A method for personalized risk self-adaptive management and control of labor dispatch personnel

Abstract

This invention relates to the field of intelligent safety management and control technology, specifically to a personalized risk adaptive management and control method for dispatched workers. The method includes: training a meta-learning model in the cloud based on historical full-volume personnel data to obtain model initialization parameters that can quickly adapt to new individuals; when the target worker performs a task, using these parameters to initialize a personalized risk prediction model at its associated edge nodes, and fine-tuning it based on its initial data; collecting multi-dimensional behavioral data of the worker in real time and inputting it into its personalized model to calculate a dynamic risk quantification value; based on this risk value, generating and executing personalized risk intervention instructions for the worker through a unary policy network. This invention rapidly builds and continuously optimizes a unique risk model for each dispatched worker using meta-learning technology, realizing a paradigm shift from generalized management and control to personalized adaptive monitoring, significantly improving the accuracy and effectiveness of safety management and control.

Description

Technical Field

[0001] This invention relates to the field of intelligent safety management and control technology, specifically to a personalized risk adaptive management and control method for dispatched workers. Background Technology

[0002] In the labor dispatch industry, especially in high-risk areas such as logistics and distribution, construction, and factory production, effectively preventing and managing workplace safety accidents among dispatched workers remains a serious challenge. Currently, mainstream intelligent safety management solutions generally employ the method of constructing a universal risk assessment model. This method integrates data from different sources (such as positioning devices, environmental sensors, and task systems) to train a unified risk prediction model applicable to all personnel, and sets fixed risk thresholds to trigger standardized early warning or intervention processes.

[0003] However, this "one-size-fits-all" general model approach has fundamental technical flaws when dealing with the high complexity and individual differences in labor dispatch scenarios. First, labor dispatch workers are highly mobile, with varying skill levels and experience, and vastly different physiological and psychological states and behavioral habits. A general model trained on common group data struggles to accurately characterize and assess the unique risk patterns of specific individuals, leading to significant biases in risk assessment, or insensitivity to high-risk individuals and over-warning of low-risk individuals. Second, newly hired or transferred employees face a severe "cold start" problem due to a lack of historical behavioral data. The general model cannot provide effective risk monitoring in the early stages of their employment, creating a safety monitoring vacuum. Third, existing intervention measures are usually pre-set and standardized, unable to generate customized safety prompts or training content based on individual risk causes, tolerance, and real-time status, resulting in limited intervention effectiveness and low employee acceptance. Furthermore, once deployed, the parameters and rules of the general model are relatively fixed, making it difficult to adaptively track the dynamic evolution of individual risk patterns as skills improve, fatigue accumulates, or job positions change, causing model effectiveness to decline over time.

[0004] Therefore, existing technologies have failed to resolve the core contradiction between "large individual differences" and "model generalization" in the safety management of dispatched workers. There is an urgent need for a new technical solution that can quickly build and continuously evolve a unique risk perception model for each individual, thereby achieving truly personalized and adaptive safety management.

[0005] Therefore, existing technologies still need further development. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a personalized risk adaptive management method for dispatched workers to solve the problems existing in the prior art.

[0007] To achieve the above technical objectives, this invention provides a personalized risk adaptive management method for dispatched workers, comprising: S1. Based on the historical work behavior data and corresponding risk tags of all dispatched workers, a meta-learning model is trained on a cloud server. The meta-learning model is used to obtain model initialization parameters that can quickly adapt to new individuals. S2. In response to the target dispatched worker starting to perform the dispatch task, based on the initialization parameters output by the meta-learning model, a lightweight personalized risk prediction model is initialized on the edge computing node associated with the target worker. S3. Collect multi-dimensional behavioral data of the target personnel in real time during the current task execution process, and input it into the personalized risk prediction model to calculate its real-time dynamic risk quantification value; S4. Based on the real-time dynamic risk quantification value, generate and execute personalized risk intervention instructions for the target personnel through a meta-policy network.

[0008] Specifically, the process of training a meta-learning model on a cloud server based on historical data of all dispatched workers' work behavior and corresponding risk tags includes: From the historical databases of multiple dispatch projects, time-series behavioral data fragments and corresponding risk assessment results of different dispatched workers under various work tasks are extracted to form a meta-training task set. Each meta-training task simulates the learning process of a new person or a new task scenario. The meta-learning model is trained using a model-independent meta-learning algorithm framework. The training objective is to enable the meta-learning model to make accurate risk predictions for samples in the query set of a task after only a small amount of gradient updates based on a small amount of support set data in a single meta-training task. The training process enables the meta-learning model to learn transferable risk pattern meta-knowledge across personnel and tasks and encode it in its model parameters. The model parameters are the initialization parameters of the model that can quickly adapt to new individuals.

[0009] Specifically, the training process of the model-independent meta-learning algorithm framework includes: In each training iteration, a batch of tasks is sampled from the meta-training task set; for each sampled task, its data is divided into a support set and a query set; the current parameters of the meta-learning model are used as initial parameters, and one or more forward and backward propagations are performed on the support set data to calculate the gradient and update the model parameters, thereby obtaining a model adapted for the sampled task; the adapted model is used to perform forward propagation on the query set data to calculate the query set loss; based on the losses of all sampled tasks on their respective query sets, the meta-gradient is calculated, and backpropagation is performed to update the initial parameters of the meta-learning model; the above process is repeated until convergence, thereby obtaining a set of meta-learning model parameters that can be quickly adapted to generalize to new people or new tasks.

[0010] Specifically, in response to the target dispatched worker starting to perform a dispatch task, based on the initialization parameters output by the meta-learning model, a lightweight personalized risk prediction model is initialized on the edge computing node associated with the target worker, including: When a target person is dispatched to the employing unit and is ready to perform a task, the cloud server sends the latest parameters of the meta-learning model to the edge computing nodes deployed on the local server of the employing unit or on the smart terminal held by the target person. On the edge computing nodes, the parameters are loaded to initialize the personalized risk prediction model. The personalized risk prediction model has the same network skeleton structure as the meta-learning model, but the number of parameters has been compressed and optimized. Using the initial behavioral data generated by the target person during the task preparation stage or at the beginning of the task, the initialized personalized risk prediction model is fine-tuned on the edge computing nodes with a small number of samples to quickly match the individual behavioral characteristics of the person and complete the cold start of the model.

[0011] Specifically, the real-time collected multidimensional behavioral data includes physiological and motion data collected by wearable devices or work vehicle sensors, real-time task attributes and progress pressure data provided by the task management system, and micro-operational environment data collected by environmental sensing devices; the personalized risk prediction model outputs a real-time dynamic risk quantification value that comprehensively reflects the probability of an accident caused by factors such as personal habits, skill level, and physiological and psychological fluctuations in the current state of the person based on the multidimensional behavioral data.

[0012] Specifically, the step of generating and executing personalized risk intervention instructions for the target personnel based on the real-time dynamic risk quantification value through a meta-policy network includes: The meta-policy network takes as input the real-time dynamic risk quantification value output by the personalized risk prediction model of the target person and the current risk causal feature vector extracted from the multi-dimensional behavioral data. Based on the input, the meta-policy network decides on one or more of the most suitable combinations of intervention measures from a variety of preset types. The types of intervention measures include, but are not limited to: pushing personalized operation guidance and risk warnings through augmented reality devices or voice devices, pushing customized safety training micro-courses through mobile applications, and sending requests to the scheduling system to adjust the task pace or suggest rest. The decision-making process incorporates feedback data on the person's historical intervention effects and continuously optimizes the decision-making strategy of the meta-policy network through an online meta-reinforcement learning mechanism, so that the generated intervention instructions match the person's risk tolerance, learning style, and current state in terms of content and form.

[0013] Specifically, the online meta-reinforcement learning mechanism continuously optimizes the decision-making strategy of the meta-policy network, including: The real-time dynamic risk quantification value change trend and safety behavior compliance data of the target personnel over a period of time after each intervention instruction are executed are used as environmental reward signals for reinforcement learning. Locally on the edge computing nodes, the reward signals are used to update the policy gradient of the meta-policy network to optimize the intervention strategy for the specific personnel. Periodically, the parameter update gradients of the meta-policy network on each edge computing node are encrypted and uploaded to the cloud server. The cloud server aggregates all uploaded gradients and performs a primary meta-update on the main model of the meta-policy network deployed in the cloud, enabling the main model to learn the policy initialization capability to quickly adapt to different personnel feedback patterns. The updated main model parameters can then be redistributed to each edge node.

