Chemical inventory management system based on industrial big data

Abstract

This invention discloses a chemical inventory management system based on industrial big data, relating to the field of intelligent warehousing technology. The system includes a sensing and acquisition layer, a data processing layer, a risk modeling layer, a decision scheduling layer, and an execution feedback layer. The sensing and acquisition layer collects temperature, gas concentration, and vibration data of the storage environment. A dynamic risk field construction module calculates the dynamic risk field within the storage space based on the collected data. A cascade reaction knowledge graph reasoning module infers the trigger probability of cascade reactions by combining a chemical reaction knowledge graph. An inventory adaptive scheduling module based on reinforcement learning generates an inventory reconstruction scheduling strategy based on the risk field and trigger probability, driving an automated guided vehicle to perform a warehouse transfer operation. This invention realizes dynamic perception of chemical storage risks, pre-deduction of cascade reactions, and adaptive adjustment of inventory layout, and can be used for chemical storage safety management.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehousing technology, specifically to a chemical inventory management system based on industrial big data. Background Technology

[0002] Chemical storage management involves well-known technological elements such as industrial Internet of Things (IoT) sensing technology, warehouse management systems, automated storage and retrieval systems (AS / RS) and automated guided vehicles (AGVs), big data stream processing frameworks, and graph database technology. These elements are used to collect storage environment data, manage storage location business processes, and automate the handling and retrieval of goods.

[0003] The closest existing technical solution to this invention is a chemical storage management method that links an environmental monitoring system and a warehouse management system. This solution pre-defines chemical storage zones according to national standards, sets fixed physical isolation distances for different types of chemicals, deploys temperature and humidity sensors and gas concentration detectors in the warehouse, collects environmental parameters at fixed intervals and uploads them to the environmental monitoring system, sets a single threshold alarm rule for the collected data, and triggers an audible and visual alarm when the value at a certain monitoring point exceeds the preset threshold, and the warehouse management system schedules the entry and exit of chemicals according to the zoning rules and batch information of the chemicals, using the first-in-first-out principle, and does not make dynamic adjustments based on environmental changes after entry. After the alarm is triggered, the on-site management personnel judge based on experience and make a manual decision on whether to transfer the chemicals.

[0004] In the aforementioned scheme, the warehouse management system and environmental monitoring system are relatively independent at the data level. Environmental monitoring exists only as an alarm tool, and its risk assessment logic is limited to static judgments based on single points and single thresholds. It lacks the ability to model and extrapolate the risks of multi-chemical interactions, and warehouse location scheduling does not dynamically adjust with environmental risks. Consequently, this scheme struggles to perceive the cumulative risk effects caused by sudden changes in local microenvironments, and it is difficult to identify the hidden cascade reaction risks in multi-chemical combination scenarios before the actual occurrence of risks. Furthermore, the warehouse location scheduling rules are independent of the environmental risk status, and post-event handling relies solely on human experience after the alarm is triggered, resulting in a significant time lag in response. Therefore, we propose... out A chemical inventory management system based on industrial big data. Summary of the Invention

[0005] The purpose of this invention is to provide a chemical inventory management system based on industrial big data.

[0006] The technical problem this invention aims to solve is that existing chemical storage management systems rely solely on the inherent properties of chemicals and static storage rules for zoning, isolation, and physical control. They are unable to perceive and assess the dynamic superimposed risks caused by sudden changes in the local microenvironment. Existing systems lack the ability to extrapolate implicit cascade reaction chains in scenarios where multiple types of chemicals are stored together, making it difficult to proactively identify secondary risks caused by the interaction of multiple chemicals before risks occur. Furthermore, existing early warning mechanisms are based on a passive triggering method using a single threshold, meaning that risk events often occur before alarms are generated, and they lack the ability to proactively reconstruct the inventory layout in the early stages of risk accumulation.

[0007] To address the aforementioned technical problems, this invention provides a chemical inventory management system based on industrial big data. The system includes a sensing and acquisition layer, a data processing layer, a risk modeling layer, a decision scheduling layer, and an execution feedback layer.

[0008] The sensing and acquisition layer includes a temperature and humidity sensor, a gas concentration detector, and a vibration sensor, which are used to collect temperature data, gas concentration data, and vibration intensity data at multiple cargo locations within the storage space.

[0009] The data processing layer is connected to the sensing and acquisition layer and is used to perform timestamp alignment, missing value interpolation, and abnormal noise filtering on the acquired data to form a structured environmental status data stream with a unified time granularity.

