Intelligent robot system and method for safety risk management and control of special operation in food and medicine industry

Abstract

The invention relates to the technical field of food and medicine product production safety risk management and control, in particular to a food and medicine industry special operation safety risk management and control intelligent robot system and method. The safety production knowledge graph comprises corresponding relations among personnel, equipment, materials, laws and regulations, environments and management systems; analyzing and judging the multi-dimensional data, and calculating a risk value of the multi-dimensional data according to the safety production knowledge graph; judging the risk level of the special operation based on the risk value; and different approval and control programs are configured according to the risk levels. According to the technical thought of the method, traditional manual and subjective approval is changed into intelligent, objective and quantitative dynamic risk assessment, the safety and approval efficiency of special operation are remarkably improved through full-process automatic closed loop from pre-event prevention, in-event monitoring to post-event tracing, and active and predictive safety risk management and control are achieved.

Description

Technical Field

[0001] This invention relates to the field of food and drug product production safety risk management technology, specifically to an intelligent robot system and method for special operation safety risk management in the food and drug industry. Background Technology

[0002] In high-risk industries such as pharmaceuticals, chemicals, and energy, safe production is the lifeline of enterprise operations. Special operations, such as hot work, confined space work, work at heights, temporary electrical work, and hoisting operations, are high-risk areas for production safety accidents due to their high potential risks, complex working environments, and strict operational requirements. To effectively control these risks and prevent accidents, enterprises generally implement a special operations permit system. This means that any special operation must undergo a rigorous application, review, and approval process before it can be carried out, and appropriate safety measures must be in place.

[0003] In existing technologies, the approval and management model for special operations has evolved from purely manual to a preliminary digital approach. The initial model relied entirely on paper work permits, which were filled out by the work application and then manually reviewed and signed by relevant personnel (such as team leaders, workshop directors, and safety officers) at each level. This method has many drawbacks:

[0004] First, the approval process is inefficient and lengthy, and the circulation, archiving, and retrieval of paper documents are extremely inconvenient.

[0005] Secondly, the approval process is highly subjective, and the approval conclusions depend heavily on the approver's personal experience and sense of responsibility. There is a lack of unified and objective quantitative standards, and key risk points are easily overlooked due to negligence or lack of experience.

[0006] Furthermore, the authenticity of the information is difficult to verify, and the approver cannot understand the actual situation at the work site in a real time and intuitively, and can only make judgments based on the text information submitted by the applicant.

[0007] With the development of information technology, many enterprises have introduced electronic safety management systems, such as EHS (Environmental, Health, and Safety) management software. These systems digitize paper work orders and facilitate online transfer between different approvers through pre-defined workflows. This improves approval efficiency and record traceability to some extent.

[0008] However, the closest existing technology to this invention, namely the EHS management system integrating electronic approval workflows, still has the following significant technical problems and drawbacks:

[0009] 1. Rigid Risk Assessment Logic and Lack of Intelligence: The core logic of such systems remains an "electronic checklist," capable of verifying whether operators have uploaded training certificates, but unable to determine the value of the training or its relevance to specific job content; it can record equipment serial numbers, but cannot automatically link them to the equipment's real-time operating status, maintenance records, and historical faults. Its risk assessment is based on a series of isolated, static Boolean logic judgments (yes / no), unable to conduct dynamic, correlated, and predictive comprehensive analysis of multiple potential risk factors such as "poor personnel condition," "equipment nearing its maintenance period," and "excessively high ambient temperature," and therefore cannot quantify the overall risk level.

[0010] 2. The problem of information silos is prominent, and the level of automation is low: Although the approval process is electronic, the data verification process is often still semi-automatic or even manual. Approver usually needs to log in to and query multiple independent systems (such as the human resources system for training records, the equipment management system for maintenance records, and the materials management system for MSDS) to cross-compare information. This is not only inefficient but also prone to errors. Although some systems have achieved certain single-point data integration, they are far from achieving cross-system, comprehensive automatic data collection and fusion.

[0011] 3. The approval process relies heavily on a single basis and lacks multimodal verification: Approval decisions are almost entirely based on structured or semi-structured data, such as text submitted by the applicant and uploaded attachments. There is a lack of objective and real-time verification methods for crucial on-site information, such as whether safety isolation has been completed as required, whether supervisors are in place, and whether protective equipment is up to standard. A significant information gap exists between the approval decision and the actual physical world of the work site.

[0012] 4. Fragmented management processes, lacking in-process monitoring and closed-loop: Traditional electronic approval systems typically end their management responsibilities the moment the work permit is approved. The system cannot monitor the work process in real time and effectively. If personnel violate regulations (such as removing safety helmets) or there are sudden environmental changes at the work site, the system cannot detect, issue warnings, or intervene, resulting in a significant "blind spot" in safety management. The entire management process is linear, lacking a feedback learning and self-optimization mechanism from "post-event results" to "pre-event prevention."

[0013] Therefore, existing technologies are significantly insufficient in achieving intelligent, automated, visualized, and closed-loop management of special operation approvals. There is an urgent need for a new method and system that can deeply integrate artificial intelligence and automation technologies to achieve proactive, predictive, and penetrating risk control. Summary of the Invention

[0014] To address the aforementioned technical problems in the existing technology, this invention provides an intelligent robot system and method for safety risk management in special operations within the food and drug industry. Specifically, this invention aims to achieve the following objectives in response to the problems identified in the background art:

[0015] 1. To address the issues of rigid risk assessment logic and reliance on human experience in existing technologies, this paper aims to provide an intelligent decision-making mechanism that can dynamically, quantitatively, and predictively assess operational risks.

