Chemical equipment safety monitoring method and system based on big data

Abstract

This invention belongs to the field of chemical equipment monitoring, and relates to a method and system for safety monitoring of chemical equipment based on big data. The method includes: Step 1, constructing a multi-source sensor monitoring network for confined spaces and collecting multi-source raw data through the network; Step 2, preprocessing the multi-source raw data to obtain a standardized dataset; Step 3, inputting the standardized dataset into a dynamic risk assessment model, performing spatiotemporal coupling processing on the multi-source data, judging the failure trend of the chemical equipment, and outputting the real-time risk value of the chemical equipment; the dynamic risk assessment model includes a Crow-AMSAA model and an improved LSTM model; Step 4, triggering an early warning mechanism and emergency response strategy based on the real-time risk value; thus achieving dynamic risk assessment of chemical equipment in confined spaces.

Description

Technical Field

[0001] This invention relates to the field of chemical equipment monitoring, and specifically discloses a method and system for chemical equipment safety monitoring based on big data. Background Technology

[0002] During maintenance, cleaning, or operation under load, confined space chemical equipment is prone to accidents such as the accumulation of toxic and harmful gases, leakage of flammable gases, fatigue cracking of equipment structures, and oxygen poisoning of personnel. Because confined spaces are typically enclosed or semi-enclosed metal structures, with limited internal space, restricted ventilation, and severe wireless signal attenuation, traditional safety monitoring methods suffer from the following shortcomings: Existing systems mostly employ single-gas detection or single-point vibration monitoring; environmental parameters, equipment status, personnel location, and historical equipment reliability data are independent, lacking a unified data fusion and correlation analysis mechanism, making multi-factor coupled risk assessment impossible. Most monitoring systems use fixed threshold alarms, failing to comprehensively assess risk levels based on dynamic factors such as equipment aging trends, gas concentration change rates, and personnel exposure time, leading to delayed warnings. Existing monitoring systems often only focus on real-time anomalies, ignoring the cumulative trend of equipment failures over time, lacking reliability analysis methods based on historical failure data, and hindering predictive maintenance. Metal casings severely shield wireless signals, making stable WiFi and cellular network coverage difficult, resulting in unstable data uploads and affecting system reliability.

[0003] In view of this, the present invention provides a method and system for safety monitoring of chemical equipment based on big data. Through multi-source sensor fusion, intelligent data preprocessing, and coupling of reliability models and deep learning models, it achieves dynamic risk assessment of chemical equipment in confined spaces. It can realize unified fusion processing of multi-source heterogeneous data; introduce equipment reliability trend parameters into risk calculation; support stable data transmission in complex environments of confined spaces; and enable early prediction of accident risks, rather than post-accident alarms. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for safety monitoring of chemical equipment based on big data, addressing the problem of how to achieve dynamic coupled assessment and early warning of environmental risks, equipment failure trends, and personnel exposure risks within a confined space, under conditions of limited communication, high data noise, and heterogeneous multi-source information. The specific solution is as follows:

[0005] The big data-based safety monitoring method for chemical equipment includes: Step 1, constructing a multi-source sensor monitoring network for confined spaces and collecting multi-source raw data through the multi-source sensor monitoring network; Step 2, preprocessing the multi-source raw data to obtain a standardized dataset; Step 3, inputting the standardized dataset into a dynamic risk assessment model, performing spatiotemporal coupling processing on the multi-source data, judging the failure trend of the chemical equipment, and outputting the real-time risk value of the chemical equipment; the dynamic risk assessment model includes the Crow-AMSAA model and an improved LSTM model; Step 4, triggering an early warning mechanism and emergency response strategy based on the real-time risk value.

[0006] Furthermore, the multi-source sensors include gas sensors, vibration sensors, and condition monitoring nodes; the multi-source raw data includes environmental parameters, equipment operation data, and equipment reliability correlation data; the environmental parameters include the concentration of toxic and harmful gases, oxygen content, temperature and humidity, and distance to obstacles; the equipment operation data includes the vibration frequency, pressure value, and spectral density characteristic parameters of the chemical equipment; and the equipment reliability correlation data includes the cumulative number of equipment failures, operating time, and maintenance records.

[0007] Furthermore, the multi-source raw data is preprocessed to obtain a standardized dataset, including: removing environmental noise from the multi-source raw data through Gaussian filtering to obtain Gaussian-filtered data; filtering out equipment vibration interference from the Gaussian-filtered data through wavelet transform to obtain wavelet-transformed data; removing outliers from the wavelet-transformed data using the RANSAC algorithm and spectral density feature analysis to obtain normal data; filling missing values ​​in the normal data using the K-nearest neighbor algorithm to obtain a complete dataset; and standardizing the complete dataset to obtain a standardized dataset.