[0010] The risk modeling layer includes a dynamic risk field construction module and a cascade reaction knowledge graph reasoning module. The dynamic risk field construction module is connected to the data processing layer and is used to identify risk source locations based on the structured environmental state data stream, calculate the risk intensity coefficient of each risk source location, and calculate the risk field intensity value of multiple spatial coordinate points within the storage space in combination with the risk diffusion radius parameter, thereby obtaining a dynamic risk field covering the storage space. The cascade reaction knowledge graph reasoning module is connected to the dynamic risk field construction module and stores a chemical reaction knowledge graph with chemicals and environmental factors as nodes and reaction relationships between chemicals as edges. When the peak intensity of the dynamic risk field reaches a preset proportion of a preset critical threshold, it is used to extract the list of chemicals stored at the risk peak location and its adjacent storage locations, retrieve candidate reaction paths in the chemical reaction knowledge graph, match the trigger condition threshold of the candidate reaction paths with the measured environmental parameters, calculate the trigger probability, and summarize to obtain the cumulative trigger probability of the cascade reaction.

[0011] The decision scheduling layer includes an inventory adaptive scheduling module based on reinforcement learning, which is connected to a cascade reaction knowledge graph reasoning module. It is used to encode the spatial intensity distribution of the dynamic risk field, the cumulative trigger probability of cascade reactions, and the chemical occupancy status of the storage location into a state vector. It estimates the value of candidate transfer actions through a deep Q-network, selects the transfer action to be executed, and generates an inventory restructuring scheduling strategy.

[0012] The execution feedback layer includes the warehouse location management module and the automated guided vehicle (AGV) scheduling interface of the warehouse management system. It is connected to the decision scheduling layer and is used to convert the inventory restructuring scheduling strategy into a transfer task instruction and send it to the AGV for execution, as well as to feed the execution result back to the data processing layer, forming a closed-loop control from environmental data collection to proactive inventory restructuring.

[0013] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0014] 1. This invention constructs a three-dimensional dynamic risk field by integrating multi-source heterogeneous real-time sensing data, thereby enabling real-time perception of the superimposed risks caused by sudden changes in the local microenvironment, and advancing the risk detection time from post-event alarm to the trend accumulation stage.

[0015] 2. This invention introduces chemical reaction knowledge graphs to deduce cascade reaction chains, enabling quantitative assessment and early warning of hidden secondary risks in multi-chemical combination scenarios;

[0016] 3. This invention achieves automated proactive inventory transfer when risks approach the critical point through a reinforcement learning-driven adaptive inventory scheduling mechanism, upgrading the inventory management mode from passive alarm and manual handling to proactive perception and self-healing scheduling, shortening risk response time and reducing the probability of accidents. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall five-layer vertical architecture of the system in this embodiment of the invention;

[0018] Figure 2 This is a schematic diagram of the vertical processing flow of the risk modeling layer in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the vertical process of decision scheduling and execution feedback closed loop in an embodiment of the present invention. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0021] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0022] I. System Overall Architecture;

[0023] The chemical inventory management system provided in this embodiment of the invention comprises five layers in its overall architecture: a perception and acquisition layer, a data processing layer, a risk modeling layer, a decision scheduling layer, and an execution feedback layer. These layers coordinate with each other through data interfaces and command channels to jointly achieve closed-loop control from environmental data acquisition to proactive inventory reconfiguration.

[0024] The sensing and acquisition layer is equipped with an array of temperature and humidity sensors, a gas concentration detector, a vibration sensor, a laser rangefinder, and RFID location tags. Each sensing device transmits raw time-series data upwards at fixed intervals or in an event-triggered manner. The laser rangefinder is used to acquire the spatial coordinate information of each location in the three-dimensional coordinate system of the warehouse, and the RFID location tags are used to record the identity information of the chemicals currently stored in each location. Together, they provide basic data support for subsequent spatial positioning of the risk field and extraction of the chemical inventory. In a specific embodiment, various sensors can be deployed in an array according to the distribution density of the locations to ensure that the spatial resolution of risk source identification meets the needs of warehouse safety management.

[0025] The data processing layer includes a data access module and a streaming computing module. The data access module uses message queue technology to cache the raw data, while the streaming computing module uses a streaming computing framework to clean, denoise, align the time, and fuse multi-source data to form a structured environmental state data stream in a unified format, which is then output to the risk modeling layer. In one specific embodiment, the data access module uses Kafka message queue to cache the raw data for high concurrency, while the streaming computing module uses the Flink streaming computing framework to perform real-time cleaning, aggregation, and computation. The two work together to ensure that the data processing layer can generate a structured environmental state data stream at a short time granularity, providing a data foundation for the real-time computation of the risk modeling layer.

[0026] The risk modeling layer includes a dynamic risk field construction module and a hidden cascade reaction knowledge graph reasoning module. The dynamic risk field construction module calculates the risk field intensity distribution of each point in the warehouse space in real time based on the fused environmental data. The hidden cascade reaction knowledge graph reasoning module infers potential cascade reaction paths and trigger probabilities based on the chemical reaction knowledge graph stored in the graph database and the current environmental state.

