Introduction

The need for intelligent monitoring of biological laboratories

Biological laboratories play a critical role as key venues and support platforms for clinical practice, diagnostics, education, and scientific research. They are essential for diagnosing, tracing, and treating emerging and re-emerging infectious diseases, as well as for research in vaccines, drugs, and testing reagents. However, safety incidents in biological laboratories are not uncommon. A study analyzing data from 150 laboratory safety accidents in China (1986–2019) and 95 typical laboratory accidents (2010–2017) found that human factors are responsible for approximately 40–50% of all accident1. It is widely acknowledged that laboratory accidents primarily result from unsafe human behavior and hazadous conditions of laboratory objects2. Moreover, accidents caused by unsafe human behavior are often the greatest damaging, with the highest mortality rates and significant social consequences3,4. Therefore, minimizing unsafe human behavior and mitigating the impact of human-environment interactions are crucial to enhancing intelligent supervision in biological laboratories5. Current research on biological laboratory supervision systems primarily focuses on three main aspects. First, supervision tends to concentrate on the operational status of equipment, laboratory operating environment, safety protection, and sample storage6,7,8,9. Second, the technological approaches to supervision are predominantly based on Internet of Things (IoT) technology, deep learning, wireless network, and digital twin technologies10,11,12,13,14,15. However, existing research mainly addresses the supervision of static elements such as samples, equipment, and environmental conditions, with limited attention given to the supervision of dynamic elements such as personnel behavior. Furthermore, the research focuses on one or a few individual elements, lacking a comprehensive approach that integrates supervision across personnel, environment, samples, systems, and equipment. Third, while data aggregation within a single laboratory or across multiple laboratories within the same unit is emphasized, secure external data communication has received comparatively less focus. Based on the deficiencies in these three aspects, the design of intelligent supervision systems for biological laboratories has advanced to include supervision of dynamic elements such as personnel behavior, integrated risk elements, and a full-process management platform. These systems ensure secure data transmission both within and outside the unit, enabling full-element coverage, full-process supervision, full-chain traceability, and dynamic real-time monitoring.

The research focuses on the realistic requirements of biological laboratory safety issues, based on the mature theoretical foundation of Internet of Things and deep learning, combined with the effectiveness of object detection algorithms such as YOLOv8 and human gesture recognition models such as OpenPose in various fields, to form the theoretical system and technical architecture of intelligent governance of biological laboratory safety, which is of great significance for preventing and reducing accidents in biological laboratories, ensuring the safety of researchers, and enhancing the ability of predicting, warning and responding to risks in laboratories. It is of great importance for preventing and reducing accidents in biological laboratories, ensuring the safety of researchers, and improving the ability to predict, warn and respond to laboratory risks. It provides new perspectives and methods for solving complex biosafety problems and also provides reliable references and application cases for the future development of intelligent control technology in biological laboratories.

Difficulties of intelligent supervision of biological laboratories

The intelligently supervision of biological laboratories faces several challenges. First, the complexity of element composition presents a significant obstacle. Biological laboratories involve numerous components, including personnel, samples, equipment, and environment, and encompass various processes such as sample reception, storage, usage, disposal, equipment setup, maintenance, and calibration. This wide range of elements increases the scope and complexity of intelligent supervision. Second, the interaction between these elements further complicate supervision. Various element, such as people and samples, people and the environment, samples and the environment, people and equipment, constantly interact, and unforeseen events can arise unexpectedly, making intelligent supervision more difficult. Third, there is a high demand for regulatory timeliness.

Basic requirements for intelligent supervision of biological laboratories

The intelligent supervision of biological laboratories should ensure full-element coverage, full-process supervision, full-chain traceability, and dynamic real-time supervision16. Specifically, it should involve the following key aspects. First, it must cover all elements of the laboratory, including personnel, environment, equipment, systems, and samples, thus enabling comprehensive control over all components—effectively managing “people, machines, materials, as well as laws and regulations”17. Second, it should provide whole-process control, from sample acquisition, transportation, storage, transfer, waste disposal, personnel access, experimental activities, and other behaviors, to the static and dynamic behaviors of accessing personnel, equipment, operation, and calibration18. Third, it should implement a full-chain closed-loop management, ensuring that every aspect of the process—inspection, evaluation, rectification, review, system, and culture—can be traced and controlled. Fourth, dynamic identification and real-time warning of personnel behavior must be achieved. This requires the development and deployment of appropriate facilities and equipment for real-time data collection, encrypted storage, edge analysis, intelligent distribution. Moreover, artificial intelligence algorithms should be employed for accurate identification and analysis of the environment, personnel, and behavioral gestures19,20. Fifth, it should ensure the integration, encryption, and secure distribution of data, guaranteeing the encryption of channels and information during data transmission, retrieval, and response21.