[0016] 2. To address the problems of severe information silos and low automation of data verification in existing technologies, the aim is to achieve automatic collection, fusion, and cross-verification of cross-system, multi-dimensional safety production information.

[0017] 3. To address the problem of existing technologies having a single approval basis that is detached from the actual situation at the work site, the aim is to introduce objective verification methods such as computer vision to achieve multimodal and authenticity verification of key safety elements.

[0018] 4. To address the issues of fragmented management processes, lack of in-process monitoring, and closed-loop learning mechanisms in existing technologies, the aim is to build a closed-loop, proactive intelligent security management system covering the entire process of "pre-approval, in-process monitoring, and post-event traceability and learning."

[0019] To achieve the above objectives, the technical solution of the present invention is as follows:

[0020] In a first aspect, the present invention provides a method for safety risk management and control in special operations in the food and drug industry, comprising:

[0021] S1. Collect and maintain multi-dimensional data, including: environmental parameters, environmental compliance data, personnel compliance data, tool compliance data, work order parsing data, personnel basic information, work equipment basic information, and material management information;

[0022] S2. Construct a safety production knowledge graph, which includes the correspondence between personnel, equipment, materials, regulations, environment, and management systems; the construction of the safety production knowledge graph specifically includes:

[0023] Ontology modeling: Define security entity nodes, which include personnel, equipment, materials, regulations, environment, and management systems, and associate the attributes of each security entity node with them;

[0024] Relationship extraction: Define the correspondence between secure entity nodes;

[0025] A safety production knowledge graph is constructed using safety entity nodes as nodes and corresponding relationships as edges.

[0026] S3. Analyze and judge multi-dimensional data, and use the safety production knowledge graph to calculate the risk value of multi-dimensional data;

[0027] S4. Based on the risk value, determine the risk level of this special operation;

[0028] S5. Configure different approval procedures according to the risk level.

[0029] Furthermore, the risk value of multi-dimensional data is calculated, specifically including:

[0030] S3.1 Extracting the work safety sub-diagram:

[0031] Based on the multi-dimensional data of this special operation collected by S1, it is mapped to the safety production knowledge graph, and the safety entity nodes and their relationships related to this special operation are extracted to form an operation safety subgraph.

[0032] S3.2, Feature vectorization:

[0033] Transform each node and its attribute information in the job safety subgraph into an initial feature vector;

[0034] S3.3 Risk Propagation and Aggregation:

[0035] By using graph neural network models to propagate information on knowledge graphs, graph-level embedding vectors containing information about neighboring nodes are formed.

[0036] S3.4 Risk Prediction and Quantification:

[0037] The graph-level embedding vectors aggregated from the graph neural network model are input into the machine learning model to calculate the risk value of multi-dimensional data.

[0038] Furthermore, calculating the risk value of multi-dimensional data also includes:

[0039] S3.5 Interpretability Analysis:

[0040] Show the risk factors in the multi-dimensional data to the approver.

[0041] Furthermore, the initial feature vectors need to be normalized before they can be used to construct the job safety subgraph.

[0042] Furthermore, step S3.3, risk propagation and aggregation, specifically includes:

[0043] Based on each initial feature vector, the neighbor node information of the corresponding secure entity node is added to form an embedding vector h. v The embedding vector h of all matched secure entity nodes v Forming graph-level embedding vectors

[0044] Furthermore, the GraphSAGE model is used to form graph-level embedding vectors layer by layer. A total of K layers of processing are performed;

[0045] At layer k, for any safe entity node v in the job safety subgraph, its updated embedding vector The calculation process includes:

[0046] Aggregation: Aggregate information from the set N(v) of neighboring nodes of node v, expressed as:

[0047]

[0048] in, It is the embedding vector of the neighbor node u in the previous layer. For the 0th layer, Let be the initial feature vector of node u;

[0049] Update: Embed the parent node of node v itself. Its aggregated neighbor information The data is then spliced ​​together, followed by a nonlinear transformation.

[0050]

[0051] Among them, W (k) It is the trainable weight matrix of the k-th layer;

[0052] b (k) It is the trainable bias vector of the k-th layer;

[0053] σ is a non-linear activation function;

[0054] CONCAT indicates a splicing operation;

[0055] After K layers of propagation, each node obtains an embedding vector that incorporates information from its K-hop neighbors.

[0056] Embedded vector h v Forming graph-level embedding vectors

[0057]

[0058] Furthermore, step S3.4, risk prediction and quantification, specifically includes:

[0059] S3.4.1 Embedding Graph-Level Vectors Inputting the data into the classifier yields the risk level probability distribution:

[0060]

[0061] Wherein, P(RiskLevels|G op ) represents the risk probability distribution; W out b represents the weights of the classifier; out This represents the classifier's bias; the Softmax function transforms the output into a probability distribution.

[0062] S3.4.2. Based on the number of risk levels, multiple risk level distributions are obtained. The probability distributions of all risk levels are then weighted and summed to obtain a continuous risk score.

[0063]

[0064] Where C represents the number of risk levels, w i It is the preset weight for each level; P(risk_level) i ) represents the risk probability distribution corresponding to the i-th risk level; R represents the risk score.

[0065] Furthermore, different approval procedures are configured according to the risk level, specifically including:

[0066] Low-risk level, configure automatic approval process: automatically provide approval results;

[0067] For medium- or high-risk levels, a manual review and approval process is configured: the special operation application form, along with the risk analysis report, is sent to the next level approver for manual review.

