Laboratory safety risk dynamic assessment method based on machine learning

Abstract

The invention relates to the technical field of laboratory safety risk assessment, and discloses a laboratory safety risk dynamic assessment method based on machine learning, and the method comprises the steps: S1, collecting the multi-source real-time data of a laboratory environment and equipment; s2, carrying out preprocessing and time sequence synchronization on the multi-source real-time data; s3, extracting time sequence features from the time sequence data based on a sliding time window; s4, inputting the feature vector into a machine learning model with an online learning capability; s5, distributing according to the risk prediction result and the current data; and S6, when the risk level reaches a preset condition, triggering laboratory safety early warning and linkage control. Incremental updating is carried out on the machine learning model by inputting feature vectors in real time, adaptive iteration adjustment is carried out on a to-be-updated parameter set based on prediction error changes of a continuous time window, risk judgment is corrected in time along with environment changes, and the real-time performance, sensitivity and stability of model risk prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laboratory safety risk assessment, in particular to a laboratory safety risk dynamic assessment method based on machine learning. BACKGROUND

[0002] Laboratory safety risk assessment is to monitor and analyze laboratory environmental parameters, equipment operating status and personnel activity, so as to identify potential safety risks and make early warning means. With the development of sensing technology, data-driven methods are gradually introduced into laboratory safety risk assessment. The laboratory safety risk dynamic assessment method based on machine learning uses the correlation between environmental and equipment data to model and analyze the laboratory operating status, so as to realize intelligent risk judgment. At present, laboratory safety risk assessment mostly relies on fixed parameter models or offline trained machine learning models.

[0003] However, in the current technology, laboratory safety risk assessment generally relies on fixed or offline trained machine learning models, which cannot automatically update model parameters according to real-time changing environmental data, resulting in lagging risk prediction and even inaccuracy. SUMMARY

[0004] In view of the shortcomings of the prior art, the present application provides a laboratory safety risk dynamic assessment method based on machine learning, which solves the problems of difficulty in automatically updating parameters according to real-time environmental changes, lagging risk prediction and insufficient accuracy.

[0005] To achieve the above purpose, the present application realizes the following technical scheme: a laboratory safety risk dynamic assessment method based on machine learning, comprising: S1, collecting multi-source real-time data of laboratory environment and equipment, the data including temperature and humidity, gas concentration, current and voltage, equipment operating status and personnel behavior data; S2, preprocessing and time sequence synchronization of the multi-source real-time data to form continuous time sequence data; S3, extracting time sequence features from the time sequence data based on a sliding time window to generate a feature vector for risk assessment; S4, inputting the feature vector into a machine learning model with online learning capability to obtain laboratory safety risk prediction results, and incrementally updating the machine learning model based on real-time input data; S5, dynamically adjusting the risk assessment threshold according to the risk prediction results and current data distribution to generate adaptive risk level division; S6, when the risk level reaches a preset condition, triggering laboratory safety warning and linkage control, and using the feedback information of the warning and control to update the machine learning model.

[0006] Through the above technical solutions, by collecting multi-source real-time data of the laboratory environment and equipment and performing preprocessing, continuous time series data is constructed. Then, based on a sliding time window, temporal features are extracted to generate feature vectors, which are input into a machine learning model with online learning capabilities to obtain the laboratory safety risk prediction results. That is, by continuously inputting data, the machine learning model is incrementally updated, and the model parameters are adaptively adjusted by combining the prediction performances of multiple consecutive time windows, enabling the model to continuously learn the latest change trends and maintain prediction accuracy. At the same time, the risk assessment threshold is dynamically adjusted according to the prediction results and the current data distribution, and when the preset risk level is reached, early warning and linkage control are triggered. The feedback information after control is further used for model update, constructing a closed-loop mechanism for risk prediction, threshold adjustment, linkage feedback, and model self-learning, thereby achieving real-time, dynamic, and accurate assessment of laboratory safety risks.