[0027] The decision scheduling layer includes an inventory adaptive scheduling engine based on reinforcement learning. This engine takes the risk field strength and cascade reaction probability as input features, combines the current warehouse location distribution and the cost of warehouse transfer, and outputs an inventory reconfiguration scheduling strategy.

[0028] The execution feedback layer includes the warehouse location management module and the automated guided vehicle (AGV) scheduling interface of the warehouse management system. It is responsible for converting the scheduling strategy into specific AGV transfer task instructions and executing them, while feeding back the execution results to the data processing layer to form a complete closed-loop control link.

[0029] II. Construction of Dynamic Risk Field;

[0030] The goal of the dynamic risk field module is to transform discrete multi-source sensor data within the warehouse into a continuous spatial risk intensity distribution in real time, in order to discover the superimposed risks caused by sudden changes in the local microenvironment. Its processing flow is as follows:

[0031] The process begins with multi-source sensor data access. After the data is preprocessed by an outlier removal and time alignment, it enters the risk source identification module. This module identifies risk source locations with current abnormal trends based on preset environmental anomaly judgment rules. Subsequently, it enters the risk field diffusion calculation module. This module calculates the risk intensity attenuation distribution in three-dimensional space with each risk source location as the center, combined with environmental diffusion parameters, and superimposes and merges the distribution fields of each risk source to generate an overall risk field. Finally, it enters the risk field assessment module to determine whether the peak value of the risk field reaches a preset proportion of a preset critical threshold. If it does, a cascade reaction deduction and scheduling decision process is triggered. If it does not, the current risk field data is archived and enters the next round of collection cycle. In a specific embodiment, the sliding time window used for risk source identification is set to 5 minutes. The anomaly judgment threshold is dynamically set based on the mean and standard deviation of historical data statistics to adapt to the normal fluctuation range under different seasons and different storage environment conditions, and to avoid misjudging normal environmental fluctuations as risk sources.

[0032] The spatial intensity distribution of the dynamic risk field is quantified using the following model:

[0033] ;

[0034] In the formula, Represents spatial coordinates At any moment The comprehensive risk field intensity value, This represents the total number of risk sources identified at the current moment. Indicates the first Each risk source is at any time The risk intensity coefficient, Indicates the first The spatial location of each risk source in the warehouse's three-dimensional coordinate system. Indicates the first Each risk source is at any time The risk diffusion radius parameter is dynamically adjusted according to environmental flow field conditions. The larger the diffusion radius, the wider the range of risk impact.

[0035] Risk intensity coefficient The calculation method is as follows:

[0036] ;

[0037] In the formula, Indicates the first Each risk source location at any time Temperature offset relative to historical baseline This represents the preset temperature offset reference value, used for normalization processing. This indicates that the point is at time [time]. Gas concentration shift relative to historical baseline, This indicates the preset gas concentration offset reference value. This indicates that the point is at time [time]. The measured value of vibration intensity This indicates the preset vibration intensity benchmark value. These represent the weighting coefficients of temperature, gas concentration, and vibration in the comprehensive risk intensity calculation, respectively. The sum of the three is 1. The specific values ​​are determined based on historical accident data or set by the experience of domain experts. In a specific embodiment, Take 0.4, Take 0.4, Take 0.2;

[0038] Taking a specific risk source location as an example, if the temperature offset at this location is 8 degrees Celsius, the preset temperature offset benchmark value is 10 degrees Celsius, the gas concentration offset is 15 mg / m³, the preset gas concentration offset benchmark value is 20 mg / m³, the measured vibration intensity value is 0.6, and the preset vibration intensity benchmark value is 1.0, then according to the above weighting coefficients, the risk intensity coefficient of this location is the sum of 0.4 multiplied by 0.8, 0.4 multiplied by 0.75, and 0.2 multiplied by 0.6, with a calculated result of approximately 0.62. Substituting this risk intensity coefficient into the spatial intensity distribution model, and combining it with the risk diffusion radius parameter corresponding to this location, the risk intensity contribution value of this risk source at each location within the three-dimensional space of the warehouse can be calculated.