[0068] Furthermore, multi-dimensional data is collected, specifically including:

[0069] Environmental parameters are collected through sensors.

[0070] By connecting the software robot cluster with the human resources system, equipment asset management system, and material management system, basic personnel information, basic equipment information, and material management information are collected.

[0071] Image recognition technology is used to collect environmental compliance data, personnel compliance data, and tool compliance data.

[0072] The natural language processing module is used to collect and parse work order data.

[0073] Furthermore, the methods for managing safety risks in special operations within the food and drug industry also include:

[0074] S6. During special operations, multi-dimensional data are continuously collected, and risk values ​​are calculated in real time. If the real-time risk value is higher than the threshold, an alarm is triggered, and the alarm information is sent to the approver and equipment manager.

[0075] Secondly, the present invention also provides an intelligent robot system for safety risk management and control in special operations in the food and drug industry, comprising:

[0076] The data acquisition and RPA execution module is used for collecting multi-dimensional data and interacting with external information.

[0077] The multimodal information perception module is used to transform the collected unstructured multidimensional data into structured data;

[0078] The AI ​​decision-making and analysis module is used to analyze and judge multi-dimensional data, calculate the risk value of multi-dimensional data using a safety production knowledge graph, determine the risk level of this special operation based on the risk value, and configure different approval procedures according to the risk level.

[0079] Furthermore, the multimodal information perception module includes:

[0080] The computer vision module is used to receive image information and video streams from the work site, and to make a preliminary judgment on whether there are any violations at the work site based on the image information and video streams.

[0081] The natural language processing module is used to generate parsed data for the work order from the input text information.

[0082] The IoT integration module is used to receive environmental parameters collected by sensors, preprocess the environmental parameters, and add timestamps.

[0083] Compared with the prior art, the present invention has the following beneficial effects:

[0084] 1. It achieves the "predictive" and "proactive" nature of risk management:

[0085] This invention introduces a dynamic risk quantification algorithm based on a safety production knowledge graph and GNN (Graph Neural Network), transforming safety approval from a "post-event" compliance check to a "pre-event" risk prediction. It can uncover complex, interconnected risks hidden behind multiple factors and present them intuitively with quantified scores, enabling safety management to shift from passive response to proactive prevention, fundamentally improving the foresight and scientific rigor of risk identification.

[0086] 2. It achieves "objectivity" and "standardization" in approval and decision-making:

[0087] This invention achieves automatic and accurate cross-system data collection through RPA (Robotic Process Automation), and obtains objective information from the field and text through CV (Computer Vision) and NLP (Natural Language Processing) technologies, completely replacing the error-prone data verification process that relies on manual labor. AI (Artificial Intelligence) evaluation and decision-making are based on a unified model and algorithm, eliminating interference from human factors, ensuring the uniformity and objectivity of approval standards, and significantly improving decision-making quality.

[0088] 3. Achieved end-to-end automation and efficiency in the entire work process: Through seamless collaboration between AI and RPA, this invention automates end-to-end processes that previously required significant manpower and time, such as information collection, risk assessment, workflow management, real-time monitoring, alarm intervention, and archiving and learning. This not only significantly shortens the work permit approval cycle and improves operational efficiency, but also frees safety management personnel from tedious administrative tasks, allowing them to focus on higher-value risk analysis and management optimization.

[0089] 4. It has achieved "all-time and all-space" and "closed-loop" security management:

[0090] This invention extends the scope of management from the traditional "approval" process to on-site verification "before operation," real-time monitoring and intervention "during operation," and data archiving and feedback learning "after operation." It breaks down the barriers between approval and execution, and between online and offline processes, forming a complete PDCA (Plan-Do-Check-Act) closed loop from prevention and control to continuous improvement. This ensures the continuity and integrity of safety management in both time and space, effectively filling the blind spots of "in-process monitoring."

[0091] 5. Enhanced compliance and traceability:

[0092] The system automatically generates digital operation files that include multimodal evidence (such as on-site photos, video snapshots, and sensor data) and AI analysis reports, providing a complete, objective, and tamper-proof record of every special operation. This not only greatly facilitates daily safety audits and compliance checks but also provides strong, traceable technical evidence in accident investigations. Attached Figure Description

[0093] Figure 1 The flowchart of the safety risk control method for special operations in the food and drug industry provided by the present invention. Detailed Implementation

[0094] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0095] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0096] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0097] Example 1

[0098] This embodiment provides an intelligent robot system for safety risk management in special operations in the food and drug industry, including:

[0099] A1. Data Acquisition and RPA Execution Module, used for multi-dimensional data acquisition and interaction with external information;

[0100] The data acquisition and RPA execution module serves as a bridge between the intelligent robot system for safety risk management in special operations within the food and drug industry and the enterprise's existing IT ecosystem. The RPA software robot cluster is the core execution unit of this module, with each RPA robot configured with a series of "skills"—scripts for accessing specific systems. For example, an "HR information acquisition robot" is granted a low-privilege, read-only account, enabling it to simulate a human logging into the HR system, navigating to the employee's profile page by entering their employee ID, and retrieving information such as training records, certificate validity periods, and years of service. Similarly, other robots are responsible for querying equipment maintenance plans from the EAM system and obtaining production area status from the MES system.