[0007] Preferably, the S1 step includes: Set multiple sensing nodes in the laboratory to collect temperature and humidity parameters, gas concentration parameters, current and voltage parameters, and equipment status parameters respectively; Collect multi-source real-time data of personnel entry and exit records and identities through an access control recognition device and send it to the server in a streaming manner.

[0008] Preferably, the S2 step includes: Perform noise filtering, outlier removal, and missing value filling on the multi-source real-time data; Align the timestamps of the multi-source real-time data based on a unified sampling benchmark; Perform normalization on the aligned various types of data to form continuous time series data.

[0009] Preferably, the S3 step includes: Set a sliding time window with a window length and a step size on the continuous time series data; Calculate the temporal feature combination of the mean, variance, change rate, and trend slope of the environmental parameters within each sliding time window and form a feature vector.

[0010] Preferably, the S4 step includes: Input the feature vector corresponding to the current time window into an online learning model to generate a risk prediction result; Determine the parameter set to be updated according to the latest feature vector; Only adjust the parameters of the parameter set to be updated to complete the incremental update.

[0011] Preferably, the incremental update is based on the prediction error change trend within multiple consecutive time windows, and the determined set of parameters to be updated is iteratively adjusted multiple times to perform adaptive updates.

[0012] Preferably, step S5: Calculate the mean, standard deviation, and skewness of the risk forecast based on the current risk forecast results; The threshold range corresponding to the risk level is adjusted based on the mean, standard deviation, and skewness. The adaptive risk level classification is generated based on the adjusted threshold.

[0013] Preferably, step S6 includes: An alarm message will be issued when the risk level reaches the preset conditions; Send ventilation control commands to the laboratory ventilation system and power control commands to the laboratory power system; Record environmental changes after the execution of linkage control and generate corresponding feedback information.

[0014] Preferably, the feedback information is labeled and used as a new training sample, and when the incremental update is performed, the new training sample is assigned a higher weight than the historical training samples.

[0015] Preferably, the dynamic adjustment of the risk assessment threshold includes: A moving average is applied to the risk prediction results across multiple consecutive time windows. The risk assessment threshold is updated based on the risk prediction results after processing with a moving average.

[0016] This invention provides a machine learning-based method for dynamic assessment of laboratory safety risks. It offers the following advantages: 1. This invention incrementally updates the machine learning model by inputting feature vectors in real time, performs adaptive iterative adjustment of the set of parameters to be updated based on the changes in prediction error over a continuous time window, and promptly corrects risk judgments as the environment changes, thereby improving the real-time performance, sensitivity, and stability of the model's risk prediction.

[0017] 2. This invention employs a sliding time window method to extract temporal features from multi-source continuous time series data. It comprehensively utilizes statistical and trend information such as mean, variance, rate of change, and trend slope to construct feature vectors, enabling the model to capture the temporal variation patterns of the laboratory environment and equipment status. This feature construction method effectively reflects transient changes and trend shifts in parameters, providing a more timely and relevant input basis for dynamic risk prediction.

[0018] 3. This invention analyzes the risk prediction results and current data distribution, dynamically adjusts the risk assessment threshold, generates an adaptive risk level classification, and triggers linkage control when the risk reaches the preset conditions. The feedback information after control is recorded for model updates, thus constructing a closed-loop mechanism from risk prediction and early warning linkage to feedback learning. This makes risk identification more consistent with the actual operation of the laboratory and improves the system's adaptive capability and overall safety management efficiency. Attached Figure Description

[0019] Figure 1 This is a flowchart of the machine learning-based dynamic assessment method for laboratory safety risks according to the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be clearly and completely described 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.

[0021] Please see the appendix Figure 1 This invention provides a machine learning-based method for dynamic assessment of laboratory safety risks, including: S1. Collect multi-source real-time data on the laboratory environment and equipment, including temperature and humidity, gas concentration, current and voltage, equipment operating status, and personnel behavior data; Furthermore, step S1 includes: Multiple sensor nodes were set up in the laboratory to collect temperature and humidity parameters, gas concentration parameters, current and voltage parameters, and equipment status parameters respectively. The access control device collects multi-source real-time data on personnel entry and exit records and identities, which is then streamed to the server.