Research design

The research addresses three main problems. Firstly, the regulatory elements are isolated and one-sided, and there is a lack of full-factor governance. Secondly, the security of data transmission is difficult to guarantee due to the absence of encryption. Thirdly, the regulation is static and lagging, which makes it difficult to regulate dynamically and in real time.The primary objectives are to improve the accuracy of personnel posture recognition and address the absence of corresponding algorithms; to resolve issues of laboratory data isolation, dispersal, and low transmission security; and to overcome challenges related to real-time monitoring, delayed alarm processing, and the lack of traceability in disposal actions. Moreover, the goal of this research is to establish comprehensive personnel management, personnel access management, personnel wear standards, personnel operational standards, accident emergencies, other prohibited behaviors, laboratory goods homing, waste disposal, and other behavioral regulations. This integrated approach aims to enhance the digital and intelligent management of biological laboratory safety, thereby significantly improving laboratory safety governance. The research employs a combination of technical methodologies, incorporating the integration of prevailing IoT technologies, the implementation of deep learning algorithms, and the optimisation of the YOLOv8 object detection algorithm and the OpenPose human posture recognition model. Additionally, the research involves the design and optimisation of the network layout and network transmission channel.

Methods

‌Research methodology‌

This study employs a ‌literature review method‌ to analyze the risk factors, key technologies, and current challenges in biosecurity laboratory regulation. Subsequently, ‌factor analysis‌ is applied to identify critical regulatory elements, frameworks, and optimized technical approaches by systematically examining human factors, material factors, and their dynamic interactions. Finally, a ‌case study method‌ is implemented in practical laboratory settings to validate improvements and synthesize a robust technical solution.

Data acquisition and training set

Data is collected through three primary channels. First, via perception terminals installed within the laboratory, including cameras, Radio Frequency Identification(RFID) tags, access control systems, environmental monitoring devices, equipment auxiliary terminals, and intelligent transport boxes. Second, from system-recorded data, which includes self-inspection and audit records, sample transportation and approval logs, sample handover documents, and personnel training files. Third, from subsequent updates, such as equipment calibration records, personnel induction training permit, laboratory system updates, and laboratory permit revisions. Early warning systems are primarily established through threshold settings to define boundaries, complemented by video surveillance and algorithms designed to monitor personnel behavior, posture, qualifications, and attire. These systems are further reinforced by both routine and random inspections to identify, address, and update early warnings.

The deep learning training dataset consists of two components: typical open-source data and data compiled from field collections. The combination of these two datasets enhances the diversity and balance of the data, with the total volume approximating 140,000 entries. The distribution of the dataset is provided in Table 1.

Table 1 Personnel wear and posture deep learning training set.
Full size table

System design

Overall architecture

The system’s architectural design is based on the Internet of Things (IoT) application framework, structured into three distinct layers: the perception layer, the network layer, and the application layer (Fig. 1). In the perception layer, terminals such as cameras, environmental monitoring systems, equipment auxiliary monitoring terminals, access control terminals, and Radio Frequency Identification (RFID) tags are deployed to collect data on personnel behavior, gestures, attire, training, and violations. The system also monitors the status of samples, including storage locations, usage, and transportation. Moreover, it evaluates environmental conditions, such as temperature, humidity, and air pressure against threshold values and tracks the calibration and status of equipment. Additionally, the system facilitates the construction, promotion, and updates of its framework while managing laboratory records, information updates, maintenance, self-inspections, inspections, and corrective actions.

The network layer is composed of two main components. The first is the laboratory’s internal communication and sensing network, which utilizes technologies such as Radio Frequency Identification (RFID), Narrow Band Internet of Things (NB-IoT), Bluetooth mesh, and Modbus to configure nodes within the laboratory. This setup enables the real-time recording of personnel, instruments, equipment, operating objects, and operating times, along with the synchronous summarization of data. The second component involves virtual private network (VPN) technology, secure tunneling, desensitization of sample information, Transport Layer Security (TLS) authentication, and communication content encryption (using the state secret SM2 level) to ensure secure data interaction and communication.To maintain data security, all laboratory data is stored locally.

