Health state early warning system and method integrated on power grid safety belt

Abstract

The invention relates to the technical field of integrated power grid safety belt early warning, in particular to a health state early warning system and method integrated on a power grid safety belt, and the system comprises a health data collection unit which comprises a photoplethysmography (PPG) sensor which is internally integrated on the inner side of a shoulder belt of the power grid safety belt, is attached to the skin of an operator and has a motion interference resisting function; the local data processing module is internally provided with a data processing and analyzing unit which is integrated on the safety belt and is used for receiving synchronous physiological data and environment data from the local data processing module, executing data preprocessing, eliminating limb movement interference in PPG data by applying a Kalman filtering algorithm, and analyzing a health state intelligent analysis model; the early warning unit is used for receiving the output result of the intelligent analysis model, and according to the health state early warning system and method integrated with the power grid safety belt, noise is actively filtered out through a Kalman filtering algorithm, secondary screening is carried out in combination with environment data, and the reliability of original data is improved.

Description

Technical Field

[0001] This invention relates to the field of integrated power grid safety belt early warning technology, specifically to a health status early warning system and method integrated into the power grid safety belt. Background Technology

[0002] In the field of power grid operation safety management, effective monitoring of workers' health status is a core element for achieving risk warning and ensuring personal safety. Currently, the application of health status early warning technology faces a series of technical bottlenecks. Existing health monitoring equipment is significantly incompatible with power grid operation scenarios. Frequent physical movements during operations, such as raising hands or bending over during high-altitude maintenance, can easily cause fluctuations and distortions in key physiological signals such as heart rate and blood oxygen saturation collected by photoplethysmography sensors. Furthermore, conventional monitoring equipment often operates independently and fails to integrate the collection of environmental factors such as temperature and humidity, making it impossible for the system to accurately analyze whether changes in physiological indicators stem from potential health risks of the workers or interference from external environmental changes.

[0003] At the data processing level, existing technologies have limited capabilities for preprocessing health data. Subtle noise from complex working environments, such as high altitudes and windy, sandy areas, is difficult to filter out effectively. Current methods typically only identify and remove significant outliers, lacking effective means to suppress persistent, low-amplitude interference signals caused by environmental changes. This results in a low signal-to-noise ratio of the raw data input to the backend analysis model, making it difficult to meet the needs of refined analysis.

[0004] The current methods for analyzing and judging health status rely too heavily on preset fixed thresholds. Power grid operations exhibit significant dynamic characteristics, with the effects of short-term high-intensity operations and long-term continuous operations on physiological indicators differing considerably. Existing models struggle to distinguish between normal physiological stress responses triggered by changes in work intensity and genuine health risk signals, easily leading to false alarms or missed alarms due to dynamic changes in work conditions. For example, a brief increase in heart rate after climbing a tower might be misjudged as a risk, while the slow deterioration of physiological indicators due to accumulated fatigue might be overlooked.

[0005] Existing analytical models typically lack self-optimization and adaptive capabilities. Their construction fails to adequately incorporate individual physiological differences among workers, such as variations in age and physical fitness levels leading to fluctuations in normal indicator ranges. This makes it difficult to apply the same judgment criteria fairly and effectively to workers with different physical characteristics. More importantly, with changes in the work environment, personnel turnover, or adjustments in work patterns, models lacking continuous learning capabilities often experience a gradual decline in accuracy over time, failing to provide reliable early warning support in the long term. These combined issues prevent existing health status early warning mechanisms from fully leveraging their risk prediction role in the complex and ever-changing working scenarios of the power grid, thus hindering the improvement of overall safety assurance levels.

[0006] To address this, we propose a health status early warning system and method integrated into the power grid safety belt. Summary of the Invention

[0007] One of the technical problems that this application aims to solve is that the existing technologies have relatively limited capabilities for preprocessing health data, the methods for analyzing and judging health status rely too heavily on preset fixed thresholds, and existing analysis models typically lack self-optimization and adaptability.

[0008] To address the aforementioned technical problems, embodiments of this application provide a health status early warning system integrated into a power grid safety belt, comprising:

[0009] The health data acquisition unit includes a photoplethysmography (PPG) sensor with motion interference resistance, which is built into the inside of the shoulder strap of the power grid safety belt and adheres to the skin of the operator. It is used to collect the operator's core physiological indicators such as heart rate, blood oxygen saturation, and heart rate variability.

[0010] Miniature temperature and humidity sensor that simultaneously collects temperature and humidity data of the working environment;

[0011] The local data processing module is built into the power grid safety belt, including a synchronization clock control unit, which controls the PPG sensor and the miniature temperature and humidity sensor to collect data synchronously at the same rhythm;

[0012] The data processing and analysis unit is used to receive synchronous physiological and environmental data from the local data processing module, perform data preprocessing, and finally input the data into the pre-trained intelligent health status analysis model for analysis.

[0013] The early warning unit is used to receive the output results of the intelligent analysis model, determine the health risk level according to the preset probability threshold, and trigger the early warning mechanism at the field end and the management end.

[0014] In some embodiments, the local data processing module is further configured to set the acquisition parameters of the PPG sensor and the temperature and humidity sensor, and control the two types of sensors to continuously acquire data at the same rhythm through a unified synchronization clock signal, so as to ensure that each set of physiological data corresponds to the environmental data of the same period.

[0015] In some embodiments, data preprocessing in the data processing and analysis unit specifically includes using a Kalman filter algorithm to remove signal fluctuation interference caused by the worker's limb movements, and simultaneously combining environmental data for secondary screening, marking physiological data segments when environmental parameters change abruptly, and determining whether to remove them in conjunction with the work behavior records of the same period.

[0016] The data processing and analysis unit performs data preprocessing, applies the Kalman filter algorithm to remove limb activity interference in PPG data, and identifies and removes confirmed environmental interference data based on changes in environmental data and concurrent work behavior records. The preprocessed physiological data and environmental data are precisely aligned in time sequence to construct a structured data frame containing acquisition time, heart rate, blood oxygen saturation, heart rate variability, and environmental temperature and humidity. The integrity of the data frame is verified and repaired, and the structured data frame is input into a pre-trained intelligent health status analysis model for analysis.

[0017] The intelligent health status analysis model uses PPG physiological data and environmental data from historical power grid operation scenarios as the training set to train a multi-dimensional temporal long short-term memory (LSTM) neural network model, which is then used to output multi-class probability distribution results representing health status.

[0018] In some embodiments, the architecture of the intelligent health status analysis model includes:

[0019] The input layer takes structured data blocks, packaged in fixed time windows and containing continuous physiological index values ​​and average environmental values, as input.

[0020] The hidden layer uses a long short-term memory (LSTM) neural network to capture long-term dependencies in time-series data, and a dropout layer is added after the LSTM layer to prevent overfitting; the output layer outputs a multi-class probability distribution of four health states: "no risk", "low risk", "medium risk" and "high risk".

