Plant safety intelligent management and control method based on heterogeneous multi-system cross-service fusion technology

Abstract

The application relates to the technical field of industrial safety, in particular to a plant safety intelligent management and control method based on a heterogeneous multi-system cross-service fusion technology, which comprises the following steps: collecting plant multi-source heterogeneous data and performing space-time alignment and fusion; constructing a risk assessment model based on a deep learning algorithm, performing quantitative risk analysis on personnel behavior, environmental parameters and equipment state; establishing a dynamic hierarchical early warning mechanism to realize alarm and automatic disposal according to the risk level; introducing risk trend prediction to analyze the risk evolution trend; and constructing a safety closed-loop optimization mechanism to realize adaptive adjustment of the management and control strategy. The method constructs a full-process and dynamically optimized intelligent safety management and control scheme, and realizes comprehensive perception, rapid response and continuous optimization of plant safety management.

Description

Technical Field

[0001] This application relates to the field of industrial safety technology, specifically a method for intelligent management and control of factory safety based on heterogeneous multi-system cross-business integration technology. Background Technology

[0002] The factory area deploys various heterogeneous business systems, including industrial IoT systems, video surveillance systems, access control systems, and environmental monitoring systems. Due to the inconsistency in data formats and interface standards among these systems, and the fact that data is stored separately on their respective business platforms, there is a lack of effective data fusion mechanisms. As a result, the value of the data has not been fully explored and utilized, creating information silos for security management.

[0003] Safety risks in industrial parks are complex and diverse, influenced by a complex interplay of factors such as personnel behavior, environmental changes, and equipment malfunctions. Single-dimensional risk assessment methods are insufficient to comprehensively reflect the safety status of the park. Furthermore, most current park safety early warning systems lack quantitative risk assessment tools, making it difficult to guarantee accuracy and real-time performance. This can lead to missed or false alarms, impacting the effectiveness of early warning systems.

[0004] In view of this, this application proposes a factory safety intelligent management and control method based on heterogeneous multi-system cross-business integration technology. Summary of the Invention

[0005] To achieve the above objectives, this invention provides a method for intelligent management and control of factory safety based on heterogeneous multi-system cross-service integration technology. The specific technical solution is as follows:

[0006] A factory safety intelligent management and control method based on heterogeneous multi-system cross-business integration technology includes:

[0007] Collect factory area data, align the collected factory area data from the time and space dimensions, fuse the aligned factory area data as factory area event fusion data and extract data features, and combine the factory area event fusion data with the extracted feature data as factory area event feature fusion data;

[0008] Based on deep learning algorithms, a risk assessment model is constructed to conduct risk assessments on personnel behavior, environmental parameters, and equipment status in the factory area by fusing data based on the characteristics of factory events.

[0009] Establish a dynamic hierarchical early warning mechanism, classify events into risk levels based on risk assessment results, and send alarm information to preset receiving terminals and trigger automatic handling processes based on the dynamic hierarchical early warning mechanism.

[0010] Construct a risk trend prediction mechanism for the plant area, and analyze the evolution trend of plant area risks based on risk assessment, combined with the fusion data of plant area events and dynamic hierarchical early warning results.

[0011] Establish a closed-loop optimization mechanism for factory safety to adaptively adjust and optimize factory safety management.

[0012] Preferably, the factory area data includes: industrial IoT terminal data, surveillance video data, access control data, and environmental data; the industrial IoT terminal data includes: equipment status, process parameters, energy consumption parameters, and fault codes; the surveillance video data is the target detection results in the surveillance video; the access control data includes: the on / off status of access control points, access control card swipe records, and access control abnormal events; the environmental data includes temperature, humidity, gas concentration, and noise.

[0013] Preferably, a unified spatiotemporal reference system is defined, and spatiotemporal alignment is performed on different types of data collected. After the spatiotemporal alignment is completed, the data from different sources are fused according to their spatiotemporal positions to obtain a unified representation of the fused data of the plant area events.

[0014] Feature extraction is performed on the fused data of factory area events, and the fused data of factory area events and the extracted feature data are used as the fused data of factory area event features.

[0015] Preferably, the factory area safety risk assessment problem is divided into a multi-classification task, namely, judging the risk coefficient of the event based on the fusion of data on the characteristics of factory area events; and constructing a risk assessment model to judge the risk of the fusion data on factory area events.

[0016] Preferably, the risk assessment model is constructed based on a multimodal deep learning algorithm, which includes: a data preprocessing layer, a feature learning layer, a modality fusion layer, and a classification output layer.