The application layer incorporates a range of advanced algorithms, including target recognition, personnel behavior recognition, edge computing, voice recognition, natural language processing, and image recognition. These algorithms enable key functions such as personnel behavior recognition, gesture recognition, authority management, sample control, equipment management, and environmental management. The collected data is synthesized into a visuallized dashboard, providing actionable insights to support intelligent decision-making.

Fig. 1
figure 1

Overall architecture of intelligent supervision system for biological laboratories.

Full size image

Regulatory elements and processing flow

The supervision system is designed around the core concept of ‘all elements, all processes, all-round’ supervision (Fig. 2). It addresses potential risk points across all supervisory elements, implements a full closed-loop process for risk management, and ensures comprehensive traceability. The system focuses on five key areas of potential risk: people, machines, materials, laws, and environment. The regulatory framework is structured into three stages: pre-approval management, daily management, and post-incident management. Pre-approval management ensures that the laboratory possesses the necessary capabilities and conditions to conduct routine experimental operations. This includes having personnel of appropriate quality, improved protection for facilities and equipment, compliance with physical space partition requirements, and maintaining an appropriate barometric environment. Daily management emphasizes the control of potential risk factors to ensure that personnel, samples, equipment, the environment, and laboratory culture remain within acceptable safety parameters. Post-incident management concentrates on tracking, processing, evaluating, self-inspecting, and updating early warning events. These elements create a closed-loop system that promotes proactive safety management in laboratories.

The system’s implementation and disposal process incorporates a pre-filing system to assess the laboratory’s readiness, a laboratory evaluation system, a sample transport management system, and an intelligent safety management system. These components collectively manage all aspects of laboratory operations and ultimately integrate data into the intelligent supervision platform for biosafety laboratories. This integration facilitates comprehensive risk management throughout all process and elements. The disposal process categorizes risks based on their magnitude and severity, assigning them to the appropriate person in charge. Designated individuals are required to address the issues within a specified timeframe, while real-time dynamic monitoring ensures the timely resolution of risks and ongoing compliance with safety standards.

Fig. 2
figure 2

Regulatory elements and processing flow.

Full size image

Network architecture

To address the critical issue of data transmission security in biological laboratories, which handle sensitive information, this study employs multiple strategies to mitigate risks associated with protocols such as Modbus (Fig. 3). First, the local laboratory storage equipment is designed with robust data encryption and integrity protection. Features such as partitioned storage, disk redundancy, and full disk encryption are implemented to categorize data into encrypted areas based on its importance. This minimizes data loss due to disk damage and ensures access is restricted to authorized users only. Second, the authentication mechanism is reinforced by following the principle of least privilege, allowing users to access solely to data within their authorized scope. The system incorporates hardware anti-tampering features, anomaly detection, and location tracking capabilities, with alerts for abnormal activities such as unauthorized access, movement, or tampering. Additionally, source authentication technology is utilized to authenticate devices, enabling point-to-point independent communication and ensuring isolation between different networks. Third, network isolation is established through SSL tunneling for secure communication between local storage devices. Each laboratory is equipped with a security gateway for external data transmission, which is conducted primarily via virtual private network (VPN) channels, secure tunneling technology, and the SM2 encryption algorithm. A dedicated encrypted communication network is established within the public network, with all cloud servers and terminals operating within the VPN. The SM2 algorithm encrypts data prior to transmission, with the original information encapsulated within another protocol’s data packet, ensuring secure and reliable data transfer.

Fig. 3
figure 3

Data transmission network architecture.

Full size image

Algorithm support

Target detection algorithm

YOLO (You Only Look Once) is a widely used object detection system in computer vision tasks. For detecting and identifying human behavior, the algorithmic demands heightened accuracy, operational velocity, and overall efficiency. Among all the single-stage network object detection algorithm.

Comparatively, the single-stage network object detection algorithm YOLOv822 exhibits superior precision, efficiency, network architecture, minimal performance demands, broad applicability, and a high degree of maturation. Therefore, YOLOv8 has been selected as the foundational algorithmic framework (Fig. 4). To meet the stringent requirements of speed, accuracy, and effectiveness, further enhancements have been applied to the algorithm. YOLOv8 is composed of three main parts: Backbone, Neck and Head.