[0101] The key to this layer lies in its "non-intrusiveness," enabling data integration without modifying existing systems. Simultaneously, this layer also acts as the "arm" for command execution. After the AI ​​decision-making and analysis module makes a decision, the RPA robot is responsible for executing a series of automated processes, such as sending emails, pushing messages via WeChat / DingTalk, and creating / updating work ticket statuses within the system.

[0102] A2. Multimodal information perception module, used to transform collected unstructured multidimensional data into structured data; it mainly includes:

[0103] A201, Computer vision module, used to receive image information and video streams from the work site, and to make a preliminary judgment on whether there are any violations at the work site based on the image information and video streams;

[0104] This module receives on-site photos uploaded from the user's mobile device or real-time video streams from fixed cameras. Internally, it can employ one or more deep learning models. For example, it might use a target detection model based on YOLOv5 or Faster R-CNN architecture, which undergoes transfer learning and fine-tuning on a self-built dataset containing hundreds of thousands of industrial safety scene images. This self-built dataset is labeled with various PPEs (such as safety helmets, safety belts, and face shields of different colors and styles), tools (such as welding torches and wrenches), and environmental conditions (such as oil stains on the ground, debris accumulation, and blocked fire exits).

[0105] After receiving image or video frames, the model outputs the bounding boxes and confidence scores for all identified objects in the image. The system then makes judgments based on preset business rules. For example, if a bounding box for flammable materials such as a "paint bucket" or "cardboard box" is detected in a hot work area, it is determined as "environmental non-compliance." For real-time monitoring, this module can continuously process video streams (e.g., 5-10 frames per second) and smooth the detection results of multiple consecutive frames to identify persistent violations, such as "personnel in a hazardous area not wearing a safety helmet for more than 5 seconds."

[0106] A202, Natural Language Processing Module, is used to generate work ticket parsing data from the input text information;

[0107] The core of this module is processing text information. When a user inputs "Welding repair is planned for tomorrow on the top platform of reactor R-101 in workshop A-2," a Named Entity Recognition (NER) model based on the BERT architecture is activated, extracting key entities such as {Time: Tomorrow, Location: Top platform of reactor R-101 in workshop A-2, Work content: Welding repair}. Subsequently, an SOP matcher based on a sentence similarity model (such as Sentence-BERT) calculates the semantic similarity between "welding repair" and all standard operating procedures in the SOP knowledge base, and returns the most matching SOP number (such as SOP-HW-003: Reactor Hot Work Procedure).

[0108] A203, IoT integration module, is used to receive environmental parameters collected by sensors, preprocess the environmental parameters, and add timestamps.

[0109] This module subscribes to all field sensor topics via an MQTT (Message Queuing Telemetry Transport) protocol broker. Sensors publish data in JSON format to the designated topics at a preset frequency (e.g., once every 10 seconds for a gas sensor). The IoT module receives and parses the data, adds a timestamp, stores it in a time-series database, and simultaneously pushes it to the AI ​​decision-making and analysis module in real time.

[0110] A3, the AI ​​decision-making and analysis module, is used to analyze and judge multi-dimensional data, use a safety production knowledge graph to calculate the risk value of multi-dimensional data; based on the risk value, it judges the risk level of this special operation; and configures different approval procedures according to the risk level.

[0111] This module serves as the "brain" of the intelligent robot system for safety risk management in special operations within the food and drug industry, responsible for performing the highest level of intelligent analysis and decision-making. Its core algorithm is a dynamic risk quantification assessment algorithm based on Generative Neural Networks (GNNs), designed to address a key challenge: how to comprehensively utilize multi-source, heterogeneous, and correlated safety information to conduct accurate, dynamic, and quantifiable risk assessments for a specific special operation.

[0112] Example 2

[0113] This invention provides a method for safety risk management in special operations in the food and drug industry, such as... Figure 1 Shown, including:

[0114] S1. Collect and maintain multi-dimensional data, including: environmental parameters, environmental compliance data, personnel compliance data, tool compliance data, work order parsing data, personnel basic information, work equipment basic information, and material management information;

[0115] When the intelligent robot system for safety risk management of special operations in the food and drug industry receives an operation application instruction, the RPA robot can automatically and concurrently log into the human resources system, equipment asset management system, material management system, environmental monitoring system, etc., and capture structured data related to the operation, including but not limited to: the identity information, qualification certificates, training records, health status, and historical violations of the operators; the serial number, maintenance plan, operating parameters, and fault history of the operating equipment; and the MSDS (Material Safety Data Sheet) information of the relevant materials.

[0116] In addition, the intelligent robot system for safety risk management in special operations in the food and drug industry also utilizes deep learning image recognition technology to analyze photos / videos of the work site uploaded by users via mobile devices, or to perform real-time analysis by connecting to fixed cameras on site, including:

[0117] Environmental compliance checks: Identify whether there are flammable materials in the hot work area, whether the scaffolding for high-altitude operations is erected in accordance with regulations, and whether warning signs are posted at the entrances to confined spaces, etc., to generate environmental compliance data.

[0118] Personnel and Clothing Identification: The system uses facial recognition to confirm the identity of workers and to identify whether they are wearing personal protective equipment (PPE) such as safety helmets, safety belts, and safety glasses as required, thus generating personnel compliance data.

[0119] Tool and equipment verification: Identify whether the tools used in the operation are compliant, and whether key equipment (such as gas detectors) is in the calibrated state, to generate tool compliance data.

[0120] It also utilizes various sensors connected via the Internet of Things integration module, such as combustible gas detectors, toxic gas sensors, temperature and humidity sensors, and anemometers, to acquire high-frequency, real-time environmental parameters.

