High-altitude operation safety early warning system based on multi-sensor fusion

Abstract

The application relates to the technical field of high-altitude operation safety, and discloses a high-altitude operation safety early warning system based on multi-sensor fusion. The system comprises data acquisition, synchronization, rule violation identification, feature processing, analysis decision and parameter adjustment components. The data acquisition component is composed of an intelligent safety helmet and an intelligent safety rope, the former acquires hat removal and silent state data, and the latter acquires end buckle state data; after receiving data, the data synchronization component processes multi-source data through time alignment and format standardization; the rule violation identification component performs preliminary rule violation judgment, and the feature processing component extracts data features and adds time marks; the analysis decision component outputs decision signals through multi-level analysis, and the parameter adjustment component adjusts analysis parameters and controls alarm devices accordingly. The system realizes comprehensive monitoring and accurate early warning of high-altitude operation safety states, and is suitable for various high-altitude operation scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of high-altitude operation safety, in particular to a high-altitude operation safety early warning system based on multi-sensor fusion. BACKGROUND

[0002] High-altitude operation is widely used in many fields such as building construction, power maintenance, bridge maintenance, etc. The safety risk faced by the operator during the operation process is high. Once a safety accident occurs, it will often cause serious personnel casualties and property losses. At present, the control means for high-altitude operation safety in the industry mainly relies on manual supervision and single device monitoring, which is difficult to achieve comprehensive and real-time monitoring of the safety state of the operator.

[0003] Under the manual supervision mode, the on-site safety inspector relies on patrol and management, but the high-altitude operation scene usually has a large range and a complex environment, and the safety inspector cannot cover all operation points at the same time, so there is a supervision blind area. Especially in high-rise building construction or outdoor power repair scenes, the operation environment often accompanies risks such as high-altitude falling and object impact. Manual supervision not only has low efficiency, but also may cause safety hazards that are not discovered in time due to human negligence.

[0004] In terms of single device monitoring, independent safety monitoring devices are used in the prior art, such as monitoring the head protection state of the operator only through an intelligent safety hat, or relying on a simple sensor on the safety rope to determine whether it is tied. This kind of single monitoring method has obvious limitations: it cannot realize the collaborative analysis of multi-dimensional data, for example, when the intelligent safety hat detects that the operator is in a hat-off state, it cannot judge whether there is a double safety hazard by combining the tying condition of the safety rope; the data collected by a single device lacks unified synchronous processing, and the data timestamps of different devices are inconsistent and the formats are not unified, which leads to the fact that the subsequent data cannot be effectively integrated and analyzed, and it is difficult to form a comprehensive safety early warning judgment.

[0005] The existing safety early warning system also has deficiencies in data analysis and decision-making. Most systems can only make a simple threshold judgment on the collected data, such as directly issuing an alarm signal when a hat-off state is detected, and lack a multi-level analysis mechanism. This simple decision-making method cannot distinguish the risk level in different scenarios, for example, a short hat-off during the preparation stage on the ground and a hat-off during the high-altitude operation process have significant differences in risk level, but the existing system often cannot make accurate distinctions, which easily leads to false positives or false negatives, affecting operation efficiency and possibly failing to provide timely warnings when real danger occurs. At the same time, the analysis parameters of the existing system are mostly fixed settings and cannot be dynamically adjusted according to the changes in the actual operation scene, for example, the judgment standard of safety risk should be different under different heights and different weather conditions, but fixed parameters cannot adapt to diversified operation environments, further reducing the accuracy and reliability of safety early warning. SUMMARY

[0006] The present application aims to provide a high-altitude operation safety warning system based on multi-sensor fusion to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides a high-altitude operation safety warning system based on multi-sensor fusion, which comprises:

[0008] a data acquisition component, a data synchronization component, a violation identification component, a feature processing component, an analysis and decision component, and a parameter adjustment component;

[0009] The data acquisition component is composed of an intelligent safety helmet and an intelligent safety rope. The intelligent safety helmet collects hat-off state and silent state data, and the intelligent safety rope collects end buckle state data. The data synchronization component receives data from the data acquisition component and processes multi-source data through time alignment and format standardization. The violation identification component makes a preliminary violation judgment. The feature processing component extracts features from the data and adds time labels. The analysis and decision component makes multi-level analysis and outputs decision signals. The parameter adjustment component adjusts analysis parameters and controls the alarm device according to the decision signals.

[0010] Preferably, the data synchronization component establishes a multi-source data stream synchronization mechanism, which includes state data stream and motion data stream, and determines the time-varying correlation between state and motion through timestamp alignment. The state data stream is constructed based on a time series discretization method, which maps state data into continuous data sequences.

[0011] Preferably, the preliminary violation judgment includes state threshold comparison, single sensor anomaly, and double verification. The state threshold comparison is used to compare whether the hat-off data, silent data, and end buckle data are within the normal state range. If the hat-off data, silent data, or end buckle data is not within the normal state range, a preliminary violation judgment is triggered. The single sensor anomaly is used to detect whether the hat-off data, silent data, and end buckle data have mutations, continuous anomalies, or signal loss. If any of the data has anomalies, a single sensor anomaly report is output, and it is checked whether the data acquisition component is working normally.

[0012] Preferably, the analysis and decision component includes a hat-off analysis unit, a silent analysis unit, and an end buckle analysis unit. The hat-off analysis unit takes hat-off features with time labels as input, monitors hat-off features at consecutive time points in real time, obtains the difference between the hat-off features at the previous time and the current time, and the difference between the hat-off features at the next time and the current time, and triggers an abnormal flag in the continuous time.

[0013] Preferably, the silence analysis unit sets a silence prediction model, predicts the next moment of silence data through the silence prediction model, and marks the silence risk through anomaly detection; the anomaly detection includes whether the difference between the predicted silence value and the actual value appears mutation, and the silence signal is processed through the time domain analysis method to identify the long-time silence section.

