Urban occupational health risk real-time supervision method and system based on data analysis

Abstract

The invention discloses an urban occupational health risk real-time supervision method and system based on data analysis, and the method comprises the steps: continuously collecting the behavior data flow of a target employee and the physical environment dynamic data flow of an associated working region, and generating an individual-environment dynamic data sequence; detecting a behavior abnormal signal based on the historical behavior mode of the target employee; in combination with the physical environment dynamic data flow at the current moment, the association strength between the physical environment dynamic data flow and the occupational health risk is evaluated, and a personalized micro-intervention instruction is generated and triggered; capturing the instant feedback response of the employee to the micro-intervention instruction in real time, updating the personal behavior baseline model of the target employee in real time, and calibrating the evaluation rule of the association strength; and dynamically generating a real-time occupational health risk portrait of the employee based on the individual behavior baseline model, the association strength evaluation rule and the individual-environment dynamic data sequence. According to the embodiment of the invention, accurate early warning and personalized intervention of occupational health risks can be realized.

Description

Technical Field

[0001] This invention belongs to the field of data analysis technology, specifically a method and system for real-time monitoring of urban occupational health risks based on data analysis. Background Technology

[0002] With the acceleration of urbanization, occupational health risk supervision faces dynamic and personalized challenges. Traditional methods, relying on periodic physical examinations or static environmental monitoring, struggle to capture the instantaneous correlation between abnormal employee behavior and environmental factors, leading to delayed risk warnings. Existing technologies, particularly behavioral data analysis, largely focus on offline assessments, lacking real-time data fusion and dynamic baseline modeling capabilities, making them ill-suited to the rapid changes in complex work scenarios. Furthermore, intervention measures often employ uniform standards, neglecting individual differences and environmental interactions, thus reducing the accuracy of risk prevention and control. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for real-time monitoring of urban occupational health risks based on data analysis, so as to overcome the shortcomings of the existing technology and achieve accurate early warning and personalized intervention of occupational health risks.

[0004] One embodiment of this application provides a method for real-time monitoring of urban occupational health risks based on data analysis, the method comprising:

[0005] Continuously collect behavioral data streams of target employees and dynamic data streams of the physical environment of related work areas, and perform real-time spatiotemporal alignment and fusion to generate individual-environment dynamic data sequences;

[0006] Based on the historical behavioral patterns of target employees, an online learning algorithm is used to dynamically maintain their personal behavioral baseline model. The behavioral data is compared with the personal behavioral baseline model to detect abnormal behavioral signals that represent potential occupational health risks.

[0007] By combining the current physical environment dynamic data stream, the abnormal behavior signal is interpreted in the context of the environment, and its correlation strength with occupational health risks is assessed. When the correlation strength exceeds the strength threshold, personalized micro-intervention instructions are generated and triggered.

[0008] Real-time capture of employees' immediate feedback responses to the micro-intervention instructions; using the immediate feedback responses, corresponding abnormal behavior signals, and environmental context data, real-time update of the target employee's personal behavior baseline model and calibration of the correlation strength assessment rules.

[0009] Based on the updated individual behavior baseline model, the calibrated correlation strength assessment rules, and the continuously input individual-environment dynamic data sequence, a real-time occupational health risk profile of the employee is dynamically generated, enabling real-time monitoring of urban occupational health risks.

[0010] Another embodiment of this application provides a real-time monitoring system for urban occupational health risks based on data analysis, the system comprising:

[0011] The fusion module is used to continuously collect behavioral data streams of target employees and dynamic data streams of the physical environment of related work areas, and perform real-time spatiotemporal alignment and fusion to generate individual-environment dynamic data sequences.

[0012] The maintenance module is used to dynamically maintain the individual behavior baseline model of the target employee based on the target employee's historical behavior patterns using an online learning algorithm, compare the behavior data with the individual behavior baseline model, and detect abnormal behavior signals that represent potential occupational health risks.

[0013] The assessment module is used to interpret the abnormal behavior signal in the context of the current physical environment dynamic data stream, assess its correlation strength with occupational health risks, and generate and trigger personalized micro-intervention instructions when the correlation strength exceeds the strength threshold.

[0014] The update module is used to capture employees' immediate feedback responses to the micro-intervention instructions in real time, and use the immediate feedback responses, corresponding abnormal behavior signals and environmental context data to update the target employee's personal behavior baseline model in real time and calibrate the evaluation rules for correlation strength.

[0015] The generation module is used to dynamically generate a real-time occupational health risk profile of an employee based on an updated individual behavior baseline model, calibrated correlation strength assessment rules, and continuously input individual-environment dynamic data sequences, thereby enabling real-time monitoring of urban occupational health risks.

[0016] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0017] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0018] Compared with existing technologies, this invention provides a data analysis-based real-time monitoring method for urban occupational health risks. It continuously collects behavioral data streams of target employees and dynamic data streams of the physical environment of their work areas to generate individual-environment dynamic data sequences. Based on the target employee's historical behavioral patterns, it detects abnormal behavioral signals. Combining the current physical environment dynamic data stream, it assesses the correlation strength between the data and occupational health risks, generating and triggering personalized micro-intervention instructions. It captures the employee's immediate feedback response to the micro-intervention instructions in real time, updates the target employee's personal behavioral baseline model in real time, and calibrates the correlation strength assessment rules. Based on the personal behavioral baseline model, the correlation strength assessment rules, and the individual-environment dynamic data sequences, it dynamically generates a real-time occupational health risk profile for the employee, thereby enabling accurate early warning and personalized intervention for occupational health risks. Attached Figure Description

[0019] Figure 1 Hardware structure block diagram of a computer terminal for a real-time monitoring method of urban occupational health risks based on data analysis, provided in an embodiment of the present invention;

[0020] Figure 2 A flowchart illustrating a real-time monitoring method for urban occupational health risks based on data analysis, provided as an embodiment of the present invention;

[0021] Figure 3 A flowchart illustrating another method for real-time monitoring of urban occupational health risks based on data analysis, provided in an embodiment of the present invention;

[0022] Figure 4 A flowchart illustrating another method for real-time monitoring of urban occupational health risks based on data analysis, provided in an embodiment of the present invention;

[0023] Figure 5 This is a schematic diagram of the structure of a real-time monitoring system for urban occupational health risks based on data analysis, provided as an embodiment of the present invention. Detailed Implementation

[0024] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0025] This invention first provides a method for real-time monitoring of urban occupational health risks based on data analysis. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0026] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a real-time monitoring method of urban occupational health risks based on data analysis, provided as an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0027] The non-volatile storage medium can store an operating system and a computer program. This computer program includes program instructions that, when executed, cause the processor to perform any data-driven, real-time monitoring method for urban occupational health risks.

[0028] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0029] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any data analysis-based method for real-time monitoring of urban occupational health risks.

[0030] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0031] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0032] See Figures 2-4 The present invention provides a method for real-time monitoring of urban occupational health risks based on data analysis, which may include the following steps:

[0033] S201 continuously collects behavioral data streams of target employees and dynamic data streams of the physical environment of related work areas, and performs real-time spatiotemporal alignment and fusion to generate individual-environment dynamic data sequences;

[0034] Specifically, wearable devices and office system logs can be used to capture raw data streams of employee behavior in real time, while a distributed environmental sensor array can be deployed to collect dynamic data streams of the physical environment, including light, noise, temperature and humidity in the work area, to obtain heterogeneous dual-modal data streams.

[0035] The system acquires data through two types of core devices:

[0036] Wearable devices: Smart wristbands worn by employees (such as wristbands equipped with a nine-axis inertial measurement unit (IMU)) capture raw behavioral data streams in real time. For example, accelerometers (range ±16g, sampling rate 50Hz) record limb movement amplitude, gyroscopes (range ±2000 degrees / second) detect joint rotation angles, and heart rate sensors (accuracy ±1bpm) monitor physiological load. Simultaneously, smart badges (integrating an ultra-wideband UWB positioning chip, positioning accuracy 30cm) track the employee's real-time position coordinates (X-axis, Y-axis, Z-axis).

[0037] Office system logs: These logs generate data streams of employee actions from the Enterprise Resource Planning (ERP) system and computer interactive software (such as keyboard and mouse event monitors). Examples include keyboard keystroke frequency (keystrokes per minute, KPM), mouse movement trajectory entropy (in bits), and duration of time spent on the software interface (in seconds).

[0038] Distributed environmental sensor array: Deploying a multi-node sensor group in the work area:

[0039] A light sensor (range 0-100,000 lux, resolution 1 lux) monitors changes in desktop illuminance;

[0040] A noise sensor (A-weighted sound pressure level, range 30-130 dB) captures instantaneous noise peaks;

[0041] A temperature and humidity integrated sensor (temperature accuracy ±0.3℃, humidity ±2%RH) collects data on local microclimate.

[0042] Carbon dioxide sensors (range 0-5000ppm) assess air quality.

[0043] All sensors upload data to the edge gateway via Bluetooth Low Energy (BLE) 5.0 or Zigbee 3.0 protocols at a sampling frequency of 1-10 times per second, forming a heterogeneous bimodal data stream (HBDS). The behavioral modality includes 20-dimensional motion features + 10-dimensional operational features, and the environmental modality includes 8-dimensional physical parameters, for a total of 38-dimensional heterogeneous data stream.