[0008] Furthermore, outlier removal is performed on the wavelet transform data using the RANSAC algorithm and spectral density feature analysis to obtain normal data. This includes: selecting multiple data points from the wavelet transform data using the RANSAC algorithm to fit a model; calculating the distance from unselected points to the fitted model and including points with a distance less than a preset threshold in the inlier set; repeating the model and inlier set generation operations to obtain the optimal fitted model, removing outliers, and obtaining the RANSAC-processed fitted data; and constructing an anomaly detection model using spectral density feature analysis to filter out anomalies in the fitted data and obtain normal data.

[0009] Furthermore, the standardized dataset is input into the dynamic risk assessment model to perform spatiotemporal coupling processing on multi-source data, and to judge the failure trend of chemical equipment, outputting the real-time risk value of the chemical equipment. This includes: enhancing the features of the standardized dataset through a dynamic ARMA model to capture the temporal features of the chemical equipment state and obtain a temporal enhanced feature vector; calculating attention weights based on the constrained space feature attention scoring function and combined with DLT extrinsic parameter optimization logic, and weighting the temporal enhanced feature vector to obtain an initial weighted feature vector; dynamically adjusting the weight coefficients in the initial weighted feature vector based on Crow-AMSAA to obtain an adjusted weighted feature vector; predicting the equipment failure trend to obtain the expected cumulative number of equipment failures; and inputting the adjusted weighted feature vector and the expected cumulative number of equipment failures into the risk value prediction model, and calculating the real-time risk value by combining data authenticity and reliability correction coefficients.

[0010] Furthermore, the dynamic ARMA model is as follows: ; in, For temporal enhancement features; This is the autoregression order index; P is the autoregression order. These are the autoregressive coefficients; Enhance historical features; Index of moving average order; The moving average order; The moving average coefficient; For historical residuals; The spatial coupling weighting coefficient is denoted by n; n represents the sensor variable. This represents the total number of sensor nodes; Spatial weights for sensors; To standardize the features of the nth sensor in the dataset at time t; The formula for dynamically adjusting the weighting coefficients is as follows: ; in, The dynamic risk weight for the i-th feature; t represents the membership degree of the i-th feature; i is the feature variable; t is the time variable; To ensure data accuracy; For the shape parameters of the Crow-AMSAA model; The expected cumulative number of failures is: ; in, Let N(t) be the expected cumulative number of faults at time t; N(t) is the total number of actual faults that occurred in the chemical equipment from the time it was put into operation to time t. For scale parameters; Real-time risk value: ; in, The risk value for time period t; t is a time variable; Let be the dynamic risk weight of the i-th feature; i is the feature variable; Let be the value of the i-th feature in the context vector at time t; To ensure data accuracy; This represents the maximum value of data accuracy. The expected cumulative number of faults; This represents the actual number of faults that have occurred. This is the threshold for the maximum number of allowed faults.

[0011] Furthermore, based on the constrained spatial feature attention scoring function, attention weights are calculated using DLT extrinsic optimization logic, and weights are assigned to the temporal augmentation feature vector to obtain an initial weighted feature vector. This includes: constructing the constrained spatial feature attention scoring function to obtain a score value; calculating attention weights based on the score value; and calculating the context vector based on the attention weights and the structure of the sensor network.

[0012] Furthermore, the scoring function for the attention given to features in a confined space is as follows: ; in, This is the score; Let be the query vector at time t; Let be the transpose of the query vector at time t; W is the optimized weight matrix. The key vector is t; time is t. Attention weights are: ; in, is the attention weight; t is the time variable; exp is the exponential function; n is the sensor variable; N is the total number of sensors; The context vector is: ; in, Let n be the context vector at time t; t is the time variable; n is the sensor variable; N is the total number of sensors. Let be the attention weight of the nth sensor at time t; This is the value vector of the nth sensor; Spatial position weight of the nth sensor.

[0013] Furthermore, based on the real-time risk value, an early warning mechanism and emergency response strategy are triggered, including: when the risk value is less than the first early warning threshold, a level one early warning is activated for routine monitoring; when the risk value is greater than or equal to the first early warning threshold but less than the second early warning threshold, ventilation equipment is activated and a risk warning message is pushed; when the risk value is greater than or equal to the second early warning threshold but less than the third early warning threshold, the operation of some equipment is stopped and an evacuation order is pushed; when the risk value is greater than or equal to the third early warning threshold but less than the fourth early warning threshold, the operation of all equipment is stopped and an evacuation order is pushed, while an emergency rescue plan is activated.