[0014] Preferably, the end buckle analysis unit sets an end buckle classification model and an end buckle identification unit; the end buckle classification model is used to process the end buckle data obtained by the end buckle identification unit, obtain the state characteristics of the end buckle through the end buckle classification model, and classify the state characteristics, including not tight, loose and normal; the end buckle identification unit is used to obtain multi-environment end buckle data and end buckle data optimization, the multi-environment end buckle includes motion influence end buckle characteristics and artificial influence end buckle characteristics, and the end buckle data optimization is used to filter and enhance the obtained end buckle data.

[0015] Preferably, the cap-off change rate, the silence gradient change rate and the end buckle state characteristics in the continuous time are obtained; the cap-off change rate in the continuous time point is used to drive the cap-off analysis unit to receive the cap-off change rate in the continuous time point, adjust the cap-off parameter analysis step length, if the cap-off data change rate per unit time is greater than the set threshold, the cap-off parameter analysis step length is shortened, if the cap-off data change rate per unit time does not exceed the set threshold, the state of waiting for adjustment is entered; the analysis step length is adjusted through the silence change acceleration of the continuous time point, if the silence change acceleration of the time point from the previous moment to the current moment rises greater than the silence change acceleration of the continuous time point from the second previous moment to the previous moment, and the acceleration rise is greater than the minimum threshold, the step length ratio is compressed; the end buckle signal is obtained through the end buckle identification unit, and the step length adjustment is activated through the comparison of the state characteristic library, if the end buckle signal is captured through the end buckle state characteristics, it is judged as a violation, and the analysis step length of the end buckle analysis unit is adjusted.

[0016] Preferably, the step length boundary limit for receiving the adjusted end buckle analysis step length, the silence compression step length ratio and the adjusted cap-off analysis step length is set, and the sensitivity level is set, the sensitivity level includes low sensitivity and high sensitivity; the step length boundary is adjusted, the redundant boundary buffer layer is triggered to temporarily tolerate the step length adjustment, if the step length temporarily exceeds the rated boundary, the actual adjusted total step length is not greater than the maximum step length of the redundant boundary buffer layer, if the buffer layer stays for more than the time threshold, it is immediately switched to the minimum safe step length.

[0017] Preferably, if the post-adjustment cap-off rate, the silent gradient change rate and the end of the buckle state feature are in the redundant boundary buffer layer stable domain, the analysis step of the analysis decision component is continuously adjusted according to the real-time data; if the total compensation number of the adjustment step of the analysis decision component is not greater than the maximum step number of the redundant boundary buffer layer, and the buffer layer stays for a time greater than the time threshold, the step freezing is triggered.

[0018] Preferably, if it is judged that the safety alarm needs to be triggered, the safety alarm triggering sound and light alarm device is started, a sound of a predetermined frequency and a flash of a predetermined intensity are emitted, and the recording unit is triggered to store event data and a time stamp.

[0019] Compared with the prior art, the beneficial effects of the present application are:

[0020] From the aspect of data acquisition, the data acquisition component is composed of an intelligent safety helmet and an intelligent safety rope, which respectively collect cap-off state, silent state and end of the buckle state data, realizing multi-dimensional monitoring of key safety indicators of high-altitude workers. Compared with traditional single device monitoring, this multi-source data acquisition method can cover multiple key safety dimensions such as head protection, body activity state and safety rope hanging condition, making the monitoring of the safety state of workers more comprehensive, avoiding the monitoring blind area caused by single data acquisition, and more completely reflecting the safety status of workers during high-altitude operation.

[0021] The setting of the data synchronization component can receive the multi-source data collected by the data acquisition component, and through time alignment and format standardization processing, solve the problem of inconsistent time stamps and non-uniform formats of different device data in traditional monitoring systems. After synchronization processing, the multi-source data has a unified time reference and data format, providing a high-quality data basis for subsequent rule violation identification, feature processing and analysis decision, so that different dimensions of data can be effectively integrated, and the data usability and subsequent analysis accuracy are improved.

[0022] The rule violation identification component can preliminarily judge the synchronized multi-source data and quickly filter out abnormal data that may have safety hazards. This preliminary filtering mechanism can identify obvious rule violations in advance, such as cap-off, end of the buckle of the safety rope not buckled, etc., laying a foundation for subsequent in-depth analysis, avoiding a large amount of non-abnormal data occupying subsequent analysis resources, improving the operation efficiency of the whole system, and enabling safety hazards to be detected in an early stage.

[0023] The feature processing component extracts features from the data and adds time markers, making the safety information contained in the data more explicit and specific. By extracting features, key information related to safety risks can be filtered from massive data, and the addition of time markers can clearly record the time nodes of safety state changes, facilitating the subsequent analysis decision component to trace the occurrence process and development trend of safety hazards, better understand the formation mechanism of safety risks, and provide detailed data support for accurate risk level judgment.

[0024] The analysis decision component performs multi-level analysis and outputs a decision signal, changing the traditional system's decision-making method of relying only on simple threshold judgment. Multi-level analysis can make comprehensive judgments combining different work scenarios and different risk factors, such as distinguishing between short-term hat removal during ground preparation and hat removal during high-altitude work, to make more accurate decisions and effectively reduce false positives and false negatives. This accurate decision-making mechanism can help workers avoid unnecessary alarm interference during normal work, while receiving timely warning signals when facing real safety risks, ensuring personnel safety and improving work efficiency.