[0014] Specifically, the method operates on a cloud-edge collaborative federated learning architecture, wherein: The training of the meta-learning model and the meta-update of the meta-policy network are performed on the cloud server, aggregating desensitized knowledge from multiple edge computing nodes without touching the original personal data; the fine-tuning of the personalized risk prediction model, the calculation of the real-time dynamic risk quantification value, and the generation and initial execution of the personalized risk intervention instructions are all completed locally on the edge computing nodes associated with the specific dispatched workers; the edge computing nodes and the cloud server only exchange model parameters, gradients, or aggregated knowledge representations, realizing personalized model evolution under privacy protection.

[0015] Specifically, this also includes the personalized model evolution steps: During the performance of tasks by the target dispatched workers, the behavioral data and risk quantification results generated by them are continuously used as new training samples and stored in the local buffer of the edge computing node. When the number of samples accumulated in the local buffer reaches a certain number or the triggering condition is met, the personalized risk prediction model is incrementally learned or periodically recalibrated on the edge computing node using these new samples. The incremental learning or recalibration process enables the personalized risk prediction model to track and adapt to the evolution of the risk pattern caused by the worker's skill improvement, habit change or state fluctuation.

[0016] Specifically, the method also includes a cross-job adaptive step: When the target dispatched workers are dispatched to perform new job tasks that are different from their previous experience, the system receives the task type identifier of the new job; based on the task type identifier, it obtains the meta-knowledge sub-module related to the task type from the cloud server, or adjusts the feature extraction weights related to the task context in the personalized risk prediction model; combined with a small amount of behavioral data in the early stage of the new job, it quickly adjusts the risk assessment focus of the personalized risk prediction model to adapt it to the risk characteristics of the new job.

[0017] Beneficial effects: The personalized risk adaptive management method and system for dispatched workers provided by this invention, by introducing a meta-learning framework, realizes a fundamental shift in the safety management paradigm from "a general model to address all" to "meta-knowledge to generate a personalized model", bringing about many significant beneficial effects.

[0018] First, this invention effectively overcomes the "cold start" challenge in risk management for new employees in the labor dispatch industry. By pre-extracting meta-knowledge about "how to quickly learn an individual's risk characteristics" from massive amounts of historical data through a cloud-based meta-learning model, the system can quickly fine-tune a personalized risk prediction model for a new employee at the edge when they start work, using only a small amount of initial data. This allows personalized safety monitoring to take effect immediately upon joining the company, significantly shortening the window period of no or weak monitoring and improving the overall safety baseline.

[0019] Secondly, this invention significantly improves the accuracy and contextual relevance of individual risk assessment. Each individual's unique personalized model deeply integrates micro-data such as physiological characteristics, behavioral habits, and skill levels, resulting in a more realistic risk quantification value that reflects the actual risk probability of the "individual" in the "current" specific task and environment, rather than the average risk of the group. This enables the system to more accurately identify potential risks caused by individual-specific factors (such as poor operating habits, latent fatigue, etc.), achieving precise early warning.

[0020] Furthermore, this invention achieves a high degree of customization and adaptive optimization of safety intervention measures. Based on a meta-reinforcement learning-based policy generation network, it can decide and generate intervention instructions (such as specific content scripts or customized micro-training) that best suit the individual's current state and cognitive characteristics, based on the risk value and causes output by the personalized model. More importantly, this strategy can continuously optimize the intervention strategy for the individual through feedback from online learners on each intervention (whether behavior improves or not), making the intervention methods "more accurate with use," significantly improving compliance with safety warnings and the effectiveness of training, and promoting behavioral improvement from the source.

[0021] Finally, this invention constructs a sustainable and evolving safety management ecosystem. Personalized models continuously learn incrementally at the edge using local data, tracking and adapting to changes in risk patterns caused by increased proficiency, fluctuating performance, or job changes, maintaining consistent assessment accuracy. Simultaneously, through a federated learning architecture, the personalized learning experiences of each edge node converge in the cloud while protecting privacy, feeding back into the evolving meta-model and meta-policy. This enables the system as a whole to possess cross-personnel and cross-task knowledge transfer and sharing capabilities, becoming increasingly intelligent with use. This provides strong technical support for addressing the dynamic management needs of the labor dispatch industry, characterized by frequent personnel turnover and diverse operational scenarios. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the personalized risk adaptive management method for dispatched workers provided in a specific embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0024] First, it's important to clarify that this method constructs a two-tiered architecture of "meta-knowledge-driven personalization." At the top level, a cloud-based meta-learning model doesn't directly predict risk; instead, it learns how to quickly build a personalized risk prediction model for any new employee. At the bottom level, each dispatched worker is associated with a lightweight, personalized model running at the edge. This model is initialized by the cloud-based meta-model and fine-tuned using the worker's local real-time data to achieve truly individualized risk assessment. The assessment results are input into a meta-policy network with similar adaptive capabilities to generate highly customized intervention measures. The entire process forms a closed loop of "learning commonalities in the cloud, practicing individuality at the edge, and seeking precision in intervention."

[0025] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0026] Please see Figure 1 This invention provides a personalized risk adaptive management method for dispatched workers, comprising: S1. Based on the historical work behavior data of all dispatched workers and their corresponding risk tags, a meta-learning model is trained on a cloud server. The meta-learning model is used to obtain model initialization parameters that can quickly adapt to new individuals.

[0027] Specifically, step S1 aims to address the problem that traditional general models cannot account for individual differences. It builds a meta-learning model in the cloud with the ability to "quickly generalize from one example to another," providing an optimal starting point for personalized models for each individual. To achieve this goal, the first step is to deeply clean and reconstruct the features of the entire historical data of the industry. The historical data on the work behavior of dispatched workers covers multiple dimensions, including basic attributes, physiological characteristics, operating habits, and environmental interactions. Corresponding risk tags accurately record whether accidents ultimately occurred under specific behaviors and environments. In the cloud server, a model-independent meta-learning framework is used for training. Specifically, the historical data is divided into multiple independent "tasks," each corresponding to a specific job type or personnel group. Within each task's internal loop, model parameters are quickly learned and updated from a small number of samples to minimize the prediction error within that task. In the external loop, the learning results from multiple tasks are combined to update the model's initialization parameters. This process iterates until the model parameters converge. The final meta-learning model parameters are the initialization parameters that can quickly fit a specific behavioral pattern using very little new individual data.

[0028] It's important to further explain that the core of step S1 lies in learning the general patterns of "how to learn a person" through massive amounts of historical data. In practice, the raw data is first preprocessed, including missing value imputation, outlier removal, and data standardization. Then, a deep neural network model structure is constructed, which includes a feature extraction layer, hidden layers, and an output layer. During training, appropriate inner and outer loop learning rates are set, and through multiple rounds of iterative optimization, the model can adapt to new personnel data with minimal computational resources and time. The training of the meta-learning model is an offline process. Once training is complete, its initialization parameters are fixed and distributed to various edge nodes as the basis for subsequent personalized model construction. This process not only significantly reduces the computational pressure on the cloud but also lays an efficient foundation for subsequent real-time risk management.

[0029] Understandably, step S1, through a meta-learning framework, successfully transforms the problem of building a general model into the problem of building an "initialization parameter generator." This innovation allows the system to move away from relying on massive real-time computing resources to process data for each new individual. Instead, it leverages pre-trained meta-knowledge, giving the system a strong cold-start capability. When a new dispatched worker joins the company, the system doesn't need to train the model from scratch. Instead, it directly uses the initialization parameters obtained in step S1 to quickly generate a personalized model prototype for that individual. This not only significantly shortens the time required to integrate new personnel into the management system but also significantly reduces the load on cloud servers. This allows the entire system to maintain efficient and stable operation even when dealing with the high-frequency personnel turnover characteristic of the industry, providing solid underlying support for subsequent accurate risk assessment.

[0030] S2. In response to the target dispatched worker starting to perform the dispatch task, a lightweight personalized risk prediction model is initialized on the edge computing node associated with the target worker based on the initialization parameters output by the meta-learning model.