[0021] In some embodiments, the system further includes an adaptive optimization module, used to: periodically collect new monitoring data from the work site and establish an incremental data pool; and use an incremental learning method with local parameter updates to adjust only the model parameters with high correlation to the newly added data features for optimization.

[0022] Establish a real-time performance monitoring mechanism, calculate performance indicators based on early warning results and actual health verification, and compare them with preset thresholds. When indicators exceed limits, conduct analysis and optimize the model by incremental training with supplementary data or resetting some parameters.

[0023] In some embodiments, a health status early warning method integrated into a power grid safety belt includes the following steps:

[0024] S1: Contextualized collection and preprocessing of health data, using a photoplethysmography sensor with anti-motion interference function, integrated into the inner side of the shoulder strap of the power grid safety belt and fixed with medical-grade silicone, to collect the core physiological indicators of the workers;

[0025] The ambient temperature and humidity are collected by a miniature temperature and humidity sensor that collects data synchronously; a unified synchronization clock is set to control the two types of sensors to continuously collect data at the same rhythm.

[0026] Kalman filtering algorithm was used to remove limb activity interference from PPG physiological data, and environmental interference data was identified, marked and removed based on sudden changes in environmental parameters and concurrent work behavior records.

[0027] S2: Integration and processing of health data and environmental data, based on timestamp information, to precisely align preprocessed physiological data and environmental data in time sequence;

[0028] Construct structured data frames that include acquisition time, heart rate, blood oxygen saturation, heart rate variability, and ambient temperature and humidity;

[0029] Perform integrity verification and repair on structured data frames;

[0030] S3: Intelligent analysis, modeling and optimization of health data. The integrated and processed structured data frames are input into the pre-trained intelligent health status analysis model for analysis. The model is a multi-dimensional temporal long short-term memory (LSTM) neural network model that is trained after constructing a multi-level annotation system based on historical PPG physiological data and environmental data in the power grid operation scenario and after expert collaborative annotation and cross-verification.

[0031] The model outputs a multi-class probability distribution representing health status;

[0032] S4: Early warning application of health data analysis results. Based on the probability distribution results output by the intelligent analysis model and the preset probability threshold, the final health risk level is determined.

[0033] If the response level is determined to be required, a multi-terminal early warning linkage between the field terminal and the management terminal will be triggered.

[0034] In some embodiments, the training optimization of the model in S3 specifically includes two-stage iterative training: in the first stage, the training weight of clearly risky samples is increased first, so that the model can master the judgment logic of high-risk scenarios.

[0035] The second stage introduces confused sample training and refines the loss function to improve the model's ability to distinguish between normal stress and suspected risk.

[0036] During training, the accuracy, false alarm rate, and false negative rate of the test set are monitored in real time, and parameters are adjusted accordingly. Independent validation datasets for different power grid operation scenarios are used to test performance differences and supplement training accordingly.

[0037] In some embodiments, S3 also includes adaptive optimization of the model: periodically collecting new health monitoring data and filtering data that includes new work scenarios, new employees, and historical false alarm / missed alarm scenarios to establish an incremental data pool;

[0038] An incremental learning approach with local parameter updates is adopted, adjusting only the model parameters that are highly correlated with the features of the newly added data;

[0039] Establish a real-time performance monitoring mechanism. When the accuracy, false positive rate, and false negative rate exceed the limits, analyze the reasons and optimize by supplementing data for incremental training or resetting some parameters.

[0040] In some embodiments, in S1, the construction of the structured data frame specifically involves: integrating the time-aligned associated data groups into a unified format that includes “collection time-heart rate-blood oxygen saturation-heart rate variability-ambient temperature and humidity”, and clarifying the definition and data format standard of each field.

[0041] In some embodiments, triggering the multi-terminal early warning linkage between the field terminal and the management terminal in S4 specifically means: if it is determined to be medium risk or high risk, the field terminal issues a graded prompt through the buzzer and LED indicator built into the safety belt;

[0042] The management system will simultaneously feed back the risk level, corresponding physiological data, and work environment information to the power grid operation safety management platform.

[0043] This invention has at least the following beneficial effects:

[0044] 1. By integrating a photoplethysmography sensor with anti-motion interference capabilities into the inside of the safety belt shoulder strap using medical-grade silicone material, the sensor maintains a stable fit against the skin even during frequent limb movements, fundamentally reducing distortion of core physiological data such as heart rate and blood oxygen saturation caused by equipment displacement. Simultaneously deployed miniature temperature and humidity sensors enable perfectly time-aligned acquisition of environmental parameters and physiological indicators, helping the system establish a direct correlation between physiological data and environmental conditions. This provides a data foundation for accurately distinguishing between the effects of environmental and health factors.

[0045] 2. In the data processing stage, the system actively filters out limb activity noise using the Kalman filter algorithm and performs secondary screening by combining the characteristics of sudden changes in environmental data and work status records, significantly improving the reliability of the raw data. Precise time-series alignment and structured data frame construction further enhance the fusion quality of physiological and environmental multi-source data, ensuring the integrity of subsequent analysis inputs. This integrated data acquisition and preprocessing approach significantly reduces the impact of interference factors on the quality of health data in complex power grid operating environments.

[0046] 3. The intelligent health status analysis model adopts a time-series modeling architecture based on long short-term memory neural networks. This effectively learns the gradual changes in physiological indicators during dynamic power grid operations, identifying subtle differences between normal work-related stress responses and actual health risks. The model's multi-level risk probability distribution, combined with the threshold judgment mechanism set by power grid safety regulations, avoids the misjudgment tendency of fixed threshold judgment systems when facing different work intensities or individual differences, thus improving the accuracy and flexibility of risk identification.

[0047] 4. The phased optimization strategy implemented during model training, by first focusing on high-risk pattern identification and then refining the training to differentiate between normal stress and potential risks, makes the model's learning path more reasonable and its recognition capabilities more precise. The continuous adaptive optimization mechanism, by regularly incorporating new scenario data, new personnel data, and historical misjudgment scenario data, updates the model's knowledge base using incremental learning of local parameters. This ensures the stability and adaptability of health risk assessment capabilities, overcoming the performance degradation issues of traditional models caused by changes in the work environment or personnel turnover.

[0048] 5. The health risk level output by the system can simultaneously trigger tiered audible and visual warnings on the local safety belt terminal and remote early warning information from the safety management platform, forming a collaborative mechanism between on-site personnel's self-control and management-level emergency response. This linkage method realizes closed-loop management of health risks from identification to response, shortens the time window for handling abnormal situations, creates technical conditions for workers to proactively avoid health risks, and provides accurate basis for management departments to promptly allocate resources.