[0025] The parameter adjustment component adjusts analysis parameters and controls the alarm device according to the decision signal, realizing dynamic optimization of system analysis parameters. The fixed analysis parameters of traditional systems cannot adapt to changes in different work environments, while the system can adjust analysis parameters in real time according to actual decision results and changes in work scenarios, such as automatically optimizing safety risk judgment standards under different heights and weather conditions, allowing the system to better adapt to diverse high-altitude work environments and further improve the reliability and applicability of safety warnings. At the same time, the parameter adjustment component controls the alarm device to trigger an alarm in a timely manner when a safety risk is confirmed, ensuring that workers and on-site managers can be aware of the danger in the first place and take appropriate protective measures to minimize the probability of safety accidents. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 A timing diagram of the high-altitude work safety warning system based on multi-sensor fusion described in the present application;

[0027] Figure 2 A flowchart of the multi-source data stream synchronization mechanism of the data synchronization component;

[0028] Figure 3 A flowchart of the hat removal analysis unit's work. DETAILED DESCRIPTION

[0029] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0030] Please refer to Figure 1 The present application provides a high-altitude operation safety early warning system based on multi-sensor fusion, which comprises a data acquisition component, a data synchronization component, a violation identification component, a feature processing component, an analysis and decision component, and a parameter adjustment component. Each component works cooperatively to complete the whole process from data acquisition to alarm triggering.

[0031] The data acquisition component is composed of an intelligent safety helmet and an intelligent safety rope. The intelligent safety helmet is used to collect the data of the hat-off state and the silent state of the operator, and the intelligent safety rope is used to collect the data of the end buckle state. These raw data are time-aligned and standardized in format by the data synchronization component to form a unified multi-source data stream. The violation identification component makes a preliminary violation judgment on the standardized data to identify possible safety risks. The feature processing component extracts key features from the data and adds time labels to provide structured input for subsequent analysis. The analysis and decision component receives the feature data, performs multi-level analysis including hat-off analysis, silent analysis, and end buckle analysis, and outputs a decision signal. The parameter adjustment component dynamically adjusts the analysis parameters according to the decision signal and controls the alarm device to perform corresponding actions, such as triggering sound and light alarms. The system improves the accuracy and real-time performance of early warning through the multi-sensor fusion mechanism and adapts to the complex environment of high-altitude operation.

[0032] Embodiment 1: Please refer to Figure 2, the intelligent safety helmet is responsible for generating off-hat state data reflecting the head protection condition of the worker and silence state data indicating the activity of the worker, and the intelligent safety rope continuously generates off- buckle state data representing whether the safety belt buckle is working normally. The motion data stream is usually provided by an additional inertial measurement unit or other motion capture sensors, which describes the position, posture, speed and acceleration change of the worker in three-dimensional space. The two types of data have natural differences in physical source, sampling frequency and data characteristics. The state data is usually discrete event type data, while the motion data is continuous analog data. Therefore, the precise synchronization of the two types of data is the premise of subsequent meaningful fusion analysis. In order to realize the precise correlation between different source data, the data synchronization component adopts a timestamp alignment method based on a hardware clock synchronization protocol as its technical basis. In the system initialization stage, the internal clocks of the intelligent safety helmet, the intelligent safety rope and the motion sensor will be calibrated through a unified synchronization signal, so as to reduce the initial time deviation between devices as much as possible. Each sensor will stamp a precise time stamp from its calibrated clock on each data point when it is generated. When these data packets with time marks are transmitted to the data synchronization component, the component will create a unified time axis to reorder and align all data points according to their time stamps. This processing method can effectively determine the time-varying correlation between state changes and motion behaviors. For example, it can accurately determine whether the motion trajectory of the worker is abnormal within a certain period of time before and after the occurrence of the off-hat state event, or whether the worker is in a state of violent movement when the off-buckle loose alarm is triggered. This correlation analysis provides key information for in-depth understanding of the context of the violation behavior.

[0033] For state data stream, which is discrete switch value itself, in order to make it seamlessly integrated and compared with continuous motion data stream, it needs to be constructed as continuous data sequence, data synchronization component completes this conversion by applying time series discretization method. This method divides the continuous time axis into a series of equal or unequal interval micro time units, for a state event such as "hat off", within the duration of the event, its corresponding state data sequence will be assigned a constant value representing "abnormal" or "hat off", while in the period of normal wearing of safety hat, it will be assigned a "normal" or "hat on" value. Through this mapping, a simple "hat on / hat off" event is converted into a signal that is continuous in time and step change in value, this form of signal is very convenient for mathematical operation and pattern recognition with continuous motion data stream, such as calculating the average motion amplitude in a certain state, or detecting the motion characteristics at the state switching moment. After receiving the multi-source data stream after synchronization and standardization, the violation identification component immediately starts the preliminary violation judgment process, which contains three core links in layers: state threshold comparison, single sensor anomaly detection and double verification. State threshold comparison is the first line of defense of the whole judgment process, which screens obvious abnormalities by comparing the real-time collected hat off data, silent data and end buckle data with the preset normal state range threshold one by one, the normal state range is usually obtained according to safety regulations and a large number of normal operation data statistics, for example, hat off state lasting more than 5 seconds may be judged as exceeding the normal range, silent state, i.e. personnel inaction time more than 1 minute, may be considered as dangerous, and end buckle state once feedback as "not locked" will immediately touch the red line. As long as any one of the data deviates from its normal range for a set period of time, the violation identification component will trigger a preliminary violation judgment flag, and mark the abnormal event as to be further verified.