[0014] The big data-based chemical equipment safety monitoring system, based on the aforementioned big data-based chemical equipment safety monitoring method, includes a data acquisition module, a preprocessing module, a risk value determination module, and an early warning module. The data acquisition module constructs a multi-source sensor monitoring network for confined spaces and collects multi-source raw data through this network. The preprocessing module preprocesses the multi-source raw data to obtain a standardized dataset. The risk value determination module inputs the standardized dataset into a dynamic risk assessment model, performs spatiotemporal coupling processing on the multi-source data, judges the failure trend of the chemical equipment, and outputs the real-time risk value of the chemical equipment. The dynamic risk assessment model includes a Crow-AMSAA model and an improved LSTM model. The early warning module triggers an early warning mechanism and emergency response strategy based on the real-time risk value.

[0015] The present invention has the following advantages and beneficial effects: This invention employs a hierarchical monitoring network structure using LoRa self-organizing networking and wired relays at key nodes, specifically addressing the wireless signal attenuation problem caused by the shielding of metal casings in confined chemical spaces. Compared to traditional WiFi and cellular networks, this networking approach achieves stable communication between sensor nodes and the external gateway, significantly reducing data transmission packet loss. It ensures real-time uploading of multi-source information such as environmental parameters and equipment operating data, providing continuous and complete data support for dynamic risk assessment, and is suitable for monitoring needs in enclosed / semi-enclosed scenarios such as storage tanks, reactors, and underground pipe corridors.

[0016] This invention effectively solves the problems of high noise, strong heterogeneity, and poor integrity in multi-source data through a full-process preprocessing scheme. Among them, the anomaly detection model constructed by spectral density feature analysis (|A|>3 to determine anomalies) can accurately identify abnormal equipment operation data, K-nearest neighbor weighted filling improves data integrity, and the standardized data provides high-quality input for subsequent model calculations, avoiding evaluation bias caused by low-quality data.

[0017] The dynamic risk assessment model integrates ARMA time-series enhancement, DLT extrinsic optimization attention mechanism, and Crow-AMSAA reliability model to achieve spatiotemporal coupling processing of multi-source data and prediction of equipment failure trends, thereby improving the accuracy of anomaly identification. Attached Figure Description

[0018] Figure 1 An exemplary flowchart of the big data-based chemical equipment safety monitoring method provided by the present invention; Figure 2 An exemplary module diagram of the big data-based chemical equipment safety monitoring system provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Figure 1 This is an exemplary flowchart of the big data-based chemical equipment safety monitoring method provided by the present invention. Figure 1 As shown, the big data-based safety monitoring method for chemical equipment proposed in this invention includes the following: Step 1: Construct a multi-source sensor monitoring network for confined spaces and collect multi-source raw data through the multi-source sensor monitoring network.

[0021] Confined spaces refer to spaces with poor natural ventilation, relatively enclosed structures, and the potential accumulation of toxic, harmful, or flammable gases, including storage tanks, reactors, and underground pipe racks. A multi-source sensor monitoring network refers to a hierarchical monitoring network structure composed of multiple types of sensor nodes, edge acquisition units, and external gateways. Specifically, sensor nodes can be distributed in critical equipment areas and dead zones. Sensor nodes can be interconnected via LoRa self-organizing networks, and critical nodes are connected to the external gateway via wired relays. Critical equipment areas can refer to structural regions subjected to high temperatures, high pressures, strong corrosion, or concentrated mechanical loads during equipment operation, including but not limited to reactor weld areas, flange connections, pipe rack connections, and heat exchanger tube sheet areas. Critical nodes can be sensor nodes located in risk-sensitive areas (e.g., near leak sources and in enclosed dead zones) and equipped with dual communication links. For example, a node located near a reactor manhole that simultaneously collects gas concentration and structural vibration data can be considered a critical node. For example, multiple sets of nodes can be arranged along the axial direction of an underground pipe rack, each set containing gas, vibration, temperature, humidity, and displacement sensors, forming a multi-source raw data matrix.

[0022] The multi-source sensors in a multi-source sensor monitoring network can include gas sensors, vibration sensors, and condition monitoring nodes. The multi-source raw data can include environmental parameters, equipment operating data, and equipment reliability correlation data. Environmental parameters can include toxic and harmful gas concentrations, oxygen content, temperature, humidity, and obstacle distances; equipment operating data can include vibration frequencies, pressure values, and spectral density characteristic parameters of chemical equipment; and equipment reliability correlation data includes the cumulative number of equipment failures, operating time, and maintenance records.