[0049] 6. This method integrates a complete technical path from high-quality data acquisition, effective interference suppression, multi-source data fusion, intelligent dynamic analysis to risk collaborative response, enabling health status early warning capabilities to be truly integrated into the power grid operation safety assurance system and reducing the risk of operational accidents caused by health problems. The system's comprehensive improvement in multiple dimensions, including worker physiological monitoring, environmental adaptability analysis, and early warning response linkage, contributes to the overall optimization of the operational efficiency of the power grid operation safety management system. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the system composition of the present invention;

[0051] Figure 2 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1, see Figure 1 This invention provides a technical solution: a health status early warning system integrated into the power grid safety belt, comprising:

[0054] The health data acquisition unit includes a photoplethysmography (PPG) sensor with motion interference resistance, which is built into the inside of the shoulder strap of the power grid safety belt and adheres to the skin of the operator. It is used to collect the operator's core physiological indicators such as heart rate, blood oxygen saturation, and heart rate variability.

[0055] Includes a miniature temperature and humidity sensor that synchronously collects the temperature and humidity of the working environment;

[0056] The PPG sensor is fixed with medical-grade silicone material;

[0057] The local data processing module is built into the power grid safety belt, including a synchronization clock control unit, which controls the PPG sensor and the miniature temperature and humidity sensor to collect data synchronously at the same rhythm;

[0058] The data processing and analysis unit is used to: receive synchronous physiological and environmental data from the local data processing module, perform data preprocessing, and finally input the data into the pre-trained intelligent health status analysis model for analysis.

[0059] The early warning unit is used to receive the output results of the intelligent analysis model, determine the health risk level according to the preset probability threshold, and trigger the early warning mechanism at the field end and the management end.

[0060] The local data processing module is further used to set the acquisition parameters of the PPG sensor and the temperature and humidity sensor, and to control the two types of sensors to continuously acquire data at the same rhythm through a unified synchronous clock signal, so as to ensure that each set of physiological data corresponds to the environmental data of the same period.

[0061] The data preprocessing in the data processing and analysis unit specifically includes using the Kalman filter algorithm to remove signal fluctuation interference caused by the limb movements of workers, and simultaneously combining environmental data for secondary screening, marking physiological data segments when environmental parameters change suddenly, and determining whether to remove them in conjunction with the work behavior records of the same period.

[0062] Data preprocessing was performed, and the Kalman filter algorithm was applied to remove limb activity interference from PPG data. Based on changes in environmental data and the records of work behavior during the same period, environmental interference data was identified, marked, and removed. The preprocessed physiological data and environmental data were precisely aligned in time to construct a structured data frame containing acquisition time, heart rate, blood oxygen saturation, heart rate variability, and environmental temperature and humidity. The integrity of the data frame was verified and repaired. The structured data frame was then input into a pre-trained intelligent health status analysis model for analysis.

[0063] The intelligent health status analysis model uses PPG physiological data and environmental data from historical power grid operation scenarios as the training set. A multi-level annotation system is constructed and expert collaborative annotation and cross-verification are performed to train a multi-dimensional temporal long short-term memory (LSTM) neural network model, which is used to output multi-class probability distribution results representing health status.

[0064] The architecture of the intelligent health status analysis model includes:

[0065] The input layer takes structured data blocks, packaged in fixed time windows and containing continuous physiological index values ​​and average environmental values, as input.

[0066] The hidden layer uses a long short-term memory (LSTM) neural network to capture long-term dependencies in time-series data, and a dropout layer is added after the LSTM layer to prevent overfitting; the output layer outputs a multi-class probability distribution of four health states: no risk, low risk, medium risk, and high risk.

[0067] The system also includes an adaptive optimization module, which is used to: periodically collect new monitoring data from the work site and establish an incremental data pool; and adopt an incremental learning approach with local parameter updates to optimize only the model parameters with high correlation to the features of the newly added data.

[0068] Establish a real-time performance monitoring mechanism, calculate performance indicators based on early warning results and actual health verification, and compare them with preset thresholds. When indicators exceed limits, conduct analysis and optimize the model by incremental training with supplementary data or resetting some parameters.

[0069] Specifically, when workers are operating in complex environments such as high altitudes or with electrical connections, frequent limb movements and environmental interference are key factors causing traditional health monitoring to fail. Therefore, the system integrates a motion-resistant PPG sensor directly into the inside of the safety belt shoulder strap using medical-grade silicone, utilizing the tension of the safety belt itself to ensure the sensor fits snugly against the skin near the collarbone. This fixation method minimizes sensor displacement when personnel raise their arms or bend over, physically ensuring the stability of core physiological signal acquisition. Simultaneously deployed miniature temperature and humidity sensors achieve millisecond-level time-aligned acquisition via a unified clock signal, establishing a precise correlation between physiological indicators and environmental parameters.

[0070] The design aims to address three technological gaps. First, there's the data quality gap. The PPG signal's susceptibility to motion artifacts and environmental abrupt changes is addressed through a dual filtering mechanism. The Kalman filter algorithm dynamically predicts and corrects regular noise generated by limb movements, preserving the true physiological trend baseline. Combined with a secondary filtering mechanism based on sudden environmental changes and work behavior logs, it specifically handles abnormal pulses caused by short-term, strong interference. Second, there's the analytical capability gap. The system employs an LSTM neural network to construct an intelligent health status analysis model, utilizing its unique temporal memory unit to capture the gradual changes in physiological indicators during risk formation, such as fatigue accumulation patterns like a sustained decrease in heart rate variability or a slow decline in blood oxygen saturation. This modeling approach effectively distinguishes between transient work stress and substantial health deterioration. Finally, there's the system's adaptability gap. The model learns normal physiological baselines for different ages and physical conditions through a multi-level expert annotation system, outputting multi-dimensional probability distribution results rather than rigid threshold judgments, and establishing an incremental learning mechanism to continuously digest new scenario data.

[0071] At the data acquisition end, the combination of medical-grade silicone fixation and motion-interference-resistant sensors significantly improves the effective signal capture rate under dynamic operation, avoiding monitoring interruptions caused by equipment slippage. Synchronous acquisition of environmental parameters provides crucial evidence for distinguishing between heat-induced heart rate increases and pathological tachycardia. The temporal alignment and structured data frame standardization in the data integration stage enable subsequent models to directly receive physiological data blocks incorporating environmental influences, reducing preprocessing computational load. The core value of the intelligent health status analysis model lies in its dynamic scene analysis capability. The LSTM network's extraction of long-term sequence features allows it to identify risk signs that fixed threshold systems cannot capture, such as delayed heart rate recovery after continuous high-altitude work. The four-level probability output, combined with safety management regulations to set response thresholds, retains the flexibility of model judgment while avoiding over-reliance on a single output value. The dual-end linkage design of the early warning mechanism forms a rapid response closed loop—on-site audio-visual prompts enable workers to instantly perceive risks and autonomously avoid them, while synchronized data from the management platform supports remote decision-making and intervention. This design ultimately achieves a seamless technical pathway from accurate identification of health risks to effective intervention, constructing an active personal protection barrier for high-risk power grid operation scenarios.