[0121] S2. Construct a safety production knowledge graph, which includes the correspondence between personnel, equipment, materials, regulations, environment, and management systems.

[0122] A safety production knowledge graph can depict the intricate internal relationships between the six key elements of "people, machine, material, method, environment, and management," elevating risk assessment from isolated, single-point inspections to a holistic, interconnected analysis. Through the safety production knowledge graph, deep associative reasoning is possible. For example, when it is discovered that applicant A is about to operate equipment B, the system can instantly traverse the graph and find that A's training records do not include any operation items for equipment B, or that equipment B is associated with a high-risk fault record "pending processing," thereby uncovering hidden risks.

[0123] Constructing a knowledge graph for safe production, specifically including:

[0124] Ontology modeling: Define security entity nodes, which include personnel, equipment, materials, regulations, environment, and management systems, and associate the attributes of each security entity node with them;

[0125] Relationship extraction: Define the correspondence between safety entity nodes, such as whether employees have received training, equipment needs maintenance, or operations involve chemicals;

[0126] A safety production knowledge graph is constructed using safety entity nodes as nodes and corresponding relationships as edges. The safety production knowledge graph contains entity nodes and basic information for each link, such as equipment models. However, if it is a variable, such as personnel information, the personnel node will be empty. The personnel information will be filled in from the subsequent multi-dimensional data collected as data for subsequent risk assessment. If there are mutual relationships between nodes, the relationships between nodes are established using edges.

[0127] S3. Analyze and assess multi-dimensional data, and calculate the risk value of the multi-dimensional data using a safety production knowledge graph; specifically including:

[0128] S3.1. Based on the multi-dimensional data of this special operation collected in S1, key safety entities (such as operators, equipment, and locations) are identified. Subsequently, in the global safety production knowledge graph, all neighboring nodes and their relationships directly related to these key safety entities and within a preset N-hop (e.g., 2-hop) range are retrieved, collectively forming a temporary, highly relevant operation safety subgraph Gop = (Vop, Eop) to this special operation, where Vop is the set of nodes and Eop is the set of edges.

[0129] S3.2, Feature vectorization:

[0130] For each node v within the job safety subgraph Gop constructed in step S3.1, its corresponding attribute information is collected from multi-dimensional data. Based on this attribute information, each node v is transformed into an initial feature vector x. v The initial feature vector contains attribute information of the secure entity nodes; the initial feature vector Where d v It is the feature dimension of this type of node.

[0131] Multidimensional data includes two types:

[0132] Standard / Static Data: Information that changes infrequently and is relatively fixed. Examples include: personnel qualification certificate numbers, equipment manufacturing dates, material MSDS information, and regulatory provisions.

[0133] Real-time / dynamic data: Information that changes frequently and reflects the current "situation". For example: real-time fatigue level of personnel (from recent working hours), real-time operating temperature of equipment, real-time gas concentration in the environment, and violations identified by on-site cameras.

[0134] Standard / static data can be included in the safety production knowledge graph and does not need to be collected repeatedly, while real-time / dynamic data needs to be collected in real time.

[0135] Personnel node v p Its eigenvector x p It may include:

[0136] The standard attribute vector includes relatively static data such as years of service, qualification level (quantified value), and total number of historical violations.

[0137] The real-time attribute vector includes dynamic data such as the average working hours over the past 7 days (fatigue index), whether sufficient rest was taken before the current task (yes = 1, no = 0), and real-time heart rate (from the smart bracelet).

[0138] Device node v e Its eigenvector x e It may include:

[0139] Equipment age, number of days since last maintenance, historical failure rate, operating temperature (real-time), operating pressure (real-time), maintenance status (0 = normal, 1 = pending maintenance, 2 = overdue), etc.

[0140] Environment node v env Its eigenvector x env It may include:

[0141] The standard attribute vector includes information such as workspace size, ventilation level, and area hazard level definition.

[0142] The real-time attribute vector includes the concentration of combustible gas (LEL%), concentration of toxic gas (PPM) collected by the sensor, real-time temperature, and the number of real-time risk items identified by the computer vision module.

[0143] All initial feature vectors are normalized, for example, by min-max normalization:

[0144]

[0145] S3.3 Risk Propagation and Aggregation:

[0146] Utilizing graph neural network models to propagate information across knowledge graphs, forming graph-level embedding vectors that contain information about neighboring nodes; specifically including:

[0147] Using the GraphSAGE model, graph-level embedding vectors are formed layer by layer. A total of K layers of processing are performed;

[0148] At layer k, for any safe entity node v in the job safety subgraph, its updated embedding vector The calculation process includes:

[0149] Aggregation: Aggregate information from the set N(v) of neighboring nodes of node v, expressed as:

[0150]

[0151] in, It is the embedding vector of the neighbor node u in the previous layer. For the 0th layer, Let be the initial feature vector of node u;

[0152] Update: Embed the parent node of node v itself. Its aggregated neighbor information The data is then spliced ​​together, followed by a nonlinear transformation.

[0153]

[0154] Among them, W (k) It is the trainable weight matrix of the k-th layer;

[0155] b (k) It is the trainable bias vector of the k-th layer;

[0156] σ is a non-linear activation function;

[0157] CONCAT indicates a splicing operation;

[0158] After K layers of propagation, each node obtains an embedding vector that incorporates information from its K-hop neighbors.

[0159] Embedded vector h v The pooling operation forms a graph-level embedding vector.