[0031] Specifically, step S2 is a crucial step in implementing the meta-knowledge trained in the cloud onto the specific execution unit. It aims to leverage the initialization parameters provided by the meta-learning model to quickly build a lightweight risk prediction model tailored to the characteristics of the target personnel at the edge. When the target dispatched worker officially begins performing their dispatch task, the system automatically triggers the model initialization process. First, the cloud server distributes the latest initialization parameters of the meta-learning model trained in step S1 to the edge computing nodes associated with the target personnel. These edge computing nodes are typically deployed on the employer's local server or in the target personnel's personal smart terminal devices, such as the computing chips embedded in industrial-grade smartphones, smart safety helmets, or dedicated AR glasses. After receiving the initialization parameters, the edge computing nodes load these parameters and, based on them, instantiate a lightweight, personalized risk prediction model with the same structure but compressed and optimized parameter count. To ensure efficient operation on resource-constrained mobile devices, this personalized model typically employs an architecture such as a deep separable convolutional neural network or a lightweight multilayer perceptron.

[0032] It's important to further explain the crucial "lightweight" design in step S2. Considering that dispatched workers are typically in dynamic work environments, and that smart terminals have limited computing power and battery capacity, the model must minimize the number of parameters and computational load while maintaining prediction accuracy. Therefore, during the model structure design phase, model compression techniques such as channel pruning and weight quantization are employed to slim down the original deep neural network. Furthermore, to ensure the model can quickly adapt to new individuals, step S2 goes beyond simply loading initialization parameters. It also incorporates initial behavioral data generated by the target personnel during task preparation or early stages to fine-tune the initialized personalized risk prediction model with a small number of samples. This process is typically completed locally on edge nodes, without uploading to the cloud, protecting personnel privacy and enabling rapid model cold start. Through this cloud-edge collaboration, the system can provide personnel with a tailored, highly efficient risk prediction model the moment they start work.

[0033] Understandably, step S2 achieves "last mile" coverage of risk management capabilities by combining the initialization parameters of the meta-learning model with the localized deployment of edge computing nodes. Traditional cloud computing models, limited by network bandwidth and latency, struggle to meet the stringent requirements of real-time risk prediction, while the introduction of edge computing perfectly solves this problem. The moment a target person begins work, the system can activate a dedicated risk prediction model on the device closest to them. This model not only inherits the powerful generalization ability of the meta-learning model but also possesses extremely high execution efficiency and personalization through lightweight design and local fine-tuning. This enables the system to respond to real-time behavioral changes of users within milliseconds, providing timely and accurate data sources for the dynamic risk assessment in the subsequent step S3, truly achieving "wherever people are, management follows," greatly improving the immediacy and effectiveness of security management.

[0034] S3. Collect multi-dimensional behavioral data of the target personnel in real time during the current task execution process, and input it into the personalized risk prediction model to calculate its real-time dynamic risk quantification value.

[0035] Specifically, step S3 is the core of the entire risk management system's perception and computation, aiming to accurately capture the dynamic risk status of target personnel during operations through high-frequency, multi-dimensional data collection and real-time inference. During the execution of dispatch tasks, the system continuously collects multi-dimensional behavioral data through the target personnel's personal or wearable smart terminal devices. These data sources are diverse, including but not limited to: limb movement amplitude, frequency, and posture data collected by inertial measurement units; physiological data such as heart rate and blood oxygen saturation collected by photoelectric volumetric sensors; real-time location, movement trajectory, and speed data collected by GPS and ultra-wideband positioning modules; operational environment data such as temperature, humidity, light intensity, and dust concentration collected by environmental sensors; and task attribute data such as current task type, workload progress, and deadline synchronized from the task management system. All these heterogeneous data are timestamped and standardized to form a continuous, multi-dimensional real-time data stream. This data stream is then input into the personalized risk prediction model initialized at the edge nodes in step S2.

[0036] It's important to further clarify that the personalized risk prediction model in step S3 is not static but a dynamic model with online learning and adaptive capabilities. As the target person's behavior continues, the model constantly receives new input data and performs calculations based on its internal weight parameters. The specific calculation process is as follows: the input layer receives multi-dimensional behavioral data, which undergoes nonlinear transformation and feature extraction in the hidden layer, ultimately generating a scalar value between 0 and 1 in the output layer—the real-time dynamic risk quantification value. This quantification value is not a simple binary judgment (safe or dangerous), but a continuous and refined risk probability indicator that intuitively reflects the target person's overall risk level at the current moment and in the current state. To further improve the model's accuracy and robustness, the system also introduces an attention mechanism, enabling the model to automatically focus on the key feature dimensions that contribute the most to the current risk. For example, in climbing operations, the model will pay more attention to posture and heart rate data; while in driving operations, it will pay more attention to speed and acceleration data.

[0037] Understandably, step S3, by constructing such a high-fidelity, multi-dimensional, real-time risk perception closed loop, completely breaks through the limitations of "hindsight" or "experience-based prediction" in traditional safety management. The system no longer relies on fixed, crude safety checklists or preset alarm thresholds, but instead dynamically assesses the risk status of individual personnel through deep learning and real-time analysis of their individual behavioral patterns. This refined assessment based on individual characteristics enables the system to keenly detect subtle anomalies easily overlooked by manual supervision, such as a slight decrease in operational precision under fatigue or a delayed response to potential dangers under distraction. The output of real-time dynamic risk quantification values ​​provides a scientific and decision-making basis for the subsequent intelligent intervention in step S4, transforming safety management from "passive response" to "proactive prevention," greatly improving the timeliness and accuracy of risk detection, and building a solid digital defense line for the life safety of dispatched workers.

[0038] S4. Based on the real-time dynamic risk quantification value, generate and execute personalized risk intervention instructions for the target personnel through a meta-policy network.

[0039] Specifically, step S4 is the final execution stage of the risk management system, moving from "perception" to "action." It aims to intelligently decide and generate the most appropriate personalized risk intervention instructions based on the real-time dynamic risk quantification value calculated in step S3, thereby proactively guiding target personnel to avoid dangerous behaviors. When the personalized risk prediction model outputs the current real-time dynamic risk quantification value, this value, along with the risk causal feature vector extracted from multi-dimensional behavioral data and the current task stage context information, is fed as input to the meta-policy network deployed on edge computing nodes. The meta-policy network is a specially trained reinforcement learning model whose core task is to learn which intervention measures can achieve the best safety management effect under specific risk states and personnel characteristics. In practice, the meta-policy network maintains a vast action space containing various types of intervention measures, such as: playing specific voice prompts through the speech synthesis module of a smart terminal; overlaying visual operation guidance or highlighting dangerous areas through augmented reality devices; sending requests to the task scheduling system to adjust the task pace, suggest short breaks, or assign auxiliary personnel; and pushing customized safety knowledge micro-training videos or graphic materials.

[0040] It's important to further clarify that the meta-policy network in step S4 doesn't simply issue mechanical alerts based on risk levels; instead, it possesses a high degree of intelligence and personalization. It comprehensively considers historical intervention feedback data for the target personnel, including their response speed to different types of prompts, compliance, and the actual decrease in risk level after intervention. Through an online meta-reinforcement learning mechanism, the meta-policy network continuously optimizes its internal decision-making strategies. For example, if historical data shows that an employee is slow to respond to voice prompts but highly sensitive to visual cues, the meta-policy network will prioritize generating visual intervention instructions when encountering that employee in a high-risk state later. Furthermore, the meta-policy network dynamically adjusts the intensity and method of intervention based on the urgency of the risk. For low-risk states, it might only generate a mild reminder; while for high-risk states, it will immediately trigger a powerful alarm and simultaneously notify the human safety officer at the remote monitoring center for intervention. The entire decision-making and generation process is completed locally at the edge node, ensuring extremely low latency and high execution efficiency.

[0041] Understandably, the S4 process, by introducing a meta-policy network, successfully upgrades traditional safety management from a standardized, one-size-fits-all response to personalized guidance tailored to each individual. This not only significantly improves the acceptance and effectiveness of safety alerts and interventions but also dramatically enhances the work experience of target personnel, avoiding "alarm fatigue" caused by frequent and inappropriate alarms. The system is no longer a cold, impersonal monitor but a smart safety assistant that understands employees and provides tailored instruction. Through continuous online learning and policy optimization, the meta-policy network becomes increasingly intelligent with use, creating safety management solutions that best suit the psychological characteristics and behavioral habits of each dispatched worker. This not only effectively curbs accidents at their source but also promotes the dual improvement of employees' safety awareness and operational skills, creating a safer, more harmonious, and efficient production environment for employers, demonstrating significant practical application value and social benefits.