[0072] The local data processing module is further used to set the acquisition parameters for the PPG sensor and the temperature and humidity sensor, and to control the two types of sensors to continuously acquire data at the same rhythm through a unified synchronous clock signal. This ensures that each set of physiological data corresponds to environmental data at the same time. The principle of this design stems from the strict requirements of time dimension consistency in health risk analysis. In power grid operations, the impact of environmental factors on physiological indicators often has immediate or short-term delay characteristics. If there is a time mismatch between the acquisition of physiological data and environmental data, it will lead to the establishment of false correlations or the omission of true correlations in subsequent analyses. By uniformly setting the sampling frequency, resolution, and other parameters of the two types of sensors through the local data processing module, and generating a hardware-level synchronous clock signal to drive the acquisition timing, it is fundamentally ensured that there is a strictly corresponding environmental data point at the time of generation of each physiological data point.

[0073] The design directly addresses the temporal disjointness problem commonly found in traditional multi-sensor systems. In systems without synchronized control, even when sensors are deployed on the same device, communication delays, sampling start-up time differences, or variations in caching mechanisms result in a time offset of hundreds of milliseconds to several seconds between the acquired physiological and environmental data. This offset can cause severe analytical distortion in highly dynamic power grid operation scenarios. Synchronized clock signal control enables the two data streams to achieve a perfectly aligned timeline, laying the technical foundation for constructing a realistic "environment-physiology" response relationship.

[0074] The core benefits of this mechanism are reflected in three aspects. First, in terms of data reliability, it eliminates the risk of misjudgment due to environmental interference caused by time differences in data collection. For example, a normal physiological response of increased heart rate after a worker enters a high-temperature area may be mislabeled as an unexplained abnormality if environmental data collection lags behind physiological data collection. Second, in terms of analysis efficiency, time-aligned data can be directly correlated and calculated in the subsequent preprocessing stage, avoiding the introduction of complex time compensation algorithms and reducing the computational load on the edge computing module. Finally, in terms of model accuracy, the synchronously collected "environment-physiology" paired data blocks enable the LSTM model to accurately learn dynamic correlation features such as heart rate fluctuation patterns during sudden changes in temperature and humidity, improving the causal relationship inference capability of health risk assessment. This millisecond-level synchronization accuracy provides irreplaceable technical support for health risk attribution analysis in complex work scenarios.

[0075] The data preprocessing in the data processing and analysis unit specifically includes using the Kalman filter algorithm to remove signal fluctuation interference caused by workers' limb movements, and simultaneously combining environmental data for secondary screening. Physiological data segments at the time of sudden changes in environmental parameters are marked and their exclusion is determined in conjunction with concurrent work behavior records. This design is based on the dual interference characteristics faced by health monitoring in power grid operations. Motion artifacts generated by workers' limb movements and physiological responses caused by sudden environmental changes present distinctly different interference patterns in the original PPG signal. Limb movement interference typically manifests as high-frequency, short-duration, and periodic fluctuations, while sudden environmental changes (such as sudden strong winds or direct sunlight) may cause physiological indicator drift with lower amplitude but longer duration. The Kalman filter algorithm establishes a state-space model of human physiological signals, uses the optimal estimate from the previous moment to predict the current signal value, and then dynamically corrects it by combining actual observations. This recursive filtering method is particularly suitable for processing regular noise caused by limb movements, effectively smoothing sudden motion interference while preserving true physiological changes such as heart rate trends.

[0076] Traditional systems often employ only motion filtering algorithms, but the impact of environmental factors on physiological indicators during power grid operations cannot be ignored. For example, when overhauling transformers in high-temperature environments, equipment heat dissipation may cause a sudden increase in local temperature exceeding 5°C. In such cases, the increased heart rate observed in the PPG signal may be a normal heat stress response rather than a health risk. Through a synchronous environmental data secondary screening mechanism, the system can identify such environmental abrupt changes and, combined with operational behavior records (such as the "operation in high-temperature equipment zone" marker in the system log), trace the source of interference. This dual verification mechanism avoids misjudging environmentally adaptive physiological changes as health abnormalities.

[0077] The core advantages of this processing flow are reflected in three dimensions. Regarding signal fidelity, the dynamic predictive characteristics of Kalman filtering make it significantly better than fixed-threshold filtering at suppressing periodic motion interference such as raising hands and bending over, ensuring that key indicators sensitive to noise, such as heart rate variability, are not incorrectly corrected. In terms of interference identification breadth, environmental data correlation analysis enables the system to capture special interference scenarios that traditional methods cannot identify, such as the drift in blood oxygen saturation measurements caused by sudden dust storms. Regarding data usability, the judgment logic combined with work behavior records significantly reduces the false rejection rate—if workers are resting during periods of sudden environmental changes, their physiological changes are more likely to reflect their true health condition rather than interference; in this case, retaining the data can actually help identify potential risks. This hierarchical, multi-chain evidence preprocessing architecture provides high signal-to-noise ratio input data for subsequent intelligent analysis.

[0078] The architecture of the intelligent health status analysis model includes: an input layer, which takes structured data blocks containing continuous physiological index values ​​and average environmental values ​​packaged in fixed time windows as input; a hidden layer, which uses a Long Short-Term Memory (LSTM) neural network to capture long-term dependencies in time-series data, and adds a dropout layer after the LSTM layer to prevent overfitting; and an output layer, which outputs a multi-class probability distribution of four health statuses: no risk, low risk, medium risk, and high risk.

[0079] The formation of real health risks often exhibits a gradual, cumulative, temporal pattern. For example, dehydration caused by continuous work in high-temperature environments may manifest as a coordinated trend of gradually increasing heart rate and slowly decreasing heart rate variability. The input layer employs a fixed-time-window packaging strategy, which essentially captures these physiological evolution processes that last for several minutes or even longer through discretization. Each data block not only contains continuous physiological indicator values ​​but also integrates the environmental average values ​​for that period, enabling the model to simultaneously perceive short-term physiological fluctuations and medium- to long-term environmental stress factors.

[0080] The design aims to address two major blind spots in traditional methods. First, there's the blind spot in dynamic scene analysis, where fixed-threshold models cannot distinguish the fundamental differences between temporary activity stress and pathological changes. Second, there's the risk of overfitting; power grid operation scenarios are complex and varied, and single models are easily misled by non-universal features specific to those scenarios. The memory unit structure of the LSTM network can autonomously identify long-distance dependencies in time series, such as capturing subtle changes in heart rate and blood oxygen recovery periods as the operation duration increases during a two-hour tower climbing operation. The subsequently added dropout layer randomly disconnects some neurons during training, forcing the model to distribute feature learning weights and reducing dependence on specific noise or sporadic patterns, for example, avoiding the generalization of physiological responses at a specific temperature (e.g., 30±2℃) to universal patterns.