[0017] The multimodal deep learning algorithm is trained using a dataset, and the trained risk assessment model is deployed to an online environment to perform risk identification on real-time collected data of fused event features in the factory area.

[0018] Preferably, based on a risk assessment model, the collected fusion data of factory area event characteristics are used to determine the risk, and the risk coefficient of each event is obtained;

[0019] Set a risk probability threshold, and classify risk levels by comparing the risk coefficient and the risk probability threshold.

[0020] Preferably, a dynamic threshold optimization mechanism is introduced, defining a sliding time window to record all events and their risk assessment results over a past period;

[0021] For each risk level, count the number of events that actually belong to the assigned level in the assessment results and the number of events with a predicted probability greater than the current threshold; calculate the warning precision and recall rate for each level:

[0022] Adjust precision and recall, and achieve a balance between precision and recall by maximizing the threshold.

[0023] Preferably, the number of events in the factory area and the average probability of each risk level are statistically analyzed, and the ARIMA model in time series analysis is used to predict the risk trend. The future trend of the risk coefficient of the factory area is predicted based on the ARIMA model.

[0024] Preferably, based on historical risk assessment results, early warning information and actual accident occurrences, the risk assessment model is periodically back-analyzed, and the multimodal deep learning model is optimized using incremental learning methods;

[0025] By combining a dynamic threshold optimization mechanism, the changes in the accuracy and recall of early warnings are monitored, and the risk level classification thresholds are dynamically adjusted based on the effectiveness feedback of historical early warnings to optimize the early warning strategy.

[0026] The intelligent management and control system for factory safety based on heterogeneous multi-system cross-business integration technology is used to implement the intelligent management and control method for factory safety based on heterogeneous multi-system cross-business integration technology, including: a data acquisition module, a risk assessment module, a hierarchical early warning module, a risk trend prediction module, and a closed-loop optimization module.

[0027] The data acquisition module is used to collect factory area data, align the collected factory area data from the time and space dimensions, fuse the aligned factory area data as factory area event fusion data and extract data features, and combine the factory area event fusion data and the extracted feature data as factory area event feature fusion data.

[0028] The risk assessment module, based on deep learning algorithms, constructs a risk assessment model and performs risk assessments on personnel behavior, environmental parameters, and equipment status in the factory area by fusing data based on the characteristics of factory events.

[0029] The graded early warning module is used to establish a dynamic graded early warning mechanism, classify the risk level of events according to the risk assessment results, send alarm information to the preset receiving end based on the dynamic graded early warning mechanism, and trigger an automatic handling process.

[0030] The risk trend prediction module is used to construct a risk trend prediction mechanism for the plant area. Based on risk assessment, it combines the fusion data of plant area events with the results of dynamic hierarchical early warning to analyze the evolution trend of plant area risks.

[0031] The closed-loop optimization module is used to construct a closed-loop optimization mechanism for plant safety, and to adaptively adjust and optimize plant safety management and control.

[0032] The beneficial effects of this invention are as follows: This application eliminates data silos and realizes data interconnection and comprehensive utilization by performing spatiotemporal alignment and fusion of heterogeneous data.

[0033] This application constructs a deep learning algorithm to fully explore the complex correlation features contained in the fused data, and realizes the quantitative assessment of multi-dimensional risk factors such as personnel, environment, and equipment.

[0034] This application implements tiered early warning based on risk assessment results, making early warnings more accurate and timely, and reducing missed and false alarms; at the same time, it triggers automatic processing, improving the timeliness of early warning response.

[0035] This application analyzes historical data and assessment results to predict the evolution trend of risks, enabling safety management to shift from passive response to proactive prevention.

[0036] This application improves the accuracy and effectiveness of safety management by continuously monitoring and providing feedback on the control process, dynamically optimizing the risk assessment model and early warning strategy, and achieving adaptive adjustment of the strategy.

[0037] This application presents an intelligent safety management closed loop that integrates data collection, risk assessment, early warning response, trend analysis, and strategy optimization. It achieves deep integration of data, algorithms, and business, significantly improving the intelligence and precision of factory safety management. This is of great significance for ensuring the stable operation of the park and preventing safety accidents. Attached Figure Description

[0038] Figure 1 The flowchart of the intelligent management and control method for factory safety based on heterogeneous multi-system cross-business integration technology provided by the present invention;

[0039] Figure 2 This invention provides a schematic diagram of the data acquisition and fusion process.