[0042] Understandably, this invention fundamentally changes the traditional, rudimentary approach of "one model for everyone" in safety management systems. By introducing a meta-learning framework, the system gains the ability to quickly build personalized risk assessments for each new hire or reassigned employee, effectively solving the "cold start" problem of models in the labor dispatch industry caused by high employee turnover and a lack of historical data. Simultaneously, the personalized model iterates using local data at the edge, protecting employee privacy while achieving precise capture and dynamic adaptation to individual behavioral habits, skill levels, and status fluctuations. Ultimately, interventions based on personalized risk assessments can be tailored to specific needs, significantly improving the effectiveness and acceptance of safety warnings and training, and preventing safety accidents caused by individual differences at the source.

[0043] Specifically, the process of training a meta-learning model on a cloud server based on historical data of all dispatched workers' work behavior and corresponding risk tags includes: From the historical databases of multiple dispatch projects, time-series behavioral data fragments and corresponding risk assessment results of different dispatched workers under various work tasks are extracted to form a meta-training task set. Each meta-training task simulates the learning process of a new person or a new task scenario. The meta-learning model is trained using a model-independent meta-learning algorithm framework. The training objective is to enable the meta-learning model to make accurate risk predictions for samples in the query set of a task after only a small amount of gradient updates based on a small amount of support set data in a single meta-training task. The training process enables the meta-learning model to learn transferable risk pattern meta-knowledge across personnel and tasks and encode it in its model parameters. The model parameters are the initialization parameters of the model that can quickly adapt to new individuals.

[0044] It's important to further explain that this step is the core of building the meta-model. First, constructing the meta-training task set is crucial. Data sources include historical databases of labor dispatch projects from multiple industries (such as construction, logistics, and factories). For each historical worker, their continuous work period (e.g., a shift) is divided into multiple data segments of fixed duration (e.g., 5 minutes). Each data segment contains multi-dimensional time-series behavioral data for that moment and corresponds to a risk label (e.g., "safe," "low risk," "high risk") determined by post-event expert judgment or incident records. When constructing the meta-task, all data segments of a historical worker under a specific task are randomly selected and randomly divided into a support set and a query set, thus forming a meta-task. It supports simulating a small amount of initial data (e.g., 1-3 segments) available from new personnel, and query sets simulating subsequent data that the model needs to predict. By constructing task sets through extensive sampling of data from different personnel and different tasks, it forces the meta-learning model to learn to capture risk patterns from a small amount of data from anyone.

[0045] Secondly, the training employs the Model-Independent Meta-Learning (MAML) algorithm. Let the parameters of the meta-learning model be... The preferred inner network structure is a three-layer fully connected neural network, with the input layer dimension matching the feature dimension (e.g., 128 dimensions), the hidden layer dimension being 64, and the output layer representing the risk probability. During training, each iteration samples a batch of tasks from the task set (batch size...). The preferred value is 4, which balances the estimation variance of the meta-gradient with computational efficiency. For each sampling task... Using its supporting set data to calculate the loss, and for Perform a gradient descent operation to obtain task-specific parameters. ,in It is a task Loss functions on the support set (such as cross-entropy loss). It is a meta-model. The inner loop learning rate is preferably 0.01 to ensure a stable adaptation process. Then, using... Calculate the loss on the query set. Ultimately, the goal of meta-optimization is to minimize the sum of the losses of all sampling tasks on their respective query sets: By optimizing this objective, It eventually converges to a set of "sensitive" initialization points. Starting from these points, good performance can be achieved for any new task with only a small number of gradient steps.

[0046] Understandably, the parameters of the meta-learning model trained in the above manner encode meta-knowledge about "how to quickly identify a person's risk characteristics from a small amount of behavioral data." This differs from training a general risk classifier. When faced with a completely new dispatched worker with no historical data, the system does not need to collect massive amounts of data from scratch to train the model. It only needs to use a small amount of behavioral data from the worker's initial period on the job (such as the first hour) as the support set, and then use the meta-model parameters... By initializing and performing the aforementioned inner loop adaptation 1-2 times, a personalized risk prediction model with a certain degree of accuracy can be obtained for that individual, greatly alleviating the data cold start problem. Furthermore, since the meta-training task covers a variety of operational scenarios, this meta-knowledge also possesses the potential for cross-task generalization.

[0047] Specifically, the training process of the model-independent meta-learning algorithm framework includes: In each training iteration, a batch of tasks is sampled from the meta-training task set; for each sampled task, its data is divided into a support set and a query set; the current parameters of the meta-learning model are used as initial parameters, and one or more forward and backward propagations are performed on the support set data to calculate the gradient and update the model parameters, thereby obtaining a model adapted for the sampled task; the adapted model is used to perform forward propagation on the query set data to calculate the query set loss; based on the losses of all sampled tasks on their respective query sets, the meta-gradient is calculated, and backpropagation is performed to update the initial parameters of the meta-learning model; the above process is repeated until convergence, thereby obtaining a set of meta-learning model parameters that can be quickly adapted to generalize to new people or new tasks.

[0048] It should be further explained that the specific steps of the MAML training process are as follows, which is crucial for obtaining the meta-model parameters: (a) Task Sampling: Let the distribution of the meta-training tasks be... In each training iteration, from Random sampling Each task constitutes a meta-batch . The optimal value is 4. Setting it too high (e.g., 16) will result in high computational cost per iteration and slow convergence, while setting it too low (e.g., 1) will lead to excessively large variance in the meta-gradient estimation and unstable training. (Each task...) Pre-divided into support sets and query set Supported set size Ideally, there should be 5 samples (i.e. 5 data fragments) to simulate the very small amount of data that new personnel can provide; the query set size includes the remaining samples for the task.

[0049] (b) Inner loop adaptation: for each sampling task Perform inner loop adaptation. Copy the current metamodel parameters. get In the mission Support set Calculate the loss. And perform a gradient descent update: .in, It's a model. The inner loop learning rate is preferably 0.01. The inner loop update frequency is preferably 1 time to simulate rapid adaptation. This yields... That is, for the task The adapted model parameters.

[0050] (c) Outer loop update: using parameters adapted for each task Each in its query set Calculate loss The meta-objective is to minimize the sum of the query set losses for all tasks: Calculate the meta-gradient The key point here is that the gradient is updated through the inner loop. Backpropagation back to initial parameters Use a meta-optimizer (such as Adam) with a meta-learning rate. (Preferably 0.001) Update meta-model parameters: .

[0051] (d) Repeated Iteration: Repeat steps (a) to (c) until the metamodel no longer shows significant performance improvement on the reserved validation task set, or reaches the preset maximum number of iterations (e.g., 50,000 iterations). The final result... These are the parameters of the trained meta-learning model.

[0052] Understandably, the essence of this training process lies in the calculation of the "meta-gradient." It optimizes not the model's performance on the training set, but rather its ability to "quickly learn new tasks." After training, the meta-model parameters... Located at a special position in the parameter space, starting from this point, moving a small step (inner loop update) in the direction of the loss gradient for any new task (new personnel) will lead to a parameter point that performs well for that task. This makes building a personalized model for each new dispatched worker extremely efficient.

[0053] Specifically, in response to the target dispatched worker starting to perform a dispatch task, based on the initialization parameters output by the meta-learning model, a lightweight personalized risk prediction model is initialized on the edge computing node associated with the target worker, including: When a target person is dispatched to the employing unit and is ready to perform a task, the cloud server sends the latest parameters of the meta-learning model to the edge computing nodes deployed on the local server of the employing unit or on the smart terminal held by the target person. On the edge computing nodes, the parameters are loaded to initialize the personalized risk prediction model. The personalized risk prediction model has the same network skeleton structure as the meta-learning model, but the number of parameters has been compressed and optimized. Using the initial behavioral data generated by the target person during the task preparation stage or at the beginning of the task, the initialized personalized risk prediction model is fine-tuned on the edge computing nodes with a small number of samples to quickly match the individual behavioral characteristics of the person and complete the cold start of the model.

[0054] It should be further explained that this step instantiates and initializes the personalized model at the edge. Firstly, there are two preferred deployment methods for edge computing nodes: for fixed workplaces (such as factory workshops and warehouses), edge nodes are deployed on the employer's local server or gateway device; for mobile or outdoor work scenarios (such as logistics delivery and exterior wall cleaning), edge nodes run as applications on smart terminals such as explosion-proof mobile phones, smart safety helmets, or special work badges held by dispatched workers. When personnel arrive at their posts, the cloud will then transfer the trained meta-model parameters... The encrypted data is then sent to the corresponding edge nodes.