[0023] Step 2: Preprocess the multi-source raw data to obtain a standardized dataset.

[0024] Environmental noise in the multi-source raw data is eliminated through Gaussian filtering to obtain Gaussian-filtered data. Equipment vibration interference in the Gaussian-filtered data is then filtered out using wavelet transform to obtain wavelet-transformed data. Outlier removal is performed on the wavelet-transformed data using the RANSAC (Random Sample Consensus) algorithm and spectral density feature analysis to obtain normal data. Missing values ​​in the normal data are imputed using the K-nearest neighbor algorithm to obtain a complete dataset. For example, missing values ​​can be imputed using sample data from historical similar operating conditions by selecting the weighted mean of K=5 neighbors based on Euclidean distance. Finally, the complete dataset is standardized using the Z-score normalization method to obtain a standardized dataset. , , and These are the standardized gas concentration characteristic value, the standardized equipment vibration characteristic value, and the standardized characteristic value corresponding to the nth type of sensor, respectively; n is the total number of standardized feature dimensions, which is determined by the number of sensor types and the number of features extracted from each type of sensor.

[0025] In some embodiments, obtaining normal data includes: The RANSAC algorithm is used to randomly select multiple points from the wavelet transform data to fit the initial fitting model, thus obtaining the fitted model.

[0026] Stable data points are selected based on the rated operating conditions of chemical equipment. For example, data points from wavelet transform data where vibration is below a preset vibration threshold and pressure fluctuation is below a preset pressure threshold can be selected as stable data points. , where k is a stable data point variable; For the time variable of stable data point k; , and These are the stable data points k at time [time]. The parameters are: the first parameter (e.g., pressure value), the second parameter (e.g., vibration value), and the nth parameter (e.g., surface temperature of the chemical equipment casing). A linear trend model is applied to each type of data based on the time variable. The fitting model can be a linear fitting model. For example, five stable points are selected within a 60-second window to fit the initial vibration trend line, resulting in the final vibration trend line.

[0027] Calculate the distance from the unselected points to the fitted model, and include points whose distance is less than a preset distance threshold into the inlier set.

[0028] The preset distance threshold refers to the maximum distance between a data point in the wavelet transform data and the fitted model. When the distance between a data point and the fitted model is greater than the preset distance threshold, the data point is discarded; when the distance between a data point and the fitted model is less than or equal to the preset distance threshold, the data point is selected as an inlier, and multiple inliers form an inlier set. For example, after fitting a vibration trend model, the measured vibration value at a certain moment is 2.1 mm / s, and the model predicted value is 1.8 mm / s, with a difference of 0.3 mm / s. If the preset distance threshold is 0.5 mm / s, then this point is an inlier; if the difference is 0.7 mm / s, it is determined to be an outlier and discarded.

[0029] Repeatedly fit the model and generate the interior point set to obtain the optimal fit model. Remove outlier data to obtain the RANSAC-processed fitted data.

[0030] The optimal fit model is the model with the largest number of inliers and the smallest mean squared error. Outliers are data points that are more than a preset distance threshold from the optimal fit model.

[0031] An anomaly detection model was constructed by analyzing spectral density features. The fitted data was then filtered for anomalies to obtain normal data.

[0032] An anomaly detection model is a standard deviation normalization model based on the statistical characteristics of the power spectral density distribution of vibration signals. The anomaly detection model is as follows: ; in, This is the normalized outlier score; The mean power spectral density of the vibration signal at time t is obtained by performing an FFT on the wavelet-transformed signal; t is the time variable. The historical average PSD (Power Spectral Density) is obtained from stable operating condition data. The historical PSD standard deviation is also obtained from stable operating condition data. When Data anomalies can be identified and removed.

[0033] Step 3: Input the standardized dataset into the dynamic risk assessment model, perform spatiotemporal coupling processing on the multi-source data, judge the failure trend of the chemical equipment, and output the real-time risk value of the chemical equipment; the dynamic risk assessment model includes the Crow-AMSAA (Crow-Army Materiel Systems Analysis Activity) model and the improved LSTM (Long Short-Term Memory) model.