[0034] The single-sensor anomaly detection module, on the other hand, focuses on identifying abnormal patterns that may exist in each individual sensor data itself, including data mutation, persistent anomaly, or signal loss, rather than directly pointing to the safety status of the worker. Data mutation refers to an unreasonable and drastic jump in sensor readings within a very short period of time, which may be caused by strong electromagnetic interference, physical collision of the sensor, or transient failure. Persistent anomaly is characterized by the fact that the readings stay in an unreasonable interval for a long time, for example, a silent sensor always outputs a "silent" signal even if the user is obviously active, which may indicate that the sensor is blocked or damaged. Signal loss is the most serious type of anomaly, which means that the data stream is interrupted and the communication link may have failed. When any type of anomaly is identified, the module outputs a detailed single-sensor anomaly report indicating the type of abnormal sensor, the classification of the abnormal pattern, and the time of occurrence. At the same time, the system triggers a self-check command to check whether the working power supply and communication connection of the data acquisition component are normal, and attempts to distinguish between transient anomalies caused by environmental interference and inherent problems of the hardware itself. The double verification mechanism is a key design to improve the robustness of the system and reduce the false alarm rate. It is not a separate step, but a comprehensive analysis and cross-verification of the results of the previous two steps. This mechanism simultaneously analyzes the results of state threshold comparison and the reports of single-sensor anomaly detection, and examines their consistency and contradictions. For example, a typical case is that state threshold comparison finds "hat-off" data anomaly, triggering a preliminary violation flag, but at the same time, the single-sensor anomaly detection module reports that the sensor inside the hat has "signal loss" or "data mutation" anomaly. At this time, the double verification mechanism does not immediately confirm the "hat-off" violation, but considers it "pending" or "suspected sensor failure", and attempts to make auxiliary judgments combined with other data, such as checking whether the motion data of this person is active at this time. If the motion data is normal, it is more likely to believe that the hat-off sensor itself is the problem, thus avoiding false alarms caused by a single sensor failure. Only when the state threshold comparison shows an anomaly and the single-sensor anomaly detection does not report a problem with this data, will the double verification mechanism finally confirm that the preliminary violation judgment is valid and pass it to the feature processing component for in-depth analysis.

[0035] Example 2: see Figure 3The off-hat analysis unit is specialized in handling continuous data streams related to the safety hat wearing status. Its input is the time-stamped off-hat feature sequence output by the feature processing component, which contains not only the simple "wearing" or "off-hat" binary status, but also the subtle changes in the contact pressure between the hat body and the head, the continuous readings of the hat band tightness signal, and the hat body stability indicators indicated by the inertial sensor. The core function of this unit is to monitor the off-hat feature dynamics at consecutive time points in real time. It tracks the feature change trajectory in the recent period of time by maintaining a sliding time window of configurable length, which contains the complete feature history from the current time back to several sampling points. During the analysis process, the unit calculates the difference between the off-hat feature at the previous time and the current time. This difference can be an Euclidean distance, a cosine similarity, or other metrics suitable for this feature space, to quantify the degree of state change between adjacent time points. Furthermore, the unit also has a lightweight prediction module built in. Based on the trend of the recent feature sequence, this module extrapolates the possible value of the off-hat feature at the next time point, and calculates the expected difference between this predicted value and the current feature value. This analysis method of correlation between the front and back makes the system able to capture abnormal patterns that do not occur instantaneously but evolve gradually, such as signs that the safety hat is about to be accidentally touched and loosened. If the actual calculated feature difference or the predicted deviation exceeds the preset tolerance threshold multiple times within a continuous time segment, the unit will trigger a persistent abnormality flag, indicating the existence of potential and continuous off-hat risk, rather than a one-time isolated sensor false alarm.

[0036] The silence analysis unit, on the other hand, focuses on processing the silence data acquired from the smart safety hat, which reflects the activity level of the worker. This unit sets up a special silence prediction model to make a forward-looking judgment on the worker's state at future time. The silence prediction model is usually built using time series analysis techniques, such as possibly based on the autoregressive integrated moving average model or its variants. This model learns from the historical silence data sequence, captures its inherent periodicity and trend rules, and generates a prediction value for the next sampling time. The silence data itself may come from the accelerometer and gyroscope inside the hat, which quantifies the worker's motion amplitude by calculating the comprehensive vector amplitude or activity count of the signal. The lower the value, the closer to the "silent" state. The predicted value output by the prediction model is compared with the real value actually collected by the sensor in real time, and the anomaly detection logic closely monitors the evolution of the difference between them, especially whether the difference suddenly and discontinuously jumps, such as the predicted value indicating that the worker should be at a moderate activity level, but the actual data returned shows a long time of zero acceleration reading. This deviation may mean that the worker has become disabled, faint or in a dangerous situation where he cannot move. In addition to the prediction-based mutation detection, this unit also runs a time-domain analysis method in parallel to process the original silence signal more deeply. Time-domain analysis does not rely on prediction, but directly examines the statistical properties of the signal on the time axis, such as calculating the mean, variance, zero-crossing rate, etc. of the signal in the sliding window. Identifying long silence segments is the focus of this analysis, and the system will set an absolute time threshold. When the silence signal continues to be below a certain activity threshold for more than this threshold, it is marked as a long silence segment, which may represent a very dangerous situation that needs to be immediately alerted. Time-domain analysis can also help filter out short and normal work breaks, such as workers stopping to look at drawings or taking short breaks. These breaks usually have a short duration and different signal characteristics from real dangerous silence.

[0037] The cap-off analysis unit and the silence analysis unit do not operate in isolation, and they have functional linkage and information interaction under the framework of the analysis decision component. This synergy enhances the system's ability to interpret complex scenarios. The length of the sliding time window relied on by the cap-off analysis unit is not fixed and can be dynamically adjusted by the upstream parameter adjustment component according to the global risk assessment level. When the system is in a high alert state, the window may be shortened to improve the sensitivity to instantaneous abnormalities. The feature difference calculation inside it may use a weighted strategy, giving higher weights to data points closer to the current time, which makes the analysis focus more on the latest state changes. The prediction model in the silence analysis unit usually needs to be trained and initialized on a stable data stream. The system may have a short learning period after startup, during which the model gradually adapts to the unique activity patterns of the current worker to individualize its prediction benchmarks. The judgment logic for long silence periods often combines absolute time and relative change. In addition to monitoring whether the silence duration exceeds the limit, it also examines whether the worker's activity pattern has an abnormal drop before the start of this silence period. This comparison of before and after helps to distinguish between planned pauses and sudden stillness. The intermediate results produced by the two units, such as the abnormal trend flag of the cap-off feature and the level signal of the silence risk, are aggregated to the higher logic layer of the analysis decision component. This logic layer will integrate these information and the output of other analysis units to make a weighted decision, and finally generate a decision signal whether to trigger an early warning. For example, a solitary and short cap-off feature anomaly may be considered a low-risk event if accompanied by normal and active silence data, and it may only be a slight touch of the safety cap. However, if the cap-off anomaly occurs simultaneously with a long silence period, it is likely to indicate a serious accident, and the system will not hesitate to trigger a high-level alarm. This multi-feature, multi-angle cross-verification mechanism greatly improves the accuracy and reliability of the early warning, effectively reducing the risk of misjudgment caused by the limitations of single sensor information.