[0055] Secondly, considering the limited computing and storage resources of edge devices, the model needs to be lightweighted. A preferred approach is knowledge distillation: train a large teacher meta-model in the cloud while simultaneously training a more streamlined student model (e.g., reducing the hidden layer dimension from 64 to 32 and the number of layers from 3 to 2), allowing the student model to mimic the teacher model's output when faced with numerous meta-tasks. The lightweight student model parameters are then distributed to the edge. Another approach is to directly prune the meta-model, removing unimportant connections.

[0056] Finally, perform cold start fine-tuning at the edge. Once personnel are on duty, the system immediately begins collecting their initial behavioral data, such as data from the first 30 minutes or the first two work cycles. This data is used as a "support set," and on the edge node, starting with the meta-model parameters, one to two gradient descent updates are performed (similar to the inner loop of MAML). The learning rate can be slightly higher than the inner loop learning rate during meta-training. For example, setting it to 0.05 can speed up the adaptation to individual characteristics. After fine-tuning, this model becomes the individual's personalized risk prediction model and begins receiving real-time data streams for prediction.

[0057] Understandably, this step "injects" general meta-knowledge from the cloud into a personalized model at the edge, and uses "first impression" data from the initial period of employment for rapid calibration, enabling the model to form a preliminary and relatively accurate understanding of the individual in a very short time. This solves the problem of the immediacy of risk assessment for new employees in labor dispatch, going from "nothing" to "something." At the same time, the lightweight model and edge computing ensure real-time performance and avoid uploading sensitive personal behavioral data to the cloud, complying with privacy protection requirements.

[0058] Specifically, the real-time collected multidimensional behavioral data includes physiological and motion data collected by wearable devices or work vehicle sensors, real-time task attributes and progress pressure data provided by the task management system, and micro-operational environment data collected by environmental sensing devices; the personalized risk prediction model outputs a real-time dynamic risk quantification value that comprehensively reflects the probability of an accident caused by factors such as personal habits, skill level, and physiological and psychological fluctuations in the current state of the person based on the multidimensional behavioral data.

[0059] It should be further explained that multidimensional behavioral data is the foundation for the model to make accurate personalized judgments, and its composition must comprehensively reflect the elements of "human-machine-environment-management". Specifically, it includes the following categories: (1) Physiological and motor data: collected through smart devices worn by personnel. This includes: a) Inertial Measurement Unit (IMU) data: from smart bracelets or badges, providing three-axis acceleration and angular velocity, used to identify abnormal movement postures (such as bending over, climbing, sudden imbalance), operation frequency and amplitude.

[0060] b) Physiological sensor data: from smart wearable devices, monitoring heart rate, heart rate variability, and skin conductance to assess fatigue, stress levels, and stress status. A heart rate consistently more than 40% above resting heart rate may indicate excessive fatigue.

[0061] c) Positioning and attitude data: from UWB or Bluetooth positioning tags, combined with helmet sensors, to determine whether the person has entered a dangerous area (such as the edge of a high-altitude operation or the range of motion of a robotic arm) or whether the posture is safe (such as the status of the safety belt hook when working at height).

[0062] Real-time task attributes and progress / pressure data: obtained from the user's task management or MES system via an interface. This includes: a) Task type and standard operating procedure (SOP) steps: Is the current task "heavy object handling" or "precision equipment adjustment", and what standard steps should be performed?

[0063] b) Task complexity and experience requirements: The skill level requirements for personnel in this task.

[0064] c) Planned time and actual progress: Whether the task is behind schedule, and the percentage of delay. The schedule pressure index can be calculated as (planned remaining time / estimated remaining time to complete). A value greater than 1.2 may indicate significant time pressure.

[0065] Micro-level operational environment data: collected by IoT devices deployed at the work sites. This includes: a) Local environmental data: acquired through temperature, humidity, light, and noise sensors. For example, high temperature and humidity environments can easily lead to heatstroke and operational errors.

[0066] b) Equipment status associated data: real-time status, such as speed, load, and alarm information, is obtained from the controllers of manually operated equipment (such as forklifts and cranes).

[0067] c) Distance between collaborating workers: Calculate the distance between workers and other nearby workers using personnel location data. Getting too close may cause a collision risk.

[0068] The personalized risk prediction model receives the multi-source fused feature vector at a fixed frequency (e.g., once per second). Internally, the model performs calculations through multiple layers of neural networks, ultimately outputting a real-valued value between 0 and 1 as a real-time dynamic risk quantification. . It's not simply environmental risk, but rather a personalized accident probability that arises from the interaction of the individual's current behavioral patterns with the aforementioned environmental and task context. For example, the same high-altitude walking task might present a different risk profile for someone with a fear of heights. The value will be significantly higher than that of other people in the model evaluation.

[0069] Understandably, by integrating multi-dimensional data such as an individual's physiological behavior, external task pressure, and the micro-environment, personalized risk prediction models can construct a highly contextualized and individualized risk perception picture. This allows risk assessment to move beyond simply relying on general rules (such as "high-altitude work is high-risk"), and instead determine "how high the individual risk is for Zhang San performing this high-altitude work under his current fatigue state and time pressure." This assessment is more accurate and can identify special risks that are difficult for general models to capture, formed by the superposition of individual states and specific situations.

[0070] Specifically, the step of generating and executing personalized risk intervention instructions for the target personnel based on the real-time dynamic risk quantification value through a meta-policy network includes: The meta-policy network takes as input the real-time dynamic risk quantification value output by the personalized risk prediction model of the target person and the current risk causal feature vector extracted from the multi-dimensional behavioral data. Based on the input, the meta-policy network decides on one or more of the most suitable combinations of intervention measures from a variety of preset types. The types of intervention measures include, but are not limited to: pushing personalized operation guidance and risk warnings through augmented reality devices or voice devices, pushing customized safety training micro-courses through mobile applications, and sending requests to the scheduling system to adjust the task pace or suggest rest. The decision-making process incorporates feedback data on the person's historical intervention effects and continuously optimizes the decision-making strategy of the meta-policy network through an online meta-reinforcement learning mechanism, so that the generated intervention instructions match the person's risk tolerance, learning style, and current state in terms of content and form.

[0071] It should be further explained that the meta-policy network is the core of achieving precise intervention. Its input state... It is a combined feature vector, including: (1) Real-time dynamic risk quantification value ; (2) Short-term trends (such as the slope over the past minute). (3) Risk causal features extracted from the current multidimensional data, such as abnormal scores for each dimension, including “high heart rate”, “excessive range of motion”, “close to the deadline”, and “excessive environmental noise”. (4) The current task stage of the personnel.

[0072] The meta-policy network itself can be a deep Q-network (DQN) or a policy gradient network. Its action space... It consists of a variety of executable intervention commands, such as: a) Action : Push a risk warning with specific content in the form of voice or text through bone conduction headphones or AR glasses, such as "Zhang San, your heart rate is detected to be too fast. Please stop your work and take three deep breaths."

[0073] b) Actions : Push a 30-second customized safety operation demonstration video to the personnel's mobile APP, focusing on the steps they are currently prone to making mistakes in.

[0074] c) Actions Generate a task adjustment suggestion and send it to the team leader's dispatch system, such as "It is suggested that Li Si suspend the current high-voltage switchgear maintenance task and rest for 15 minutes."

[0075] d) Actions No proactive intervention is taken; only log entries are recorded.

[0076] The network output is in the current state. The probability distribution of each action selection or the Q value of each action The system samples or selects the action with the highest Q value based on the output probability.

[0077] Feedback on the intervention's effectiveness is crucial for optimizing the strategy. (Execution action) Afterwards, the system will remain in a window for a subsequent period of time. (For example, within the next 5 minutes) monitor two indicators: one is the real-time dynamic risk quantification value of personnel. The evaluation criteria include: first, whether the number of people decreased and the extent of the decrease; second, whether the personnel exhibited the expected safe behaviors (such as following instructions to take a break or performing actions according to video guidance). Rewards are calculated based on these criteria. ,For example: ,in It is the risk value before intervention. It is the risk value after the window period. It is an indicator function; it is 1 if safe behavior is observed, and 0 otherwise. and It is a weighting coefficient, preferred. It places greater emphasis on reducing actual risk. This reward, along with status and actions, constitutes experience. The experience replay buffer of the edge node is stored and used to update the parameters of the meta-policy network online.

[0078] Understandably, meta-policy networks, through learning, can map different risk states and causes to the most effective interventions. For example, for individuals with high risk due to lack of skills, the policy might tend to push instructional videos; for those with high risk due to fatigue, the policy might suggest rest. More importantly, through online reinforcement learning, the policy can be personalized based on each individual's actual response (feedback reward) to a specific intervention. For example, for individuals who are not sensitive to voice prompts but are sensitive to vibration feedback, the policy will gradually learn to choose to trigger device vibration rather than voice prompts in subsequent similar situations, thus achieving "increasing accuracy with use" in intervention methods.