[0034] The dynamic risk assessment model is a composite model that integrates time series forecasting, feature attention allocation, and equipment reliability trend analysis. The dynamic risk assessment model can include an ARMA (AutoRegressive Moving Average Model) time series enhancement layer, an attention-weighted layer, a Crow-AMSAA reliability layer, and an LSTM risk prediction layer. The ARMA time series enhancement layer processes the standardized dataset to obtain a time series feature vector; the attention-weighted layer processes the time series feature vector to obtain attention-weighted features; the Crow-AMSAA reliability layer processes the attention-weighted features to obtain reliability trend parameters; and the LSTM risk prediction layer processes the reliability trend parameters to obtain a risk value.

[0035] Spatiotemporal coupling processing refers to simultaneously considering the spatial location weights of sensors and their temporal variation trends. For example, gas sensors closer to the leak risk point have a higher weight than those further away. Fault trend judgment refers to determining whether the equipment has entered an accelerated deterioration phase based on the rate of change of risk values ​​and the direction of change of the reliability parameter β within multiple time windows. For example, if the slope of the risk value is >0.02 / min and β>1 for 10 consecutive minutes, the equipment is judged to have entered a fault growth trend phase. For example, if the gas concentration at the bottom of the reactor is detected to rise to 2.3 at a certain moment, and the vibration characteristic Z value is 1.8, the spatial weights are applied and input into the model; the Crow-AMSAA model calculates β=1.12, indicating an increase in the failure rate; the LSTM predicts and outputs a risk value R(t)=0.72, and the system enters a level three warning phase.

[0036] By using a dynamic ARMA model to enhance the features of a standardized dataset over time, the temporal characteristics of chemical equipment are captured, resulting in a time-enhanced feature vector.

[0037] Feature time series enhancement refers to extracting the trend, autocorrelation, and periodicity of signals through time series models. Time series features of chemical equipment status are used to characterize the vibration growth rate, temperature drift rate, and gas concentration accumulation rate of chemical equipment over time. The time series enhancement vector contains a high-dimensional feature representation of historical time series influencing factors.

[0038] A dynamic ARMA model with spatial weights is constructed. The formula for calculating the temporal feature enhancement of the dynamic ARMA model is as follows: ; in, For temporal enhancement features; This is the autoregression order index; P is the autoregression order. These are the autoregressive coefficients; Enhance historical features; Index of moving average order; The moving average order; The moving average coefficient; For historical residuals; These are the spatial coupling weight coefficients; Index the sensor numbers; This represents the total number of sensor nodes; The spatial weight of the sensor is inversely proportional to the distance from the risk source; To standardize the features of the nth sensor in the standardized dataset.

[0039] Based on the constrained space feature attention scoring function, attention weights are calculated by combining DLT (Direct Linear Transformation) extrinsic parameter optimization logic, and weights are assigned to the temporal enhancement feature vector to obtain the initial weighted feature vector; Based on the attention scoring function of constrained space features, and combined with DLT extrinsic parameter optimization logic, attention weights are calculated, including: The query vector is obtained based on the hidden state of the LSTM. The hidden state of the LSTM contains historical time-step information about the device's operation. For example, historical features such as vibration trends and gas concentration changes over the past 10 minutes. By using the hidden state of the LSTM as the query vector, the model can match the sensor features at the current moment with the historical operating trend of the device as a reference standard. For example, if the historical hidden state shows that the device vibration has been stable, the query vector will focus more on the current features that match the stable trend. The key vector is determined based on sensor feature embedding. Since the sensor data of chemical equipment is multi-source and heterogeneous, sensor feature embedding is needed to map these heterogeneous data to the same high-dimensional feature space, obtaining a feature vector with a unified format. Using the sensor feature embedding as the key vector allows the key vector and the query vector (LSTM hidden state) to calculate their similarity in the same space.

[0040] DLT extrinsic optimization logic refers to constructing a spatial projection relationship using the three-dimensional installation position coordinates of the sensor and the reference coordinates of the risk source, and optimizing the weight matrix by minimizing the projection error to make the attention weight consistent with the real spatial geometric relationship.

[0041] Construct a scoring function for the attention given to features in a restricted space, and obtain the scoring value; the scoring function for the attention given to features in a restricted space is: ; in, This is the score; Let t be the query vector at time t, which comes from the hidden state of the LSTM. Let W be the transpose of the query vector at time t; W is the optimized weight matrix, obtained through DLT extrinsic optimization and nonlinear iterative solution. is the key vector, the feature embedding vector from the nth sensor; t is the time variable.

[0042] Attention weights are calculated based on the ratings. The attention weights are: ; in, is the attention weight; t is the time variable; exp is the exponential function; n is the sensor variable; N is the total number of sensors.