[0022] Specifically, multiple sensor nodes are set up in different areas within the laboratory to collect parameters such as temperature and humidity, gas concentration, current and voltage, and equipment status. Each sensor node acquires monitoring data according to a set sampling frequency and generates time-stamped data records. This method continuously reflects changes in the laboratory environment and the operational status of equipment, providing raw input for subsequent processing. In addition, access control devices installed at the laboratory entrance collect personnel entry and exit records and personnel identity information. These access control devices can use methods such as card swiping or facial recognition to generate corresponding records when personnel enter or leave the laboratory, thereby obtaining information on personnel activities within the laboratory and associating personnel identity with entry and exit times. In terms of data transmission, the data collected by each sensor node and access control device is sent to the server in a streaming manner. The data transmission module can use a message queue to upload data, so that laboratory environmental parameters, equipment status data and personnel behavior data can be continuously and in real time transmitted to the server, providing a continuous data stream for subsequent steps.

[0023] For example, in a specific instance, when the gas concentration detector detects a current gas concentration of 10 ppm, this data is immediately sent to the server via the data transmission module. Simultaneously, the access control device records a person entering the laboratory, and their identity information and entry / exit time are also uploaded synchronously. After receiving these two types of data, the server incorporates them into the subsequent processing flow in chronological order.

[0024] S2. Preprocess and synchronize multi-source real-time data to form continuous time series data; Furthermore, step S2 includes: Perform noise filtering, outlier removal, and missing value completion on multi-source real-time data; The timestamps of multi-source real-time data are aligned based on a unified sampling benchmark; Normalization is performed on the aligned data to form continuous time series data.

[0025] Specifically, the first step is to perform noise filtering, outlier removal, and missing value completion on the multi-source real-time data to make the monitoring data more stable and reduce noise interference in subsequent modeling. Subsequently, the timestamps of the multi-source real-time data are aligned based on a unified sampling benchmark. Since the data sampling frequencies of different types of sensor nodes may differ, in order to ensure the consistency of data in the time dimension, this embodiment uses a unified time step for data alignment. Based on the preset sampling interval, the most recent valid data of each sensor node at each time point is selected as the alignment result, so that each data item corresponds to a unique time label, thereby forming consistent time series data. After completing timestamp alignment, normalization is performed on the aligned data. This is because there are differences in the numerical ranges of environmental parameters, equipment parameters, and personnel behavior parameters. For example, temperature is generally in the range of tens of degrees Celsius, while current values ​​may be between several amperes and tens of amperes. If these values ​​are directly used for subsequent feature extraction and model training, they may cause an unbalanced impact on the model weights. Therefore, normalization is used, such as linearly mapping the data according to its maximum and minimum values, so that different types of data are mapped to the same numerical range, thereby forming continuous time series data with a uniform scale. For example, in a specific instance, the temperature sensor collects data of 25.3℃ at a certain time point, while the gas concentration detector collects a value of 8ppm and the current acquisition device collects a value of 1.6A. Personnel behavior data is missing at this time point. After the personnel behavior data is supplemented by interpolation, all data are normalized under a unified time reference, converting them into small values ​​in the range of 0 to 1 to form a continuous time series input suitable for subsequent use. This can effectively improve the integrity, continuity and usability of sensor data.

[0026] S3. Extract time series features from time series data based on a sliding time window to generate feature vectors for risk assessment; Furthermore, step S3 includes: Set the window length and step size for a sliding time window on continuous time series data; Within each sliding time window, the time-series characteristic combination of the mean, variance, rate of change, and trend slope of environmental parameters is calculated and formed into a feature vector.