[0160]

[0161] S3.4 Risk Prediction and Quantification:

[0162] The graph-level embedding vectors aggregated from the graph neural network model are input into a machine learning model to calculate the risk value of multi-dimensional data. Specifically, this includes:

[0163] S3.4.1 Embedding Graph-Level Vectors Inputting the data into a multilayer perceptron (MLP) classifier yields the risk level probability distribution:

[0164]

[0165] Wherein, P(RiskLevels|G op ) represents the risk probability distribution; W out b represents the weights of the classifier; out This represents the classifier bias; the Softmax function transforms the output into a probability distribution; the output could be P(low) = 0.1, P(medium) = 0.6, P(high) = 0.3.

[0166] S3.4.2. Based on the number of risk levels, multiple risk level probability distributions are obtained. The probability distributions of all risk levels are then weighted and summed to obtain a continuous risk score.

[0167]

[0168] Where C represents the number of risk levels, w i This refers to the preset weights for each level; these weights can be non-linearly set by the company's safety management experts based on the severity of the risk consequences. For example, the weight of a high-risk level can be set several times that of a medium-risk level to amplify the impact of high-risk events in the final score, reflecting the principle of risk aversion. 低 =10,w 中 =50,w 高 =90. P(risk_level) i Let represent the risk probability distribution corresponding to the i-th risk level; R represents the risk score. Then the risk score of the above probability distribution is R = 10 × 0.1 + 50 × 0.6 + 90 × 0.3 = 1 + 30 + 27 = 58.

[0169] S3.5 Interpretability Analysis:

[0170] Show the risk factors in the multi-dimensional data to the approver.

[0171] By combining interpretable AI technologies such as LIME or SHAP, the key factors leading to a high-risk score can be presented to the approver (e.g., "Risk score 82 points, main reasons: 1. High risk of operator fatigue driving (based on recent overtime data); 2. On-site flammable gas concentration is close to the alarm threshold"), assisting the approver in making a final decision.

[0172] Before deploying the model for real-time risk assessment (i.e., performing step S3), a risk assessment model (containing an MLP classifier) ​​needs to be built. This model must undergo an initial training phase to learn and acquire its intelligent decision-making capabilities. This training phase ensures that the model parameters (including W(k) and b(k) of the GNN part and W of the final MLP classifier) ​​are properly configured. out 、b out The effectiveness of the training is as follows:

[0173] Construction of the training dataset:

[0174] First, a large number of special operation records from the company's history are collected as raw data. For each historical operation case, the system automatically extracts its corresponding multi-dimensional data and constructs an "operation safety subgraph" and its corresponding "initial feature vector set" according to the methods in steps S3.1 and S3.2. Simultaneously, based on the actual result of the operation (e.g., safe completion, near miss, resulting in a safety accident), safety experts assign a clear risk level label (i.e., "high," "medium," or "low") to it. This forms a data pair of "(subgraph + feature vector) - risk label," creating a training dataset.

[0175] End-to-end model training:

[0176] Supervised learning is employed to train the entire risk assessment model end-to-end. During the training process:

[0177] Input the "subgraph + feature vector" from the training dataset into the model.

[0178] After risk propagation and aggregation in S3.3 and risk prediction in S3.4.1, the model outputs a predicted probability distribution of risk level.

[0179] The probability distribution predicted by the model is compared with the true risk level label, and the “error” between the two is calculated using a loss function (e.g., the cross-entropy loss function suitable for classification tasks).

[0180] The system employs a backpropagation algorithm and optimizer (such as the Adam optimizer) to adjust and optimize all trainable parameters in the model simultaneously and in one go based on this "error," including the weights W(k) and bias b(k) of the GNN and the weights W of the MLP classifier. out Bias b out .

[0181] Model validation and deployment:

[0182] Repeat the end-to-end model training steps, iterating multiple times on all training data until the model reaches a preset accuracy standard on an independent validation set. After training, this model, which has accurate risk assessment capabilities, can be deployed to the system to perform real-time risk calculations for new and unseen special operation applications.

[0183] This initial training process forms the basis for the "predictive" safety management of this invention.

[0184] S4. Based on the risk value, determine the risk level of this special operation;

[0185] S5. Configure different approval procedures based on risk levels; specifically including:

[0186] For low-risk levels, configure an automatic approval process: automatically generate approval results, and let RPA handle the subsequent generation and distribution of work tickets;

[0187] For medium- or high-risk levels, a manual review and approval process is configured: the special operation application form is linked to the risk analysis report and pushed to the next level approver for manual review. The approver can make a decision of "approval", "approval after additional control measures" or "rejection" based on AI suggestions.

[0188] S6. During special operations, multi-dimensional data are continuously collected, and risk values ​​are calculated in real time. If the real-time risk value is higher than the threshold, an alarm is triggered, and the alarm information is sent to the approver and equipment manager.

[0189] Once the work permit is approved, the system enters real-time monitoring mode. The computer vision module and the IoT integration module continuously monitor the site. Once an anomaly is detected, an alarm is immediately triggered, and the RPA executes intervention commands to achieve dynamic risk control.

[0190] Upon completion of the task, the person in charge submits a completion report via the terminal. All process data, including approval records, monitoring logs, and any recorded near misses, are fully archived. This new data will serve as training samples, used periodically to retrain and optimize the risk assessment model for AI decision-making and analysis, thereby continuously improving the system's risk identification capabilities over time and forming a complete management and technology closed loop.

[0191] The following is a specific application example, in a pharmaceutical clean area, where hot work welding is being performed on pipelines.