Lightweight Multi-scale Convolution Model (LMCM): This model improves detection efficiency by reducing parameter requirements, decreasing computational load, enriching feature information flow, and simplifying the complexityof the model. Fasterblock Module: To bolster detection efficiency and effectiveness without compromising accuracy, the Fasterblock module has been integrated. This module is designed to reduce computation demands, thereby accelerating the algorithm’s operation. GhostConv Module: This plug-and-play module facilitates channel reduction through simple transformations, significantly lowering parameter and computation demands while maintaining model performance. Bottleneck Module with Dilation-wise Residual (DWR), the input feature data is first extracted by a 3 × 3 convolution layer, followed by batch normalisation and rectified linear activation to increase the simplicity of the feature map. The output is a series of feature maps of different sizes.The features are grouped according to their scales, and dilation-based convolutions with different expansion rates are used for feature processing. The fusion steps aggregate the multi-scale features extracted from the input to form a stronger feature representation. This is then fed into an inference network to perform semantic segmentation of the image.SPPF (Spatial Pyramid Pooling Feature) module combined with DLCM (Lightweight Convolution Module), replacing standard convolution in SPPF with DLCM.The SPPF module has been optimized and combined with the DLCM to further minimize parameter and computation needs. This integration enhances the algorithm’s performance and operational speed(Table 2).

These enhancements enable our target detection algorithm to accurately and efficiently monitor human behavior within biological laboratories, significantly contributing to the overall effectiveness of the intelligent supervision system.

Fig. 4
figure 4

Structure of YOLOv8 Optimisation Model.

Full size image
Table 2 Pseudo-code for the target detection algorithm.
Full size table

Human posture detection algorithm

Our human posture detection algorithm is built on the highly accurate OpenPose model, which is pivotal for recognizing human movements and extracting relevant features. The algorithm is underpinned by a multi-stage Convolutional Neural Network (CNN), which progressively extracts and synthesizes features such as image edges, corner points, and joint points through multiple convolutional and pooling layers. Subsequently, it performs a correlation analysis of key points, such as the head, shoulders, elbows, wrists, hips, knees, and ankles, to determine the relationships and construct a comprehensive posture profile.

To achieve faster processing speeds, we have replaced the original VCG-19 network with the MobileNet V3 network (Fig. 5). This modification significantly improves both the accuracy and the operational speed of the algorithm. We have further optimized the architecture by integrating Automated Mobile Neural Architecture Search and Aware Neural Network Adaptation for Mobile Applications, enabling efficient retrieval and selection of modules and network layers.(Table 3).

Additionally, we have integrated the nonlinear Hard Version of the Swish activation function, which delivers notable improvements in accuracy, albeit with a slight increase in latency. By combining these advanced techniques, our human posture detection algorithm ensures precise and efficient analysis of human behavior within the laboratory environment.

Fig. 5
figure 5

Structure of the openpose optimisation model.

Full size image
Table 3 Pseudo-code for the human posture detection algorithm.
Full size table

Algorithm detection process

The detection of personnel wear and behavior primarily relies on video decoding from cameras. The video stream is first hardware-decoded into a sequence of image frames. Since the quality of captured images directly affects the accuracy of machine learning algorithms, it is essential to screen and evaluate these images for inclusion in the machine learning training set. This evaluation is conducted across four key aspects:

Image Clarity: We analyze the image ratio, edge clarity, and frequency domain features. Techniques such as gradient analysis or frequency domain transforms (e.g., Fourier transform) are utilized to evaluate image clarity and determine suitability for behavior recognition algorithm training and recognition. Image Brightness: This involves analyzing the image’s brightness, light uniformity, and shadow conditions. Tools like brightness histograms, light distribution, and shadow coverage area are employed to gauge optimal brightness conditions. Image Noise: The noise condition is assessed by examining the noise distribution, texture characteristics, and signal-to-noise ratio. These indicators help us measure the image quality and the extent of interference. Personnel Posture and Occlusion: We evaluate whether the target’s posture appears natural and whether occlusions are present. These factors are crucial for achieving accurate behavior analysis.

Images that meet the quality criteria are then processed using the YOLOv8 model, which has been trained via deep learning. The detection results include bounding boxes for human bodies and dressed objects (e.g., masks, protective clothing, gloves). However, factors such as occlusion or incorrect posture may lead to false alarms. To address this problem, we employ the OpenPose human posture detection model, which identifies 17 key points on the human body. Using these key points, we calculate the frontal posture angle relative to the camera. If the calculated posture angle falls within plus or minus 30 degrees from the front and no dressed target is detected, an alarm is triggered; otherwise, no alarm is generated (Fig. 6).

Fig. 6
figure 6

Schematic diagram of the algorithm detection process.