[0043] Based on attention weights and the structure of the sensor network, a context vector is calculated to assign weights to the temporal augmentation feature vector. The structure of the sensor network refers to the spatial topology and communication connections of the sensor nodes. The context vector is: ; in, Let be the context vector at time t, i.e., the initial weighted feature vector; t is the time variable; n is the sensor variable; N is the total number of sensors; Let be the attention weight of the nth sensor at time t; is the value vector of the nth sensor, derived from sensor feature embedding, and is the core feature representation after mapping heterogeneous data of that sensor; The spatial location weight of the nth sensor is determined by the distance between the sensor and the risk source (such as weld seam or pipe interface) and the equipment failure risk value. The closer the distance and the higher the risk, the greater the weight.

[0044] Based on the Crow-AMSAA reliability parameter β, the weight coefficients in the weighted feature vector are dynamically adjusted to obtain the adjusted weighted feature vector.

[0045] The Crow-AMSAA reliability parameter β is obtained through logarithmic regression of historical failure times. For example, linear regression is performed on the historical cumulative failure curves ln(N(t)) and ln(t) to obtain the slope β. This invention uses dynamically adjusted weighted feature vectors obtained by adjusting the weight coefficients of the reliability parameter β to reflect the equipment failure growth trend, thereby reducing the risk weight of high-reliability equipment.

[0046] The formula for dynamically adjusting the weighting coefficients is as follows: ; in, The dynamic risk weight for the i-th feature; t represents the membership degree of the i-th feature, calculated by a fuzzy function, reflecting the correlation between the feature and the risk category; i is the feature variable; t is the time variable. To determine data accuracy, reflecting the reliability of feature data, it can be calculated as signal integrity rate multiplied by signal-to-noise ratio. The shape parameters of the Crow-AMSAA model reflect the overall reliability trend of the equipment (β<1 indicates improved reliability, β=1 indicates no change, and β>1 indicates decreased reliability). An adjusted weighted feature vector is constructed based on the one-to-one product relationship between the dynamic risk weights and the feature components of the context vector.

[0047] The expected cumulative number of equipment failures is obtained by predicting equipment failure trends using the Crow-AMSAA model. The expected cumulative failure count represents the total number of failures expected to occur before time t, and is used to assess the rate of equipment health degradation.

[0048] The formula for calculating the expected cumulative number of equipment failures in the Crow-AMSAA model is: ; in, Let N(t) be the expected cumulative number of faults at time t; N(t) is the total number of actual faults that occurred in the chemical equipment from the time it was put into operation to time t. β is the scaling parameter, reflecting initial reliability, which can be obtained through failure statistics during the early operation phase of chemical equipment. β is the shape parameter of the Crow-AMSAA model, and t is the current time.

[0049] The weighted feature vector and the expected input risk value prediction model of the cumulative number of equipment failures will be adjusted, and the real-time risk value will be calculated by combining the data authenticity and reliability correction coefficients.

[0050] A risk value prediction model is a model used to predict the risks of chemical equipment. Risk value prediction models can include multi-layer LSTM and fully connected layer networks. Real-time risk values ​​can be obtained through risk value prediction models. ; in, The risk value for time period t; t is a time variable; Let be the dynamic risk weight of the i-th feature; i is the feature variable; Let be the value of the i-th feature in the context vector at time t; To ensure data accuracy; This represents the maximum data accuracy, for example, the highest accuracy among all features. The expected cumulative number of faults; This represents the actual number of faults that have occurred. This is the threshold for the maximum number of allowed faults.

[0051] The model parameters are iteratively optimized using a nonlinear optimization algorithm to output the final risk value for the chemical equipment. For example, the Adam algorithm is used for iterative training to minimize the error between the predicted risk and the historically labeled risk.

[0052] Step 4: Trigger early warning mechanisms and emergency response strategies in stages based on real-time risk values ​​and shape parameters of the Crow-AMSAA model.

[0053] Level 1 Warning: Risk value less than the first warning threshold (low risk, R<0.3): Routine monitoring begins; the robot patrols along a preset route, uploading data every 10 seconds. If the Crow-AMSAA model β<0.8, the existing equipment operation status is maintained. Level 2 Warning: Risk value greater than or equal to the first warning threshold but less than the second warning threshold (medium-low risk, 0.3≤R<0.6): Ventilation equipment continues to operate; risk warning information is pushed to workers; the robot increases the monitoring frequency of key areas to once / second. If 0.8≤β<1.0, preventative equipment maintenance is initiated. Level 3 Warning: Risk value greater than or equal to the second warning threshold but less than the third warning threshold (medium-high risk, 0.6≤R<0.8): Non-critical equipment operation is stopped; evacuation preparation instructions are pushed to workers and managers; the robot adjusts its trajectory to avoid high-risk areas. If β≥1.0, an equipment failure warning is triggered, and a shutdown inspection is arranged. Level 4 warning, risk value greater than or equal to the third warning threshold and less than the fourth warning threshold (high risk, R≥0.8): immediately stop all equipment operation, push emergency evacuation instructions, activate emergency rescue plan, robot lock personnel location and plan evacuation route; if E[N(t)]≥N_max, simultaneously trigger equipment emergency shutdown process.