[0039] The specific implementation of the dynamic risk field construction is as follows: The system receives raw time-series data collected in real time from temperature and humidity sensors, gas concentration detectors, and vibration sensors via industrial IoT communication protocols. It then performs timestamp alignment, missing value interpolation, and abnormal noise filtering to form a structured environmental state data stream with a unified time granularity. The system performs statistical analysis on the pre-processed data stream according to a preset sliding time window, calculating the rate of change of values ​​at each monitoring point within the window. When the rate of change of temperature, gas concentration, or vibration intensity at a certain point exceeds a preset anomaly judgment threshold, that point is marked as a risk source at the current moment, and its spatial coordinate information is recorded. For each risk source, the system calculates the risk intensity coefficient using the aforementioned method, integrating the temperature offset, gas concentration offset, and vibration intensity at that point. The system calculates the risk intensity coefficient of the risk source at the current moment based on the measured value. Combining the current warehouse ventilation speed, temperature gradient and other environmental flow field parameters, the system dynamically adjusts the risk diffusion radius parameter of each risk source. According to the above spatial intensity distribution model, the system calculates the risk intensity contribution value of the risk source at each location in the three-dimensional space of the warehouse. The system superimposes and sums the spatial intensity contribution values ​​of all risk sources in the same spatial coordinate system to obtain the dynamic risk field distribution of the warehouse as a whole at the current moment. The system extracts the peak intensity and its corresponding spatial location in the dynamic risk field and compares the peak intensity with the preset critical threshold. When the peak intensity reaches the preset proportion of the critical threshold, the subsequent implicit cascade reaction inference process is triggered. Otherwise, the current risk field data is archived and the above processing is repeated in the next acquisition cycle.

[0040] III. Reasoning based on the knowledge graph of implicit cascade reactions;

[0041] To enable advance simulation of hidden secondary risks in multi-chemical combination scenarios, this embodiment of the invention introduces a chemical reaction knowledge graph based on a graph database and combines it with current dynamic risk field data to perform sand table simulation of cascade reaction chains.

[0042] The chemical reaction knowledge graph is stored in the form of a graph database. Its node types include chemical nodes and environmental factor nodes. Chemical nodes contain attribute information such as chemical name, hazard category, and physicochemical properties, while environmental factor nodes contain attribute information such as temperature range and humidity range. Edges represent the possible reaction relationships between chemical nodes or between chemical nodes and environmental factor nodes under specific conditions. The attributes of the edges include reaction type, trigger condition threshold, and historical trigger probability statistics. The graph data comes from the structured transformation of chemical safety data sheets, historical accident case databases, and chemical incompatibility databases. This knowledge graph can be incrementally updated based on the subsequent addition of chemical safety data sheets, newly occurring accident case data, and updates to the chemical incompatibility database to ensure that the deduced cascade reaction paths can continuously reflect the latest understanding of chemical reaction relationships.

[0043] The reasoning process begins with trigger condition access. The system receives the chemical list information of high-risk locations and adjacent storage locations output by the dynamic risk field module. Then, it enters the graph query module. This module uses the chemicals stored at the high-risk locations and adjacent storage locations as the query starting point and searches for candidate reaction paths with these chemicals as nodes in the chemical reaction knowledge graph. Next, it enters the trigger condition matching module to determine whether the trigger condition of each candidate reaction path matches the measured environmental parameters of the current dynamic risk field. Then, it enters the cascade probability calculation module to calculate the trigger probability of each successfully matched reaction path and summarizes and calculates the cumulative probability of the overall cascade reaction. Finally, it enters the result output module to transmit the calculated cumulative cascade reaction probability and specific reaction path information to the inventory adaptive scheduling module.

[0044] The cumulative trigger probability of a cascade reaction is calculated as follows:

[0045] ;

[0046] In the formula, This represents the cumulative probability of a cascade reaction occurring under current environmental conditions. This represents the total number of candidate reaction paths retrieved from the knowledge graph whose triggering conditions match the current environmental parameters. Indicates the first The individual trigger probability of each candidate reaction path under the current environmental conditions;

[0047] Single reaction path trigger probability The correction calculation method is as follows:

[0048] ;

[0049] In the formula, Indicates the first The historical trigger probability of each reaction path is derived from the statistical values ​​in the edge attributes of the knowledge graph. Indicates time Measured environmental parameter values ​​related to the triggering conditions of this reaction pathway. This indicates the trigger condition threshold corresponding to this reaction path. This represents the amplification factor of the degree to which environmental parameters deviate from the threshold on the trigger probability. It is used to reflect the causal relationship that the closer the environmental parameters are to or exceed the threshold, the higher the probability of triggering the reaction. In a specific embodiment, Take 0.5;

[0050] Taking a candidate reaction path as an example, if the historical basic trigger probability of this path is 0.3, the temperature corresponding to the trigger condition threshold is 60 degrees Celsius, and the current measured temperature is 66 degrees Celsius, then the trigger probability of this path at the current moment is approximately 0.315. If two candidate reaction paths with matching trigger conditions are found in the knowledge graph, and the trigger probability of the other path is 0.2, then according to the cumulative trigger probability calculation method of cascade reaction, the cumulative trigger probability of cascade reaction in this scenario is approximately 0.452.