[0044] The data stream transmission architecture adopts a layered design:

[0045] Edge layer preprocessing: Each sensor node incorporates a lightweight filtering algorithm (such as sliding window mean filtering, with a window size of 500 milliseconds ms) to remove signal glitches. For example, the raw accelerometer data is smoothed by Kalman filtering (process noise parameter Q = 0.01, observation noise parameter R = 0.1) and then outputs the three-axis composite vector magnitude (VM).

[0046] Communication protocol adaptation: Wearable devices transmit behavioral data via the ANT+ protocol (1Mbps bandwidth), environmental sensors publish environmental data via the MQTT protocol (QoS = Level 1), and edge gateways uniformly convert to Apache Avro format (Schema version 2.3) and label it with the device ID (e.g., Device_ID = SN2024_053).

[0047] Real-time performance guarantee mechanism: A priority queue (PQL) is set up for data acquisition. Physiological data (e.g., heart rate > 120 bpm) is marked as PQL = 1 (transmission latency < 100 ms), and environmental data is marked as PQL = 2 (latency < 500 ms). Finally, a raw bimodal data packet (RBDP) with a timestamp (Unix timestamp, precision milliseconds) is generated on the edge server. Its structure includes:

[0048] Packet header: Timestamp (1567830452000) + Device ID (SN2024_053) + Data type (Behavior / Env);

[0049] Data volume: Accelerometer X / Y / Z (1.23g, -0.87g, 9.81g) + Heart rate (78bpm) + Noise level (65.3dB), etc.;

[0050] This design ensures processing 5000+ data packets per second with a packet loss rate of less than 0.1%.

[0051] Data quality control is achieved through triple verification:

[0052] Integrity verification: The CRC-16 algorithm (polynomial 0x8005) is used to check the integrity of data packets. If the verification fails, a retransmission mechanism is triggered (maximum number of retries = 3).

[0053] Outlier interception: Invalid data is filtered based on dynamic threshold (DT). For example, data with a body temperature >42℃ or humidity >100%RH is marked as invalid, triggering the sensor self-test program (Self-Test Code, STC=0x5A).

[0054] Data traceability mechanism: Each data packet is appended with metadata, including the sensor calibration date (CD = 20240501), battery voltage (BV = 3.7V), and signal strength (RSSI = -65dBm). The final output is a heterogeneous dual-modal data stream (HBDS) conforming to the ISO / IEC 30128 standard, which serves as the input source for subsequent processing.

[0055] An adaptive clock synchronization protocol is used to align the timestamps of heterogeneous dual-modal data streams, and a dynamic time warping algorithm is used to compensate for the sampling rate differences between devices, generating a primary fused data stream with millisecond-level time alignment.

[0056] The clock synchronization protocol uses a modified version of the IEEE 1588 Precision Time Protocol (PTP):

[0057] The Grandmaster Clock uses GPS-disciplined rubidium atomic clocks (accuracy ±10 nanoseconds) and broadcasts standard time to the entire network via the PTPv2 protocol (announce message every 2 seconds).

[0058] The edge gateway acts as a boundary clock (BC) to calculate the path delay (PD). The formula is: *Delay compensation value = ((t2-t1)+(t4-t3)) / 2*, where t1 is the request sending time, t2 is the request receiving time, t3 is the response sending time, and t4 is the response receiving time. For example, if t1 = 1567830452000, t2 = 1567830452015, t3 = 1567830452020, and t4 = 1567830452030, then PD = (15+10) / 2 = 12.5ms.

[0059] Adaptive mechanism: The synchronization interval (SI) is dynamically adjusted based on network jitter. When jitter > 50ms, the SI is shortened from the default 1 second to 0.5 seconds; when jitter < 10ms, the SI is extended to 2 seconds. Ultimately, the timestamp deviation of all devices is controlled within ±5ms.

[0060] Dynamic Time Warping (DTW) handles sampling rate differences:

[0061] Define the Reference Sequence (RS) and the Test Sequence (TS). For example, wearable data sampled at 50Hz is RS, and temperature and humidity data sampled at 5Hz is TS.

[0062] Construct a cumulative cost matrix (CM), where each element CM[i,j] represents the Euclidean distance (ED) between point i in RS and point j in TS. For example, the VM value (1.2g) of RS and the temperature value (25.3℃) of TS have different dimensions and need to be standardized to Z-scores (e.g., VM_z = (1.2-0.8) / 0.3 = 1.33, Temp_z = (25.3-24.5) / 0.5 = 1.6), before calculating ED = √(1.33-1.6). 2 =0.27.

[0063] Warping Path (WP): Generate a path from CM[1,1] to CM[n,m] according to the principle of minimum cumulative cost (Local Constraint = Slope Weight = 1.5). For example, when RS has 100 points and TS has 10 points, WP maps 1 point of TS to the interval of 5-15 points of RS.

[0064] Resampling output: Based on WP, the TS data is linearly interpolated to 50Hz. For example, the j-th point of TS (temperature = 25.3℃) is mapped to the RS timestamp [1567830452000, 1567830452020], and 10 temperature values ​​(25.30℃, 25.31℃...25.39℃) are generated evenly in this interval.

[0065] Data alignment and fusion:

[0066] Timestamp alignment: Sort the normalized data according to the master clock timestamp (Sort Key = Timestamp). For example:

[0067] Time points 1567830452000: Behavioral data (VM = 1.2g, heart rate = 78bpm) + environmental data (temperature = 25.30℃, noise = 65.3dB);

[0068] Time points 1567830452010: Behavioral data (VM = 1.3g, heart rate = 79bpm) + environmental data (temperature = 25.31℃, noise = 65.2dB).

[0069] Generate a Primary Fused Data Stream (PFDS): Stored in Apache Arrow format (Version 1.0), each record contains:

[0070] Timestamp (1567830452000);

[0071] Behavioral feature vector (VM:1.2, HR:78, KPM:45...);

[0072] Environmental feature vector (Temp:25.30, Noise:65.3, Lux:350...).

[0073] Quality metrics: time deviation of aligned data ≤ 1ms, feature missing rate < 0.05%, 10,000 records processed per second.

[0074] Based on the workstation space topology map, a geofencing-enhanced spatial indexing engine is used to map the primary fused data stream to three-dimensional coordinate points, generating a spatiotemporal correlated data stream with spatial location labels;

[0075] Workstation space topology map construction process:

[0076] Laser scanning of the office generates a point cloud map (PCM) with an accuracy of ±2 cm. The RANSAC algorithm (iterations = 1000, interior point threshold = 5 cm) is used to fit the plane and identify workstation boundaries (e.g., rectangular areas: coordinates (10.2m, 5.6m) to (11.0m, 6.4m)).

[0077] Establish a topology graph (TG): Nodes represent workstations (e.g., Node_ID = WS_203), and their attributes include area (Area = 6.4㎡) and height (Height = 2.8m); edges represent connectivity (e.g., the path length from WS_203 to the water dispenser = 8.7m).

[0078] Dynamic update mechanism: When a change in workstation layout is detected (such as furniture movement distance > 30cm), an immediate remap reconstruction is triggered (Remap Trigger, RT = 1).

[0079] Geofencing Enhanced Spatial Indexing Engine Workflow:

[0080] Geofencing definition: Create an electronic fence for each risk area. For example:

[0081] High noise zone (Noise>85dB): Circular fence (center (15.3m, 7.8m), radius 2.5m);

[0082] Screen glare area (Lux>1000): Polygonal fence (vertices (12.1m, 3.4m), (12.9m, 3.4m), (12.9m, 4.1m));

[0083] Spatial index construction: R* tree index is used, with a node capacity of 50. The office is divided into 1m×1m grid cells, and each grid cell is associated with an environmental sensor ID (e.g., Cell(12,7)→Sensor_ID=ENV_47).

[0084] Real-time coordinate mapping:

[0085] Employee location: UWB location data (13.2m, 8.7m, 1.0m) is used to query the corresponding cell (13,8) using an R* tree;

[0086] Matching environmental data: Retrieve the Sensor_ID = ENV_53 associated with Cell(13,8) and bind the current noise value of the sensor (63.5dB);

[0087] Fence status detection: Calculate the distance from point (13.2, 8.7) to the center of the high-noise zone = √((15.3-13.2) 2 +(7.8-8.7) 2 If 2.3m < 2.5m, mark "Enter_Noise_Zone".

[0088] Generate spatiotemporal correlated data streams:

[0089] Data augmentation: Appending fields to the primary fused data stream PFDS:

[0090] Spatial location label: 3D coordinates (X=13.2, Y=8.7, Z=1.0);

[0091] Geofencing status: Noise_Zone = 1 (1 indicates within the area);

[0092] Workstation ID: WS_205 (achieved through collision detection between coordinates and workstation polygon).

[0093] Output structure example:

[0094] Timestamp: 1567830452000; Location: (13.2, 8.7, 1.0); Behavior: {VM: 1.2, HR: 78, KPM: 45}; Environment: {Noise: 63.5dB, Temp: 25.3℃, Lux: 420}; Fence: {Noise_Zone: 1, Glare_Zone: 0}; Workstation: WS_205.

[0095] By using a multimodal feature fusion encoder, behavioral feature vectors and environmental feature vectors in spatiotemporal correlated data streams are compressed into a multidimensional unified tensor sequence, outputting a standardized individual-environment dynamic data sequence.