[0040] Figure 3 A schematic diagram illustrating the risk assessment model construction process provided by this invention;

[0041] Figure 4 This is a schematic diagram of the dynamic hierarchical early warning process provided by the present invention;

[0042] Figure 5 This is a schematic diagram of the risk trend prediction process provided by the present invention;

[0043] Figure 6 The structural diagram of the intelligent management and control system for factory safety based on heterogeneous multi-system cross-business integration technology provided by the present invention. Detailed Implementation

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention can also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0047] Example 1

[0048] Reference Figures 1 to 5 This is the first embodiment of the present invention, which provides a method for intelligent management and control of factory safety based on heterogeneous multi-system cross-business integration technology.

[0049] Step 1: Collect factory area data, align the collected factory area data from the time and space dimensions, merge the aligned factory area data as factory area event fusion data and extract data features, and combine the factory area event fusion data with the extracted feature data as factory area event feature fusion data.

[0050] Define the data collected by the i-th IoT terminal at time t as an I-dimensional vector. Where I represents the number of data types collected by the terminal.

[0051] The types of data collected by IoT terminals include: device status. i (t)∈{0,1}, process parameters Energy consumption parameters and fault codes Where M p M e M c These are the dimensions for each data type. If the matrix is ​​sparse, then the data of the i-th terminal can be represented as: The data dimension is M = 1 + M p +M e +M c .

[0052] There are K surveillance video cameras. The original video frame captured by the k-th camera at time t is: Where H, W, and G are the image height, width, and number of channels, respectively.

[0053] Perform object detection on the original video frames to extract L targets of interest (e.g., people, vehicles, equipment, etc.). The l-th target is represented by a bounding box vector. It means that (x) l ,y l ) represents the center coordinates of the bounding box, w l and h l This represents the width and height of the frame.

[0054] The target detection result is represented as Where L k Let be the number of targets detected by the k-th camera at time t.

[0055] There are D access control controllers. The state of the d-th controller at time t is defined as follows: Where N d The number of access control points managed by this controller. This indicates the on / off state of the nth access control point.

[0056] Access control card swipe records are defined as a triple r j (t)= j ,d j ,t>, where u j For the card-swiping user ID, d j This is the controller number for the card reader.

[0057] Access control anomalies (such as doors not being closed, unauthorized entry, etc.) are defined as e q (t)= <d q ,c q ,t>, where d q The controller number where the event occurred, c q Encode the event type.

[0058] There are E environmental sensors. The data collected by the e-th sensor at time t is: Data from different types of sensors include: temperature tmp(t), humidity hmd(t), gas concentration gas(t), and noise db(t).

[0059] The data collected by the sensors is uniformly represented as an E×4 dimensional vector:

[0060]

[0061] Since the collected multi-source heterogeneous data may be inconsistent in time and space, spatiotemporal alignment processing is required.

[0062] ​Define a unified spacetime reference frame, with the time axis starting at t0 and ending at t. end The time granularity is Δt; the origin of the spatial coordinate system is O(0,0,0).

[0063] For each data acquisition device, which can be any one of the following: industrial IoT terminal data, surveillance video data, access control data, and environmental data, the spatiotemporal acquisition domain of the data acquisition device is defined as D. s (t,x,y,z) represents the data collected at time t and spatial location (x,y,z).

[0064] Different spatiotemporal alignment methods are used for different data types:

[0065] For IoT terminal data Since the terminal location is fixed, alignment is only required along the time dimension; the collected time series is mapped to a unified time axis to obtain the aligned data. [t0,t end ].

[0066] For surveillance video data For each detection target Transform its bounding box coordinates to a spatial coordinate system, denoted as Then, the target data from different cameras are aligned along the time dimension to obtain...

[0067] For access control data Map each access controller and the location of the event to a spatial coordinate system, denoted as . The alignment method in the time dimension is the same as that for IoT terminals.

[0068] For environmental data Map the position of each sensor to a spatial coordinate system, denoted as . The alignment method in the time dimension is the same as that for IoT terminals.

[0069] After completing the spatiotemporal alignment, data from different sources are fused according to their spatiotemporal location to obtain a unified representation of fused data for factory events.

[0070] Define a factory area event as a quintuple. Where (t,x,y,z) are the spatiotemporal coordinates of the event. It is an event feature vector, containing multi-source heterogeneous data including IoT terminal data, surveillance video data, access control data, and environmental data.

[0071] For each spatiotemporal acquisition domain D s (t,x,y,z) is used to extract heterogeneous data features and construct a local event representation. By fusing local events at the same spatiotemporal location, a complete representation of factory area events can be obtained: Here, ∪ represents the feature concatenation or fusion operation, which can be implemented by vector concatenation, weighted averaging, or max pooling.