[0081] LSTM's temporal modeling capabilities effectively analyze the gradual abnormal evolution of physiological indicators, such as identifying the stepwise decline in heart rate variability (HRV) during fatigue accumulation. This feature capture mechanism enables the model to identify potential risks that a fixed threshold system would inevitably miss—after a high-intensity operation, even if the HRV value remains within the traditional safety threshold, a significant increase in the recovery period compared to the previous one can trigger a low-risk warning. The multi-class probability output structure provides the system with flexible judgment capabilities, outputting probability distributions for subsequent decision-making references in ambiguous states (such as early heatstroke symptoms being similar to ordinary dehydration physiological indicators). The output layer uses a four-level rather than continuous value classification, avoiding both the oversensitivity of regression models to absolute values ​​and the arbitrariness of dichotomous methods in determining intermediate states. Based on this refined hierarchical mechanism, the system can significantly reduce the false alarm rate while maintaining early risk detection capabilities. The model generalization improvement introduced by the dropout layer ensures that the warning rules are applicable to various power grid operation scenarios, such as maintaining stable performance in high-altitude substations and coastal transmission line operations.

[0082] The system also includes an adaptive optimization module, used for: periodically collecting new monitoring data from the work site and establishing an incremental data pool; employing an incremental learning approach with local parameter updates, adjusting only model parameters with high correlation to newly added data features for optimization; and establishing a real-time performance monitoring mechanism to calculate performance indicators based on early warning results and actual health verification, comparing them with preset thresholds. When indicators exceed limits, analysis is performed, and the model is optimized through incremental training with supplementary data or partial parameter reset. This design principle is based on the core contradiction faced by the long-term operation of the power grid health monitoring system: the conflict between the dynamic changes in the work environment, personnel turnover, and the emergence of new risk patterns, and the inherent limitations of static analysis models. The design essence of the adaptive optimization module is to build a metabolic mechanism for continuous model evolution, and its operation follows the principle of "minimal intervention and precise optimization." The incremental data pool strategy acknowledges the uneven value density of new scenario data—model updates are only triggered when specific feature dimensions (such as data from high-altitude low-temperature environments) accumulate to the verification threshold, avoiding fragmented data contaminating the model's knowledge system.

[0083] The power grid operation system exhibits significant spatiotemporal migration characteristics, with systematic differences in physiological response patterns between humid spring environments and high-temperature summer environments; the baseline resting heart rate of workers of different ages can differ by more than 20 bpm. An incremental learning approach with local parameter updates is employed, focusing on identifying the correlation dimensions between new data and the existing knowledge base. For example, when the system adds high-altitude operation data, only the weights of the LSTM layer related to ambient oxygen content are adjusted, while the baseline heart rate analysis capability is retained. This targeted optimization model both absorbs new knowledge and prevents the risk of performance degradation in existing scenarios due to overtraining.

[0084] The real-time performance monitoring module establishes a quantitative evaluation system for model health by continuously tracking indicators such as early warning accuracy and false alarm rate. When the system detects a continuous increase in the false alarm rate under specific environmental scenarios (such as sandstorm weather), it can automatically mark the data gap for that scenario and trigger a targeted data collection request. At the model optimization level, local parameter updates reduce the computational resource consumption for retraining by about 70%, enabling edge computing devices to undertake quarterly model maintenance tasks. More importantly, the partial parameter reset function provides the model with fault-tolerant self-healing capabilities—when a certain type of abnormal input (such as outliers caused by sensor failure) causes distortion of the weights of specific neurons, the system can recover to a stable state through a weight rollback mechanism. This layered and progressive adaptive architecture ultimately enables the health risk analysis capability to break through the timeliness boundaries of traditional models, maintaining long-term stability of early warning accuracy against the backdrop of continuously changing power grid operating environments.

[0085] "Unified rhythm" refers to forcing two types of sensors to operate under the exact same timing reference through a hardware-level clock synchronization mechanism. All sensor sampling times are strictly bound to the central clock pulse, eliminating inherent latency differences between devices. For example, a global acquisition command is triggered every 200 milliseconds to ensure that the timestamp error between the PPG signal and the ambient temperature and humidity readings is less than 5 milliseconds. This microsecond-level alignment allows the physiological response of sweat evaporation during high-temperature operations to be precisely correlated with the peak values ​​of environmental parameters. A fixed time window (e.g., 2 seconds / acquisition) is used for the acquisition cycle, forming a continuous and uniform data stream. In scenarios such as tower climbing operations, this avoids data density fluctuations caused by changes in the intensity of movement, ensuring the consistency of the time dimension of the input tensor during subsequent LSTM network processing. When a specific sensor triggers an abnormal threshold (e.g., PPG signal loss), the synchronization mechanism pauses all acquisition channels and records the breakpoint. After the signal is restored, the system automatically aligns the time series before and after the interruption to prevent asymmetric data gaps. For example, if a safety belt buckle is accidentally released, the entire data stream will maintain a complete timeline.

[0086] Example 2, see Figure 2 A health status early warning method integrated into the power grid safety belt, characterized by comprising the following steps:

[0087] S1: Contextualized collection and preprocessing of health data, using a photoplethysmography sensor with anti-motion interference function, integrated into the inner side of the shoulder strap of the power grid safety belt and fixed with medical-grade silicone, to collect the core physiological indicators of the workers;

[0088] The ambient temperature and humidity are collected by a miniature temperature and humidity sensor that collects data synchronously; a unified synchronization clock is set to control the two types of sensors to continuously collect data at the same rhythm.

[0089] Kalman filtering algorithm was used to remove limb activity interference from PPG physiological data, and environmental interference data was identified, marked and removed based on sudden changes in environmental parameters and concurrent work behavior records.

[0090] S2: Integration and processing of health data and environmental data, based on timestamp information, to precisely align preprocessed physiological data and environmental data in time sequence;

[0091] Construct structured data frames that include acquisition time, heart rate, blood oxygen saturation, heart rate variability, and ambient temperature and humidity;

[0092] Perform integrity verification and repair on structured data frames;

[0093] S3: Intelligent analysis, modeling and optimization of health data. The integrated and processed structured data frames are input into the pre-trained intelligent health status analysis model for analysis. The model is a multi-dimensional temporal long short-term memory neural network model that is trained after constructing a multi-level annotation system based on historical PPG physiological data and environmental data in the power grid operation scenario and after expert collaborative annotation and cross-verification.

[0094] The model outputs a multi-class probability distribution representing health status;

[0095] S4: Early warning application of health data analysis results. Based on the probability distribution results output by the intelligent analysis model and the preset probability threshold, the final health risk level is determined.

[0096] If the response level is determined to be required, a multi-terminal early warning linkage between the field terminal and the management terminal will be triggered.