[0096] Multimodal feature fusion encoder architecture:

[0097] Input layer:

[0098] Behavioral feature branch: 20-dimensional vector (VM, HR, KPM...) input to fully connected layer FC1 (number of neurons = 64, activation function ReLU);

[0099] Environmental feature branch: 8-dimensional vector (Noise, Temp, Lux...) input to FC2 (number of neurons = 32, activation function ReLU);

[0100] Spatial feature branch: 3D coordinates (X,Y,Z) + 4D fence state (Noise_Zone,Glare_Zone...) input to FC3 (number of neurons = 16).

[0101] Feature cross layer:

[0102] Behavioral and environmental features are multiplied by the outer product to generate 160-dimensional interaction features (such as VM×Noise, HR×Temp);

[0103] Weighted by attention mechanism (Attention Weight, AW): Calculate the attention score of behavioral features to the environment (Score = Softmax(FC1·FC2^T / √8)), and output 32-dimensional context features after weighting.

[0104] Feature compression and standardization:

[0105] Feature concatenation: The outputs of FC1 (64-dimensional), FC2 (32-dimensional), interactive features (32-dimensional), and FC3 (16-dimensional) are merged into a 144-dimensional vector;

[0106] Dimensionality reduction: Compression via the bottleneck layer:

[0107] FC4 (number of neurons = 72, activation function Swish);

[0108] FC5 (number of neurons = 36, activation function Swish);

[0109] Output layer: Generates a 12-dimensional unified tensor.

[0110] Standardization: Using the Moving Z-score algorithm:

[0111] Mean update: μ_t = 0.99 × μ_{t-1} + 0.01 × X_t;

[0112] Variance update: σ 2 _t=0.99×σ 2 _{t-1}+0.01×(X_t-μ_t)^2;

[0113] Standardization: X_{norm}=(X_t-μ_t) / (σ_t+ε)(ε=1e-7 to prevent division by zero).

[0114] Dynamic data sequence generation:

[0115] Serialization process: Using a 5-second time window, the tensor (12-dimensional × 60) of 60 consecutive timestamps is reassembled into a 720-dimensional time vector.

[0116] Output structure:

[0117] Header information: Employee ID (Emp_No = 10087) + time window start stamp (1567830452000);

[0118] Data volume: tensor sequence [T1,T2...T60], where T1 = [0.23,-1.07,0.95...] (12 floating-point numbers);

[0119] Quality label: Data integrity score (Completeness Score = 0.98);

[0120] Performance metrics: Feature fusion latency <200ms, data compression rate 90% (original 38 dimensions → 12 dimensions), supports processing 1000+ employee data streams per second.

[0121] This step utilizes wearable devices and an environmental sensor network to simultaneously collect employee behavioral data (such as posture and operation frequency) and work environment data (such as lighting and noise). A spatiotemporal alignment algorithm eliminates time differences and spatial misalignments between devices, forming a precisely matched sequence of individual-environment interaction data. The data fusion process employs multimodal coding technology to transform heterogeneous data into a unified time series format, establishing a precise foundation for individual-environment correlation data and providing high-quality input for subsequent risk identification. Spatiotemporal alignment ensures the traceability of causal relationships between behavioral and environmental data, avoiding misjudgments caused by collection delays or locational deviations.

[0122] S202, Based on the historical behavioral patterns of target employees, use online learning algorithms to dynamically maintain their personal behavioral baseline model, compare the behavioral data with the personal behavioral baseline model, and detect abnormal behavioral signals that represent potential occupational health risks.

[0123] Specifically, behavioral data sequences of employees over the past few days can be extracted, and a typical behavioral state transition matrix can be established through an incremental hidden Markov model to generate an initial behavioral baseline model.

[0124] The system first extracts a 30-day (configurable number of days N=30) sequence of behavioral data from the target employee's historical database. This data originates from wearable devices (such as smart bracelets and employee badge sensors) and office system logs, including action types (such as sitting duration, standing frequency, and keyboard typing intensity), physiological indicators (such as heart rate variability (HRV) and skin conductance response (GSR), and work interaction events (such as screen viewing duration and document switching frequency). After preprocessing, the raw data is discretized into a finite number of behavioral states (BS), for example: BS1 represents "high-intensity focused work" (accompanied by high keyboard typing rate and low body movement), BS2 represents "intermittent rest" (accompanied by periodic standing and low screen interaction), and BS3 represents "abnormal body tension" (accompanied by persistently high GSR and abnormal posture).

[0125] An incremental hidden Markov model (IHMM) is used to model the temporal dependencies between states. During model initialization, the initial state transition matrix (STM) is trained using data from the first day. This matrix is ​​a K×K two-dimensional probability table (K is the total number of behavioral states, e.g., K=5), where the matrix element STM[i][j] represents the probability of transitioning from state BSi to BSj. For example, if the data from the first day shows that the probability of an employee transitioning from BS1 (focused work) to BS2 (rest) is 0.3, then STM[1][2] = 0.3. Subsequent incremental learning occurs daily: when new data is input, the model recalculates the transition frequency using a forward-backward algorithm and updates the probability values ​​in the STM. For example, if new data indicates an increase in the frequency of the BS1→BS2 transition, then the probability values ​​in STM[1][2] are adjusted accordingly.

[0126] The final generated Initial Behavioral Baseline Model (IBBM) consists of two core components:

[0127] Behavioral State Transition Matrix (STM): Completely describes the probability transition rules between each state;

[0128] Emission Probability Distribution (EPD): Defines the statistical distribution (e.g., Gaussian distribution with parameters of mean μ and standard deviation σ) of sensor observations (such as heart rate range and posture angle) for each behavioral state.

[0129] For example, the model might learn that, in the BS1 state, heart rate values ​​follow a normal distribution with μ = 75 bpm (heart beats per minute) and σ = 5. This model serves as a statistical benchmark for employees' normal behavior, providing a reference for subsequent anomaly detection.

[0130] For the initial behavioral baseline model, a streaming Bayesian inference algorithm is used to take the newly input behavioral data sequence as the observation value, update the behavioral state transition probability distribution in real time, and output a dynamically evolving personal behavioral baseline model.

[0131] After the initial model deployment, the system continuously receives real-time behavioral data streams (e.g., a set of sensor readings per second). The Streaming Bayesian Inference (SBI) algorithm processes new data in sliding windows (e.g., window size T = 60 seconds). The observation sequence within each window (e.g., [HR = 72, GSR = 0.5, posture = sitting]) serves as input, and the algorithm dynamically updates the model parameters based on Bayes' theorem. The core mechanism treats the current state transition probability as a prior distribution, new observation data as likelihood evidence, and uses Bayesian calculation of the posterior distribution to update the probability values.

[0132] The update process consists of two steps:

[0133] State sequence decoding: The Viterbi algorithm is used to decode the most likely state path of the current window (e.g., [BS1,BS1,BS2]).

[0134] Parameter update: Count the frequency of state transitions in the path (e.g., BS1→BS1 occurs 2 times, BS1→BS2 occurs 1 time), and combine it with the prior transition probabilities (e.g., original STM[1][1]=0.6, STM[1][2]=0.3), and adjust the probabilities according to Bayesian rules. For example, if the prior hypothesis is that the transition frequency follows a Dirichlet distribution, then the posterior probability is: new probability = (prior frequency + actual observation frequency) / (total prior frequency + total observation frequency).

[0135] Assuming that the equivalent frequency of BS1→BS1 in the prior is 10 times and 2 new observations are made, then the updated STM[1][1]=(10+2) / (10+3)≈0.923 (denominator 13 is the total equivalent frequency).

[0136] This process continues, allowing the model parameters to adapt to changes in employee behavior habits (such as adapting to new work tasks). The final output, the Dynamic Behavioral Baseline Model (DBBM), always reflects the employee's latest behavioral patterns. For example, if an employee has recently extended their focus due to project pressure, the model will automatically increase the self-transfer probability of BS1 (STM[1][1]) and decrease the transition probability to a resting state (STM[1][2]).

[0137] Input the current behavior feature vector into the individual behavior baseline model, decode the most likely behavior state sequence through the Viterbi algorithm, and output the predicted behavior state vector;

[0138] The Current Behavioral Feature Vector (CBFV) is a multi-dimensional data point collected in real time, such as [time = 14:30:05, heart rate = 80 bpm, posture angle = 15°, keystroke rate = 2.1 times / second]. After this vector is input into the latest version of DBBM, the system needs to decode its most likely corresponding hidden state sequence under the current model. Because employee behaviors are temporally correlated (e.g., the state of the previous moment affects the state of the next moment), the Viterbi Algorithm is used to solve for the optimal path.

[0139] The algorithm consists of three steps:

[0140] Initialization: Calculate the probability (calculated via EPD) that the observed value CBFV1 belongs to each state BSi at the first time point t=1, and record the path.

[0141] Recursion: For subsequent time points t = 2, 3, ..., T, calculate the joint probability of transitioning from any state BSj at time t-1 to state BSi at time t. The formula is: Path probability = Optimal path probability of state j at the previous moment × Transition probability STM[j][i] × Observation probability P(CBFV) t |BSi).

[0142] Select the predecessor state j that maximizes the probability and update the optimal path to state i at time t.

[0143] Termination and backtracking: At t=T, select the state with the highest probability and backtrack the entire optimal path in reverse.

[0144] The final output, the Predicted Behavioral State Vector (PBSV), is a time series, such as [BS1, BS1, BS1, BS2], representing the most likely state evolution of an employee under the current behavioral pattern, as predicted by the model. For example, if observations for four consecutive seconds all conform to the statistical characteristics of BS1, the predicted vector will be four consecutive BS1 states.