[0027] Specifically, by employing a sliding time window, the model reflects the changes in laboratory environmental parameters over time, enabling it to assess risk based on continuous trends. First, a sliding time window with a window length and step size is set on the continuous time series data. The window length determines the time range covered by one feature extraction, and the step size controls the sliding interval between adjacent windows. Using a sliding window approach allows the time series data to form multiple time-continuous window segments, facilitating independent analysis of the state of each time period. After setting the sliding time window, the mean, variance, rate of change, and trend slope of the environmental parameters are calculated in each window to form a combination of time-series features of the window. This combination can reflect the overall level, fluctuation, magnitude of change, and trend of the environmental parameters within the window, enabling the model to fully understand the changing patterns of the laboratory environment and equipment status during that period. After feature extraction within the window is completed, the features are combined in a preset order to form the feature vector corresponding to the window. Each feature vector is used to describe the statistical characteristics and changing trends of the laboratory environmental parameters within the window, serving as structured data for model input.

[0028] For example, in one specific embodiment, the window length is set to 60 seconds and the step size to 10 seconds. When the temperature sampling values ​​within a certain window show an upward trend, the mean value of that window is slightly higher, the variance is lower, the rate of change is positive, and the trend slope is positive. Combining these features forms the feature vector of that window, thereby accurately reflecting the continuous upward trend of temperature within that window and providing a valid basis for the model's risk assessment.

[0029] S4. Input the feature vector into a machine learning model with online learning capabilities to obtain laboratory safety risk prediction results, and incrementally update the machine learning model based on real-time input data; Furthermore, step S4 includes: Input the feature vector corresponding to the current time window into the online learning model to generate risk prediction results; The set of parameters to be updated is determined based on the latest feature vector; Only the set of parameters to be updated is adjusted to complete the incremental update.

[0030] Specifically, feature vectors are input into a machine learning model with online learning capabilities to generate risk prediction results, and the model is incrementally updated based on the latest input data, enabling the model to dynamically adapt to changes in the environment and improve the real-time performance and accuracy of risk prediction. In one embodiment, a machine learning model with online learning capabilities may be used, including any one of an online random forest model, an incremental gradient model, or a neural network model with a recurrent structure. The above models can adjust local parameters after receiving new feature vectors, thereby achieving incremental learning. First, the feature vector corresponding to the current time window is input into the online learning model to generate risk prediction results. The online learning model can obtain a predicted value reflecting the current risk level of the laboratory by performing forward calculation on the input feature vector based on gradient update, incremental update, or local parameter update mechanism. For example, when the input feature vector reflects a continuous increase in temperature and a small fluctuation in gas concentration, the model may output a higher risk score to indicate potential hazards. Subsequently, the set of parameters to be updated is determined based on the latest feature vector. Since online learning models need to gradually adjust their parameters based on the latest data to enhance their adaptability to changes over time, this embodiment determines the set of parameters to be updated by analyzing the sensitivity of the current feature vector to the model output. For example, parameters that have a significant impact on the model output can be identified based on gradient information, and these parameters can be included in the set to be updated, avoiding the need to adjust all model parameters, thereby reducing computation and improving model response speed. After determining the set of parameters to be updated, only the parameters in this set are adjusted to complete the incremental update. Parameter adjustment can be based on the update rules of the online learning model, for example, for a set containing parameters... The model can be updated using the following formula: ; in, For the j-th parameter in the parameter set before the t-th update, For learning rate, This is the gradient value calculated based on the latest eigenvectors. This formula is only applied to parameters in the set to be updated, while parameters in the model not included in the set remain unchanged, thus achieving local incremental updates to the model. Through this update method, the model can maintain overall stability while rapidly responding to recent data changes, improving the adaptability of risk prediction. This enables real-time prediction based on sliding time window input data and incremental updates of model parameters, allowing the online learning model to continuously adapt to changes in the laboratory environment and equipment status, providing dynamic, stable, and real-time support for laboratory safety risk assessment.

[0031] Furthermore, the incremental update is based on the trend of prediction error changes within multiple consecutive time windows, and the determined set of parameters to be updated is iteratively adjusted multiple times to perform adaptive updates.