[0054] Figure 2 This is an exemplary module diagram of the big data-based chemical equipment safety monitoring system provided by the present invention. Figure 2As shown, the big data-based chemical equipment safety monitoring system includes a data acquisition module, a preprocessing module, a risk value determination module, and an early warning module. The data acquisition module is used to construct a multi-source sensor monitoring network for confined spaces and collect multi-source raw data through the multi-source sensor monitoring network. The preprocessing module is used to preprocess the multi-source raw data to obtain a standardized dataset. The risk value determination module is used to input the standardized dataset into a dynamic risk assessment model, perform spatiotemporal coupling processing on the multi-source data, judge the failure trend of the chemical equipment, and output the real-time risk value of the chemical equipment. The dynamic risk assessment model includes the Crow-AMSAA model and an improved LSTM model. The early warning module is used to trigger early warning mechanisms and emergency response strategies based on the real-time risk value.

[0055] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for safety monitoring of chemical equipment based on big data, characterized in that, include: Step 1: Construct a multi-source sensor monitoring network for confined spaces, and collect multi-source raw data through the multi-source sensor monitoring network; Step 2: Preprocess the multi-source raw data to obtain a standardized dataset; Step 3: Input the standardized dataset into the dynamic risk assessment model, perform spatiotemporal coupling processing on the multi-source data, judge the failure trend of the chemical equipment, and output the real-time risk value of the chemical equipment. Dynamic risk assessment models include the Crow-AMSAA model and the improved LSTM model; Step 4: Trigger the early warning mechanism and emergency response strategy based on the real-time risk value.

2. The method for safety monitoring of chemical equipment based on big data according to claim 1, characterized in that, Multi-source sensors include gas sensors, vibration sensors, and condition monitoring nodes; Multi-source raw data includes environmental parameters, equipment operation data, and equipment reliability-related data; Environmental parameters include the concentration of toxic and harmful gases, oxygen content, temperature and humidity, and distance to obstacles; Equipment operation data includes the vibration frequency, pressure value, and spectral density characteristic parameters of chemical equipment; Equipment reliability-related data includes the cumulative number of equipment failures, runtime, and maintenance records.

3. The method for safety monitoring of chemical equipment based on big data according to claim 1, characterized in that, Preprocessing of multi-source raw data yields a standardized dataset, including: Environmental noise in the multi-source raw data is eliminated by Gaussian filtering to obtain Gaussian filtered data; Wavelet transform data is obtained by filtering out equipment vibration interference in Gaussian filtered data using wavelet transform. Outliers were removed from the wavelet transform data using the RANSAC algorithm and spectral density feature analysis to obtain normal data. The missing values ​​in the normal data are filled using the K-nearest neighbor algorithm to obtain a complete dataset; The complete dataset is standardized to obtain a standardized dataset.

4. The method for safety monitoring of chemical equipment based on big data according to claim 3, characterized in that, Outlier removal was performed on wavelet transform data using the RANSAC algorithm and spectral density feature analysis to obtain normal data, including: The RANSAC algorithm is used to select multiple data points from the wavelet transform data for fitting, and a fitting model is obtained. Calculate the distance from the unselected points to the fitted model, and include points whose distance is less than a preset distance threshold into the inlier set; Repeatedly fit the model and generate the interior point set to obtain the optimal fit model, remove outlier data, and obtain the RANSAC-processed fit data. An anomaly detection model was constructed by analyzing spectral density features. The fitted data was then filtered for anomalies to obtain normal data.