[0072] Feature extraction is performed on the fused data of factory events to obtain higher-level feature representations for subsequent risk analysis and early warning; a set of feature extraction functions {φ1,φ2,...,φ} are defined. K The different attributes of events in the factory area are encoded. For example, a time-series feature extraction function extracts the statistical quantities of event duration and frequency; a spatial feature extraction function extracts the geometric quantities of event spatial distribution and cluster density; after applying the feature extraction functions, the event feature vector is obtained. Where N1 is the dimension of the feature vector.

[0073] The spatiotemporally aligned and fused plant area event fusion data and the extracted feature data are represented as plant area event feature fusion data {W}:

[0074]

[0075] Where M is the total number of events. These are the original features and extracted features of the m-th event, respectively.

[0076] See Figure 2 This is a schematic diagram of the data acquisition and fusion process.

[0077] This step eliminates data silos by aligning and fusing heterogeneous data in time and space, enabling data interconnection and comprehensive utilization. This provides comprehensive and accurate data support for subsequent risk assessment, early warning, and optimization, and increases the data value density.

[0078] Step 2: Based on deep learning algorithms, construct a risk assessment model and conduct risk assessments on personnel behavior, environmental parameters, and equipment status in the factory area by fusing data based on the characteristics of factory events.

[0079] Define risk level set Where r C For risk levels, C represents the number of risk levels; W represents the risk level for each event. m Corresponding risk coefficient

[0080] By training the risk assessment model f θ Risk coefficient prediction is performed on the fusion data of new plant area event characteristics; where θ is the model parameter.

[0081] A multimodal deep learning algorithm is used as the risk assessment model to fully utilize the complementary information of heterogeneous data. The multimodal deep learning algorithm includes: a data preprocessing layer, a feature learning layer, a modality fusion layer, and a classification output layer.

[0082] The data preprocessing layer is used to process the raw event features. Perform standardization and normalization operations to map them to a unified feature space; denoted as the processed features.

[0083] For feature extraction Different encoding methods are used depending on the feature type (numerical or categorical), such as embedding encoding or one-hot encoding; the encoded feature is denoted as...

[0084] The feature learning layer includes: using independent sub-networks to learn features from data of different modalities.

[0085] For structured data such as IoT terminal data and environmental data, a multilayer perceptron (MLP) model is used. Perform nonlinear transformation: It is a structured data characteristic of IoT terminal data and environmental data.

[0086] For surveillance video data, a 3D convolutional neural network (3D-CNN) model is used to extract spatiotemporal features: These are characteristics of surveillance video data.

[0087] For discrete data on access control card swipes and access control anomaly events, a recurrent neural network (RNN) is used to capture temporal dependencies: These are discrete data characteristics of access control card swiping and access control abnormal events.

[0088] For feature extraction Automatically learn the importance weights of different features using an attention mechanism: These are features extracted by the attention mechanism; among them For attention parameters, d s ,d v ,d d ,d f Output the dimensions for the subnetwork.

[0089] The modality fusion layer is used to fuse features learned from different modalities to obtain a comprehensive representation of the event. Fu(·) can be any of the functions of concatenation, summation, and attention fusion.

[0090] The classification output layer includes: event-based representation. Predict its risk coefficient:

[0091]

[0092] in For classifier parameters, Δ C-1 Let C be a C-1 dimensional simplex, representing the predicted probability distribution. For the predicted risk coefficient, Indicates event W m The predicted probability of belonging to risk coefficient of type c. Indicates from Select the category with the highest probability.

[0093] Obtain historical datasets of the factory and manually label them. Based on the labeled historical event datasets... Using cross-entropy loss function The risk assessment model was trained using the backpropagation algorithm.

[0094] The Adam optimizer and mini-batch gradient descent method are used to minimize the loss function and update the risk assessment model parameters θ until the model converges or reaches the preset number of iterations.

[0095] The risk coefficient is assessed by using a trained risk assessment model to integrate real-time collected data on the characteristics of factory events.

[0096] See Figure 3 This is a schematic diagram illustrating the risk assessment model construction process in this step.

[0097] This step utilizes deep learning algorithms to fully explore the complex correlation features contained in the fused data, enabling quantitative assessment of multi-dimensional risk factors such as personnel, environment, and equipment. It overcomes the limitations of traditional methods and provides comprehensive and accurate risk identification results.

[0098] Step 3: Establish a dynamic hierarchical early warning mechanism. Based on the risk assessment results, classify the risk level of the event, send alarm information to the preset receiving end based on the dynamic hierarchical early warning mechanism, and trigger the automatic handling process.