[0032] Specifically, to enable the online learning model to continuously optimize according to environmental changes, an adaptive update based on the prediction error trend across multiple consecutive time windows is adopted. This means that after the model completes its prediction for the current window, the error between the model's prediction and the actual feedback is recorded, and the error trend is analyzed over multiple consecutive time windows. When a continuous upward trend or increased fluctuation in the prediction error is detected, it indicates that the model's current parameters are less adaptable to recent data. Therefore, based on this trend information, the previously determined set of parameters to be updated is iteratively adjusted multiple times. In each iteration, the parameters in the set are locally updated based on the input features and error information of the latest window, allowing the model to gradually reduce prediction bias and restore stability. Thus, this iterative adjustment method based on error trends achieves adaptive model updates, enabling the model to maintain accurate risk level prediction capabilities even when the time series is continuously changing.

[0033] S5. Based on the risk prediction results and the current data distribution, dynamically adjust the risk assessment threshold to generate an adaptive risk level classification. Further, S5 steps: Calculate the mean, standard deviation, and skewness of the risk forecast based on the current risk forecast results; The threshold range corresponding to the risk level is adjusted based on the mean, standard deviation, and skewness; An adaptive risk level classification is generated based on the adjusted threshold.

[0034] Specifically, based on the output risk prediction results and combined with the current data distribution, the risk assessment threshold is dynamically adjusted so that the risk level classification can be updated in real time with changes in the laboratory environment and equipment status, thereby improving the adaptability of the risk assessment results. First, the mean, standard deviation, and skewness of the risk prediction values ​​are calculated based on the risk prediction results over a current period. Since the online learning model generates multiple risk prediction values ​​within a continuous time window, the above statistics can be calculated to reflect the central tendency, fluctuation range, and distribution skewness of the current risk prediction results. The statistical information can provide a basis for subsequent threshold adjustments, so that the risk level classification is no longer fixed but can respond to changes in the prediction results. After obtaining the above statistics, the threshold range corresponding to the risk level is adjusted based on the mean, standard deviation, and skewness. In one embodiment, if the risk prediction results show a concentrated upward trend in the near future (mean increases), the threshold between low risk and medium risk is increased accordingly to enhance the system's sensitivity to potential risks. If the prediction results fluctuate more (standard deviation increases), the range of each risk level can be appropriately expanded to avoid misjudgment due to fluctuations. At the same time, when the distribution is significantly skewed (skewness changes), the threshold corresponding to the high risk level can be fine-tuned to make the risk level classification more consistent with the current data distribution, thereby ensuring that the risk level evaluation system can dynamically adapt to changes in the laboratory's operating status. After adjusting the threshold, an adaptive risk level classification is generated based on the adjusted threshold. The newly generated risk level can be used to guide early warning triggering and can be continuously updated in subsequent time windows. By adopting a dynamic threshold mechanism, false alarms or missed alarms caused by fixed thresholds in situations of drastic environmental changes can be effectively avoided, thereby maintaining the stability and real-time performance of the risk assessment system.

[0035] S6. When the risk level reaches the preset conditions, trigger laboratory safety early warning and linkage control, and use the feedback information of the early warning and control to update the machine learning model.

[0036] Furthermore, step S6 includes: An alarm message will be issued when the risk level reaches the preset conditions; Send ventilation control commands to the laboratory ventilation system and power control commands to the laboratory power system; Record environmental changes after the execution of linkage control and generate corresponding feedback information.