[0145] Calculate the Mahalanobis distance between the current behavior feature vector and the predicted behavior state vector, and generate the behavior deviation coefficient;

[0146] The system compares the real-time observed behavioral feature vector CBFV (e.g., [heart rate = 85, GSR = 0.7]) with the ideal observation distribution corresponding to the predicted state vector PBSV at the same time point. For example, if PBSV predicts the state as BS1 at t = 5 seconds, then the ideal observation distribution of BS1 is heart rate N (μ = 75, σ = 5) and GSR follows N (μ = 0.4, σ = 0.1).

[0147] The Mahalanobis distance (MD) measures how much a CBFV deviates from an ideal distribution. Its calculation depends on two parameters:

[0148] The covariance matrix (CM) of the current state describes the correlation between sensor dimensions (such as the covariance between heart rate and GSR); the difference between the observed vector and the ideal mean vector.

[0149] The calculation process is as follows: MD 2 =(CBFV-μ) T ×CM⁻¹×(CBFV-μ). For example, if CBFV=[85,0.7], μ=[75,0.4], and CM⁻¹ is the inverse covariance matrix, then MD 2 It is a scalar value (e.g., 6.2).

[0150] The output Behavioral Deviation Coefficient (BDC) is the Mahalanobis distance value MD (or MD). 2 This coefficient eliminates the differences in dimensions between different sensors and takes into account dimensional correlation. For example, BDC = 2.5 means that the current observation deviates from the ideal distribution by 2.5 standard deviations (generalized distance), and the larger the value, the higher the degree of anomaly.

[0151] When the behavior deviation coefficient exceeds the preset deviation threshold for a consecutive preset number of sampling periods, the behavior feature vector of the corresponding time period is marked as a behavior abnormal signal.

[0152] The system sets two key threshold parameters:

[0153] Deviation Threshold (DT): For example, DT = 3.0 means that a Mahalanobis distance greater than 3 is considered a significant deviation;

[0154] Consecutive Period Threshold (CPT): For example, CPT=3 requires that the anomaly lasts for at least three sampling periods (e.g., 3 seconds).

[0155] The BDC value of each sampling point is monitored in real time. If all BDC values ​​are greater than DT within a consecutive CPT period (e.g., seconds t, t+1, t+2), an anomaly flag is triggered. For example:

[0156] t = 10s: BDC = 3.2 > DT (3.0) → Anomaly count = 1;

[0157] t = 11s: BDC = 3.5 > DT → Anomaly count = 2;

[0158] t = 12s: BDC = 3.1 > DT → Anomaly count = 3, reaching CPT = 3.

[0159] At this point, the system will label all behavioral feature vectors (i.e., CBFV) from t=10s to t=12s. 10 ,CBFV 11 ,CBFV 12 The Behavioral Anomaly Signal (BAS) is a signal that includes raw sensor data, timestamps, and an inferred type of anomaly (such as "persistent high heart rate and physical stress") for subsequent environmental correlation analysis. For example, a continuous 5-second abnormal sitting posture (excessive spinal curvature) would be flagged as a "potential musculoskeletal risk signal".

[0160] By analyzing long-term employee behavioral data using incremental learning algorithms, a personalized database of routine behavioral patterns is constructed. Current behavioral characteristics are compared in real-time with a baseline model, and a statistical deviation detection algorithm is used to identify abnormal behavioral segments (such as persistent head-down posture or high-frequency repetitive movements). These anomalies may indicate occupational health risks such as musculoskeletal strain or visual fatigue, enabling personalized early risk warnings and avoiding false alarms caused by generic thresholds. The dynamic baseline model can adapt to the natural evolution of employee behavioral habits, improving the accuracy of anomaly detection.

[0161] S203, combining the current physical environment dynamic data stream, interpret the abnormal behavior signal in the environmental context, assess its correlation strength with occupational health risks, and when the correlation strength exceeds the strength threshold, generate and trigger personalized micro-intervention instructions.

[0162] Specifically, it may include: S2031, extracting dynamic physical environment data for the time period corresponding to the abnormal behavior signal, fusing light fluctuation gradient, temperature and humidity change rate, and noise spectrum characteristics to generate an environmental risk feature tensor;

[0163] When the system detects that an employee's behavioral deviation coefficient continuously exceeds a preset threshold (e.g., exceeding the threshold of 2.5 for five consecutive sampling periods), it will lock that time period as a behavioral anomaly signal window (e.g., 10:05:30 AM to 10:07:45 AM). At this time, the distributed environmental sensor array synchronously captures the raw physical environment data stream of the work area within this window:

[0164] Light intensity gradient extraction: A light intensity sensor (sampling frequency 20 Hz) deployed at the top of the workstation records light intensity values ​​(in lux) every millisecond. The system calculates the change in light intensity (ΔLux) between adjacent sampling points and then generates a light gradient vector (LGV) through first-order derivative calculation. For example, if the light intensity is detected to drop sharply from 500 Lux to 200 Lux within 0.5 seconds, its gradient value is marked as -600 Lux / s. Such drastic fluctuations may cause visual fatigue risks.

[0165] Temperature and humidity change rate calculation: A combined temperature and humidity sensor (accuracy ±0.3℃) collects data 10 times per second. The system applies a sliding window difference algorithm (window size 1 second) to the temperature series, outputting a temperature change rate series (℃ / s); the same method is used to generate a humidity change rate series (%RH / s). For example, if the temperature rises from 22℃ to 28℃ in 30 seconds, the change rate is marked as 0.2℃ / s, which may indicate a risk of heat stress due to an air conditioning system malfunction.

[0166] Noise Spectrum Characteristic Analysis: A wideband noise sensor (monitoring range 20Hz-10kHz) acquires sound pressure level data in real time. The system converts the time-domain sound wave into a frequency-domain signal using Fast Fourier Transform (FFT) and extracts the energy of three key sub-bands: the energy proportion of the low-frequency band (20-250Hz) (reflecting mechanical vibration), the peak intensity of the mid-frequency band (250-2kHz) (correlated with human voice interference), and the duration of the high-frequency band (2kHz-10kHz) (indicating sharp noise). These three parameters constitute the Noise Spectrum Triplet (NST).

[0167] The three types of features (LGV, temperature and humidity change rate, and NST) are aligned with millisecond-level timestamps and then input into the multimodal feature fusion engine. This engine employs a weighted concatenation strategy: the illumination gradient is assigned a weight of 0.4 (due to high visual sensitivity), the temperature and humidity change rate is assigned a weight of 0.2, and the noise spectrum triplet is assigned a weight of 0.4. The concatenated multidimensional vector is then normalized (e.g., Z-score standardization) to finally generate an Environmental Risk Tensor (ERT). This tensor has the dimension of [time step × number of features]. For example, a 2-minute window of behavioral anomalies will generate a 120×7 tensor (5 sampling points per second × 7 feature classes).

[0168] S2032 inputs the behavioral anomaly signal and the environmental risk feature tensor into the pre-trained risk correlation assessment model and outputs the real-time correlation strength coefficient in the 0-1 interval;

[0169] The Risk Correlation Assessment Model (RCAM) is a classifier based on a deep neural network. Its training data comes from historical occupational health accident reports and corresponding environmental behavior logs.

[0170] Model Architecture: A dual-channel input design is employed. Channel 1 processes abnormal behavioral signals, receiving an 8-dimensional behavioral feature vector containing posture angles (e.g., cervical curvature) and operation frequency (e.g., keyboard typing speed). Channel 2 processes the Environmental Risk Tensor (ERT), extracting local environmental patterns through a one-dimensional convolutional layer (kernel size 3). The dual-channel outputs are concatenated in a fusion layer, then passed through a fully connected layer (128 neurons) and a sigmoid activation function to output the correlation strength coefficient.

[0171] Real-time inference process: When new data is input, the model performs three calculations:

[0172] Behavioral feature embedding: Mapping current behavioral anomaly signals (such as continuous head tilt angle > 30 degrees) to latent space vectors.

[0173] Environmental feature convolution: Temporal convolution is performed on the ERT tensor to capture key patterns (such as the temporal overlap between abrupt changes in illumination gradient and cervical curvature).

[0174] Correlation strength calculation: After fusing the above results, the real-time correlation coefficient (RCC) between 0 and 1 is output. For example, when an employee is continuously looking down, if an illumination gradient < -300 Lux / s is detected at the same time, the model may output RCC = 0.82 (high correlation), while the same behavior under stable illumination will only output RCC = 0.15.

[0175] Dynamic confidence mechanism: The model is equipped with a confidence calibration module. When there are missing input environmental data (such as a faulty temperature and humidity sensor), the weights are automatically reduced and a confidence interval (e.g., RCC = 0.75 ± 0.05) is output. A coefficient exceeding a preset strength threshold (e.g., 0.7) is considered a valid risk association.

[0176] S2033, when the correlation strength coefficient exceeds the preset strength threshold, the optimal micro-intervention instruction template is retrieved from the intervention strategy knowledge graph based on the behavioral abnormality signal type and environmental risk feature tensor.

[0177] If RCC > 0.7 (threshold is configurable), the system activates the intervention strategy knowledge graph engine:

[0178] Knowledge graph structure: A knowledge graph consists of three layers of nodes:

[0179] Risk layer nodes: Define 12 types of occupational health risks (such as "visual fatigue" and "muscle strain"), and associate them with abnormal behavior types through historical data (such as "excessive screen time" corresponding to visual fatigue).