[0097] The training optimization of the model in S3 specifically includes two-stage iterative training: the first stage prioritizes increasing the training weight of clearly risky samples so that the model can master the judgment logic of high-risk scenarios;

[0098] The second stage introduces confused sample training and refines the loss function to improve the model's ability to distinguish between normal stress and suspected risk.

[0099] During training, the accuracy, false alarm rate, and false negative rate of the test set are monitored in real time, and parameters are adjusted accordingly. Independent validation datasets for different power grid operation scenarios are used to test performance differences and supplement training accordingly.

[0100] S3 also includes adaptive optimization of the model: regularly collect new health monitoring data, and filter data including new work scenarios, new employees, and historical false alarm / missed alarm scenarios to build an incremental data pool;

[0101] An incremental learning approach with local parameter updates is adopted, adjusting only the model parameters that are highly correlated with the features of the newly added data;

[0102] Establish a real-time performance monitoring mechanism. When the accuracy, false positive rate, and false negative rate exceed the limits, analyze the reasons and optimize by supplementing data for incremental training or resetting some parameters.

[0103] In S1, the construction of structured data frames is specifically as follows: the time-aligned associated data groups are integrated in a unified format that includes "acquisition time-heart rate-blood oxygen saturation-heart rate variability-ambient temperature and humidity", and the definition and data format standard of each field are clarified.

[0104] In S4, the multi-terminal early warning linkage between the field terminal and the management terminal is triggered as follows: if the risk level is determined to be medium or high, the field terminal will issue a graded warning through the buzzer and LED indicator built into the safety belt.

[0105] The management system will simultaneously feed back the risk level, corresponding physiological data, and work environment information to the power grid operation safety management platform.

[0106] Specifically, the design principle of this early warning method follows the basic law of health risk identification in high-risk power grid operation scenarios—real health hazards usually manifest as a temporal process of the coordinated evolution of multiple physiological indicators, and this process is always deeply coupled with environmental pressure and work behavior. The scenario-based acquisition design in the S1 stage profoundly addresses the signal acquisition dilemma under dynamic working conditions. The medical-grade silicone fixation solution ensures that the PPG sensor continuously acquires effective contact when personnel move frequently, avoiding signal interruption caused by slippage. Dual-sensor synchronous clock control eliminates correlation distortion caused by environmental response delays at the source. For example, the baseline drift of the PPG signal caused by sweat evaporation in high-temperature environments must accurately correspond to the temperature and humidity jump point to have analytical value. Kalman filtering and secondary screening based on work logs form a hierarchical logic for interference elimination. The former suppresses transient limb movement noise, while the latter solves the systematic offset caused by sudden environmental changes. This combined strategy significantly reduces the interference false retention rate while preserving the integrity of the real physiological trend.

[0107] The design goal consistently revolves around building an end-to-end risk identification pathway. The temporal alignment and data frame structure conversion in stage S2 are not simply format processing; their essence is establishing standardized analysis units for the interaction between "environment and physiology." Fixed-position storage of parameters within structured data frames allows subsequent models to obtain physiological evolution sequences incorporating environmental context through a unified interface, avoiding parsing interruptions due to missing fields or formatting issues. The integrity verification mechanism specifically protects against data packet anomalies in extreme operating scenarios (such as areas with strong electromagnetic interference), maintaining data flow continuity through neighboring data interpolation or specific pattern backfilling. The core value of stage S3 lies in overcoming the scenario adaptability bottleneck of traditional static models. During model construction, a multi-level expert annotation system incorporates grid-specific knowledge—such as the difference in HRV fluctuation thresholds caused by psychological load differences in live-line work at different voltage levels—while a cross-validation mechanism ensures that training labels are free from individual subjective bias. The unique gating unit structure of the LSTM network allows it to memorize physiological state characteristics from earlier operating phases. When it identifies a cross-time-period pattern of slowly decreasing heart rate variability accompanied by a stepwise decline in blood oxygen saturation, it can trigger an early warning even if all indicators are still within the traditional safety range. The multi-class probability output structure essentially adopts the idea of ​​fuzzy judgment to avoid artificial binary segmentation of complex physiological states.

[0108] From the acquisition end, the synchronized "physiological-environmental" data stream enables the system to distinguish between normal operational stress and potential risk signs. For example, a heart rate of 140 bpm during tower climbing is considered high-risk in a 15°C environment, while the same value may be within a controllable range in a 32°C environment. Layered interference management in the preprocessing stage ensures the availability of key indicators; the HRV waveform after Kalman filtering clearly shows the characteristics of the fatigue recovery period, avoiding misjudgments of recovery capacity due to noise masking. The time-series modeling capability in the model analysis stage extends the risk identification window from instantaneous states to the entire operation process, typically identifying a heatstroke precursor pattern of a 0.8°C increase in body temperature accompanied by a 10-second delay in heart rate recovery after two hours of continuous work. Four-level probability output combined with a dual-end early warning mechanism forms a graded response closed loop. Strong vibration alarms at the field end are suitable for emergency avoidance in high-risk situations, while long-term risk statistical reports pushed by the management end support work scheduling optimization. The entire method ultimately achieves a dual leap from real-time protection to system improvement, evolving the traditional single protective function of safety belts into an active health protection system.

[0109] The training optimization of the S3 model specifically includes a two-stage iterative training process: The first stage prioritizes increasing the training weight of explicitly risky samples, enabling the model to master the logic for judging high-risk scenarios; the second stage introduces training with obfuscated samples and refines the loss function to improve the model's ability to distinguish between normal stress and suspected risks. During training, the accuracy, false alarm rate, and false negative rate on the test set are monitored in real time, and parameters are adjusted accordingly. Independent validation datasets for different power grid operation scenarios are used to test performance differences and supplement training accordingly. Real high-risk events are sparse samples in the overall operation data, but the cost of misjudging them is extremely high; while the boundary between health and risk (such as dehydration in the early stages of high-temperature operations) occurs frequently. The two-stage iterative training architecture addresses this dual challenge of uneven sample distribution and blurred boundaries. Increasing the weight of explicitly risky samples in the first stage essentially prioritizes establishing the ability to recognize key risk patterns within a limited training period, such as abnormal PPG waveform features caused by premature ventricular contractions or typical heart rate / blood oxygen symmetry patterns in the precursors of heatstroke. This strategy ensures that the model has a basic sensitivity to serious threats, avoiding the submersion of key features by massive amounts of normal data.

[0110] In power grid operation scenarios, confused samples often stem from the combined effect of normal physiological stress and non-pathological environmental responses. For example, when working in strong winds at an altitude of 40 meters, fluctuations in heart rate and blood oxygen may simultaneously include normal physiological responses to wind pressure stimulation and stress responses induced by potential fear of heights. The second stage introduces these confused samples and refines the loss function, essentially guiding the model to learn more precise discrimination boundaries—by modifying the loss function to assign differential penalty weights to different categories of errors. For example, significantly increasing the loss coefficient for missed risk samples and moderately increasing the loss value for false positives due to normal stress, forcing the model to achieve a reasonable balance between risk identification sensitivity and specificity.