[0037] Specifically, when the risk level reaches the preset conditions, a laboratory safety warning is triggered and corresponding linkage control operations are executed. At the same time, the feedback information after the linkage control is implemented is used for model updates. When potential risks are detected, timely measures can be taken, and the online learning model can be continuously optimized based on the actual execution results, thereby achieving a dynamic closed loop of risk assessment. In one embodiment, when the generated risk level reaches the preset warning conditions, an alarm message is immediately issued. The alarm message can be sent to the laboratory management personnel through sound and light alarms, display terminal interfaces or mobile terminal notifications to alert them of the potential risks. By issuing timely warnings, the management personnel can understand the laboratory status as soon as possible and reduce the possibility of safety accidents. After issuing an alarm, a ventilation control command is sent to the laboratory ventilation system and a power control command is sent to the laboratory power system based on the current risk level to execute the linkage control. The ventilation control command can be used to start the ventilation equipment in the workshop, thereby reducing the concentration of combustible or harmful gases in the air. The power control command can be used to cut off the power to some or all of the experimental equipment to avoid further safety risks that may be caused by the continued operation of the equipment. This allows for rapid intervention when the risk level rises, ensuring the safety of laboratory personnel and equipment. After the linkage control is executed, the changes in environmental parameters, including indicators such as gas concentration, temperature and humidity, current and voltage, and equipment operating status, are recorded. This information is then used to generate feedback information and combined with the online learning model update process to evaluate the actual effect of the linkage control. For example, after ventilation control is executed, the effectiveness of the control measures can be judged by monitoring whether the gas concentration drops significantly. By using this feedback information to update the model, the online learning model can better adapt to actual environmental changes in subsequent predictions, thereby improving the accuracy and stability of risk assessment. For example, if the risk level is determined to be "high risk" within a certain time window, an alarm will be issued immediately, and a start command will be sent to the ventilation system and a partial power cut command will be sent to the power control system. After the control is executed, the sensor detects that the concentration of combustible gas has dropped from 15 ppm to 8 ppm, and the equipment status will switch from running to standby. These changes will be recorded as feedback information to improve the model's ability to learn the actual effects of ventilation and power cut measures during the next model update, thereby enhancing the overall safety and intelligence of the laboratory operation.

[0038] Furthermore, the feedback information is labeled and used as new training samples, and during incremental updates, the new training samples are assigned a higher weight than the historical training samples.

[0039] Specifically, to improve the online learning model's ability to learn from actual control effects, feedback information is labeled after it is obtained. The input feature vector corresponding to the feedback information is combined with the actual risk results to form new training samples. When performing incremental updates, the new training samples are given a higher weight than historical training samples. When the model updates local parameters, the new training samples are included in the current update batch, so that the model focuses more on reflecting the real-time environmental changes represented by the new samples when calculating parameter adjustment amounts, thereby enhancing the model's adaptability to the latest risk change trends. For example, in one embodiment, when the system issues ventilation and power-off commands at a high-risk moment, and detects a significant decrease in the concentration of combustible gas, this change is recorded as feedback information and marked as "effective handling." In the next incremental update, the model assigns a higher weight to this new sample, enabling the model to learn the "causal relationship between ventilation measures and risk reduction" more quickly, thus exhibiting higher accuracy and stability in subsequent predictions. Through the above method, this embodiment can achieve feedback-based reinforcement learning, enabling the model to continuously improve its ability to identify and judge changes in risk during continuous operation.

[0040] Dynamic adjustment of risk assessment thresholds includes: A moving average is applied to the risk prediction results across multiple consecutive time windows. The risk assessment threshold is updated based on the risk prediction results after processing with a moving average.

[0041] Specifically, to ensure that risk level classification adapts to real-time changes in laboratory risk status, the system performs a moving average process on risk prediction results obtained over multiple consecutive time windows. This smooths short-term fluctuations and extracts the overall trend. Subsequently, the risk assessment threshold is updated based on the risk prediction values ​​after the moving average process. This allows the threshold to dynamically reflect recent overall changes in risk levels. For example, when risk prediction values ​​show a gradually increasing trend over multiple consecutive time windows, the comprehensive risk value obtained through the moving average also increases accordingly. This automatically raises the threshold between low and medium risk, making the risk level classification more accurate for potential risks. Sensitive to changes, if the continuous prediction results change steadily at a low level, the moving average will approach a stable state. In this case, the thresholds for each risk level remain unchanged or are only slightly adjusted. For example, when the model prediction values ​​for the most recent five time windows are 0.52, 0.55, 0.58, 0.60 and 0.62 respectively, the moving average reflects the overall upward trend of risk. Accordingly, the medium-risk threshold is appropriately increased so that the subsequent risk level judgment is more in line with the current laboratory risk level. Through the above method, its dynamic threshold adjustment mechanism can reduce the false alarms or false negatives that may be caused by fixed thresholds, making the risk level classification more stable and adaptable in real time.