[0038] The entire analysis process relies on accurate time labeling, and the timestamp stamped by the feature processing component for each feature data point is the cornerstone of ensuring the correctness of the timing logic, so that the sequence of doffing events and the sequence of silent activities can be accurately compared on a unified time axis. The analysis decision component may maintain a state machine inside to depict the safety state transition of the work personnel, and the output of the doffing analysis unit and the silent analysis unit is an important input condition for driving the state machine to switch. All intermediate data, trigger flags, risk levels and final decision signals generated during the analysis process are recorded in detail in the non-volatile memory of the system along with their timestamps to form a complete audit log, which is indispensable for analyzing the causes of accidents, optimizing algorithm parameters and evaluating the performance of the system. Through this structured and multi-level in-depth analysis, the system realizes the monitoring of the safety state of the work personnel from appearance to connotation, surpassing simple yes / no judgment and moving towards intelligent perception of behavior intention and potential risks.

[0039] In embodiment 3, the doffing analysis unit is composed of two core modules: a doffing classification model and a doffing recognition unit. The doffing recognition unit serves as a front-end data interface and is responsible for directly obtaining the original doffing state signals from the sensor array of the intelligent safety rope. These signals are usually multi-dimensional and may include microswitch signals representing the position of the carabiner latch, strain gauge readings reflecting the internal tension of the carabiner, and vibration sensor data indicating whether the carabiner has been subjected to accidental impact. The multi-environment doffing data obtained fully considers the complexity of the actual work scene, and the motion-influenced doffing features mainly refer to the dynamic mechanical load on the carabiner due to body swings and rope traction during normal movement, climbing or tool operation of the work personnel. This load can cause regular fluctuations in sensor readings, while human-influenced doffing features involve abnormal operation behaviors such as the carabiner being considered locked without being fully inserted into the carabiner (undocked state) or the carabiner being released due to accidental or intentional touching or manipulation of the release mechanism during work. In the face of inevitable noise interference in these raw data, the doffing data optimization module performs filtering and enhancement processing. The filtering stage may use algorithms based on moving average or low-pass filtering to smooth out high-frequency jitter, and the enhancement stage may use signal amplification or feature value normalization to improve the signal-to-noise ratio of the effective signal, so that the essence of state changes can be more clearly captured in subsequent analysis.

[0040] The optimized end buckle data is sent to the end buckle classification model for deep processing. The model is a pre-trained machine learning model, and its core task is to extract state features that can effectively distinguish different safety states from the input data stream. These state features are not simple raw readings, but higher-level abstract representations obtained through feature engineering or internal model layer conversion, such as the statistical distribution of tension readings over a period of time, the spectral features of vibration signals, or the correlation coefficients between multiple sensor readings. The end buckle classification model uses these extracted state features to classify the current end buckle state in real time, and the output result is usually assigned to several explicit categories: not fastened, loose, and normal. The not fastened state usually means that the buckle has been partially connected but has not reached the fully locked safety position, and its characteristics may be that the tension value is below the safety threshold and the position signal is unstable; the loose state means that the buckle has completely disconnected, and its characteristics are that the tension disappears and the position signal indicates disconnection; the normal state corresponds to the buckle being fully locked and the force being within the safety range. The classification decision of the model provides direct and objective judgment basis for the system about the physical connection state of the safety belt.

[0041] The system continuously monitors the time series indicators from different analysis units, i.e., the hat-off change rate, the silence gradient change rate, and the end buckle state features in continuous time, which together form the basis for dynamically adjusting the analysis strategy. The hat-off change rate quantifies the change speed of the hat-off feature per unit time, which is a continuously calculated value that drives the analysis rhythm within the hat-off analysis unit. When the system detects that the hat-off data change rate per unit time increases sharply and exceeds a certain preset sensitivity threshold, it indicates that the state of the safety hat may be undergoing rapid and unstable changes, and there is a higher risk. At this time, the hat-off analysis unit will actively shorten its analysis step length. The analysis step length refers to the time window interval or data sampling interval used for continuous time point difference calculation and prediction, and shortening the step length means increasing the monitoring frequency to capture possible fleeting danger signs with more intensive sampling and analysis. Conversely, if the hat-off data change rate remains at a low level and does not exceed the set threshold, it indicates that the state is relatively stable, and the hat-off analysis unit will enter a state of adjustment, maintaining the current analysis step length to save computing resources while remaining vigilant.

[0042] The silent gradient rate of change goes a step further, focusing on the acceleration of the rate of change of silent data itself, i.e., the rate of change of the rate of change. It adjusts the granularity of the analysis by examining the trend of silent changes over consecutive time points. Specifically, the system compares the acceleration over two consecutive time periods: one from two moments ago to the previous moment, and the other from the previous moment to the current moment. If the analysis finds that the acceleration value in the latter time period is significantly greater than that in the former, and this increase exceeds a minimum threshold set by the system, it indicates that the activity level of the workers is experiencing a sharp, non-linear decline, and the risk is rapidly accumulating. In this situation, the silent analysis unit adopts a strategy of compressing the step size ratio, that is, significantly reducing the analysis step size to track the subtle evolution of the silent signal with extremely high temporal resolution, striving to make the most timely response before danger occurs.

