithms and machine learning models; constructing dietary pattern impact analysis results based on the animal state assessment model and the execution data of the execution control through time-series data analysis and correlation calculations; generating a personalized dietary pattern optimization scheme based on the dietary pattern impact analysis results and preset research objectives through a multi-objective optimization algorithm and dynamic sparse training; and outputting a personalized time-limited feeding intervention strategy based on the personalized dietary pattern optimization scheme through a rule transformation engine and parameter mapping. Specifically, the multi-objective monitoring data stream... The algorithm for optimizing personalized dietary patterns and dynamic sparse training generates personalized dietary pattern optimization schemes, including: based on the analysis results of the dietary pattern impact, extracting a joint feature representation containing time-related features, physiological response features, and behavioral pattern features to form an embedding representation in a high-dimensional vector space; based on the embedding representation, calculating an importance score matrix through feature importance evaluation to identify key feature dimensions; due to limitations in computing resources, performing feature sparsification on the key feature dimensions, pruning features whose importance is below a preset threshold to obtain pruned key feature dimensions; and performing hierarchical allocation on the pruned key feature dimensions to form a lightweight optimization model; based on the lightweight optimization model and guided by a preset research objective, performing a multi-objective optimization search to generate personalized dietary pattern optimization schemes. Based on the multimodal monitoring data stream, the execution data of the execution control, and the personalized time-limited feeding intervention strategy, experimental data management and analysis are realized through the data acquisition and storage module, the data processing and analysis module, and the visualization display interface.

2. The method according to claim 1, characterized in that, The acquisition of multimodal monitoring data streams of experimental animal status and behavior includes: The food intake monitoring sensor, which includes a weight sensor and a miniature camera, monitors food consumption in real time and filters out interference signals to obtain food intake data. Animal movement patterns are identified and analyzed using the animal activity monitoring sensor containing an infrared activity detector or an accelerometer to obtain activity classification data; Based on the physiological parameter monitoring sensor, the physiological parameters of the experimental animals are monitored to obtain physiological parameter data; The environmental parameter monitoring sensor, which includes temperature, humidity, and air pressure sensors, collects and processes environmental parameters to obtain environmental change data. Based on the food intake data, activity classification data, physiological parameter data, and environmental change data, time alignment and multi-source data fusion are performed to generate the multimodal monitoring data stream.

3. The method according to claim 1, characterized in that, The implementation of experimental data management and analysis includes: Based on the multimodal monitoring data stream, the execution data of the execution control, and the personalized time-limited feeding intervention strategy, the multi-channel data is traceably stored through a time-series database and a data synchronization mechanism to obtain the stored multi-channel data. For the stored multi-channel data, a preprocessed dataset is generated through noise filtering, data standardization, and missing value handling. The preprocessed dataset is processed using various data analysis methods to extract experimental results and pattern features; these methods include statistical analysis, machine learning, and time series analysis algorithms. Based on the experimental results and pattern characteristics, experimental data management and analysis are achieved through interactive charts and multidimensional data visualization techniques.

4. The method according to claim 1, characterized in that, Also includes: Based on the execution results of the personalized time-limited feeding intervention strategy and the multimodal monitoring data stream, a comprehensive effect score is calculated using an effect evaluation algorithm to generate an intervention effect evaluation result. The effect evaluation algorithm includes: calculating the real-time health status score of the experimental animal using the animal status assessment model based on the execution results of the personalized time-limited feeding intervention strategy; comparing the real-time health status score with the health status score before the intervention to calculate the comprehensive effect score. When the intervention effect evaluation result is lower than the preset effect threshold, the feedback adjustment mechanism is activated to conduct multi-angle investigation and generate adjustment suggestions. Based on the aforementioned adjustment recommendations, the animal status assessment model and optimization strategy are updated to achieve iterative optimization of the intervention strategy.

5. The method according to claim 1, characterized in that, Before acquiring the multimodal monitoring data stream of experimental animal status and behavior, the following steps are also included: Based on experimental requirements and animal types, separate rearing spaces and dedicated passage areas were designed, and the basic structure of rearing boxes that meet SPF-level cleanliness standards was constructed. Based on the basic structure of the feeding box, a sliding door gate driven by a micro servo motor and a position sensor are integrated to construct the electronically controlled gate channel; Based on the basic structure of the feeding box and the electronically controlled gate channel, the food intake monitoring sensor, the animal activity status monitoring sensor, the physiological parameter monitoring sensor and the environmental parameter monitoring sensor are deployed to construct the multimodal sensor array; Based on the multimodal sensor array and the electronically controlled gating channel, an embedded processor, a motor drive circuit, and a high-precision analog-to-digital converter module are integrated to construct an intelligent control unit.