5. The method for safety monitoring of chemical equipment based on big data according to claim 1, characterized in that, A standardized dataset is input into a dynamic risk assessment model to perform spatiotemporal coupling processing on multi-source data, determine the failure trend of chemical equipment, and output the real-time risk value of the chemical equipment, including: By using a dynamic ARMA model to enhance the features of a standardized dataset over time, the temporal characteristics of chemical equipment are captured, resulting in a time-enhanced feature vector. Based on the constrained space feature attention scoring function, attention weights are calculated by combining DLT extrinsic optimization logic, and weights are assigned to the temporal enhancement feature vector to obtain the initial weighted feature vector. Based on Crow-AMSAA, the weight coefficients in the initial weighted feature vector are dynamically adjusted to obtain the adjusted weighted feature vector; Predict equipment failure trends to obtain the expected cumulative number of equipment failures; The weighted feature vector and the expected input risk value prediction model of the cumulative number of equipment failures will be adjusted, and the real-time risk value will be calculated by combining the data authenticity and reliability correction coefficients.

6. The method for safety monitoring of chemical equipment based on big data according to claim 5, characterized in that, The dynamic ARMA model is as follows: ； in, For temporal enhancement features; This is the autoregression order index; P is the autoregression order. These are the autoregressive coefficients; Enhance historical features; Index of moving average order; The moving average order; The moving average coefficient; For historical residuals; The spatial coupling weighting coefficient is denoted by n; n represents the sensor variable. This represents the total number of sensor nodes; Spatial weights for sensors; To standardize the features of the nth sensor in the dataset at time t; The formula for dynamically adjusting the weighting coefficients is as follows: ； in, The dynamic risk weight for the i-th feature; t represents the membership degree of the i-th feature; i is the feature variable; t is the time variable; To ensure data accuracy; For the shape parameters of the Crow-AMSAA model; The expected cumulative number of failures is: ； in, Let N(t) be the expected cumulative number of faults at time t; N(t) is the total number of actual faults that occurred in the chemical equipment from the time it was put into operation to time t. For scale parameters; Real-time risk value: ； in, The risk value for time period t; t is a time variable; Let be the dynamic risk weight of the i-th feature; i is the feature variable; Let be the value of the i-th feature in the context vector at time t; To ensure data accuracy; This represents the maximum value of data accuracy. The expected cumulative number of faults; This represents the actual number of faults that have occurred. This is the threshold for the maximum number of allowed faults.

7. The method for safety monitoring of chemical equipment based on big data according to claim 5, characterized in that, Based on the constrained space feature attention scoring function, and combined with DLT extrinsic parameter optimization logic, attention weights are calculated, and weights are assigned to the temporal enhancement feature vector to obtain an initial weighted feature vector, including: Construct a scoring function for the attention given to features in a constrained space to obtain the score value; Calculate attention weights based on the rating values; Context vectors are calculated based on attention weights and the structure of the sensor network.

8. The method for safety monitoring of chemical equipment based on big data according to claim 7, characterized in that, The scoring function for the attention given to features in a confined space is: ； in, This is the score; Let be the query vector at time t; Let be the transpose of the query vector at time t; W is the optimized weight matrix. The key vector is t; time is t. Attention weights are: ； in, is the attention weight; t is the time variable; exp is the exponential function; n is the sensor variable; N is the total number of sensors; The context vector is: ； in, Let n be the context vector at time t; t is the time variable; n is the sensor variable; N is the total number of sensors. Let be the attention weight of the nth sensor at time t; This is the value vector of the nth sensor; Spatial position weight of the nth sensor.

9. The method for safety monitoring of chemical equipment based on big data according to claim 5, characterized in that, The early warning mechanism and emergency response strategy are triggered based on real-time risk values, including: When the risk value is less than the first warning threshold, a level one warning is activated for routine monitoring. When the risk value is greater than the first warning threshold but less than the second warning threshold, the ventilation equipment will be activated and a risk warning message will be sent. When the risk value is greater than or equal to the second warning threshold and less than the third warning threshold, some equipment will be shut down and an evacuation order will be sent. When the risk value is greater than or equal to the third warning threshold but less than the fourth warning threshold, all equipment will be shut down and an evacuation order will be sent out, while the emergency rescue plan will be activated.

10. The big data-based chemical equipment safety monitoring system according to any one of claims 1-9, characterized in that, It includes a data acquisition module, a preprocessing module, a risk value determination module, and an early warning module; The data acquisition module is used to build a multi-source sensor monitoring network for confined spaces and to collect multi-source raw data through the multi-source sensor monitoring network; The preprocessing module is used to preprocess multi-source raw data to obtain a standardized dataset; The risk value determination module is used to input standardized datasets into the dynamic risk assessment model, perform spatiotemporal coupling processing on multi-source data, judge the failure trend of chemical equipment, and output the real-time risk value of chemical equipment. Dynamic risk assessment models include the Crow-AMSAA model and the improved LSTM model. The early warning module is used to trigger early warning mechanisms and emergency response strategies based on real-time risk values.

Images