[0043] After acquiring and optimizing the last-deduction signal, the last-deduction identification unit compares its extracted state features with a pre-established state feature library in real time. This library contains feature patterns corresponding to various typical last-deduction states. The comparison process not only triggers violation judgments but also activates the analysis step size adjustment mechanism. Once the last-deduction signal clearly points to the highest risk level of "loosening" through the matching of last-deduction state features, the system immediately judges it as a serious violation. Simultaneously, the analysis step size of the last-deduction analysis unit is adjusted accordingly, typically shortening the step size to increase monitoring frequency and data collection density in emergency situations, providing more detailed state information for potential subsequent emergency response procedures. This dynamic adjustment process demonstrates the system's ability to adaptively configure monitoring resources based on real-time risk levels, enabling the system to remain highly efficient during stable periods and become extremely sensitive during periods of heightened risk.

[0044] The classification process for the final state features can be achieved using a discriminant function that comprehensively considers scores across multiple feature dimensions. Its decision logic can be expressed as:

[0045]

[0046] in: The output represents the category determined by the model. This indicates the category that maximizes the summation term within the parentheses. . The candidate categories are "normal", "not fastened" and "loose" in the set. This represents the total number of state features used. Representing the The numerical value of each state feature is the result of optimization and standardization. Then is a weight parameter, which represents the first... The importance degree of each feature for distinguishing the category The weights are learned from historical data in the model training phase. The model completes the state classification by calculating the weighted sum of each category and selecting the category corresponding to the maximum value. This comprehensive evaluation based on weighted features is more robust than relying on a single threshold, and can better handle complex and variable environmental disturbances.

[0047] Embodiment 4: The step boundary limit sets the upper and lower limits of the analysis step length for each analysis unit, for example, the rated boundary of the cap-off analysis step length can be set between 0.1 seconds and 5 seconds, the rated boundary of the quiet compression step length ratio can be limited between 50% and 150% of the original step length, and the rated boundary of the end-off analysis step length can be set between 0.2 seconds and 10 seconds. These boundary values are not fixed, they are directly related to the sensitivity level of the system configuration, including low sensitivity and medium-high sensitivity two main modes. In the conventional, low sensitivity operation environment with controllable risk, the system will adopt a more relaxed step boundary, allowing the analysis step length to adaptively adjust within a larger range to optimize resource consumption. In the medium-high sensitivity mode under high-risk or abnormal conditions, the step boundary will be tightened, for example, the lower limit of the cap-off analysis step length can be increased from 0.1 seconds to 0.05 seconds, and the upper limit can be reduced from 5 seconds to 2 seconds, forcing the system to monitor and analyze at a higher frequency, thereby improving the capture ability of dangerous signs.

[0048] The redundant boundary buffer layer is a key design in the step boundary limit mechanism, which provides a temporary tolerance area for step adjustment to prevent unnecessary drastic oscillation of system analysis rhythm caused by transient jitter or temporary disturbance of data. The buffer layer defines a soft buffer area outside the rated boundary, for example, assuming that the rated upper boundary of the cap-off analysis step length is 5 seconds, the buffer layer can set a temporary tolerance area between 5 seconds and 5.5 seconds. When the parameter adjustment component calculates a new analysis step length based on the recommendations of each unit, it will first check whether the step length falls within the rated boundary. If the calculated step length temporarily exceeds the rated boundary but has not broken through the maximum limit of the buffer layer, the system will not immediately take forced reset measures, but will allow the step length to run temporarily within the buffer area for a period of time. The system will record the time the step length stays in the buffer layer, and use it together with the total step length number as a basis for judgment. The total step length number can be understood as the cumulative effect of step length adjustment instructions within a period of time. If the step length stays in the buffer layer for less than the pre-set time threshold, and the total calculation load increment brought by the step length adjustment does not exceed the maximum step length number that the redundant boundary buffer layer can accommodate, the system determines that this out-of-boundary is an acceptable temporary fluctuation, allowing it to continue running in the current state and observing.

[0049] The analysis decision component continuously assesses the dynamic performance of the three key indicators, adjusted decap rate, silence gradient rate, and end-off state feature, to observe whether they are within a "stable domain" defined by the redundant boundary buffer layer. This stable domain is a sub-region within the buffer layer where, although the step size may slightly exceed the nominal boundary, the core monitoring indicators show that the system state is relatively stable and risks are controllable. For example, it can be defined that when the decap rate is less than 70% of its high-risk threshold, the silence gradient rate is negative or close to zero, and the end-off state feature continuously shows "normal", the system is considered to be in the stable domain. When these conditions are met, the parameter adjustment component will continue to fine-tune the analysis step size of the analysis decision component according to the continuous input of real-time data, trying to make it slowly and smoothly return to within the nominal boundary. This gradual adjustment strategy helps to maintain the continuity of system monitoring, see Table 1.

[0050] Table 1: Step size adjustment and buffer layer state determination

[0051]

[0052] If the step size adjustment calculated by the analysis decision component results in the total compensation amount exceeding the maximum step size amount set by the redundant boundary buffer layer, and the residence time of the step size in the buffer layer also exceeds the set time threshold, the system will trigger the step size freezing mechanism. Exceeding the total compensation amount means that the system has made too frequent step size adjustments in a short period of time, consuming too many computing resources, and there is a risk of over-response to data noise. Exceeding the threshold of residence time indicates that the step size anomaly is not a short-term phenomenon, but may be caused by persistent interference or system parameter mismatch. Step size freezing is a protective measure. Once triggered, the analysis step size of each unit of the analysis decision component will no longer be adaptively adjusted according to real-time data, but will be fixed at the value at the current time. This frozen state will continue until the system confirms through self-checking that external interference has been eliminated, or manual reset is performed by the operation and maintenance personnel, so as to ensure that the system can still maintain the basic safety monitoring function at a certain and predictable pace in abnormal conditions, and avoid complete failure due to parameter oscillation.