Full size image

Laboratory evaluation

The laboratory evaluation index system is based on eight key aspects: organizational management, facilities and equipment, personnel management, management of bacterial (toxic) strains or infectious samples, management of experimental waste, laboratory housekeeping and labeling, fire safety and security confidentiality management, and other daily management. In accordance with legal regulations, standards, and norms, secondary and tertiary sub-indicators are developed. Using the Analytic Hierarchy Process (AHP), 30 experts from various sectors, including medical institutions, disease control, higher education institutions, customs, and enterprises, are invited to provide scores. Subsequently, these evaluations are then processed using Deng’s association degree method to generate overall laboratory risk evaluation results, as well as regional risk assessments.

The correlation is calculated using the formula:

$$\varepsilon _{{0i}} (k) = \frac{{\min _{i} \min _{k} |x_{0} (k) - x_{i} (k)| + \rho \max _{i} \max _{k} |x_{0} (k) - x_{i} (k)|}}{{|x_{0} (k) - x_{i} (k)| + \rho \max _{i} \max _{k} |x_{0} (k) - x_{i} (k)|}}$$

where ξ0i(k) denotes the correlation coefficient between the reference sequence and the i-th comparison sequence at the k-th moment. \(\left| {x_{0} (k) - x_{i} (k)} \right|\)denotes the sequence x0 with the absolute value of xi at point k.

Results and applications

Algorithm accuracy test

The results of the optimization evaluation of the target detection algorithms (Table 4) show that the SPPF model combined with the DLCM models significantly reduces computation and improves inference speed without sacrificing accuracy.

Table 4 Results of the optimisation evaluation of the target detection algorithms.
Full size table

The results of the optimization evaluation of the human posture detection algorithm (Table 5) show that the MobileNet V3 maintains relatively high accuracy compared to VCG-19 under limited computational resources, and is more parameter dense to reduce the amount of computation, making it suitable for laboratories with high computational efficiency requirements and limited resources.

Table 5 Results of the optimisation evaluation of the human posture detection algorithm.
Full size table

The performance of different algorithms (Table 6) indicates the strengths and weaknesses of the models. Specifically, the improved model demonstrates a 1.7% increase in accuracy compared to the benchmark YOLOv8 model, along with a 3.4% improvement in recall sensitivity. Moreover, it achieves a 3% improvement in average precision at a threshold of 0.5, and a 1.74% enhancement in average precision at 0.95. Additionally, the leakage rate is reduced by 0.9%. These results indicate that the improved YOLOv8 model significantly enhances precision and sensitivity while effectively reducing the leakage detection rate, even in complex environments characterized by varying lighting conditions and occlusions.

Table 6 Performance of different algorithms.
Full size table

In real-world application tests, our personnel wear and posture recognition algorithm demonstrated high accuracy (Table 7). The detection rates for protective gear such as clothing, masks, gloves, goggles, and similar items exceeded 90%, with a false alarm rate of approximately 1%. For recognizing various personnel postures and behavioral states, including actions and interactions such as chatting, the detection rate was over 95%, while the false alarm rate remained generally lower.

Table 7 Accuracy of person wear and posture recognition algorithms.
Full size table

However, the detection performance for behaviors related to experimental operations was comparatively lower, with a higher false alarm rate. This can be attributed to the limited sample size and the complexity of the processes involved in these specific operations.

Laboratory evaluation results (Table 8) indicate that experts consider personnel management and bacterial (toxic) species or infectious samples management as the most important factor in laboratory risk control, followed by experimental waste management, organisational management, facilities and equipment. In contrast, areas such as fire safety and security confidentiality management, laboratory housekeeping and signage and marking, and other day-to-day management are regarded as routine standardised management, which accounts for a relatively low percentage to overall risk control.

Table 8 Weights of elements of laboratory evaluation.
Full size table

Laboratory management

As of September 2024, our intelligent supervision technology and system have been successfully deployed in over 2000 laboratories across Zhejiang Province. This initiative includes the pre-approval management of more than 6000 laboratories, over 4000 of which are BSL-2 laboratories. The system manages over 10,000 strains of Class I pathogenic microorganisms, such as the monkeypox and yellow fever viruses, as well as Class II pathogens like HIV and the novel coronavirus. Additionally, it oversees more than 10,000 strains of Class III pathogenic microorganisms, including the Hepatitis B and Hepatitis C viruses, and facilitates the approval of experimental activities.

The sample transport management system has been applied in over 1000 instances for sample transport approvals, identifying more than 100 instances of abnormal information. These cases include undisclosed sample details, deviations in transport routes, incomplete handover records, discrepancies in handover personnel information, and anomalies during unpacking.