[0051] The specific implementation of implicit cascade reaction knowledge graph reasoning is as follows: The system receives the spatial coordinates of the peak risk point output by the dynamic risk field module. Combined with the location-goods correspondence data of the warehouse management system, it extracts the chemical list of all locations within the preset neighboring range of the point. Using the extracted chemical list as the query starting point, the system performs a graph traversal query in the chemical reaction knowledge graph to retrieve all candidate reaction paths with the chemicals in the list as nodes and the path length not exceeding the preset hop limit. For each candidate reaction path retrieved, the system reads the trigger condition threshold recorded on each edge of the path and compares it with the measured environmental parameters of the point output by the current dynamic risk field module. When the measured parameters reach the trigger threshold, the system will trigger the trigger condition. When the preset proportion of the trigger condition threshold is reached, the path is marked as a successfully matched path. For each reaction path marked as a successfully matched path, the system calculates the trigger probability of the path at the current moment by combining the historical basic trigger probability of the path with the degree of deviation of the current environmental parameters from the threshold according to the above-mentioned single reaction path trigger probability correction calculation method. The system summarizes and calculates the individual trigger probabilities of all successfully matched paths according to the above-mentioned cascade reaction cumulative trigger probability calculation method to obtain the overall cumulative probability of the cascade reaction occurring at the current moment. The system transmits the calculated cumulative cascade reaction probability, the specific chemicals involved and the reaction path information to the inventory adaptive scheduling module as the input basis for subsequent scheduling decisions.

[0052] IV. Adaptive Inventory Scheduling;

[0053] After obtaining the dynamic risk field distribution and cascade reaction simulation results, this embodiment of the invention generates an adaptive inventory reconstruction strategy through a reinforcement learning algorithm to guide the automated guided vehicle to perform proactive inventory transfer operations, thereby realizing the transformation from passive alarm to proactive defense.

[0054] The state space of the reinforcement learning model consists of dynamic risk field distribution data, cascade reaction cumulative probability and chemical occupancy status of each storage location, while the action space consists of a set of transfer actions that the automated guided vehicle can execute. Each action corresponds to moving a specified item from the current storage location to a candidate storage location. The reward function is used to comprehensively evaluate the combined effect of the transfer actions on reducing risk and controlling scheduling costs.

[0055] The process begins with the state observation module, which encodes the current dynamic risk field distribution, cascade reaction probability, and chemical distribution information at each location into a reinforcement learning state vector. Next, it enters the action selection module, which estimates the value of each candidate transfer action in the current state based on a deep Q-network and selects the action to execute according to an exploration-utilization balance strategy. Then, it enters the action execution and environmental feedback module, where the selected transfer action is sent to the automated guided vehicle (AGV) via the warehouse management system interface. After execution, the system re-collects environmental data to calculate the new risk field state and transfer cost, and calculates the reward value accordingly. Next, it enters the network parameter update module, which uses the observed state transition samples to update the deep Q-network parameters using gradients. Finally, it determines whether the scheduling cycle termination condition has been met. If not, it returns to the state observation module to continue the next round of decision-making; if it has, it outputs the final inventory restructuring strategy for the current cycle.

[0056] The reward function is calculated as follows:

[0057] ;

[0058] In the formula, Indicates at time The immediate reward value obtained after executing a scheduling action. This represents the peak intensity of the overall dynamic risk field of the warehouse at the next moment after the action is performed. This represents the handling costs incurred in performing this warehouse transfer operation, a comprehensive measure including factors such as the distance traveled by the automated guided vehicle and the operation time. This represents the cumulative probability of cascading reactions being triggered in the next moment after an action is performed. These represent the weighting coefficients of risk peak, transfer cost, and cascading reaction probability in the reward function, respectively. A larger weighting value indicates a higher degree of influence of that factor on the scheduling decision. In a specific embodiment... Take 0.5, Take 0.2, Take 0.3;

[0059] In a sample inventory transfer decision, if the peak risk intensity is 0.35, the handling cost metric is 0.1, and the cumulative cascade reaction trigger probability is 0.2 after executing a candidate inventory transfer action, then according to the above weights, the immediate reward value corresponding to this action is approximately -0.255. If the peak risk intensity is 0.5, the handling cost is 0.05, and the cumulative cascade reaction trigger probability is 0.3, then the reward value corresponding to this action is approximately -0.35. Since the reward value of the former is higher than that of the latter, the system selects the former as the inventory transfer action to be executed in this round from the two candidate actions.