[0099] Based on risk assessment model f θ Risk coefficient assessment is performed on the collected fusion data of factory event characteristics {W} to obtain the risk coefficient of each event in the fusion data of factory event characteristics. Set risk probability thresholds τ1, τ2, ..., τ C The condition 0 < τ1 < τ2 < ... < τ C-1 <τ C ≤1; then event W mRisk level L m Determined as:

[0100]

[0101] If event W m risk factor If it falls within a certain risk threshold range, then event W will be... m The risk level is classified into this risk level, and the risk probability threshold {τ} c It can be configured according to your needs.

[0102] To adapt to changing scenarios and improve the sensitivity of early warnings, a dynamic threshold optimization mechanism is introduced; a sliding time window Δht is defined to record all events and their risk assessment results within the past Δht time period.

[0103] For each risk level r c Statistics in The number N of events that actually belong to this level c And the predicted probability is greater than the current threshold τ c Number of events M1; Calculation level r c Early warning accuracy P c and recall rate R c :

[0104]

[0105] in M is an indicator function that takes the value 1 when the condition is met and 0 otherwise. c It is the evaluation result The number of events in the data.

[0106] Through F 1c Values ​​are used to balance precision and recall:

[0107]

[0108] By maximizing F 1c To adjust the threshold τ c This allows for a balance between precision and recall in the early warning results:

[0109]

[0110] The threshold is recalculated and updated at regular intervals to achieve adaptive optimization of the early warning strategy.

[0111] For example, based on the risk level L of the event m This triggers the corresponding early warning and response procedures:

[0112] For low-risk events (L m =r1), generate early warning information and send it to on-site management personnel to remind them to pay attention to monitoring;

[0113] For medium-risk events (L m =r2), in addition to the warning information, it also automatically triggers the video surveillance system to capture and record the relevant area, and notifies the management personnel to come to the scene for investigation;

[0114] For high-risk events (L m ≥r3), based on early warning and image capture, the following measures will be further taken:

[0115] The system automatically triggers nearby access control systems to lock down the area and control personnel access; sends shutdown commands to relevant production equipment to prevent endangerment of personal and property safety; issues evacuation or shelter instructions through the emergency broadcast system to guide personnel to evacuate safely; notifies the emergency rescue team to respond quickly and arrive at the scene for handling; the warning information includes: the time and location of the incident, the personnel / equipment / area involved, the risk level, and the handling requirements, and is sent to relevant personnel and systems in the form of text, voice, and images.

[0116] After issuing an alert, track the subsequent status of the event in real time until the event is properly handled;

[0117] Define event state set Including "Under Warning", "Under Response", and "Resolved"; each event W m In addition to the risk level attribute L m In addition, state attributes are also established. Initialize to S1;

[0118] For example, after an alert is issued, relevant personnel and the system provide feedback on the progress of the handling and update the event status; when management personnel arrive at the scene, they will... i Updated to "In Progress"; once the event is under control and processed, S will be... i Update to "Resolved"; for events that have been in an intermediate state for a long time, urge relevant personnel to handle them as soon as possible, and raise the warning level if necessary.

[0119] See Figure 4 This is a flowchart illustrating the dynamic hierarchical early warning process for this step.

[0120] This step enables tiered early warning based on risk assessment results, making warnings more accurate and timely, reducing missed and false alarms; at the same time, it triggers automatic handling, improving the timeliness of early warning response, reducing the possibility of accidents, and ensuring the safe and stable operation of the park.

[0121] Step 4: Construct a risk trend prediction mechanism for the plant area. Based on the risk assessment, combine the fusion data of plant area events with the results of dynamic hierarchical early warning to analyze the evolution trend of plant area risks.

[0122] Count the number of events in the factory area N t and each risk level r c The percentage of events w tc :

[0123]

[0124] in Indicator function; time series of risk distribution And to calculate the average probability of each risk level within that time period:

[0125]

[0126] Time series of risk probabilities The risk time series dataset is represented as follows: Where T is the total number of time steps; the risk time series dataset contains the dynamic changes in the risk distribution and risk probability of the plant area.

[0127] Using the ARIMA model in time series analysis for risk trend prediction, the ARIMA model is an autoregressive moving average model. The ARIMA model combines the characteristics of autoregressive (AR), difference (I), and moving average (MA) models, and can fit and predict non-stationary time series well.