The laboratory evaluation management system has supported over 10,000 risk self-inspections and expert inspections, achieving comprehensive self-inspection coverage for more than 6000 laboratories. It has facilitated over 1000 expert inspections, over 100 review and rectification processes, and addressed more than 20,000 risk points. Furthermore, this system has enabled the completion of overall risk management capability assessments for biological laboratories across all 11 prefectural-level cities in Zhejiang Province. The intelligent security box management system has locally stored over 300,000 records, emcompassing personnel access, induction qualifications, sample usage, destruction, equipment utilization and maintenance, and air pressure. It has issued over 10,000 early warning notifications and has managed the processing, rectification, and closed-loop disposal of these warnings more than 3000 times. In terms of personnel posture and behavior identification, including attire and abnormal behavior, our system has accurately identified various postures in over 100,000 instances, achieving a warning accuracy rate of over 93%.

The intelligent supervision platform for biological laboratories has demonstrated its capability to comprehensively address various aspects of laboratory management, including:

Pre-incident Completeness Review: Conducting thorough reviews to ensure that all necessary protocols and preparations are in place before incident occurs. Real-time Supervision and Management: Providing continuous oversight and management during incidents to ensure rapid response and effective handling of emergencies. Post-incident Closed-loop Disposal: Facilitating systematic processing and rectification of issues after incidents to prevent recurrence. Complete Processing Flow: Establishing a seamless processing flow encompassing the push, processing, rectification, and feedback of early warning information within specified timeframes. This integrated approach significantly enhances the intelligent governance capacity for the safety of biological laboratories in Zhejiang Province, ensuring that laboratories operate with higher standards of safety, efficiency and reliability.

Conclusion

The development of intelligent regulatory technology and systems for biological laboratories has effectively addressed challenges associated with complex components, difficult supervision, and stringent time constraints. Driven by practical needs, this study integrates existing components, protocols, and algorithms to innovate and deliver a comprehensive set of feasible technical solutions. It addresses the issues of comprehensive, full-process, and dynamic supervision of laboratories, alongside ensuring the security of data transmission. These solutions have been validated through practical implementation, significantly improving the efficiency and quality of laboratory safety management. This advancement promotes the intelligent development of biological laboratories and provides valuable insights and guidance for future research, offering significant reference value. The key innovations are as follows:

Closed-loop process development: A comprehensive closed-loop process has been established, covering pre-approval, in-process management, and post-disposal stages. This system ensures that laboratories are equipped with the necessary capabilities and conditions to operate safely, effectively supervise operational status, and address early warnings in a timely manner.

Intelligent technology for sensitive issues: Intelligent technology has been leveraged to address critical issues such as information security. The development of an intelligent security box has ensured local information collection and storage security. A dedicated network, encryption algorithms, and secure tunnels have been implemented to safeguard the transmission of data. Additionally, advanced personnel posture recognition algorithms have been developed to monitor and mitigate the highest risks associated with human behavior, enhancing safety protocols within the laboratory environment.

Clarification of responsibilities and management: Intelligent methods have been used to define primary responsibilities and implement tiered and classified management. Risk classification facilitates targeted risk treatment and tracking. A combination of self-inspection and external inspection encourages laboratories to adopt standardized and systematic construction. This approach prompts laboratories to initiate self-inspection and self-correction, preventing risks and addressing issues promptly.

Looking ahead, the future of intelligent regulation in biosafety laboratories will focus on three key areas: People-Oriented Innovation: The core concept driving innovation in regulatory technology will shift from traditional, mandatory standardization to a more proactive and dynamic approach. Technologies such as digital twins, virtual reality (VR), and augmented reality (AR) will be leveraged to create virtual network labs, enhancing personnel experience, well-being, and sense of belonging. This shift will strengthen virtual interactions and embed safety culture deeply within the emotional fabric of the workforce. Continuous Iteration of Algorithms and Technology: Intelligent supervision technology in biological laboratories will continue to evolve, incorporating advanced capabilities such as visual analysis, automated event response, intelligent prediction and early warning, and decision-making support. The integration of cloud computing and Internet of Things (IoT) applications will optimize algorithms, further improving the accuracy, efficiency, and intelligence of supervision. Standardization and Governance of Data: Future innovations in regulatory technology will require standardized data elements, consistent collection methods, and more advanced metadata mining applications. The breadth of data collection, storage, integration, analysis, and application trhoughout the entire management process will be more extensive, ensuring higher standards of governance and enhanced regulatory oversight.