[0053] The buffer layer mechanism demonstrates its value in handling a specific exceptional scenario, assuming a sudden strong wind blows at the job site, causing the sensors on the smart safety hat to be subjected to continuous wind impact, resulting in dramatic fluctuations in the off- hat feature data. The off-hat analysis unit perceives an extremely high off-hat change rate, and quickly shortens the analysis step length to capture potential risks, the new step length calculated can be reduced from the normal 3 seconds to 0.2 seconds. However, the rated lower limit of the boundary is 0.1 seconds, 0.2 seconds is still within the boundary, but then the wind causes the data to be abnormal for a long time, the step length is further calculated to be 0.08 seconds, at this time the step length falls below the lower limit of 0.1 seconds, entering the buffer layer area. The system starts the buffer layer timing, and monitors the total step length number change. If this wind passes in a few seconds, the data returns to normal, and the step length soon returns to within the boundary, then the system cancels the alarm. But if the strong wind continues, the step length stays at 0.08 seconds for more than the time threshold, and frequent step length adjustment instructions make the total step length number compensation value exceed the maximum limit, the system will decisively trigger the step length freeze, lock the analysis step length at 0.1 seconds or another safe value, stop this excessive response to wind noise, and instead rely on the data of the silent analysis unit and the off-hat analysis unit for comprehensive judgment to prevent false positives.

[0054] In the case that the analysis decision component outputs a decision signal that requires triggering a safety alarm by fusing and analyzing multi-dimensional data such as hat-off, silence, and end-docking, this signal is transmitted to the parameter adjustment component in real time. The parameter adjustment component, as the execution control center of the system, is responsible for converting the decision signal into specific control instructions for the alarm device and recording unit. The generation of the decision signal is not triggered by a single condition, but based on a set of multi-level, weighted evaluation logic. For example, the system may detect continuous hat-off anomalies, long silence states, and "loose" signals output by the end-docking classification model at the same time. The superposition of these multiple high-risk indicators generates a high-priority alarm decision signal. On the other hand, if there is only a slight anomaly in a single indicator, a low-level warning or prompt signal may be generated. After receiving the alarm decision signal, the parameter adjustment component immediately drives the sound and light alarm device to start. The sound and light alarm device is usually integrated on the smart safety hat or at a key location in the work site, aiming to arouse the awareness of the workers themselves and their surrounding guardians through strong sensory stimulation. The alarm sound is not a single and unchanging beep, but a sound pattern with a predetermined frequency. This pattern may be dynamically adjusted according to the level of the alarm. For example, for the highest level of emergency alarm, the system triggers a high-frequency, intermittent, and urgent screeching sound. This sound has strong penetrating power in noisy industrial environments. For lower-level warnings, a slightly lower frequency and slower-paced beep may be used to distinguish but avoid unnecessary panic. The flashing device is also started at the same time, emitting a predetermined intensity of light. The intensity of the flash is designed to ensure that it can be clearly perceived even in strong sunlight outdoor high-altitude environments. The flash pattern is usually synchronized with the sound frequency, using high-frequency flashing to enhance the warning effect. This sound and light combined alarm method fully utilizes the auditory and visual channels of humans, greatly improving the reliability and timeliness of information transmission.

[0055] At the same time of triggering the alarm, the parameter adjustment component sends a trigger signal to the recording unit, instructing it to immediately store the current event data and related context information. The recording unit is usually composed of non-volatile memory, which has the ability to save data even in the case of power failure. The stored event data package is a structured collection of information, which not only contains the simple conclusion of "alarm triggered", but also includes all the original data that led to the alarm and the snapshot of the analysis process. The specific content may include but is not limited to: the specific violation type that triggered the alarm, the original sensor reading fragment returned from the intelligent safety helmet and intelligent safety rope, the key feature values extracted by the feature processing component, the intermediate judgment results of each analysis unit in the analysis decision component, and the alarm parameters finally adopted by the parameter adjustment component. Each data package is accurately time-stamped, and this time stamp is strictly aligned with the unified time axis maintained by the data synchronization component, ensuring that the entire event chain from data collection to alarm triggering is completely traceable. This detailed record is crucial for post-accident review, responsibility definition, and optimization iteration of system algorithms.

[0056] Consider a specific application scenario: a high-altitude worker is maintaining the outer wall, and the intelligent safety rope's end buckle classification model continuously outputs "normal" state, and the silent analysis unit also shows that its activity is normal, but the safety helmet's contact pressure feature is continuously abnormal in the last three analysis periods, and the off-helmet change rate is rising, indicating that the safety helmet may be gradually loosening. The analysis decision component integrates these information and judges that there is a significant risk of safety helmet falling off, and generates a high-level alarm decision signal. After receiving the signal, the parameter adjustment component immediately starts the sound and light alarm, and the alarm on the safety helmet emits urgent buzzing and flashing red light, reminding the worker to pay attention and re-tighten the safety helmet. At the same time, the recording unit starts working, it saves the data window of about 30 seconds from the beginning of the pressure feature anomaly to the alarm trigger, including the original waveform of the pressure sensor, the calculated off-helmet change rate curve, and the logic trajectory of the system's final decision, all data are attached with millisecond-level time stamp. If the worker responds to the alarm and puts on the safety helmet within a few seconds, the pressure feature returns to normal, and the system will cancel the alarm, but all the data of this event has been completely archived. If the system does not get a response, for example, the worker may not notice the alarm for some reason, the off-helmet state continues, and even the silent analysis unit starts reporting anomalies, the system may upgrade the alarm level after a preset time, such as increasing the alarm volume and flashing frequency, or starting the linkage alarm with the remote monitoring center.

[0057] The control logic of the alarm device has certain intelligence, for example, after triggering the continuous alarm, the system will intermittently monitor whether the risk indicators are eliminated due to the response action of the personnel. If the hat removal feature is detected to be restored to normal and the undocking state is confirmed to be locked, the alarm device will automatically stop after a minimum period of continuous sound and light alarm to avoid long ringing interference with normal work order. However, regardless of whether the alarm is automatically stopped, the recording unit will record the complete timeline of the entire event - from the risk appearing, the alarm triggering, to the state returning to normal. For extreme cases that require immediate evacuation or intervention, the system can be pre-set to be unable to automatically mute, and must be manually confirmed and reset by the on-site safety officer or through remote control.