[0180] Environment layer node: Stores environment feature patterns (e.g., "high frequency noise > 70dB" associated with "attention distraction").

[0181] Intervention layer node: Stores 200+ micro-intervention instruction templates (MIT), each containing attributes such as instruction content, triggering conditions, and expected feedback.

[0182] Retrieval logic: Employs a multi-path reasoning algorithm.

[0183] Step 1: Match risk layer nodes based on the type of abnormal behavior signal (such as "excessive mouse operation force") to locate the "carpal tunnel syndrome" risk cluster.

[0184] Step 2: Analyze the Environmental Risk Tensor (ERT). If a temperature change rate > 0.15℃ / s is detected, expand the risk cluster along the "heat stress → muscle stiffness" path.

[0185] Step 3: Execute the cost-benefit scoring algorithm to select the optimal template from the associated MITs. For example:

[0186] Template A: "Vibrating Wrist Pad Reminder" (Cost score: 0.2, Expected effect score: 0.8)

[0187] Template B: "Play wrist stretching animation" (Cost score 0.4, Expected effect score 0.9)

[0188] When the ERT displays a noise level > 65dB, the system prioritizes visual template B (to avoid masking audio commands).

[0189] Template instantiation: The retrieved MIT template (e.g., ID_MIT_07) contains placeholders. The system injects real-time data into the template.

[0190] Environmental parameter input: Enter "Current temperature 28℃" into the prompt message: "High temperature environment detected, please reduce the intensity of operation".

[0191] Customized behavior parameters: Adjust animation speed based on mouse usage frequency. Generate the final personalized micro-intervention set (PMIS).

[0192] S2034 instantiates the micro-intervention instruction template into a micro-intervention instruction set containing text prompts and visual guidance, and triggers it to the terminal through a low-latency message middleware.

[0193] The generated PMIS needs to be efficiently delivered to employee terminals:

[0194] Multimodal instruction construction:

[0195] Text prompts: A hierarchical message system is used. Level 1 prompts are brief warnings (e.g., "Beware of wrist strain!"), while level 2 prompts contain detailed suggestions (e.g., "Please stretch the radial wrist flexor muscles as shown in the diagram"). The text length is limited to 40 Chinese characters, and the font is automatically adjusted according to the ambient light (dark gray for >500 Lux, bright yellow for <200 Lux).

[0196] Visual guidance: Dynamically generate SVG vector animations. For example, for forward head posture, the animation demonstrates the trajectory of the head's center of gravity shifting backward, with key skeletal points calibrated in real time according to the employee's actual tilt angle (error < 3°). The animation frame rate is matched to the terminal performance (default 30fps, reduced to 15fps on older devices).

[0197] Message middleware transmission:

[0198] Protocol optimization: Lightweight MQTT protocol (message body < 10KB) is adopted, and QoS is set to Level 1 for transmission guarantee.

[0199] Routing strategy: Intelligent routing based on terminal status:

[0200] When employees wear AR glasses: commands are transmitted directly to the glasses (delay < 50ms);

[0201] When only carrying a mobile phone: the animation resolution is compressed and then transmitted via 5G slices (latency <100ms);

[0202] Offline status: Commands are temporarily stored on the edge server and pushed to the terminal once it is online.

[0203] Terminal execution control:

[0204] Multi-device collaboration: When PMIS contains compound commands (such as "adjust seat height + play instruction video"), the middleware synchronously triggers the office equipment controller (adjusting the electric seat via Zigbee protocol) and the mobile terminal (playing video);

[0205] Feedback channel pre-activation: When the instruction is sent, the terminal microphone and camera are activated (permissions need to be granted in advance) to prepare to capture the employee's subsequent actions (such as whether to perform stretching) as the starting point of the feedback data stream.

[0206] This approach involves multi-dimensional correlation analysis between behavioral abnormalities and environmental parameters (such as frequent eye rubbing in low light conditions), quantifying risk levels using a pre-trained risk assessment model. When the risk value exceeds a threshold, a customized intervention plan is automatically generated (such as screen brightness adjustment prompts + eye exercise guidance), avoiding over-intervention due to isolated judgments of behavioral abnormalities and improving the accuracy of early warnings through environmental context validation. Personalized design of micro-intervention instructions enhances employee acceptance and execution effectiveness.

[0207] S204, Real-time capture of employees' immediate feedback responses to the micro-intervention instructions, and real-time update of the target employee's personal behavior baseline model and calibration of the correlation strength assessment rules using the immediate feedback responses, corresponding abnormal behavior signals and environmental context data;

[0208] Specifically, this may include: S2041, collecting employees' immediate feedback responses to micro-intervention instructions and constructing a multi-dimensional feedback feature vector;

[0209] When the system pushes personalized micro-intervention instructions (such as "Please adjust your posture" or "It is recommended to wear noise-canceling headphones") to employees via terminal devices (such as smart bracelets, AR glasses, or mobile phone pop-ups), it immediately activates a multi-source sensor collaborative monitoring mechanism. Wearable devices worn by employees (such as wrist sensors) collect changes in physiological indicators in real time.

[0210] Galvanic Skin Response (GSR) monitors fluctuations in skin conductivity (unit: micro-Siemens μS), reflecting the excitability of the autonomic nervous system;

[0211] A 3-axis accelerometer captures the amplitude of limb movements (unit: gravitational acceleration g) and quantifies posture adjustment compliance;

[0212] Heart rate variability (HRV) is calculated as the standard deviation of the interval between adjacent heartbeats (in milliseconds ms) to assess the effectiveness of stress relief.

[0213] Meanwhile, environmental sensors (such as workstation cameras) use computer vision algorithms to analyze employees' head turning angles (in degrees) and gaze duration (in seconds) to determine their attentional response to visual guidance instructions (such as screen flashing prompts). All raw data is synchronized to edge computing nodes at a frequency of ten times per second.

[0214] The data preprocessing module performs key feature extraction:

[0215] Physiological response index: The slope of GSR rise (GSR_Slope) within five seconds after the command is triggered is calculated. If it exceeds 0.05 microsiemens per millisecond (0.05 μS / ms), it is marked as significant stress.

[0216] Behavioral compliance index: The similarity between the accelerometer waveform and the preset "ideal response action template" is matched by the Dynamic Time Warping (DTW) algorithm, and the compliance matching score (compliance_Score, range 0-100 points) is output.

[0217] Attention metric: Eye gaze direction is detected based on the YOLOv4-tiny lightweight model. If the gaze is focused on the prompt area within two seconds after the instruction is given, the attention flag (Attention_Flag) is set to 1; otherwise, it is set to 0.

[0218] The final result is a feedback feature vector (FFV) with six dimensions: [GSR_Slope, HRV_Change_Rate, Compliance_Score, Attention_Flag, Posture_Correction_Angle, Response_Latency].

[0219] To eliminate the influence of individual differences, the system introduces baseline normalization. For example, the current GSR_Slope is divided by the employee's average stress slope (Baseline_GSR_Slope) from the previous week to generate a standardized stress coefficient (Normalized_Stress_Index). After all dimensional data are standardized by Z-score, the feature selector (using random forest importance ranking) retains the three dimensions with the highest contribution to form a streamlined feedback feature vector (Optimized_FFV) for downstream model updates.

[0220] S2042, Generate a feedback effect weight matrix based on the deviation of each dimension in the feedback feature vector from the expected feedback;

[0221] The system has a preset expected response template (Expected_Response_Template, ERT), which dynamically sets target values ​​based on different command types:

[0222] For the "posture correction" instruction, the ideal Compliance_Score is >85 points and Posture_Correction_Angle is >15°;

[0223] For the "Relief Tip" instruction, the target HRV_Change_Rate needs to be increased by 20%.

[0224] The Deviation Calculator compares the actual value with the expected value dimension by dimension:

[0225] The Absolute Relative Error (ARE) formula is used: Deviation = |Actual Value - Expected Value| / Expected Value. For example, if the actual Compliance Score is 70 points and the expected value is 85 points, then the deviation is 0.176.

[0226] Generate the initial weight matrix (Initial_Weight_Matrix) based on the deviation:

[0227] Design an inverse exponential decay weight allocation strategy: for every 0.1 increase in deviation, the weight of that dimension decreases by 15%;

[0228] Introducing an environmental confidence factor: When the ambient noise exceeds 65 dB, the reliability weight of the Attention_Flag dimension is automatically multiplied by a decay factor of 0.8;

[0229] Add time decay compensation: Data that takes more than five seconds to respond after an instruction is issued will have its Response_Latency dimension weight multiplied by an additional 0.7.

[0230] Finally, a weight matrix (WMC) with environment and time correction is generated, with the same number of dimensions as the feedback feature vector.

[0231] The weight matrix needs to pass a consistency check:

[0232] If the weight of a certain dimension is below 0.3 three times in a row, a sensor calibration alarm will be triggered.

[0233] The entropy weight method is used to rebalance the weights, avoiding subjective allocation bias. Specifically, the information entropy (Entropy_Value) of each dimension's data is calculated; the smaller the entropy value, the higher the weight. Finally, a standardized feedback weight matrix (Normalized_Feedback_Weight_Matrix, NFWM) is output, ensuring that the sum of the weights for each dimension is 1.