[0042] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based dynamic assessment method for laboratory safety risks, characterized in that, include: S1. Collect multi-source real-time data on the laboratory environment and equipment, including temperature and humidity, gas concentration, current and voltage, equipment operating status, and personnel behavior data; S2. Preprocess and synchronize the multi-source real-time data to form continuous time series data; S3. Extract time-series features from the time-series data based on a sliding time window to generate a feature vector for risk assessment; S4. Input the feature vector into a machine learning model with online learning capabilities to obtain laboratory safety risk prediction results, and incrementally update the machine learning model based on real-time input data; S5. Based on the risk prediction results and the current data distribution, dynamically adjust the risk assessment threshold to generate an adaptive risk level classification; S6. When the risk level reaches the preset conditions, trigger laboratory safety early warning and linkage control, and use the feedback information of the early warning and control to update the machine learning model.

2. The method for dynamic assessment of laboratory safety risks based on machine learning according to claim 1, characterized in that, Step S1 includes: Multiple sensor nodes were set up in the laboratory to collect temperature and humidity parameters, gas concentration parameters, current and voltage parameters, and equipment status parameters respectively. The access control device collects multi-source real-time data on personnel entry and exit records and identities, which is then streamed to the server.

3. The method for dynamic assessment of laboratory safety risks based on machine learning according to claim 1, characterized in that, Step S2 includes: The multi-source real-time data is subjected to noise filtering, outlier removal, and missing value completion processing. The timestamps of the multi-source real-time data are aligned based on a unified sampling benchmark. Normalization is performed on the aligned data to form continuous time series data.

4. The method for dynamic assessment of laboratory safety risks based on machine learning according to claim 1, characterized in that, Step S3 includes: A sliding time window with a window length and step size is set on the continuous time series data; Within each sliding time window, the time-series characteristic combination of the mean, variance, rate of change, and trend slope of environmental parameters is calculated and formed into a feature vector.

5. The method for dynamic assessment of laboratory safety risks based on machine learning according to claim 1, characterized in that, The S4 step includes: Input the feature vector corresponding to the current time window into the online learning model to generate risk prediction results; The set of parameters to be updated is determined based on the latest feature vector; The incremental update is completed by adjusting the parameters only on the set of parameters to be updated.

6. The method for dynamic assessment of laboratory safety risks based on machine learning according to claim 5, characterized in that, The incremental update is based on the prediction error change trend within multiple consecutive time windows, and the determined set of parameters to be updated is iteratively adjusted multiple times to perform adaptive updates.

7. The method for dynamic assessment of laboratory safety risks based on machine learning according to claim 1, characterized in that, Step S5: Calculate the mean, standard deviation, and skewness of the risk forecast based on the current risk forecast results; The threshold range corresponding to the risk level is adjusted based on the mean, standard deviation, and skewness. The adaptive risk level classification is generated based on the adjusted threshold.

8. The method for dynamic assessment of laboratory safety risks based on machine learning according to claim 1, characterized in that, Step S6 includes: An alarm message will be issued when the risk level reaches the preset conditions; Send ventilation control commands to the laboratory ventilation system and power control commands to the laboratory power system; Record environmental changes after the execution of linkage control and generate corresponding feedback information.

9. The method for dynamic assessment of laboratory safety risks based on machine learning according to claim 1, characterized in that, The feedback information is labeled and used as new training samples. When the incremental update is performed, the new training samples are assigned a higher weight than the historical training samples.

10. The method for dynamic assessment of laboratory safety risks based on machine learning according to claim 1, characterized in that, The dynamic adjustment of the risk assessment threshold includes: A moving average is applied to the risk prediction results across multiple consecutive time windows. The risk assessment threshold is updated based on the risk prediction results after processing with a moving average.

Images