[0079] Specifically, the online meta-reinforcement learning mechanism continuously optimizes the decision-making strategy of the meta-policy network, including: The real-time dynamic risk quantification value change trend and safety behavior compliance data of the target personnel over a period of time after each intervention instruction are executed are used as environmental reward signals for reinforcement learning. Locally on the edge computing nodes, the reward signals are used to update the policy gradient of the meta-policy network to optimize the intervention strategy for the specific personnel. Periodically, the parameter update gradients of the meta-policy network on each edge computing node are encrypted and uploaded to the cloud server. The cloud server aggregates all uploaded gradients and performs a primary meta-update on the main model of the meta-policy network deployed in the cloud, enabling the main model to learn the policy initialization capability to quickly adapt to different personnel feedback patterns. The updated main model parameters can then be redistributed to each edge node.

[0080] It should be further explained that the online meta-reinforcement learning mechanism adopts a two-layer architecture of "federated learning + meta-learning" to optimize the meta-policy network, achieving policy personalization and continuous evolution while protecting privacy. Specific design features include: (1) Local personalized strategy optimization (edge ​​side): On each edge node associated with a dispatched worker, maintain a meta-policy network copy exclusive to that worker, with initial parameter values. Meta-policy network master model parameters delivered from the cloud During the operation, the replica learns online based on locally collected state-action-reward experiences. Specifically, it can adopt a single-worker form using the Asynchronous Advantage Actor-Critic (A3C) algorithm. Each edge node runs an agent independently, periodically (e.g., every 32 experiences collected) sampling a mini-batch of data from the local experience replay buffer, calculating the policy gradient to update the local network parameters. The learning rate can be set to 0.0001 to maintain the stability of policy updates. This process allows the policy to become increasingly adapted to the individual's unique feedback patterns.

[0081] (2) Global Meta-Policy Knowledge Aggregation and Distribution (Cloud Side): Set an aggregation period, such as every 24 hours or every 100 edge nodes that have completed local updates. Each edge node participating in the aggregation will update its local policy network parameters. With initial parameters (i.e., the one issued last time) The difference between (i.e., parameter update amount) The data is encrypted and uploaded to the cloud server. The cloud server collects a batch of updates. Then, perform secure aggregation (e.g., using the FedAvg algorithm): Then, use this average update amount to update the master model parameters in the cloud: ,in It is the global aggregate learning rate, preferably 1.0. This update process is essentially meta-learning: the main model in the cloud. What is learned is a set of initial policy parameters, which are then distributed to any edge node as... Afterwards, each node can quickly adapt to its personalized strategy through local learning (corresponding to the inner loop of MAML). Updated It can be used as a new round of initialization parameters to be sent to all or some edge nodes.

[0082] Understandably, this mechanism creates a continuously evolving ecosystem of security intervention strategies. At the edge, each person's strategy is fine-tuned in real time based on their individual feedback, achieving a high degree of personalization. At the cloud, by aggregating the personalized learning experiences of thousands of dispatched workers, the meta-policy network master model extracts meta-knowledge about "how to quickly develop effective intervention strategies for different types of personnel." When a new employee joins, the system distributes a strategy network that has already incorporated the experience of many and possesses good "start-up capabilities." This network can more quickly find the optimal strategy suitable for the new employee through local learning, thereby improving the overall efficiency and effectiveness of the system's security interventions.

[0083] Specifically, the method operates on a cloud-edge collaborative federated learning architecture, wherein: The training of the meta-learning model and the meta-update of the meta-policy network are performed on the cloud server, aggregating desensitized knowledge from multiple edge computing nodes without touching the original personal data; the fine-tuning of the personalized risk prediction model, the calculation of the real-time dynamic risk quantification value, and the generation and initial execution of the personalized risk intervention instructions are all completed locally on the edge computing nodes associated with the specific dispatched workers; the edge computing nodes and the cloud server only exchange model parameters, gradients, or aggregated knowledge representations, realizing personalized model evolution under privacy protection.

[0084] It should be further explained that the entire system architecture strictly adheres to the principles of cloud-edge collaboration and federated learning, clearly defining the computing and data boundaries between the cloud and the edge. The cloud server, acting as the "brain," is responsible for the extraction and distribution of global knowledge, specifically hosting two core models: a meta-learning model for initializing personalized risk prediction models, and a meta-policy network main model for initializing personalized intervention strategies. The training and updating of these two models rely on encrypted knowledge (model parameter updates) aggregated from edge nodes, rather than raw behavioral data.

[0085] Furthermore, edge computing nodes, acting as "peripheral nerves," are deployed in close proximity to the data source, with each node associated with one or a group of dispatched workers. Each node runs a personalized risk prediction model instance and a personalized meta-policy network instance for that individual. All processes involving that individual's sensitive data—including reading raw data from sensors, feature extraction, running personalized models for risk inference, generating intervention instructions based on policies, and fine-tuning / learning the two personalized models using local interactive feedback—are completed locally within a closed loop at the edge node. The raw behavioral data never leaves the edge device.

[0086] Furthermore, communication between the cloud and the edge is conducted via an encrypted channel. The uplink primarily transmits encrypted model parameter updates, while the downlink primarily transmits updated cloud-based meta-model parameters. This design ensures that highly sensitive information such as personnel's physiological data, behavioral trajectories, and operating habits are always retained locally. From a technical architecture perspective, it meets the requirements of regulations such as the Personal Information Protection Law regarding data minimization, de-identification, and secure transmission, while also reducing bandwidth and storage pressure on the cloud.

[0087] Understandably, this architecture perfectly aligns with the dual high requirements of data privacy and real-time response in labor dispatch scenarios. On one hand, employing units and dispatch companies cannot directly access employees' detailed personal data, avoiding privacy risks and legal disputes, and enhancing the system's feasibility. On the other hand, risk calculation and intervention decisions are completed locally in milliseconds, independent of network latency, ensuring the real-time nature of safety warnings and alerts—crucial for scenarios with extremely high real-time requirements, such as high-altitude operations and heavy machinery operation. Meanwhile, the cloud, through federated meta-learning, achieves continuous evolution and sharing of global knowledge without physically accessing the data.

[0088] Specifically, this also includes the personalized model evolution steps: During the performance of tasks by the target dispatched workers, the behavioral data and risk quantification results generated by them are continuously used as new training samples and stored in the local buffer of the edge computing node. When the number of samples accumulated in the local buffer reaches a certain number or the triggering condition is met, the personalized risk prediction model is incrementally learned or periodically recalibrated on the edge computing node using these new samples. The incremental learning or recalibration process enables the personalized risk prediction model to track and adapt to the evolution of the risk pattern caused by the worker's skill improvement, habit change or state fluctuation.

[0089] It should be further explained that the risk characteristics of dispatched workers are dynamic. For example, new employees improve their skills through training and practice, thus reducing operational risks; long-term repetitive work may lead to fatigue accumulation or decreased attention, increasing risks; personal life events may cause emotional fluctuations, affecting work performance. Therefore, the system is designed with a continuous personalized model evolution mechanism, specifically including: (a) Data accumulation: Edge nodes continuously accumulate processed feature vectors and model-predicted risk values. And post-verification labels (if any) as samples. Store the data in a fixed-size First-In-First-Out (FIFO) buffer. The buffer size is preferably 1000 samples, which is sufficient to cover behavioral patterns over a period of time, but not too large to avoid consuming excessive storage.

[0090] (b) Triggering conditions: The evolutionary process can be triggered by two conditions, and execution will proceed if either condition is met: 1. Quantity Trigger: When the number of new samples in the buffer reaches a threshold. When the value is (preferably 100), an incremental learning is triggered.

[0091] 2. Time-triggered: Triggered at fixed time intervals. (For example, every 24 hours), triggering a periodic recalibration.

[0092] (c) Evolutionary Approach: For incremental learning, online learning or mini-batch stochastic gradient descent is employed. Each time, a mini-batch (e.g., 32 samples) is randomly sampled from the buffer, the loss (e.g., cross-entropy loss) is calculated, and backpropagation is performed to update the parameters of the personalized risk prediction model. The learning rate is set to a small value (e.g., 0.001) to ensure the model smoothly adapts to new knowledge and avoids "catastrophic forgetting." For periodic recalibration, all (or most) samples in the buffer are used for a complete training epoch, equivalent to a comprehensive retraining of the model using recent data.