[0234] S2043, after weighting the abnormal behavior signals according to the feedback effect weight matrix, they are injected as new samples into the online learning algorithm to trigger the parameter adjustment of the personal behavior baseline model;

[0235] The Behavior Anomaly Signal (BAS) is retrieved from the cache queue. Essentially, it's a time-series segment containing all behavioral characteristics (such as keystroke frequency and mouse movement trajectory curvature) during the abnormal period. The weighted fusion engine then performs the following operations:

[0236] Multiply each feature dimension in BAS by its corresponding weight in NFWM. For example, the weight of "abnormal neck tilt angle" is 0.6, and the weight of "excessive continuous working time" is 0.3.

[0237] The weighted features are subjected to sliding window averaging filtering (window width = five seconds) to eliminate instantaneous noise;

[0238] Add timestamps and environmental context labels (such as "light intensity less than 500 lux") to form weighted reinforcement learning samples (Weighted_Anomaly_Sample, WAS).

[0239] The online learning algorithm employs a streaming Bayesian updating mechanism:

[0240] Prior distribution update: The "sitting posture transition probability matrix" in the individual behavior baseline model is treated as a prior distribution. For example, the probability of transitioning from the original state of "upright sitting posture" to "leaning forward" is 0.2;

[0241] Likelihood function construction: Using WAS as the observed data, calculate its likelihood value. If the sample shows a forward tilt probability of 0.5 under insufficient lighting, the peak of the likelihood function appears at 0.5.

[0242] Posterior distribution calculation: The prior distribution and likelihood function are fused using Bayes' theorem to output the posterior transition probability (Posterior_Probability). This process uses variational inference to accelerate the calculation, processing a new sample every 200 milliseconds.

[0243] The effectiveness of model adjustments is ensured through an online validation mechanism.

[0244] A / B testing module: Temporarily store 10% of the weighted samples in the shadow model to compare the anomaly detection differences between the old and new models for the same input;

[0245] Drift Detector: Monitors the standard deviation of state transition probabilities. If the standard deviation exceeds the fluctuation threshold of 0.05 for three consecutive updates, it triggers a full model retraining.

[0246] Finally, stable posterior parameters are written into the dynamic behavior baseline library to complete real-time updates.

[0247] S2044 uses environmental context data as conditional variables, abnormal behavior signals as observed variables, and feedback feature vectors as reinforcement signals. It dynamically updates the weight parameters of the risk correlation assessment model through a Bayesian network.

[0248] A three-layer Bayesian Network (BN) is constructed as the core for evaluating association strength:

[0249] Conditional variable layer (environmental context): Nodes include noise level, illuminance, and temperature and humidity combination index (THI_Index);

[0250] Observation variable layer (behavioral anomalies): Nodes integrate keyboard error rate (Key_Error_Rate), blink frequency (Blink_Freq), and posture stability score (Posture_Stability_Score);

[0251] Target variable layer: The output node is the risk correlation strength coefficient (Risk_Correlation_Coefficient, RCC).

[0252] The dynamic update process consists of three stages:

[0253] Evidence Propagation: When a new data packet arrives, environmental data (such as Noise_Level = 72 dB) is input into the conditional variable layer, and the intermediate layer probability is updated through Belief Propagation;

[0254] Likelihood correction: Inject abnormal behavior signals (such as a sudden increase in Key_Error_Rate to 25 times per minute) into the observation layer and calculate their likelihood contribution to RCC;

[0255] Reinforcement learning feedback loop: The feedback feature vector (e.g., Compliance_Score = 40 in FFV) is used as the reward signal. If the risk persists due to the employee's failure to comply with the intervention, the causal weight of the "noise-keyboard error" path is reduced (e.g., adjusted from 0.8 to 0.7).

[0256] Parameter calibration mechanisms ensure network reliability:

[0257] Conditional Probability Table Adjustment (CPT Adjustment): When the actual risk occurrence rate is 20% higher than the prediction in a certain environmental scenario (e.g., THI_Index>28), the conditional probability value of the abnormal behavior node to RCC under this condition is increased.

[0258] D-Separation Test: Periodically verify the independence of nodes. If a spurious correlation is found between ambient light and posture stability due to the introduction of new equipment, an isolation node is inserted.

[0259] Monte Carlo sampling verification: After each update, five thousand simulation data are generated using the Markov Chain Monte Carlo (MCMC) method to test the KL divergence between the RCC prediction value and the actual accident record. If the divergence exceeds 0.01, the weights are rolled back.

[0260] Monitor employee responses to interventions (e.g., whether suggested actions are taken, whether abnormalities are alleviated), and dynamically adjust the sensitivity and risk assessment weights of the baseline model through reinforcement learning mechanisms. For example, for employees who frequently ignore brightness adjustment prompts, automatically increase the threshold for judging light-related abnormalities, forming a closed-loop optimization of "monitoring-intervention-feedback" to enable the system to continuously adapt to user characteristics. This avoids the warning fatigue or missed detection problems caused by static models.

[0261] S205 dynamically generates a real-time occupational health risk profile for an employee based on an updated individual behavior baseline model, calibrated correlation strength assessment rules, and continuously input individual-environment dynamic data sequences, enabling real-time monitoring of urban occupational health risks.

[0262] Specifically, the behavioral deviation coefficient, environmental risk feature tensor, and correlation strength coefficient can be aligned along the time axis and fused into a comprehensive risk feature vector through a gating attention mechanism.

[0263] The system first establishes a unified Time Baseline Axis (TBA) to synchronize three key data sources with millisecond-level precision: Behavioral Deviation Coefficient (BDC), Environmental Risk Feature Tensor (ERFT), and Association Strength Coefficient (ASC). BDC, derived from the behavioral baseline model comparison results, is a scalar value (range 0-1) reflecting the degree of deviation between employee actions and normal patterns. ERFT is a multidimensional array (e.g., a 5-dimensional tensor) fused from environmental sensor data (such as Luminance Fluctuation Gradient (LFG), Temperature-Humidity Change Rate (THCR), and Noise Spectrum Feature (NSF)). ASC is a scalar (range 0-1) output from the risk association assessment model, representing the causal probability between behavioral abnormalities and environmental risks. The time alignment process employs an improved version of the Dynamic Time Warping (DTW) algorithm: it matches the timestamp offsets of three sets of data using a sliding window, and leverages the path constraint function of the DTW algorithm to force alignment to a unified time grid. For example, when the environmental sensor sampling rate is 10Hz and the wearable device's is 20Hz, the system automatically interpolates and completes the low-frequency data to ensure that each time slice (e.g., every 100 milliseconds) contains a complete BDC, ERFT, and ASC triplet.

[0264] The aligned triples are used as input for feature fusion via a gated attention mechanism (GAM). This mechanism comprises three core components:

[0265] Feature encoding layer: The two scalars BDC and ASC are expanded into vectors of the same dimension as ERFT (e.g., 5-dimensional) through a fully connected layer to eliminate the dimensionality difference;

[0266] Attention Weight Generation Layer: Calculates the interaction weights of the three types of features within each time slice. Taking the temperature and humidity change rate (THCR) in ERFT as an example, if its value exceeds the threshold (e.g., temperature change rate > 0.5℃ / minute), then Thermal Risk Attention (TRA) is triggered, and the system automatically increases the priority of THCR in weight allocation.

[0267] Gated fusion layer: The sigmoid function is used to generate a gating value (GV) in the 0-1 range to dynamically adjust the contribution ratio of the three types of features. For example, when BDC > 0.8 (highly abnormal behavior) and ASC > 0.7 (strong environmental association), GV will amplify the weight of the noise spectrum feature NSF in ERFT (if instantaneous noise > 85 dB is detected), and finally output a weighted fusion integrated risk feature vector (IRFV). This vector has a fixed dimension (e.g., 8 dimensions) and includes fields such as timestamp, behavioral risk intensity, environmental risk type weight, and comprehensive association confidence.

[0268] To enhance the robustness of the fusion process, the system introduces an Environmental Context Verification Module (ECVM). For example, when ERFT detects an abnormal Light Fluctuation Gradient (LFG) (e.g., flicker > 10Hz) but ASC < 0.3, ECVM will check the office system logs to see if it falls within a device debugging period. If it is confirmed to be a planned activity, the LFG weight is reduced to 0.1 through a gating mechanism to avoid false positives. All fusion processes are completed on edge computing nodes, with latency controlled within 50 milliseconds to ensure real-time performance.

[0269] The comprehensive risk feature vector is input into a long short-term memory network to predict the probability of risk occurrence within a preset time period in the future.

[0270] The Long Short-Term Memory Network (LSTM) employs a three-layer hidden layer structure (128-64-32 units), with the input sequence being an IRFV sequence of 60 consecutive time slices (i.e., 6-second windows). The network first filters invalid historical features through a Forget Gate (FG): for example, if the behavioral risk intensity is <0.2 in 5 consecutive IRFVs, the FG reduces the weight of that time segment to 0.05. The Input Gate (IG) assigns a weight greater than 0.9 to burst features in the current ERFT (e.g., THCR > 1℃ / minute). The Memory Cell (MC) accumulates risk trends; for example, when the ASC (Advanced Risk Scale) is >0.6 for 3 consecutive minutes, the MC output value increases by 0.3 times.

[0271] Risk prediction is divided into two channels:

[0272] Short-term prediction channel (0-30 minutes): Outputs the risk probability value (RPV) for the next 6 time intervals (every 5 minutes). For example, when a persistent forward head posture (BDC>0.7) is detected in an employee and the workstation lighting is <300 lux (ERFT lighting dimension abnormality), the predicted risk probability of muscle strain within 30 minutes is RPV = 0.88.