[0192] Step 1: Job Application and Automatic Information Collection

[0193] Time: 9:00 AM the day before the assignment.

[0194] Operation: A maintenance worker named Wang submitted a "Special Operation Application" via a mobile app. He filled in the form with the operation type as "Hot Work," the location as "Clean Room 301, Building A2, Pipeline P-3B," and the operation content as "Repair welding of the P-3B pipe flange." He also took and uploaded a photo of the surrounding environment as required.

[0195] System Response: Upon receiving the request, the system immediately triggers the data acquisition and RPA execution module.

[0196] RPA Robot 1 accessed the HR system and found Wang's employee number, senior welder certificate (valid until June next year), and hot work safety training (just completed last week with a score of 95).

[0197] RPA Robot 2 accessed the EAM system and found that the material of pipe P-3B is 316L stainless steel, and the last pressure test was six months ago, and the condition is normal.

[0198] RPA robot 3 accessed the MES system and confirmed that there were no production tasks in workshop 301 of building A2, so maintenance work could be carried out.

[0199] Step 2: Multimodal Validation and Intelligent Evaluation

[0200] System response: All information is aggregated into the AI ​​decision-making and analysis module.

[0201] The computer vision module analyzed the photos uploaded by Wang Gong and identified several uncleaned cardboard boxes within 1 meter of the work site, classifying them as "general environmental risks".

[0202] The natural language processing module parses the job content and matches it with the internal "SOP-HW-002: Clean Area Hot Work Procedures".

[0203] The AI ​​decision-making and analysis module constructs a subgraph for this task, which includes nodes such as Engineer Wang, Pipeline P-3B, Workshop 301, Cardboard Box Risk, and SOP-HW-002, as well as their relationships.

[0204] The risk quantification algorithm begins calculation:

[0205] Engineer Wang has good performance characteristics (training just completed, all qualifications are complete).

[0206] The pipeline node characteristics are good.

[0207] There is a general risk at the environmental node (the cardboard box identified by CV).

[0208] After information propagation and aggregation analysis, the GNN model outputs a risk score of 45. The interpretability report states: "Main risk contributing factor: Presence of flammable materials (cardboard boxes) at the work site."

[0209] Step 3: Tiered Approval and Process Execution

[0210] System Decision: The system's preset automatic approval threshold is 40, and the manual review threshold is 70. The current risk score is 45, falling between 40 and 70. The system determines that manual review is required.

[0211] System Response: The RPA robot immediately pushed an approval form containing all collected information, CV analysis results (the cardboard box is marked with a red box in the photo), and AI risk report (risk score 45, main risk factors) to Manager Li, the on-duty security supervisor, via DingTalk.

[0212] Step 4: Manual review and issuance of permits

[0213] Time: 9:15 AM.

[0214] Procedure: Manager Li received the approval form on the service terminal (such as DingTalk). He saw the complete AI analysis, which was very clear. He did not approve it, but instead added a mandatory requirement to the approval comments: "Before the work begins, all cardboard boxes on site must be cleared to a safe distance of 5 meters, and photos must be taken and uploaded for review." Then he clicked "Conditional Approval".

[0215] System response: The RPA robot received the "conditional approval" instruction and sent the updated work order (including Manager Li's additional requirements) back to Engineer Wang.

[0216] Step 5: Continuous monitoring and dynamic intervention during the process

[0217] Time: 2:00 PM on the day of the assignment.

[0218] Procedure: After Wang cleaned the cardboard boxes and uploaded photos for verification, he began the work. He wore all required PPE.

[0219] System Response: The fixed cameras deployed in Workshop 301 began sending video streams to the computer vision module. At 2:30 PM, the computer vision module detected that Engineer Wang's welding mask had been raised for more than 10 seconds (possibly to observe the weld seam), while the welding arc was still burning. The system immediately identified this as a "serious violation" and initiated the following intervention: Engineer Wang's smart bracelet emitted a "beep" vibration alarm.

[0220] Manager Li's phone immediately received a highlighted notification: "[Level 1 Alert] Serious violation occurred during hot work in Workshop 301: Mask removed during operation! Please handle immediately!" The on-site audible and visual alarms flashed yellow.

[0221] Action taken: Upon realizing the alarm, Engineer Wang immediately lowered his welding mask, eliminating the risk. Manager Li also verbally warned Engineer Wang via walkie-talkie.

[0222] Step 6: Completion Confirmation and Closed-Loop Learning

[0223] Time: 4:00 PM.

[0224] Operation: The task is completed, and Engineer Wang submits the completion confirmation via the APP.

[0225] System Response: The RPA robot packages all records from this operation (from the application form, AI report, Manager Li's approval comments, Engineer Wang's review photos, to screenshots of violations and alarm records from in-process monitoring) into a complete PDF file and stores it in the safety knowledge base. More importantly, this complete data chain of "violations occurring during the operation, but timely intervention preventing an accident" is marked as a new training sample. In the next model iteration training, the AI ​​will learn the association between this specific pattern and potential risks, and may make more sensitive and accurate judgments about risks when encountering similar situations in the future.

[0226] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for safety risk management and control in special operations in the food and drug industry, characterized in that, include: S1. Collect and maintain multi-dimensional data, including: environmental parameters, environmental compliance data, personnel compliance data, tool compliance data, work order parsing data, personnel basic information, work equipment basic information, and material management information; S2. Construct a safety production knowledge graph, which includes the correspondence between personnel, equipment, materials, regulations, environment, and management systems. S3. Analyze and judge multi-dimensional data, and use the safety production knowledge graph to calculate the risk value of multi-dimensional data; S4. Based on the risk value, determine the risk level of this special operation; S5. Configure different approval procedures according to the risk level.