[0111] A closed-loop optimization mechanism is formed by real-time monitoring of test set metrics. When an abnormal increase in the false alarm rate is detected in a certain scenario (such as tower climbing operations in low-temperature environments), the learning rate or batch size can be adjusted immediately to correct the optimization direction. Independent validation datasets are used to examine performance differences in different operation scenarios, effectively building a mapping mechanism for the model's capabilities. For example, if the false alarm rate is significantly higher in live-line operation scenarios than in line inspection scenarios, it indicates that the model is oversensitive to physiological fluctuations under electromagnetic interference. In this case, targeted supplementary training will strengthen its anti-electromagnetic interference capabilities. The final model possesses triple adaptability: high-precision capture of clear risks enables an emergency state recognition rate of over 99% for signs of electric shock; fine discrimination of confused samples reduces the false alarm rate of non-pathological abnormalities such as hyperventilation to below 5%; and cross-scenario robustness ensures that performance fluctuations remain below 8% in different environments such as 500kV substations and distribution network lines. This phased, focused, and multi-dimensional optimization training system maximizes model performance improvement from limited labeled data.

[0112] S3 also includes adaptive optimization of the model: It periodically collects new health monitoring data, filters data including new work scenarios, new employees, and historical false alarm / missed alarm scenarios to establish an incremental data pool; it adopts an incremental learning approach with local parameter updates, adjusting only model parameters highly correlated with the features of the new data; and it establishes a real-time performance monitoring mechanism, analyzing the reasons when accuracy, false alarm rate, and missed alarm rate exceed limits, and optimizing through incremental training with supplementary data or partial parameter reset. The principle of this adaptive optimization design is rooted in the core contradiction of the long-term operation of the power grid operation health monitoring system—the conflict between the static analysis model and the continuous evolution of dynamic real-world scenarios. After model deployment, it inevitably faces three variables: new work scenarios (such as high-altitude operations with new insulated bucket trucks), differences in personnel physiological characteristics (such as workers with different physical conditions), and special working conditions not covered by the original model (such as complex physiological responses during sandstorms). This module, by constructing a continuously evolving closed-loop learning mechanism, enables the model to adapt to environmental changes. The selection logic of the incremental data pool reflects the principle of data value density. Only when data of a specific feature dimension (such as a high-altitude low-temperature environment) accumulates to a statistically significant threshold will it be included in the training sequence, thus avoiding fragmented data from interfering with the model's knowledge structure.

[0113] Previous model updates often employed a periodic global retraining approach, which not only consumed enormous computational resources but, more critically, could potentially disrupt the existing knowledge base. Our local parameter update strategy utilizes a feature correlation filtering mechanism. For example, when the system introduces high-altitude operational data, only the weight parameters of the LSTM layer, which are strongly correlated with ambient oxygen levels, are adjusted, while maintaining the stability of the basic heart rate analysis module. This precise, surgical optimization effectively preserves the integrity of existing core capabilities while absorbing new knowledge.

[0114] The real-time performance monitoring module establishes an indicator early warning system (accuracy threshold set at 92%, false alarm rate capped at 5%), serving as a barometer of model health. When the system detects an abnormal surge in the false alarm rate in a specific scenario (e.g., a false alarm rate exceeding 7% in a humidity-sensitive scenario during the rainy season), it automatically traces the source to the corresponding environmental parameter channel for diagnosis. Incremental training employs a feature-driven parameter update strategy, consuming only 30% of the computational resources required for global retraining, enabling edge computing units to handle quarterly model maintenance tasks. A partial parameter reset function endows the system with fault tolerance and self-healing capabilities. In typical scenarios, such as abnormal inputs caused by temporary sensor malfunctions leading to local neuron weight distortions, a rollback mechanism can restore a stable state. More importantly, this architecture forms a data-driven problem discovery-optimization closed loop: when data from newly hired personnel shows a deviation in the resting heart rate characteristic distribution from the traditional pattern, the system automatically marks the data gap for that personnel type and triggers a targeted data collection command. This self-evolutionary capability ultimately allows health risk early warning to break through the timeliness limitations of traditional models, maintaining long-term stability of core performance indicators amidst continuous changes in the power grid operating environment.

[0115] In S1, the construction of the structured data frame involves integrating time-aligned associated data groups into a unified format containing "collection time - heart rate - blood oxygen saturation - heart rate variability - ambient temperature and humidity," clearly defining the definition and data format standard for each field. The principle behind this structured data frame design directly addresses the parsing problem of multi-source heterogeneous data in complex power grid environments. Physiological and environmental data generated at the work site are essentially asynchronous heterogeneous streams, with the collection time serving as the unique deterministic time stamp and becoming the cornerstone of data fusion. The arrangement logic of the field sequence "collection time - heart rate - blood oxygen saturation - heart rate variability - ambient temperature and humidity" aligns with the feature priority of risk analysis—timestamps construct the analysis benchmark, heart rate and blood oxygen constitute core vital signs, heart rate variability maps to neural regulatory states, and temperature and humidity explain environmental stressors. This sorting method allows subsequent model processing to extract feature vectors according to the strength of risk correlation.

[0116] Clearly defining field definitions and format standards establishes a language system for cross-departmental collaboration. For example, "heart rate" is strictly defined as "the number of ventricular contractions per minute (integer)," and "ambient temperature and humidity" are limited to "air parameters within 1 cm of the sensor contact point (temperature in degrees Celsius retained to one decimal place, humidity percentage as an integer)." This standardized processing effectively prevents data misreading incidents, such as mistakenly recording equipment surface temperature as ambient temperature, or treating instantaneous heart rate peaks as continuous outliers. Format constraints also ensure the comparability of data generated by equipment deployed in different years.

[0117] A unified format enables lossless data transmission between acquisition points, edge computing units, and cloud analytics platforms, avoiding timestamp misalignment or numerical truncation caused by format conversion. During model training, the fixed field structure allows for direct reuse of preprocessing pipelines, and new batches of data can be input into the LSTM network without additional format cleaning. More importantly, when the system performs fault backtracking analysis, structured data frames can quickly correlate multi-dimensional state parameters at specific points in time—for example, a false alarm can be traced back to outliers in the central rate variability field caused by sweat interference in a high-temperature environment, rather than the actual risk. This standardized architecture also minimizes the impact of on-site equipment upgrades; new sensor models only need to be adapted to the basic field specifications to be integrated into the existing analysis system, greatly reducing system expansion costs. The resulting data assets have long-term reuse value; the standardized data frames accumulated over five years can be directly used for training next-generation models without undergoing time-consuming data normalization.