[0273] Long-term prediction channel (8 hours): Generates a Cumulative Risk Index (CRI). For example, if the noise spectrum characteristics NSF show a continuous 65 dB low-frequency noise, combined with the frequency of employees looking down during operations per hour, the predicted hearing loss CRI after 8 hours is 0.75.

[0274] The prediction results are quantified using Monte Carlo Dropout (MCD) technology, and the standard deviation (SD) is calculated by sampling 50 times. If SD > 0.15, a review process is initiated, and the local model is retrained using historical similar scene data.

[0275] The predictive model is dynamically updated using Online Incremental Learning (OIL). When an actual risk event occurs (such as an employee triggering a health alert), the system uses the IRFV sequence from the 30 minutes prior to the event as new samples and updates the LSTM weights with a learning rate of 0.001. Simultaneously, a Catastrophic Forgetting Suppression (CFS) algorithm is employed to limit the weight changes of key neurons to ±5%, ensuring that historical knowledge is not lost.

[0276] Based on graph neural networks, a causal relationship topology between abnormal behavioral signals and environmental risks is constructed, and key risk transmission paths are marked to form a risk propagation tree;

[0277] The topology of a Graph Neural Network (GNN) contains three types of nodes:

[0278] Abnormal behavior nodes (red): such as "mouse operation frequency suddenly increased by 300%" (BDC=0.92);

[0279] Environmental risk points (blue): such as "CO2 concentration at workstation > 1000 ppm" (ERFT air quality dimension anomaly);

[0280] Health event nodes (yellow): such as "Visual fatigue report" (from the medical system interface); edge weights are initialized to ASC coefficients (0-1), for example, "screen glare → forward head posture" edge weight = 0.78.

[0281] Causal relationship mining employs a combination of Graph Attention Network (GAT) and Random Walk Algorithm (RWA):

[0282] GAT computing node influence: For example, the influence score of the temperature and humidity node TH on the "headache" event = 0.63;

[0283] RWA generates a transmission path: starting from the "excessive keystroke force" node, passing through the "abnormal wrist flexion angle" node (propagation probability 0.81), and terminating at the "carpal tunnel syndrome" node (cumulative probability 0.69).

[0284] The system automatically labels critical risk propagation paths (CRPP): selects paths with a cumulative probability > 0.7 and a number of hops ≤ 3. For example, the path "insufficient lighting → screen close proximity → abnormal cervical curvature → neck and shoulder pain" is labeled as CRPP-01.

[0285] The Risk Propagation Tree (RPT) is visualized as a tree structure: the root node represents the original environmental risk (e.g., a noise source), and the leaf nodes represent potential health events (e.g., hearing loss). Each path is labeled with: conduction delay (e.g., noise → irritability: 2.3 minutes); probability decay coefficient (e.g., decay of 0.15 per hop); and intervention sensitivity (e.g., wearing noise-canceling earplugs can reduce the path probability by 58%). The tree structure is updated every 5 minutes, retaining only the top 10 subtrees with the highest sum of probabilities.

[0286] By integrating the probability of risk occurrence with the risk propagation tree, an interactive digital twin is generated that includes a 3D risk heat map, an intervention effect trend curve, and a risk tracing path, enabling real-time visualization of risk profiles.

[0287] A 3D risk heatmap is constructed in the digital twin space: X / Y axis: workstation plane coordinates (accuracy 0.1 meters); Z axis: risk intensity value (0-100%); voxel coloring rules: based on the RPV value output by LSTM, >80% is red, 30-80% is yellow, and <30% is green.

[0288] For example, at an assembly line workstation, a fall risk value of 85% is caused by oil stains on the ground (ERFT slip risk dimension), and a red cube is generated at the corresponding coordinates (7.2, 3.5, 0), with the transparency changing dynamically with the risk value.

[0289] The Intervention Effect Trend Curve (IETC) consists of two axes: the primary axis represents the real-time risk value (taken from IRFV); the secondary axis represents the historical intervention effectiveness rate (such as microinstruction execution rate).

[0290] After the system sends the "eye exercises every hour" instruction, the curve shows that the predicted risk of visual fatigue in the next 30 minutes decreases from 72% to 45% (blue dashed line), compared with the uninterrupted scenario (red solid line). Key inflection points are marked on the curve, such as the point where the maximum decrease (Δ = 27%) occurs 8 minutes after the instruction is issued.

[0291] Interactive Digital Twin (IDT) enables triple-penetration analysis:

[0292] Click on a heatmap block: Drill down to display historical risk events at that location (e.g., 7 high-temperature warnings occurred within 3 days at coordinates (5.1, 2.3));

[0293] Hovering over the propagation tree node: Expanding the details of the causal chain (e.g., the path "noise source → distraction → operational error" contains 12 sensor evidences);

[0294] Drag the timeline to replay the risk evolution process (e.g., retracing the 2-hour process of CO2 concentration rising from 800ppm to 1200ppm).

[0295] The system automatically generates a Risk Traceability Path Report (RTPR). For example, for a current lower back pain risk value of 85%, the report lists the primary cause as "insufficient lumbar support in the chair" (contribution 42%) and the secondary cause as "incorrect lifting posture" (contribution 31%), along with a 3D animation demonstration of improvement solutions.

[0296] Integrating multidimensional data generates a visualized profile including current risk level, historical trends, and major risk sources. Digital twin technology simulates risk propagation paths, supporting managers in macro-level decision-making (such as workplace modifications) and individual care (such as health guidance), providing a comprehensive regulatory view from the individual to the city level, and enabling preventative health management. The interpretability of the risk profile helps develop targeted improvement measures to reduce the overall incidence of occupational injuries and illnesses.

[0297] As can be seen, by continuously collecting behavioral data streams of target employees and dynamic data streams of the physical environment of related work areas, an individual-environment dynamic data sequence is generated; based on the target employee's historical behavioral patterns, abnormal behavioral signals are detected; combined with the current dynamic data stream of the physical environment, the correlation strength between the behavior and occupational health risks is assessed, and personalized micro-intervention instructions are generated and triggered; the employee's immediate feedback response to the micro-intervention instructions is captured in real time, the target employee's personal behavioral baseline model is updated in real time, and the correlation strength assessment rules are calibrated; based on the personal behavioral baseline model, the correlation strength assessment rules, and the individual-environment dynamic data sequence, a real-time occupational health risk profile of the employee is dynamically generated, thereby enabling accurate early warning and personalized intervention for occupational health risks.

[0298] Another embodiment of the present invention provides a real-time monitoring system for urban occupational health risks based on data analysis, see [link to relevant documentation]. Figure 5 The system may include:

[0299] The fusion module 501 is used to continuously collect the behavioral data stream of the target employee and the dynamic data stream of the physical environment of the associated work area, and perform real-time spatiotemporal alignment and fusion to generate an individual-environment dynamic data sequence.

[0300] The maintenance module 502 is used to dynamically maintain the personal behavior baseline model of the target employee based on the target employee's historical behavior pattern using an online learning algorithm, compare the behavior data with the personal behavior baseline model, and detect abnormal behavior signals that represent potential occupational health risks.

[0301] The assessment module 503 is used to interpret the abnormal behavior signal in the context of the current physical environment dynamic data stream, assess its correlation strength with occupational health risks, and generate and trigger personalized micro-intervention instructions when the correlation strength exceeds the strength threshold.

[0302] The update module 504 is used to capture the employee's immediate feedback response to the micro-intervention command in real time, and use the immediate feedback response, the corresponding abnormal behavior signal and environmental context data to update the target employee's personal behavior baseline model in real time and calibrate the evaluation rules of the correlation strength.

[0303] The generation module 505 is used to dynamically generate a real-time occupational health risk profile of the employee based on the updated individual behavior baseline model, the calibrated correlation strength assessment rules, and the continuously input individual-environment dynamic data sequence, so as to realize real-time monitoring of urban occupational health risks.

[0304] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0305] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps:

[0306] S201 continuously collects behavioral data streams of target employees and dynamic data streams of the physical environment of related work areas, and performs real-time spatiotemporal alignment and fusion to generate individual-environment dynamic data sequences;

[0307] S202, Based on the historical behavioral patterns of target employees, use online learning algorithms to dynamically maintain their personal behavioral baseline model, compare the behavioral data with the personal behavioral baseline model, and detect abnormal behavioral signals that represent potential occupational health risks.

[0308] S203, combining the current physical environment dynamic data stream, interpret the abnormal behavior signal in the environmental context, assess its correlation strength with occupational health risks, and when the correlation strength exceeds the strength threshold, generate and trigger personalized micro-intervention instructions.

[0309] S204, Real-time capture of employees' immediate feedback responses to the micro-intervention instructions, and real-time update of the target employee's personal behavior baseline model and calibration of the correlation strength assessment rules using the immediate feedback responses, corresponding abnormal behavior signals and environmental context data;

[0310] S205 dynamically generates a real-time occupational health risk profile for an employee based on an updated individual behavior baseline model, calibrated correlation strength assessment rules, and continuously input individual-environment dynamic data sequences, enabling real-time monitoring of urban occupational health risks.