[0060] The value function of a deep Q-network is updated as follows:

[0061] ;

[0062] In the formula, Indicates the state Next action The value estimate, The learning rate is used to control the step size for updating the value function. This represents a discount factor, used to measure the importance of future rewards in current decision-making. Indicates the next state The maximum value estimate among all available actions. Indicates the candidate action for the next state;

[0063] The specific implementation of the reinforcement learning-based adaptive inventory scheduling is as follows: The system normalizes and encodes the spatial risk intensity distribution output by the dynamic risk field module, the cumulative trigger probability output by the cascade reaction knowledge graph inference module, and the current location occupancy information of the warehouse management system to form the state vector of the reinforcement learning model. Based on the current high-risk locations and the list of hazardous chemicals derived from the knowledge graph, combined with the information of currently available and low-risk candidate locations in the warehouse, the system generates a set of candidate transfer actions for this round of decision-making. The system inputs the state vector into a deep Q-network to obtain the value estimate corresponding to each action in the candidate action set, and selects between the maximum value action and random actions according to the exploration rate parameter to determine the transfer action to be executed in this round. The system converts the selected transfer action into a specific automated guided vehicle (AGV) scheduling instruction and sends it to the AGV control system through the warehouse management system interface. The AGV then executes the actual chemical location transfer operation. After the action is completed, the system re-collects warehouse environmental data and, according to the aforementioned dynamic... The system recalculates the new risk field distribution using the state risk field construction method and recalculates the new cumulative trigger probability using the aforementioned cascade reaction knowledge graph reasoning method. Then, it calculates the reward value of the current action according to the aforementioned reward function calculation method. The system stores the sample consisting of the current state, action, reward, and new state into the experience replay pool and randomly samples a batch of samples from the experience replay pool. The system updates the parameters of the deep Q network according to the aforementioned value function update method. The system determines whether the current scheduling cycle has reached the preset termination condition. If it has not reached the condition, it continues the next round of decision-making based on the new state. If it has reached the condition, it outputs the final inventory reconstruction strategy formed in the current cycle and synchronizes the strategy to the warehouse management system to update the storage location information. During the model training phase, the system can construct an offline training sample set based on historical warehousing environment data and historical transfer operation records to pre-train the deep Q network. During the actual operation of the system, it continuously collects new state transition samples to update the network parameters online, so as to gradually improve the adaptability of the scheduling strategy in different warehousing scenarios.

[0064] V. Overall System Collaboration Workflow;

[0065] The perception and acquisition layer continuously sends raw data to the data processing layer, which in turn transmits fused data to the dynamic risk field module. When the dynamic risk field module detects that the risk is approaching the critical value, it sends a trigger request to the knowledge graph reasoning module. The knowledge graph reasoning module transmits the cascade reaction inference results to the reinforcement learning scheduling module. The scheduling module generates scheduling instructions and issues them to the automated guided vehicle execution layer. After the automated guided vehicle completes its execution, it feeds back the results to the perception and acquisition layer to form a closed loop. This collaborative workflow fully reflects the closed-loop collaborative relationship of the entire chain from data perception to proactive defense decision-making and then to execution feedback in this system.

[0066] VI. Application Scenarios Examples;

[0067] Taking a peroxide chemical storage area in a chemical warehouse as an example, the specific operation of this invention is explained. This storage area is divided into multiple storage locations in a three-dimensional coordinate system. Each storage location is equipped with a temperature and humidity sensor, a gas concentration detector, and a vibration sensor. When a forklift in the warehouse generates local vibration during handling operations, accompanied by a slight increase in temperature in that area, after data cleaning and time alignment at the data processing layer, the risk source identification module determines, based on the rate of change of values ​​within a sliding time window, that both the vibration intensity change rate and the temperature change rate in that area exceed a preset anomaly threshold. The storage location is then marked as a risk source, and its spatial coordinates are recorded. The dynamic risk field construction module combines the temperature offset, gas concentration offset, and measured vibration intensity of the risk source to obtain its risk intensity coefficient according to the risk intensity coefficient calculation method. Combined with the risk diffusion radius parameter determined by the current warehouse ventilation speed, the dynamic risk field distribution covering the storage area is calculated. When the peak intensity of the risk field reaches a preset proportion of a preset critical threshold, the system triggers a cascade reaction simulation flow. The process involves the following steps: The cascade reaction knowledge graph reasoning module extracts the risk peak point and the list of chemicals stored in adjacent locations. It then searches the chemical reaction knowledge graph for candidate reaction paths with the chemicals in the list as nodes. The trigger condition thresholds of each candidate path are matched with the currently measured temperature and vibration parameters. The trigger probability of the successfully matched path is calculated and the cumulative trigger probability of the cascade reaction is obtained. The inventory adaptive scheduling module encodes the above risk field distribution, the cumulative trigger probability of the cascade reaction, and the chemical occupancy status of the location into a state vector. This vector is input into a deep Q network to estimate the value of candidate transfer actions. The module selects to transfer the chemicals stored in the high-risk location to a candidate location with a lower risk level, generates an inventory restructuring scheduling strategy, and issues a transfer task instruction to the automated guided vehicle (AGV) through the warehouse management system interface. After the AAV completes the transfer, the system re-collects environmental data. The peak intensity of the dynamic risk field and the cumulative trigger probability of the cascade reaction are both reduced before the transfer, verifying the active defense effect of this invention against the superimposed risks caused by local microenvironmental changes and the risks of hidden cascade reactions.