[0128] set up This is a time series of risk levels for the factory area, where μ mt Indicates event W m The risk coefficient at time t; the ARIMA(A,g,B) model can be expressed as:

[0129]

[0130] Where A is the order of the autoregressive term, g is the order of the difference, B is the order of the moving average term, and Ω is the lag operator, defined as Ω a μ mt =μ mt-a ;δ a σ is the autoregressive coefficient. b ε is the moving average coefficient. t The term is a random disturbance, which is usually assumed to follow a normal distribution with a mean of 0.

[0131] The process of building and applying an ARIMA model is as follows:

[0132] To test the stationarity of the ARIMA model, use the ADF test or KPSS test to determine the original risk sequence {μ}. mt If the sequence is not stationary, perform a difference operation until it becomes stationary; denoted as {μ′}. mt}

[0133] Determine the order of the ARIMA model and observe the sequence {μ′ mt The values ​​of A and B are initially determined based on the tailing and truncation of the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots of the model. The models of different orders are scored, and the combination of (A, g, B) with the best score is selected.

[0134] To estimate the parameters of the ARIMA model, the maximum likelihood estimation method or the conditional least squares method can be used to estimate the model parameters δ. a and σ b .

[0135] ARIMA model diagnostics involves performing residual analysis on the fitted model to check whether the residual sequence is white noise, i.e., whether there is significant autocorrelation. If the model passes the diagnostic, it can be used for prediction; otherwise, the model order needs to be reselected.

[0136] Using the fitted ARIMA(A,g,B) model, predict the risk coefficients for the next h time points:

[0137]

[0138] in and For the estimated model parameters, The prediction error is usually set to 0;

[0139] Future risk levels predicted by the ARIMA model Based on pre-set risk level thresholds, corresponding risk warning information is generated.

[0140] For example, a representation of risk coefficients showing an upward trend can be defined as follows: If This indicates that the risk factor in the factory area will increase within the next h time points, requiring preventative measures and strengthened monitoring and control; if This indicates that the risk factor will increase rapidly within the next h time intervals, requiring immediate activation of the emergency plan and preparation for emergency response.

[0141] Early warning information can be pushed to relevant managers and decision-makers in the form of charts, text, voice, etc., to help them to promptly understand risk trends, optimize resource allocation, and improve risk prevention and emergency response capabilities.

[0142] See Figure 5This is a schematic diagram of the risk trend prediction process for this step.

[0143] This step, based on the ARIMA time series model, constructs a technical solution for predicting and warning of risks in the plant area. It expands the time dimension of risk assessment and early warning, enabling it to make forward-looking judgments and responses to risk evolution over a future period, which has significant practical implications.

[0144] Step 5: Construct a closed-loop optimization mechanism for plant safety to adaptively adjust and optimize plant safety management.

[0145] Based on historical risk assessment results, early warning information, and actual accident occurrences, the risk assessment model is periodically back-analyzed. Incremental learning methods are used to optimize the multimodal deep learning model, enabling it to continuously adapt to the dynamic changes in the plant environment and improve the accuracy of risk identification.

[0146] By combining a dynamic threshold optimization mechanism, the changes in the accuracy and recall of early warnings are monitored. Through reinforcement learning, the risk level classification threshold is dynamically adjusted based on the effectiveness feedback of historical early warnings, thereby optimizing the early warning strategy and ensuring that the early warning results achieve the best balance between accuracy and coverage.

[0147] Based on risk assessment and trend prediction, an intelligent decision-making mechanism is further constructed to realize the intelligent and automated handling of safety incidents in the plant area. According to the risk level judgment results, plant incidents are classified, such as personnel violations, equipment abnormalities, and environmental exceedances. Combined with the plant safety management standards, the best handling plan is automatically matched, such as warning reminders, human-machine collaborative handling, or automatic control strategies.

[0148] For high-risk events, automated emergency response procedures can be triggered, such as automatically cutting off power, closing valves, and activating fire protection systems. Low-risk events can be intervened through intelligent prompts or management personnel review to reduce interference caused by false alarms.

[0149] Through the above optimization mechanisms, a closed-loop security management system is ultimately formed, consisting of "data collection → data fusion → risk assessment → early warning triggering → trend prediction → intelligent decision-making → automatic handling → feedback optimization".

[0150] A closed-loop safety management system can continuously learn and optimize during the factory safety management process, realizing intelligent, automated and continuous improvement of safety management, thereby improving the accuracy and response speed of factory safety management and ensuring the stable operation of the factory.

[0151] Example 2

[0152] Reference Figure 6This is the second embodiment of the present invention, which provides a factory safety intelligent management and control system based on heterogeneous multi-system cross-business integration technology.