[0311] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0312] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0313] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0314] S201 continuously collects behavioral data streams of target employees and dynamic data streams of the physical environment of related work areas, and performs real-time spatiotemporal alignment and fusion to generate individual-environment dynamic data sequences;

[0315] S202, Based on the historical behavioral patterns of target employees, use online learning algorithms to dynamically maintain their personal behavioral baseline model, compare the behavioral data with the personal behavioral baseline model, and detect abnormal behavioral signals that represent potential occupational health risks.

[0316] S203, combining the current physical environment dynamic data stream, interpret the abnormal behavior signal in the environmental context, assess its correlation strength with occupational health risks, and when the correlation strength exceeds the strength threshold, generate and trigger personalized micro-intervention instructions.

[0317] S204, Real-time capture of employees' immediate feedback responses to the micro-intervention instructions, and real-time update of the target employee's personal behavior baseline model and calibration of the correlation strength assessment rules using the immediate feedback responses, corresponding abnormal behavior signals and environmental context data;

[0318] S205 dynamically generates a real-time occupational health risk profile for an employee based on an updated individual behavior baseline model, calibrated correlation strength assessment rules, and continuously input individual-environment dynamic data sequences, enabling real-time monitoring of urban occupational health risks.

[0319] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A real-time supervision method for urban occupational health risks based on data analysis, characterized in that, The method comprises: continuously collecting behavior data streams of target employees and physical environment dynamic data streams of associated work areas, and performing real-time space-time alignment and fusion to generate individual-environment dynamic data sequences; based on the historical behavior patterns of the target employees, dynamically maintaining their personal behavior baseline models using online learning algorithms, comparing the behavior data with the personal behavior baseline models, and detecting behavior anomaly signals representing potential occupational health risks; combining the physical environment dynamic data streams at the current moment, interpreting the behavior anomaly signals in the environmental context, evaluating their correlation strength with occupational health risks, and generating and triggering personalized micro-intervention instructions when the correlation strength exceeds a strength threshold; real-time capturing of the immediate feedback responses of employees to the micro-intervention instructions, real-time updating of the personal behavior baseline models of the target employees and calibration of the correlation strength evaluation rules using the immediate feedback responses, corresponding behavior anomaly signals and environmental context data; based on the updated personal behavior baseline models, calibrated correlation strength evaluation rules and continuously input individual-environment dynamic data sequences, dynamically generating real-time occupational health risk portraits of the employees to realize real-time supervision of urban occupational health risks.

2. The method of claim 1, wherein, The method comprises: real-time capturing of employee behavior raw data streams through wearable devices and office system logs, simultaneously deploying a distributed environmental sensor array to collect physical environment dynamic data streams including work area illumination, noise, temperature and humidity, obtaining heterogeneous bimodal data streams; aligning the timestamps of the heterogeneous bimodal data streams using an adaptive clock synchronization protocol, compensating for the sampling rate differences between devices using a dynamic time warping algorithm, and generating a primary fusion data stream with millisecond-level time alignment; based on the workstation spatial topology map, using a geofence-enhanced spatial index engine to map the primary fusion data stream to three-dimensional coordinate points, generating a space-time correlation data stream with spatial location labels; using a multi-modal feature fusion encoder, compressing the behavior feature vectors and environmental feature vectors in the space-time correlation data stream into a multi-dimensional unified tensor sequence, and outputting a standardized individual-environment dynamic data sequence.

3. The method of claim 2, wherein, The method comprises: extracting behavior data sequences of employees for a number of days in the past, establishing a typical behavior state transition matrix through an incremental hidden Markov model, and generating an initial behavior baseline model; for the initial behavior baseline model, using a streaming Bayesian inference algorithm, inputting new behavior data sequences as observation values, real-time updating behavior state transition probability distribution, and outputting a dynamically evolving personal behavior baseline model; inputting the current behavior feature vector into the personal behavior baseline model, decoding the most likely behavior state sequence through the Viterbi algorithm, and outputting a predicted behavior state vector. calculating Mahalanobis distance between the current behavior feature vector and the predicted behavior state vector to generate a behavior deviation degree coefficient; when the behavior deviation degree coefficient exceeds a preset deviation degree threshold in continuous preset sampling periods, marking the behavior feature vector of the corresponding period as a behavior anomaly signal.

4. The method of claim 3, wherein, combining the physical environment dynamic data at the current time with the behavior anomaly signal to perform environmental context interpretation and evaluate the correlation strength with the occupational health risk, and when the correlation strength exceeds a strength threshold, generating and triggering a personalized micro-intervention instruction, including: extracting the physical environment dynamic data of the period corresponding to the behavior anomaly signal, fusing the illumination fluctuation gradient, temperature and humidity change rate, and noise spectrum feature to generate an environmental risk feature tensor; inputting the behavior anomaly signal and the environmental risk feature tensor into a pre-trained risk correlation degree evaluation model to output a real-time correlation strength coefficient in the interval of 0-1; when the correlation strength coefficient exceeds a preset strength threshold, retrieving an optimal micro-intervention instruction template from an intervention strategy knowledge graph based on the behavior anomaly signal type and the environmental risk feature tensor; instantiating the micro-intervention instruction template into a micro-intervention instruction set containing a text prompt and a visual guide, and triggering the micro-intervention instruction set to the terminal through a low-latency message middleware.

5. The method of claim 4, wherein, real-time capturing the immediate feedback response of the employee to the micro-intervention instruction, and using the immediate feedback response, the corresponding behavior anomaly signal and the environmental context data to update the personal behavior baseline model of the target employee in real time and calibrate the evaluation rule of the correlation strength, including: collecting the immediate feedback response of the employee to the micro-intervention instruction to construct a multi-dimensional feedback feature vector; generating a feedback effect weight matrix according to the deviation degree of each dimension in the feedback feature vector and the expected feedback; injecting the behavior anomaly signal weighted by the feedback effect weight matrix into an online learning algorithm as a new sample to trigger parameter adjustment of the personal behavior baseline model; using the environmental context data as a conditional variable, the behavior anomaly signal as an observation variable, and the feedback feature vector as a reinforcement signal to dynamically update the weight parameters of the risk correlation degree evaluation model through a Bayesian network.

6. The method of claim 5, wherein, based on the updated personal behavior baseline model, the calibrated correlation strength evaluation rule and the continuously input individual-environment dynamic data sequence, dynamically generating a real-time occupational health risk portrait of the employee to realize real-time supervision of urban occupational health risks, including: aligning the behavior deviation degree coefficient, the environmental risk feature tensor and the correlation strength coefficient on the time axis, and fusing them into a comprehensive risk feature vector through a gated attention mechanism; inputting the comprehensive risk feature vector into a long short-term memory network to predict the risk occurrence probability in a future preset period; constructing a causal relationship topology of the behavior anomaly signal and the environmental risk based on a graph neural network, labeling a key risk transmission path to form a risk propagation tree; integrating the risk occurrence probability and the risk propagation tree to generate an interactive digital twin containing a three-dimensional risk heat map, an intervention effect trend curve and a risk traceability path, and realizing real-time risk portrait visualization.

7. A real-time supervision system for urban occupational health risks based on data analysis, characterized in that, the system comprises: a fusion module, configured to continuously collect a behavior data stream of a target employee and a physical environment dynamic data stream of an associated work area, and perform real-time spatio-temporal alignment and fusion to generate an individual-environment dynamic data sequence; a maintenance module, configured to dynamically maintain a personal behavior baseline model of the target employee based on a historical behavior pattern of the target employee, compare the behavior data with the personal behavior baseline model, and detect a behavior anomaly signal representing a potential occupational health risk; an evaluation module, configured to perform environmental context interpretation on the behavior anomaly signal in combination with a physical environment dynamic data stream at a current time, evaluate a correlation strength of the behavior anomaly signal with the occupational health risk, and generate and trigger a personalized micro-intervention instruction when the correlation strength exceeds a strength threshold; an update module, configured to capture an immediate feedback response of the employee to the micro-intervention instruction in real time, and update the personal behavior baseline model of the target employee and calibrate an evaluation rule of the correlation strength in real time by using the immediate feedback response, the corresponding behavior anomaly signal, and environmental context data; a generation module, configured to dynamically generate a real-time occupational health risk portrait of the employee based on the updated personal behavior baseline model, the calibrated evaluation rule of the correlation strength, and the continuously input individual-environment dynamic data sequence, and realize real-time supervision of urban occupational health risks.

8. The system of claim 7, wherein, The fusion module is specifically configured to: capture employee behavior raw data streams in real time through wearable devices and office system logs, simultaneously deploy a distributed environmental sensor array to collect physical environment dynamic data streams containing work area illumination, noise, temperature and humidity, and obtain heterogeneous bimodal data streams; align time stamps of the heterogeneous bimodal data streams by using an adaptive clock synchronization protocol, compensate for sampling rate differences between devices by using a dynamic time warping algorithm, and generate a primary fusion data stream with millisecond-level time alignment; based on a work station space topology map, map the primary fusion data stream to three-dimensional coordinate points by using a geo-fence enhanced spatial index engine, and generate a spatio-temporal correlation data stream with a spatial position label; compress behavior feature vectors and environment feature vectors in the spatio-temporal correlation data stream into a multi-dimensional unified tensor sequence by using a multi-modal feature fusion encoder, and output a standardized individual-environment dynamic data sequence.

9. A storage medium, characterized by The storage medium stores a computer program, wherein the computer program is configured to execute the method in any one of claims 1-6 when running.

10. An electronic device comprising a memory and a processor, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method in any one of claims 1-6 when running.

Images