[0058] While embodiments of the present application have been shown and described, it is to be understood that the embodiments described are merely divergences, modifications, replacements and variations of the embodiments, which can be made by those of ordinary skill in the art without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A high-altitude operation safety early warning system based on multi-sensor fusion, characterized in that, It includes data acquisition components, data synchronization components, violation detection components, feature processing components, analysis and decision-making components, and parameter adjustment components; The data acquisition component consists of a smart helmet and a smart safety rope. The smart helmet collects data on the helmet-off state and the silent state, while the smart safety rope collects data on the unbuckled state. The data synchronization component receives data from the data acquisition component and processes multi-source data through time alignment and format standardization. The violation identification component makes a preliminary violation judgment. The feature processing component extracts features from the data and adds time stamps. The analysis and decision component performs multi-level analysis and outputs decision signals. The parameter adjustment component adjusts the analysis parameters and controls the alarm device according to the decision signals. The preliminary violation judgment includes state threshold comparison, single sensor anomaly, and dual verification; the state threshold comparison is used to compare whether the hat-removal data, silent data, and last-deduction data are within the normal state range. If the hat-removal data, silent data, or last-deduction data are not within the normal state range, the preliminary violation judgment is triggered. The single sensor anomaly is used to detect whether there are sudden changes, continuous anomalies, or signal loss in the cap removal data, silent data, and last-button data. If any data is abnormal, a single sensor anomaly report is output, and the data acquisition component is checked to see if it is working properly. The analysis and decision-making components include a hat removal analysis unit, a silent analysis unit, and a last-button analysis unit. The hat removal analysis unit takes hat removal features with time stamps as input, monitors hat removal features in real time within consecutive time points, obtains the difference between the hat removal features at the previous moment and the current moment, and predicts the difference between the hat removal features at the next moment and the current moment, triggering anomaly flags within the continuous time period. The silence analysis unit sets up a silence prediction model, predicts the silence data for the next moment through the silence prediction model, and marks the silence risk through anomaly detection; the anomaly detection includes whether there is a sudden change in the difference between the predicted silence value and the actual value, and processes the silence signal through time domain analysis methods to identify long silence periods. The final buckle analysis unit is equipped with a final buckle classification model and a final buckle recognition unit. The final buckle classification model is used to process the final buckle data obtained by the final buckle recognition unit, obtain the final buckle status characteristics through the final buckle classification model, and classify the status characteristics, including unfastened, loose, and normal. The final cut recognition unit is used to acquire multi-environment final cut data and optimize the final cut data. The multi-environment final cut includes motion-affected final cut features and human-affected final cut features. The final cut data optimization is used to filter and enhance the acquired final cut data. Obtain the rate of change of hat removal, the rate of change of silent gradient, and the features of the last-hook state within a continuous time period; The hat removal analysis unit is driven to receive the hat removal change rate within a continuous time point and adjust the hat removal parameter analysis step size; the analysis step size is adjusted by the silent gradient change rate within a continuous time point; the last buckle recognition unit obtains the last buckle signal and activates the step size adjustment by comparing it with the state feature library.

2. The high-altitude operation safety early warning system based on multi-sensor fusion according to claim 1, characterized in that: The data synchronization component establishes a multi-source data stream synchronization mechanism, which includes a state data stream and a motion data stream. The time-varying correlation between state and motion is determined by timestamp alignment. The state data stream is constructed based on a time series discretization method, which maps the state data into a continuous data sequence.

3. The high-altitude operation safety early warning system based on multi-sensor fusion according to claim 2, characterized in that: The step of driving the hat removal analysis unit to receive the hat removal change rate within a continuous time point and adjust the hat removal parameter analysis step size by using the hat removal change rate within a continuous time point includes: if the hat removal data change rate per unit time is greater than a set threshold, then shorten the hat removal parameter analysis step size; if the hat removal data change rate per unit time does not exceed the set threshold, then enter the waiting adjustment state. The step size adjustment analysis based on the rate of change of silent gradient at continuous time points includes: if the rate of change of silent gradient at the time point from the previous moment to the current moment increases more than the rate of change of silent gradient at the continuous time point from the second moment before the previous moment, and the rate of change increases more than the minimum threshold, then the step size ratio is compressed. The step of obtaining the final buckle signal through the final buckle identification unit and activating step size adjustment by comparing it with the state feature library includes: if the final buckle signal is captured as loose through the final buckle state feature, it is judged as a violation, and the analysis step size of the final buckle analysis unit is adjusted.

4. The high-altitude operation safety early warning system based on multi-sensor fusion according to claim 3, characterized in that: The step size boundary is set to receive the adjusted final buckle analysis step size, silent compression step size ratio, and adjusted decapping analysis step size, and a sensitivity level is set, including low sensitivity and medium-high sensitivity. By adjusting the step size boundary, the redundant boundary buffer layer is triggered to temporarily tolerate the step size adjustment. If the step size briefly exceeds the rated boundary, the actual adjusted total number of steps will not be greater than the maximum number of steps of the redundant boundary buffer layer. If the buffer layer stays for more than the time threshold, it will immediately switch to the minimum safe step size.

5. The high-altitude operation safety early warning system based on multi-sensor fusion according to claim 4, characterized in that: If the adjusted rate of change of the hat removal, the rate of change of the silent gradient, and the feature of the last button state are in the stable domain of the redundant boundary buffer layer, the analysis step size of the analysis decision component is further adjusted according to real-time data; if the total compensation amount of the adjustment step size of the analysis decision component is not greater than the maximum step size of the redundant boundary buffer layer, and the buffer layer dwell time is greater than the time threshold, then step size freezing is triggered.

6. The high-altitude operation safety early warning system based on multi-sensor fusion according to claim 5, characterized in that: If it is determined that a safety alarm needs to be triggered, the safety alarm triggers the sound and light alarm device to start, emitting a sound of a predetermined frequency and a flash of a predetermined intensity, and triggers the recording unit to store event data and timestamps.

Images