[0093] (d) Model Rollback and Evaluation: To prevent model performance degradation due to noisy data or abnormal states during evolution, edge nodes save the current model's checkpoint before each evolution. After evolution, the new model's performance is evaluated using a reserved validation set (such as the last 20% of data in the buffer). If performance significantly degrades (e.g., accuracy drops by more than 5%), the model is rolled back to the previous checkpoint.

[0094] Understandably, this evolutionary step makes the personalized risk prediction model a "living" model, capable of evolving alongside the growth and changes of dispatched workers. It can not only capture the long-term trend of reduced risk due to increased skill, but also sensitively identify short-term risk spikes caused by fatigue, distraction, etc., thus maintaining high accuracy in risk assessment. This overcomes the shortcomings of traditional static models or models requiring periodic manual retraining that cannot adapt to dynamic individual changes, achieving truly continuous personalized safety monitoring.

[0095] Specifically, the method also includes a cross-job adaptive step: When the target dispatched workers are dispatched to perform new job tasks that are different from their previous experience, the system receives the task type identifier of the new job; based on the task type identifier, it obtains the meta-knowledge sub-module related to the task type from the cloud server, or adjusts the feature extraction weights related to the task context in the personalized risk prediction model; combined with a small amount of behavioral data in the early stage of the new job, it quickly adjusts the risk assessment focus of the personalized risk prediction model to adapt it to the risk characteristics of the new job.

[0096] It's important to further clarify that temporary workers often move between different positions, and the risk factors for these positions vary significantly. For example, warehouse pickers primarily face risks related to handling posture and vehicle collisions, while electrical assemblers face risks from equipment malfunctions and static electricity. This step enables the personalized model to quickly adapt to these job changes, and the specific design includes: (a) Task Context Awareness: A task type knowledge base is maintained on the cloud server. Each task type (e.g., "high-bay rack picking," "indoor painting," "equipment inspection") is associated with a "task context vector." And an optional "task adaptation meta-knowledge submodule" Task context vector It is a multidimensional code that represents the weights of the main risk dimensions of the task, such as the weight distribution of physical requirements, delicate operation requirements, high-altitude risks, electrical risks, etc. This vector can be defined by domain experts or learned from historical data.

[0097] (b) Rapid model adaptation: When personnel switch to a new role At that time, its edge nodes obtain the corresponding data from the cloud. and / or There are two preferred implementation methods for the adaptation process: 1. Feature weighting method: Weighting the task context vector... The feature vectors of the personalized risk prediction model's input layer are multiplied element-wise (or used as attention weights) to amplify features strongly correlated with the new job at the model input, while suppressing irrelevant features. This is equivalent to soft-selecting the input features based on the task type.

[0098] 2. Module plug-and-play method: If the cloud provides a task adaptation metadata submodule... This submodule is typically a small neural network layer. It is inserted into a layer of a personalized risk prediction model (e.g., after the feature extraction layer) and works in conjunction with the original model parameters to transform the features for that task type.

[0099] (c) Fine-tuning with a small number of samples: After completing the above structural adjustments or parameter imports, use behavioral data generated by personnel in the initial period (e.g., the first hour) in their new positions to quickly fine-tune the adjusted model. The learning rate for fine-tuning can be set to 0.005, and the number of iterations should be controlled within 10 to prevent the forgetting of risk patterns corresponding to existing skills. This process can be regarded as a "small sample learning" for the new task, due to the existence of task-specific meta-knowledge. or With guidance, their adaptation speed is much faster than learning from scratch.

[0100] Understandably, the cross-job adaptive process greatly enhances the system's flexibility and practicality. It allows a personalized model trained for a specific job (such as a porter) to quickly shift its risk assessment focus to the key risk points of the new job when the employee changes positions (e.g., to an electrician), without waiting for extensive data accumulation and retraining in the new role. This solves the model "skills" transfer problem caused by frequent job turnover in the labor dispatch industry, ensuring that personnel receive accurate, timely, and personalized risk assessments and protection regardless of the task they perform.

[0101] It should be further explained that, in order to enable those skilled in the art to more clearly understand how the present invention works, a specific operational example is provided below. The protagonist of the example is a fictitious dispatched worker named Zhang San, whose position is a logistics deliveryman.

[0102] (a) Scenario and Data Preparation: Zhang San is a newly hired deliveryman, assigned to his first delivery task. The meta-learning model parameters have been pre-trained in the cloud. Heyuan policy network main model parameters Its smart terminals (as edge nodes) have been downloaded from the cloud. and After the task began, the terminal collected Zhang San's multidimensional behavioral data at a frequency of once per second through its built-in sensors and business system interface. During the 5-minute period tracked in this case, the system collected data including: motion amplitude values ​​calculated via IMU. (Normalized to [0,1]), heart rate value obtained through a heart rate sensor (Times / minute, normalized), progress pressure index obtained from the order system (Calculated as remaining time / estimated time, rounded down to 1.2 if greater than 1.2, then normalized), plus the current road segment complexity score obtained through mobile phone GPS and map services. (For example, 0.2 for ordinary roads and 0.8 for school zones, normalized). Each time step The feature vector is .

[0103] (b) Cold Start and Model Initialization: During the first 5 minutes of the task, the system uses the first 5 valid data segments (10 seconds each, 5 samples in total) as the support set. The corresponding initial risk label is temporarily generated by general rules in the cloud (e.g., Marked as risk 0.7, the rest are marked as 0.3). At edge nodes, using... Initialize Zhang San's personalized risk prediction model Execute the inner loop adaptation once: calculate exist Cross-entropy loss Perform a gradient descent operation to obtain the initial personalized model parameters for Zhang San. The inner loop learning rate Set the value to 0.01. This completes the cold start and yields the model. .

[0104] (c) Real-time risk assessment: After a cold start, the system begins to calculate risks in real time. For example, at a certain moment... Data collected This indicates a large range of motion, a high heart rate, pressure to keep up with the pace, and being on a high-complexity road segment. Input this vector. The model undergoes forward propagation (assuming it's a 2-layer network, with weights already changed from...). (Confirm), output scalar This is the real-time dynamic risk quantification value at this moment, indicating a high risk.

[0105] (d) Personalized intervention and strategy learning: Meta-policy network (initial parameters are...) Received status , include Risk Causes and Characteristics And the "in progress" phase of the mission. The network operates according to its current strategy. (by parameters) (Definition), calculate the probability of performing each action. Assume it chooses an action... The system will execute a voice prompt played through the headphones: "Zhang San, your current speed is too high and you are in a school zone. Please slow down and pay attention to safety."

[0106] (e) Feedback and Strategy Update: During the next 60-second window Inside, the system monitors Zhang San's behavior. Assume that, due to a prompt, Zhang San slows down, and the range of motion... Heart rate dropped to 0.3. It dropped to 0.6, despite schedule pressure. The value remains at 1.1, indicating road segment complexity. It remains at 0.8. At that time, the model According to new data Calculate the new risk value Meanwhile, the system uses IMU data to determine that Zhang San performed a deceleration action (a safe behavior). The reward is calculated based on this. .in Preset weights. Experience The experiences are stored in the local experience replay buffer. Later, when the buffer accumulates more than 32 experiences, the edge nodes sample a mini-batch from it, calculate the policy gradient, and update the local meta-policy network parameters. (from (beginning), making the future similar Select in status The tendency to perform actions is influenced by positive rewards. And enhanced.

[0107] (f) Model evolution: Zhang San continues to work. New data is continuously stored in his terminal's local buffer. Data pairs. When the number of new samples reaches a threshold. At that time, incremental learning is triggered. The system randomly samples 32 samples from the buffer and calculates the current... The prediction loss on these samples is updated by performing a gradient descent operation. The learning rate is set to 0.001. For example, after multiple incremental learning iterations, the model may more accurately identify Zhang San's risk response pattern under the combination of "high schedule pressure + ordinary road", thereby fine-tuning its risk assessment.