[0068] VII. Other implementation methods;

[0069] In the spatial diffusion calculation of dynamic risk fields, as an alternative implementation method, the Gaussian spatial decay model mentioned above can be abandoned. Instead, a numerical approximation solution of the hydrodynamic diffusion partial differential equation can be performed using a physical information neural network. By training the neural network, the diffusion law of risky substances in the warehouse space can be directly learned, thereby obtaining a higher accuracy risk field distribution result in warehouse scenarios with complex ventilation conditions and uneven airflow distribution. This implementation method remains unchanged in the overall system architecture and knowledge graph inference and reinforcement learning scheduling.

[0070] In the adaptive inventory scheduling algorithm, as an alternative implementation method, the deep Q-network mentioned above can be omitted for action value estimation. Instead, a scheduling method based on a combination of rule engine and heuristic search algorithm can be used. That is, a comprehensive scoring function is designed in advance based on the risk field strength, cascade reaction probability and transfer cost. Heuristic search is used to find the transfer scheme with better score in the candidate storage space. This implementation method does not require offline training and online updating of reinforcement learning model. It has a faster deployment capability in warehousing scenarios with limited computing resources or insufficient historical training data. The remaining risk field construction and knowledge graph inference parts remain unchanged.

[0071] The above 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 through the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention, and such modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

[0072] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A chemical inventory management system based on industrial big data, characterized in that, It includes a perception and acquisition layer, a data processing layer, a risk modeling layer, a decision scheduling layer, and an execution feedback layer; The sensing and acquisition layer includes a temperature and humidity sensor, a gas concentration detector, and a vibration sensor, which are used to collect temperature data, gas concentration data, and vibration intensity data at multiple cargo points within the storage space. The data processing layer is connected to the sensing and acquisition layer and is used to perform timestamp alignment, missing value interpolation, and abnormal noise filtering on the acquired temperature data, gas concentration data, and vibration intensity data to form a structured environmental state data stream. The risk modeling layer includes a dynamic risk field construction module and a cascade reaction knowledge graph reasoning module; The dynamic risk field construction module is connected to the data processing layer and is used to identify risk source locations based on the structured environmental state data stream, calculate the risk intensity coefficient of each risk source location, and calculate the risk field intensity value of multiple spatial coordinate points in the storage space based on the risk intensity coefficient and the risk diffusion radius parameter, so as to obtain a dynamic risk field covering the storage space. The cascade reaction knowledge graph reasoning module is connected to the dynamic risk field construction module and stores a chemical reaction knowledge graph. The chemical reaction knowledge graph uses chemicals and environmental factors as nodes and the reaction relationships between chemicals as edges. The cascade reaction knowledge graph reasoning module is used to extract the list of chemicals stored at the corresponding location of the risk peak and its adjacent storage locations when the peak intensity of the dynamic risk field reaches a preset proportion of a preset critical threshold. It then retrieves candidate reaction paths with the chemicals in the chemical list as nodes in the chemical reaction knowledge graph, matches the trigger condition threshold of the candidate reaction path with the measured environmental parameters of the corresponding location of the dynamic risk field, calculates the trigger probability of the successfully matched candidate reaction path, and summarizes the cumulative trigger probability of the cascade reaction. The decision scheduling layer includes an inventory adaptive scheduling module based on reinforcement learning, which is connected to the cascade reaction knowledge graph reasoning module. It is used to encode the spatial intensity distribution of the dynamic risk field, the cumulative trigger probability of the cascade reaction, and the chemical occupancy status of the storage location into a state vector, estimate the value of candidate transfer actions through a deep Q network, select the transfer action to be executed, and generate an inventory restructuring scheduling strategy. The execution feedback layer includes the warehouse location management module and the automated guided vehicle (AGV) scheduling interface of the warehouse management system. It is connected to the decision scheduling layer and is used to convert the inventory restructuring scheduling strategy into a transfer task instruction and send it to the AGV for execution, as well as to feed back the execution result to the data processing layer.

2. The chemical inventory management system based on industrial big data according to claim 1, characterized in that: When calculating the risk intensity coefficient, the dynamic risk field construction module multiplies the ratio of the temperature offset of the risk source point to the preset temperature offset benchmark value, the ratio of the gas concentration offset to the preset gas concentration offset benchmark value, and the ratio of the measured vibration intensity value to the preset vibration intensity benchmark value by preset temperature weighting coefficients, gas concentration weighting coefficients, and vibration weighting coefficients, respectively, and then sums them to obtain the risk intensity coefficient of the risk source point. The sum of the temperature weighting coefficient, the gas concentration weighting coefficient, and the vibration weighting coefficient is 1.