6. An animal experimental timed feeding system, characterized in that, include: Feeding enclosures are used to house laboratory animals; An electrically controlled gate is installed on the feeding box to control the entry of laboratory animals into the feeding area; A multimodal sensor array, including food intake monitoring sensors, animal activity status monitoring sensors, physiological parameter monitoring sensors, and environmental parameter monitoring sensors, is used to acquire multimodal monitoring data streams of laboratory animal status and behavior; The intelligent control unit, connected to the electronically controlled gating channel and the multimodal sensor array, includes: The data acquisition module is used to acquire multimodal monitoring data streams of experimental animal status and behavior through the multimodal sensor array containing food intake monitoring sensors, animal activity status monitoring sensors, physiological parameter monitoring sensors and environmental parameter monitoring sensors; The feeding control module is used to implement the execution control of the time-limited feeding mode based on the multimodal monitoring data stream and preset time-limited feeding parameters, through a feeding time window control algorithm, a dynamic adjustment algorithm, and a safety monitoring mechanism, and to obtain the execution control data. The implementation of the time-limited feeding mode execution control includes: determining the feeding time window based on the multimodal monitoring data stream and the preset time-limited feeding parameters through a time triggering mechanism, and generating a feeding time window switching event table based on state monitoring feedback; controlling the switching state of the electronically controlled gate channel based on the feeding time window switching event table using a state machine model; identifying potential health risks and abnormal behaviors using an abnormal state detection algorithm based on the multimodal monitoring data stream, and triggering an emergency response mechanism when a potential health risk or abnormal behavior is identified; and based on the emergency response mechanism, combined with the... The event table for the feeding time window and the on / off status of the electronically controlled gating channel are used to generate and prioritize gating execution instructions to achieve execution control of the time-limited feeding mode. The step of using an abnormal state detection algorithm to identify potential health risks and abnormal behaviors includes: performing multi-scale processing on the multi-modal monitoring data stream to construct a multi-scale time representation; using a cross-scale attention module to interact with features at different scales of the multi-scale time representation to form multi-time-scale features; calculating health status indicators and their uncertainties using a Bayesian neural network based on the multi-time-scale features to identify abnormal types; wherein the abnormal types include high-confidence abnormalities, high-uncertainty abnormalities, and multi-modal conflicts; and selecting different response strategies based on the abnormal types. An adaptive adjustment module is used to output personalized time-limited feeding intervention strategies based on the execution data of the execution control and the multimodal monitoring data stream, through an animal state assessment model, a diet pattern optimization algorithm, and an intervention strategy generator. Specifically, it includes: constructing an animal state assessment model based on the multimodal monitoring data stream using feature extraction algorithms and machine learning models; constructing a diet pattern impact analysis result based on the animal state assessment model and the execution data of the execution control, through time-series data analysis and correlation calculation; generating a personalized diet pattern optimization scheme based on the diet pattern impact analysis result and preset research objectives, through a multi-objective optimization algorithm and dynamic sparse training; and outputting a personalized time-limited feeding intervention strategy based on the personalized diet pattern optimization scheme, through a rule transformation engine and parameter mapping. The method for generating personalized dietary pattern optimization schemes through multi-objective optimization algorithms and dynamic sparse training includes: extracting a joint feature representation containing time-related features, physiological response features, and behavioral pattern features based on the analysis results of the dietary pattern impact, forming an embedded representation in a high-dimensional vector space; calculating an importance score matrix based on the embedded representation to identify key feature dimensions through feature importance evaluation; performing feature sparsification on the key feature dimensions due to computational resource limitations, pruning features whose importance is below a preset threshold to obtain pruned key feature dimensions; and performing computational hierarchy allocation on the pruned key feature dimensions to form a lightweight optimization model; and performing multi-objective optimization search based on the lightweight optimization model and guided by preset research objectives to generate personalized dietary pattern optimization schemes. The data management module is used to manage and analyze experimental data based on the multimodal monitoring data stream, the execution data of the execution control, and the personalized time-limited feeding intervention strategy, through a data acquisition and storage module, a data processing and analysis module, and a visualization interface.

Images