[0108] It is understandable that this specific operational case clearly demonstrates the entire process of the method of the present invention, from the cold start of employee onboarding, real-time personalized risk assessment, to intelligent intervention and strategy learning, and finally to the self-evolution of the model. All calculation steps in the case (gradient descent, forward propagation, reward calculation) are based on the publicly disclosed formulas and optimized parameter values ​​in the specific implementation. This case demonstrates that Zhang San, as a new employee, has his personalized model quickly initialized with a very small amount of data (5 samples) and immediately becomes effective in actual work, accurately assessing his high risk caused by comprehensive factors. Instead of using generic alarms, the system generates personalized voice prompts based on the specific risk causes (high speed, school area) through a meta-policy network. These prompts produced positive results (the risk value dropped significantly to 0.20, accompanied by safer behavior). The system used this positive feedback to optimize the intervention strategy for Zhang San online through reinforcement learning. Simultaneously, Zhang San's personalized risk prediction model continuously evolved with data accumulation. This case powerfully demonstrates that the method described in this invention can effectively solve key technical problems in personalized risk management of dispatched workers, such as cold start, accurate assessment, dynamic intervention, and continuous adaptation.

[0109] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the personalized risk adaptive management method for dispatched workers. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0110] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0111] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0112] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0113] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A personalized risk adaptive management method for dispatched workers, characterized in that, include: S1. Based on the historical work behavior data and corresponding risk tags of all dispatched workers, a meta-learning model is trained on a cloud server. The meta-learning model is used to obtain model initialization parameters that can quickly adapt to new individuals. S2. In response to the target dispatched worker starting to perform the dispatch task, based on the initialization parameters output by the meta-learning model, a lightweight personalized risk prediction model is initialized on the edge computing node associated with the target worker. S3. Collect multi-dimensional behavioral data of the target personnel in real time during the current task execution process, and input it into the personalized risk prediction model to calculate its real-time dynamic risk quantification value; S4. Based on the real-time dynamic risk quantification value, generate and execute personalized risk intervention instructions for the target personnel through a meta-policy network.

2. The personalized risk adaptive management method for dispatched workers according to claim 1, characterized in that, The process of training a meta-learning model on a cloud server based on historical data of all dispatched workers' work behavior and corresponding risk tags includes: From the historical databases of multiple dispatch projects, time-series behavioral data fragments and corresponding risk assessment results of different dispatched workers under various work tasks are extracted to form a meta-training task set. Each meta-training task simulates the learning process of a new person or a new task scenario. The meta-learning model is trained using a model-independent meta-learning algorithm framework. The training objective is to enable the meta-learning model to make accurate risk predictions for samples in the query set of a task after only a small amount of gradient updates based on a small amount of support set data in a single meta-training task. The training process enables the meta-learning model to learn transferable risk pattern meta-knowledge across personnel and tasks and encode it in its model parameters. The model parameters are the initialization parameters of the model that can quickly adapt to new individuals.

3. The personalized risk adaptive management method for dispatched workers according to claim 2, characterized in that, The training process of the model-independent meta-learning algorithm framework specifically includes: In each training iteration, a batch of tasks is sampled from the meta-training task set; for each sampled task, its data is divided into a support set and a query set; the current parameters of the meta-learning model are used as initial parameters, and one or more forward and backward propagations are performed on the support set data to calculate the gradient and update the model parameters, thereby obtaining a model adapted for the sampled task; the adapted model is used to perform forward propagation on the query set data to calculate the query set loss; based on the losses of all sampled tasks on their respective query sets, the meta-gradient is calculated, and backpropagation is performed to update the initial parameters of the meta-learning model; the above process is repeated until convergence, thereby obtaining a set of meta-learning model parameters that can be quickly adapted to generalize to new people or new tasks.

4. The personalized risk adaptive management method for dispatched workers according to claim 1, characterized in that, In response to the target dispatched worker starting to perform a dispatch task, based on the initialization parameters output by the meta-learning model, a lightweight personalized risk prediction model is initialized on the edge computing node associated with the target worker, including: When a target person is dispatched to the employing unit and is ready to perform a task, the cloud server sends the latest parameters of the meta-learning model to the edge computing nodes deployed on the local server of the employing unit or on the smart terminal held by the target person. On the edge computing nodes, the parameters are loaded to initialize the personalized risk prediction model. The personalized risk prediction model has the same network skeleton structure as the meta-learning model, but the number of parameters has been compressed and optimized. Using the initial behavioral data generated by the target person during the task preparation stage or at the beginning of the task, the initialized personalized risk prediction model is fine-tuned on the edge computing nodes with a small number of samples to quickly match the individual behavioral characteristics of the person and complete the cold start of the model.

5. The personalized risk adaptive management method for dispatched workers according to claim 1 or 4, characterized in that, The real-time collected multidimensional behavioral data includes physiological and motion data collected by wearable devices or work vehicle sensors, real-time task attributes and progress pressure data provided by the task management system, and micro-operational environment data collected by environmental sensing devices; the personalized risk prediction model outputs a real-time dynamic risk quantification value that comprehensively reflects the probability of an accident caused by factors such as personal habits, skill level, and physiological and psychological fluctuations in the current state of the person based on the multidimensional behavioral data.

6. The personalized risk adaptive management method for dispatched workers according to claim 1, characterized in that, The process of generating and executing personalized risk intervention instructions for the target personnel based on the real-time dynamic risk quantification value through a meta-policy network includes: The meta-policy network takes as input the real-time dynamic risk quantification value output by the personalized risk prediction model of the target person and the current risk causal feature vector extracted from the multi-dimensional behavioral data. Based on the input, the meta-policy network decides on one or more of the most suitable combinations of intervention measures from a variety of preset types. The types of intervention measures include, but are not limited to: pushing personalized operation guidance and risk warnings through augmented reality devices or voice devices, pushing customized safety training micro-courses through mobile applications, and sending requests to the scheduling system to adjust the task pace or suggest rest. The decision-making process incorporates feedback data on the person's historical intervention effects and continuously optimizes the decision-making strategy of the meta-policy network through an online meta-reinforcement learning mechanism, so that the generated intervention instructions match the person's risk tolerance, learning style, and current state in terms of content and form.

7. The personalized risk adaptive management method for dispatched workers according to claim 6, characterized in that, The online meta-reinforcement learning mechanism continuously optimizes the decision-making strategy of the meta-policy network, including: The real-time dynamic risk quantification value change trend and safety behavior compliance data of the target personnel over a period of time after each intervention instruction are executed are used as environmental reward signals for reinforcement learning. Locally on the edge computing nodes, the reward signals are used to update the policy gradient of the meta-policy network to optimize the intervention strategy for the specific personnel. Periodically, the parameter update gradients of the meta-policy network on each edge computing node are encrypted and uploaded to the cloud server. The cloud server aggregates all uploaded gradients and performs a primary meta-update on the main model of the meta-policy network deployed in the cloud, enabling the main model to learn the policy initialization capability to quickly adapt to different personnel feedback patterns. The updated main model parameters can then be redistributed to each edge node.

8. The personalized risk adaptive management method for dispatched workers according to claim 1, characterized in that, The method operates on a cloud-edge collaborative federated learning architecture, wherein: The training of the meta-learning model and the meta-update of the meta-policy network are performed on the cloud server, aggregating desensitized knowledge from multiple edge computing nodes without touching the original personal data; the fine-tuning of the personalized risk prediction model, the calculation of the real-time dynamic risk quantification value, and the generation and initial execution of the personalized risk intervention instructions are all completed locally on the edge computing nodes associated with the specific dispatched workers; the edge computing nodes and the cloud server only exchange model parameters, gradients, or aggregated knowledge representations, realizing personalized model evolution under privacy protection.

9. The personalized risk adaptive management method for dispatched workers according to claim 1, characterized in that, It also includes personalized model evolution steps: During the performance of tasks by the target dispatched workers, the behavioral data and risk quantification results generated by them are continuously used as new training samples and stored in the local buffer of the edge computing node. When the number of samples accumulated in the local buffer reaches a certain number or the triggering condition is met, the personalized risk prediction model is incrementally learned or periodically recalibrated on the edge computing node using these new samples. The incremental learning or recalibration process enables the personalized risk prediction model to track and adapt to the evolution of the risk pattern caused by the worker's skill improvement, habit change or state fluctuation.

10. The personalized risk adaptive management method for dispatched workers according to claim 1, characterized in that, The method also includes a cross-job adaptive step: When the target dispatched workers are dispatched to perform new job tasks that are different from their previous experience, the system receives the task type identifier of the new job; based on the task type identifier, it obtains the meta-knowledge sub-module related to the task type from the cloud server, or adjusts the feature extraction weights related to the task context in the personalized risk prediction model. By combining a small amount of behavioral data from the initial stage of the new position, the risk assessment focus of the personalized risk prediction model can be quickly adjusted to adapt it to the risk characteristics of the new position.

Images