[0153] The system includes: a data acquisition module, a risk assessment module, a tiered early warning module, a risk trend prediction module, and a closed-loop optimization module.

[0154] The data acquisition module is used to collect factory area data, align the collected factory area data from the time and space dimensions, fuse the aligned factory area data as factory area event fusion data and extract data features, and combine the factory area event fusion data and the extracted feature data as factory area event feature fusion data.

[0155] The risk assessment module, based on deep learning algorithms, constructs a risk assessment model and performs risk assessments on personnel behavior, environmental parameters, and equipment status in the factory area by fusing data based on the characteristics of events in the factory area.

[0156] The graded early warning module is used to establish a dynamic graded early warning mechanism, classify events into risk levels based on risk assessment results, send alarm information to preset receiving terminals based on the dynamic graded early warning mechanism, and trigger automatic handling procedures.

[0157] The risk trend prediction module is used to construct a risk trend prediction mechanism for the plant area. Based on risk assessment, it combines fused data of plant area events with dynamic hierarchical early warning results to analyze the evolution trend of plant area risks.

[0158] The closed-loop optimization module is used to construct a closed-loop optimization mechanism for plant safety, and to adaptively adjust and optimize plant safety management and control.

[0159] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0160] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0161] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0162] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments without departing from the spirit and scope of protection of the present invention, and all such changes and variations are within the protection scope of the present invention.

Claims

1. A plant safety intelligent management and control method based on heterogeneous multi-system cross-service fusion technology, characterized in that, The method comprises the following steps: Collecting plant area data, aligning the collected plant area data in time and space dimensions, fusing the aligned plant area data as plant area event fusion data and extracting data features, taking the plant area event fusion data and the extracted feature data as plant area event feature fusion data; Based on a deep learning algorithm, a risk assessment model is constructed, and personnel behavior, environmental parameters and equipment state of the plant area are assessed according to the plant area event feature fusion data: A multi-modal deep learning algorithm is used as the risk assessment model, which includes a data preprocessing layer, a feature learning layer, a modal fusion layer and a classification output layer; The data preprocessing layer is configured to perform normalization and standardization operations on original event features to map the original event features to a unified feature space, and the processed features are denoted as . ​ For extracting features , different encoding methods are adopted according to feature types, and the encoded features are denoted as ; The feature learning layer includes using independent sub-networks to learn features of data in different modalities; For structured data of the Internet of Things terminal data and environmental data, a multi-layer perception model is adopted to perform nonlinear transformation: , represents a multi-layer perception feature extractor, is a structured data feature of the Internet of Things terminal data and environmental data; is an encoded original event feature of the first event; represents a feature space, represents a feature dimension of the structured data of the terminal data and environmental data;​ For monitoring video data, a 3D convolutional neural network model is used to extract spatio-temporal features: , is a feature of the monitoring video data; represents a three-dimensional convolutional network for extracting spatio-temporal features; represents a video feature space; represents the dimension of the video feature; For the discrete data of access control card swiping and access control abnormal events, a recurrent neural network is used to capture the time sequence dependence: , is the discrete data feature of access control card swiping and access control abnormal events; represents the time sequence feature extracted by the recurrent neural network; is the feature space of the event, is the feature dimension of the event; For extracting features , an attention mechanism is adopted to automatically learn the importance weights of different features: , is the feature extracted by the attention mechanism; wherein and are attention parameters, is the output dimension of the subnetwork; The modal fusion layer is configured to fuse features learned by different modalities to obtain a comprehensive representation of the event. ; wherein may be any one of splicing, summation, and attention fusion function; represents a space of the comprehensive representation of the event, is a dimension of the space of the comprehensive representation. The classification output layer comprises: based on the event representation , predicting its risk coefficient: ； ； wherein and are classifier parameters, are simplices in the dimensionality of the feature space, representing the predicted probability distribution, is the predicted risk score, represents the event belongs to the class risk score, represents the class with the highest probability is selected from the set of classes. A factory historical data set is acquired, and the historical data set is labeled by a human, and a risk assessment model is trained based on the labeled historical event data set , The event quantity is the cross-entropy loss function and a back propagation algorithm A dynamic hierarchical early warning mechanism is established, the risk level of the event is divided according to the risk assessment result, and alarm information is sent to the preset receiving end based on the dynamic hierarchical early warning mechanism and the automatic disposal process is triggered; A plant area risk trend prediction mechanism is constructed, which analyzes the evolution trend of the plant area risk based on the risk assessment, combined with the plant area event fusion data and the dynamic hierarchical early warning result; A plant area safety closed-loop optimization mechanism is constructed to adaptively adjust and optimize the plant area safety control.