2. The method for safety risk control in special operations in the food and drug industry according to claim 1, characterized in that, Constructing a knowledge graph for safe production, specifically including: Ontology modeling: Define security entity nodes, which include personnel, equipment, materials, regulations, environment, and management systems, and associate the attributes of each security entity node with them; Relationship extraction: Define the correspondence between secure entity nodes; A safety production knowledge graph is constructed using safety entity nodes as nodes and corresponding relationships as edges.

3. The method for safety risk control in special operations in the food and drug industry according to claim 1, characterized in that, Calculating the risk value of multi-dimensional data specifically includes: S3.1 Extracting the work safety sub-diagram: Based on the multi-dimensional data of this special operation collected by S1, it is mapped to the safety production knowledge graph, and the safety entity nodes and their relationships related to this special operation are extracted to form an operation safety subgraph. S3.2, Feature vectorization: Transform each node and its attribute information in the job safety subgraph into an initial feature vector; S3.3 Risk Propagation and Aggregation: By using graph neural network models to propagate information on knowledge graphs, graph-level embedding vectors containing information about neighboring nodes are formed. S3.4 Risk Prediction and Quantification: The graph-level embedding vectors aggregated from the graph neural network model are input into the machine learning model to calculate the risk value of multi-dimensional data.

4. The method for safety risk control in special operations in the food and drug industry according to claim 3, characterized in that, Calculating the risk value of multidimensional data also includes: S3.5 Interpretability Analysis: Show the risk factors in the multi-dimensional data to the approver.

5. The method for safety risk control in special operations in the food and drug industry according to claim 3, characterized in that, Step S3.3 Risk propagation and aggregation specifically includes: Based on each initial feature vector, the neighbor node information of the corresponding secure entity node is added to form an embedding vector h. v The embedding vector h of all matched secure entity nodes v Forming graph-level embedding vectors Using the GraphSAGE model, graph-level embedding vectors are formed layer by layer. A total of K layers of processing are performed; At layer k, for any safe entity node v in the job safety subgraph, its updated embedding vector The calculation process includes: Aggregation: Aggregate information from the set N(v) of neighboring nodes of node v, expressed as: in, It is the embedding vector of the neighbor node u in the previous layer. For the 0th layer, Let be the initial feature vector of node u; Update: Embed the parent node of node v itself. Its aggregated neighbor information The data is then spliced ​​together, followed by a nonlinear transformation. Among them, W (k) It is the trainable weight matrix of the k-th layer; b (k) It is the trainable bias vector of the k-th layer; σ is a non-linear activation function; CONCAT indicates a splicing operation; After K layers of propagation, each node obtains an embedding vector that incorporates information from its K-hop neighbors. Embedded vector h v Forming graph-level embedding vectors 6. The method for safety risk control in special operations in the food and drug industry according to claim 5, characterized in that, Step S3.4 Risk Prediction and Quantification specifically includes: S3.4.1 Embedding Graph-Level Vectors Inputting the data into the classifier yields the risk level probability distribution: Wherein, P(RiskLevels|G op ) represents the risk probability distribution; W out b represents the weights of the classifier; out This represents the classifier's bias; the Softmax function transforms the output into a probability distribution. S3.4.

2. Based on the number of risk levels, multiple risk level distributions are obtained. The probability distributions of all risk levels are then weighted and summed to obtain a continuous risk score. Where C represents the number of risk levels, w i It is the preset weight for each level; P(risk_level) i ) represents the risk probability distribution corresponding to the i-th risk level; R represents the risk score.

7. The method for safety risk control in special operations in the food and drug industry according to claim 1, characterized in that, Different approval procedures are configured according to the risk level, specifically including: Low-risk level, configure automatic approval process: automatically provide approval results; For medium- or high-risk levels, a manual review and approval process is configured: the special operation application form, along with the risk analysis report, is sent to the next level approver for manual review.

8. The method for safety risk control in special operations in the food and drug industry according to claim 1, characterized in that, The methods for managing safety risks in special operations in the food and drug industry also include: S6. During special operations, multi-dimensional data are continuously collected, and risk values ​​are calculated in real time. If the real-time risk value is higher than the threshold, an alarm is triggered, and the alarm information is sent to the approver and equipment manager.

9. An intelligent robot system for safety risk management in special operations of the food and drug industry, which executes the method for safety risk management in special operations of the food and drug industry as described in any one of claims 1-8, characterized in that, include: The data acquisition and RPA execution module is used for collecting multi-dimensional data and interacting with external information. The multimodal information perception module is used to transform the collected unstructured multidimensional data into structured data; The AI ​​decision-making and analysis module is used to analyze and judge multi-dimensional data, and to calculate the risk value of multi-dimensional data using a safety production knowledge graph. Based on the risk value, determine the risk level of this special operation; Different approval procedures are configured according to the risk level.

10. The intelligent robot system for safety risk management and control in special operations of the food and drug industry according to claim 9, characterized in that, The multimodal information sensing module includes: The computer vision module is used to receive image information and video streams from the work site, and to make a preliminary judgment on whether there are any violations at the work site based on the image information and video streams. The natural language processing module is used to generate parsed data for the work order from the input text information. The IoT integration module is used to receive environmental parameters collected by sensors, preprocess the environmental parameters, and add timestamps.

Images