Claims

1. A health status early warning system integrated into a power grid safety belt, characterized in that, include: The health data acquisition unit includes a photoplethysmography sensor built into the inside of the shoulder strap of the power grid safety belt and attached to the skin of the operator, used to collect the operator's core physiological indicators such as heart rate, blood oxygen saturation, and heart rate variability. Miniature temperature and humidity sensor that simultaneously collects temperature and humidity data of the working environment; The local data processing module is built into the power grid safety belt, including a synchronization clock control unit, which controls the PPG sensor and the miniature temperature and humidity sensor to collect data synchronously at the same rhythm; The data processing and analysis unit is used to receive synchronous physiological and environmental data from the local data processing module, perform data preprocessing, and finally input the data into the pre-trained intelligent health status analysis model for analysis. The early warning unit is used to receive the output results of the intelligent analysis model, determine the health risk level according to the preset probability threshold, and trigger the early warning mechanism at the field end and the management end.

2. The health status early warning system integrated into the power grid safety belt according to claim 1, characterized in that: The local data processing module is further used to set the acquisition parameters of the PPG sensor and the temperature and humidity sensor, and to control the two types of sensors to continuously acquire data at the same rhythm through a unified synchronous clock signal, so as to ensure that each set of physiological data corresponds to the environmental data of the same period.

3. The health status early warning system integrated into the power grid safety belt according to claim 1, characterized in that: The data preprocessing in the data processing and analysis unit specifically includes using the Kalman filter algorithm to remove signal fluctuation interference caused by the workers' limb movements, and simultaneously combining environmental data for secondary screening, marking physiological data segments when environmental parameters change abruptly, and determining whether to remove them based on concurrent work behavior records; The data processing and analysis unit performs data preprocessing, applies the Kalman filter algorithm to remove limb activity interference in PPG data, and identifies and removes confirmed environmental interference data based on changes in environmental data and concurrent work behavior records. The preprocessed physiological data and environmental data are precisely aligned in time sequence to construct a structured data frame containing acquisition time, heart rate, blood oxygen saturation, heart rate variability, and environmental temperature and humidity. The integrity of the data frame is verified and repaired, and the structured data frame is input into a pre-trained intelligent health status analysis model for analysis. The intelligent health status analysis model uses PPG physiological data and environmental data from historical power grid operation scenarios as the training set to train a multi-dimensional temporal long short-term memory (LSTM) neural network model, which is used to output multi-class probability distribution results representing health status.

4. The health status early warning system integrated into the power grid safety belt according to claim 3, characterized in that, The architecture of the intelligent health status analysis model includes: The input layer takes structured data blocks, packaged in fixed time windows and containing continuous physiological index values ​​and average environmental values, as input. The hidden layer uses a long short-term memory neural network to capture long-term dependencies in time-series data, and a dropout layer is added after the LSTM layer to prevent overfitting. The output layer outputs a multi-class probability distribution of four health states: no risk, low risk, medium risk, and high risk.

5. A health status early warning system integrated into a power grid safety belt according to claim 4, characterized in that, The system also includes an adaptive optimization module, which is used to periodically collect new monitoring data from the work site and establish an incremental data pool; An incremental learning approach with local parameter updates is adopted, adjusting only the model parameters with high correlation to newly added data features for optimization. Establish a real-time performance monitoring mechanism, calculate performance indicators based on early warning results and actual health verification, and compare them with preset thresholds. When indicators exceed limits, conduct analysis and optimize the model by incremental training with supplementary data or resetting some parameters.

6. A health status early warning method integrated into a power grid safety belt, applied to the system described in any one of claims 1-5, characterized in that, Includes the following steps: S1: Contextualized collection and preprocessing of health data, using a photoplethysmography sensor with anti-motion interference function, integrated into the inner side of the shoulder strap of the power grid safety belt and fixed with medical-grade silicone, to collect the core physiological indicators of the workers; The ambient temperature and humidity are collected by a miniature temperature and humidity sensor that collects data synchronously; a unified synchronization clock is set to control the two types of sensors to continuously collect data at the same rhythm. Kalman filtering algorithm was used to remove limb activity interference from PPG physiological data, and environmental interference data was identified, marked and removed based on sudden changes in environmental parameters and concurrent work behavior records. S2: Integration and processing of health data and environmental data, based on timestamp information, to precisely align preprocessed physiological data and environmental data in time sequence; Construct structured data frames that include acquisition time, heart rate, blood oxygen saturation, heart rate variability, and ambient temperature and humidity; Perform integrity verification and repair on structured data frames; S3: Intelligent analysis, modeling and optimization of health data. The integrated and processed structured data frames are input into the pre-trained intelligent health status analysis model for analysis. The model is a multi-dimensional temporal long short-term memory (LSTM) neural network model that is trained after constructing a multi-level annotation system based on historical PPG physiological data and environmental data under the power grid operation scenario and after expert collaborative annotation and cross-verification. The model outputs a multi-class probability distribution representing health status; S4: Early warning application of health data analysis results. Based on the probability distribution results output by the intelligent analysis model and the preset probability threshold, the final health risk level is determined. If the response level is determined to be required, a multi-terminal early warning linkage between the field terminal and the management terminal will be triggered.

7. The health status early warning method integrated into the power grid safety belt according to claim 6, characterized in that: The training optimization of the model in S3 specifically includes two-stage iterative training: the first stage prioritizes increasing the training weight of clearly risky samples so that the model can master the judgment logic of high-risk scenarios; The second stage introduces confused sample training and refines the loss function to improve the model's ability to distinguish between normal stress and suspected risk. During training, the accuracy, false alarm rate, and false negative rate of the test set are monitored in real time, and parameters are adjusted accordingly. Independent validation datasets for different power grid operation scenarios are used to test performance differences and supplement training accordingly.

8. A health status early warning method integrated into a power grid safety belt according to claim 6, characterized in that... The S3 also includes adaptive optimization of the model: regularly collecting new health monitoring data, filtering data that includes new work scenarios, new employees, and historical false alarm / missed alarm scenarios to establish an incremental data pool; An incremental learning approach with local parameter updates is adopted, adjusting only the model parameters that are highly correlated with the features of the newly added data; Establish a real-time performance monitoring mechanism. When the accuracy, false positive rate, and false negative rate exceed the limits, analyze the reasons and optimize by supplementing data for incremental training or resetting some parameters.

9. A health status early warning method integrated into a power grid safety belt according to claim 6, characterized in that: In S1, the construction of the structured data frame specifically involves: integrating the time-aligned associated data groups into a unified format that includes "collection time - heart rate - blood oxygen saturation - heart rate variability - ambient temperature and humidity", and clarifying the definition and data format standard of each field.

10. A health status early warning method integrated into a power grid safety belt according to claim 6, characterized in that: The specific steps for triggering multi-terminal early warning linkage between the field terminal and the management terminal in S4 are as follows: if the risk level is determined to be medium or high, the field terminal will issue a graded warning through the buzzer and LED indicator built into the safety belt. The management system will simultaneously feed back the risk level, corresponding physiological data, and work environment information to the power grid operation safety management platform.

Images