2. The plant safety intelligent management and control method based on the heterogeneous multi-system cross-service fusion technology according to claim 1, characterized in that, The plant area data includes industrial Internet of Things terminal data, monitoring video data, access control data and environmental data; the industrial Internet of Things terminal data includes equipment state, process parameters, energy consumption parameters and fault codes; the monitoring video data is the target detection result in the monitoring video; the access control data includes the opening and closing state of the access control point, access control card record and access control abnormal event; the environmental data includes temperature, humidity, gas concentration and noise.

3. The plant safety intelligent management and control method based on the heterogeneous multi-system cross-service fusion technology according to claim 2, characterized in that, A unified space-time reference system is defined, and different types of collected data are respectively aligned in time and space; After completing the space-time alignment, the data from different sources are fused according to the space-time position to obtain unified plant area event fusion data representation; Feature extraction is performed on the plant area event fusion data, and the plant area event fusion data and the extracted feature data are taken as plant area event feature fusion data.

4. The plant safety intelligent management and control method based on the heterogeneous multi-system cross-service fusion technology according to claim 3, characterized in that, The plant safety risk assessment problem is divided into a multi-classification task, that is, the risk coefficient of the event is judged according to the plant area event feature fusion data; a risk assessment model is constructed to judge the risk of the plant area event feature fusion data.

5. The plant safety intelligent management and control method based on the heterogeneous multi-system cross-service fusion technology according to claim 4, characterized in that, Based on the risk assessment model, the risk of the collected plant area event feature fusion data is judged to obtain the risk coefficient of each event; A risk probability threshold is set, and the risk level is divided by comparing the risk coefficient and the risk probability threshold.

6. The plant safety intelligent management and control method based on the heterogeneous multi-system cross-service fusion technology according to claim 5, characterized in that, A dynamic threshold optimization mechanism is introduced, a sliding time window is defined, and all events and their risk assessment results in the past time are recorded; For each risk level, the number of events that actually belong to the divided level in the evaluation result and the number of events whose prediction probability is greater than the current threshold are counted; The warning accuracy and recall rate of the level are calculated: The accuracy and recall rate are adjusted, and the threshold is maximized to balance the warning result between the accuracy and recall rate.

7. The plant safety intelligent management and control method based on the heterogeneous multi-system cross-service fusion technology according to claim 6, characterized in that, The number of plant events and the average probability of each risk level are counted, an ARIMA model in time series analysis is used to predict the risk trend, and the trend of the future risk coefficient of the plant is predicted according to the ARIMA model.

8. The plant safety intelligent management and control method based on the heterogeneous multi-system cross-service fusion technology according to claim 7, characterized in that, Based on historical risk assessment results, early warning information and actual accident occurrence, the risk assessment model is periodically backtracked and analyzed, and the incremental learning method is used to optimize the multi-modal deep learning model. Combined with the dynamic threshold optimization mechanism, the change of the monitoring and early warning accuracy and recall rate is monitored, the risk level division threshold is dynamically adjusted according to the effectiveness feedback of historical early warning, and the early warning strategy is optimized.

9. A plant safety intelligent management and control system based on heterogeneous multi-system cross-service fusion technology, which is used to implement the plant safety intelligent management and control method based on heterogeneous multi-system cross-service fusion technology according to any one of claims 1 to 8, characterized in that, It includes: A data acquisition module, a risk assessment module, a hierarchical early warning module, a risk trend prediction module and a closed-loop optimization module. The data acquisition module is used for collecting plant data, aligning the collected plant data from time and space dimensions, fusing the aligned plant data as plant event fusion data and extracting data features, and taking the plant event fusion data and the extracted feature data as plant event feature fusion data. The risk assessment module is based on a deep learning algorithm to build a risk assessment model, and conducts risk assessment on personnel behavior, environmental parameters and equipment status of the plant according to the plant event feature fusion data. The hierarchical early warning module is used to establish a dynamic hierarchical early warning mechanism, divide the event into a risk level according to the risk assessment result, send an alarm information to a preset receiving end based on the dynamic hierarchical early warning mechanism and trigger an automatic disposal process. The risk trend prediction module is used to build a plant risk trend prediction mechanism, analyze the evolution trend of the plant risk based on the risk assessment, the plant event fusion data and the dynamic hierarchical early warning result. The closed-loop optimization module is used to build a plant safety closed-loop optimization mechanism to adaptively adjust and optimize the plant safety control.

Images