3. A chemical inventory management system based on industrial big data according to claim 1, characterized in that: When calculating the trigger probability of a successfully matched candidate reaction path, the cascade reaction knowledge graph reasoning module performs a correction calculation based on the historical basic trigger probability of the candidate reaction path and the degree of deviation between the measured environmental parameter value corresponding to the candidate reaction path and the trigger condition threshold. The greater the deviation of the measured environmental parameter value from the trigger condition threshold, the higher the corrected trigger probability.

4. A chemical inventory management system based on industrial big data according to claim 1, characterized in that: The reward function of the inventory adaptive scheduling module is calculated based on the peak intensity of the dynamic risk field of the storage space after the transfer action, the handling cost generated by the transfer action, and the cumulative trigger probability of the cascading reaction after the transfer action. The lower the peak intensity, the lower the handling cost, and the lower the cumulative trigger probability of the cascading reaction, the higher the corresponding reward value.

5. A chemical inventory management system based on industrial big data according to claim 1, characterized in that: The sensing and acquisition layer also includes a laser ranging device and an RFID cargo location tag. The laser ranging device is used to collect the spatial coordinate information corresponding to the cargo location, and the RFID cargo location tag is used to record the identity information of the chemicals stored in the cargo location. The spatial coordinate information and the chemical identity information are used to support the cascade reaction knowledge graph reasoning module in extracting the list of chemicals stored in cargo locations adjacent to the risk peak point.

6. A chemical inventory management system based on industrial big data according to claim 1, characterized in that: The execution feedback layer is also used to feed back the storage location occupancy information after the automated guided vehicle completes the storage transfer task instruction to the data processing layer, so that the dynamic risk field construction module can update the dynamic risk field in the next collection cycle.

7. A chemical inventory management method based on industrial big data, characterized in that, include: Collect temperature data, gas concentration data, and vibration intensity data at multiple cargo points within the storage space; The temperature data, gas concentration data, and vibration intensity data are processed by timestamp alignment, missing value interpolation, and abnormal noise filtering to form a structured environmental state data stream; Based on the structured environmental state data stream, risk source locations are identified, risk intensity coefficients for each risk source location are calculated, and risk field intensity values ​​for multiple spatial coordinate points within the storage space are calculated based on the risk intensity coefficients and risk diffusion radius parameters, thereby obtaining a dynamic risk field covering the storage space. When the peak intensity of the dynamic risk field reaches a preset proportion of a preset critical threshold, the list of chemicals stored at the corresponding location of the risk peak and its adjacent locations is extracted. Candidate reaction paths with the chemicals in the chemical list as nodes are retrieved in the pre-constructed chemical reaction knowledge graph. The trigger condition threshold of the candidate reaction path is matched with the measured environmental parameters of the corresponding location of the dynamic risk field. The trigger probability of the successfully matched candidate reaction path is calculated and the cumulative trigger probability of the cascade reaction is obtained. The spatial intensity distribution of the dynamic risk field, the cumulative trigger probability of the cascade reaction, and the occupancy status of chemicals in the storage location are encoded into state vectors. The value of candidate transfer actions is estimated through a deep Q-network, and the transfer action to be executed is selected to generate an inventory restructuring scheduling strategy. The inventory restructuring and scheduling strategy is transformed into a transfer task instruction and sent to the automated guided vehicle for execution. The execution results are then fed back to update the dynamic risk field for the next collection cycle.

8. The chemical inventory management method based on industrial big data according to claim 7, characterized in that: The calculation of the risk intensity coefficient for each risk source location includes: multiplying the ratio of the temperature offset to the preset temperature offset benchmark value, the ratio of the gas concentration offset to the preset gas concentration offset benchmark value, and the ratio of the measured vibration intensity value to the preset vibration intensity benchmark value at the risk source location by preset temperature weighting coefficients, gas concentration weighting coefficients, and vibration weighting coefficients, respectively, and then summing them to obtain the risk intensity coefficient for that risk source location. The sum of the temperature weighting coefficient, the gas concentration weighting coefficient, and the vibration weighting coefficient is 1.

9. A chemical inventory management method based on industrial big data according to claim 7, characterized in that: Calculating the trigger probability of a successfully matched candidate response path includes: based on the historical basic trigger probability of the candidate response path, and combined with the deviation of the measured environmental parameter value corresponding to the candidate response path from the trigger condition threshold, a correction calculation is performed. The greater the deviation of the measured environmental parameter value from the trigger condition threshold, the higher the corrected trigger probability.

10. A chemical inventory management method based on industrial big data according to claim 7, characterized in that: The value estimation of the deep Q-network is determined based on a reward function, which is calculated based on the peak intensity of the dynamic risk field of the storage space after the transfer action, the handling cost generated by the transfer action, and the cumulative trigger probability of the cascade reaction after the transfer action. The lower the peak intensity, the lower the handling cost, and the lower the cumulative trigger probability of the cascade reaction, the higher the corresponding reward value.

Images