A warehouse digital collaborative management method and system based on artificial intelligence

Abstract

The present application relates to the technical field of grain depot supervision, and particularly relates to a kind of grain storehouse digital collaborative management method and system based on artificial intelligence, comprising: the multi-dimensional data collected by sensor node is preprocessed, features are extracted using lightweight federated learning algorithm, the correlation degree of features is calculated through parameter exchange mechanism, and risk data with correlation is screened out;Find out the potential link of different dimensions by mining algorithm, and determine the correlation factors, input the preprocessed multi-dimensional data into the Bayesian network model, obtain the posterior probability of single-dimensional risk event, obtain the enhanced posterior probability by total weight enhancement formula, and determine the risk level;Actor network matches the disposal strategy from the strategy knowledge base according to the risk level and type, Critic network evaluates the strategy value based on real-time resource state, feeds back to Actor network, and generates the optimal disposal strategy.The present application improves the risk judgment accuracy through multi-dimensional correlation evaluation.

Description

Technical Field

[0001] This invention relates to the field of grain depot supervision technology, and in particular to a digital collaborative management method and system for grain depots based on artificial intelligence. Background Technology

[0002] In the grain storage industry, safe grain storage and risk control are key links in ensuring food security. With the expansion of storage scale and the influence of complex environments, traditional and existing technical systems have exposed many technical problems, making it difficult to meet the needs of efficient and accurate risk management.

[0003] From the perspective of data acquisition and processing, existing technologies mostly rely on single-type sensors or manual assistance for data collection, resulting in incomplete data coverage and poor real-time performance. Furthermore, data from different sources (such as temperature, images, and water levels) exhibits significant format differences, lacking unified cleaning and fusion standards, leading to inconsistent data quality and failing to provide reliable foundational data support for risk assessment. At the risk assessment level, current technologies often focus on single-dimensional risks (such as monitoring only grain temperature or personnel violations), neglecting cross-dimensional risks such as "abnormal flood control water levels affecting grain temperature" or "improper personnel operation causing open flames." Moreover, risk probability calculations often rely on fixed thresholds or simple statistical models, making it difficult to dynamically adapt to changes in the grain storage environment, resulting in significant biases in risk level determination and a high likelihood of missed or incorrect assessments. Regarding the generation and optimization of disposal strategies, existing solutions mostly employ preset template strategies, unable to flexibly adjust according to real-time grain storage resource status (such as personnel skills and qualifications, equipment availability), and lack effective feedback mechanisms for disposal effects. After strategy execution, it is difficult to perform targeted optimization of model parameters and strategy libraries, leading to low disposal efficiency, frequent resource waste, and untimely disposal problems.

[0004] In addition, regarding data security and privacy protection, some technologies require centralized transmission and storage of data from multiple nodes, which poses a risk of data leakage. Meanwhile, the application of edge computing and distributed processing technologies faces challenges such as insufficient computing power and insufficient model lightweighting, which further restricts the implementation and application of intelligent risk management technologies. Summary of the Invention

[0005] This invention improves the accuracy of risk assessment through multi-dimensional correlation evaluation.

[0006] The technical solution proposed in this invention is: a digital collaborative management method for grain warehouses based on artificial intelligence, the method comprising:

[0007] The multidimensional data collected by the sensor nodes is preprocessed, and a lightweight federated learning algorithm is used to extract features from the preprocessed multidimensional data. The feature correlation degree between the sensor nodes is calculated through a parameter exchange mechanism to filter out risky data that are correlated.

[0008] By mining algorithms to identify potential connections in different dimensions, and combining them with correlated risk data, the correlation factors are identified. The preprocessed multidimensional data is then input into a Bayesian network model to obtain the posterior probability of a single-dimensional risk event. By combining the weight coefficients and influence functions of the correlation factors, the enhanced posterior probability is obtained through the total weight enhancement formula. The risk level is then determined based on the enhanced posterior probability.

[0009] The Actor network initially matches disposal strategies from the strategy knowledge base based on risk level and type. The Critic network evaluates the value of the strategies based on real-time resource status and feeds it back to the Actor network. The Actor network then adjusts and generates the optimal disposal strategy based on the feedback results.

[0010] Preferably, the specific process for extracting features from the preprocessed multidimensional data is as follows:

[0011] By using convolutional neural networks to locate personnel work areas and identify behavior types, and combining historical records to statistically analyze the number of violations and total frequency, behavioral weight coefficients are introduced to construct the probability characteristics of personnel work behavior violations. Based on abnormal grain temperature data, grain temperature gradients are calculated, and combined with historical abnormality records to statistically analyze the frequency of abnormalities under different gradients, introducing a normal gradient benchmark value and influence coefficient to construct the probability characteristics of abnormal grain temperature gradients. Temporal analysis of personnel location trajectory data identifies activity areas, dwell time, and paths. Combined with the functional division of grain storage areas, it determines whether there are violations such as entering restricted areas or exceeding the permitted stay time, extracting the compliance characteristics of personnel location trajectory areas. Water level data is extracted from the flood control water level deviation table, and deviation values ​​are calculated. Combined with historical flood season records, the probability characteristics of water level deviation overflow risk are extracted. The location and duration of open flames are extracted from the open flame verification data table. Combined with the distribution of flammable areas, the fire risk characteristics of open flame location and duration are extracted. Combining the node feature fusion results, multi-dimensional data associations are mined to construct a risk association network.

[0012] Preferably, the specific process for filtering out the associated risk data is as follows:

[0013] Each node processes the pre-processed single-dimensional features extracted using an asymmetric encryption algorithm, encrypts them with a public key, and sends them to the associated nodes. The receiving nodes decrypt them using their private keys. For the decrypted multi-dimensional feature parameters, core risk feature combinations are selected. The correlation strength between different risk features is calculated using a correlation calculation formula, combined with multiple sets of historical data from the same period, to obtain the correlation degree. A preset correlation degree threshold is set, and the calculated correlation degree is compared with the threshold. If the correlation degree is greater than the threshold, it is determined that the two sets of risk data are related; otherwise, it is determined that there is no correlation. Historical risk association cases stored in the dynamic knowledge graph are called to verify the initially screened associated risk data. If the current associated risk data is consistent with historical patterns, the association is determined to be valid. If there are contradictions, the data is further checked or the correlation threshold is adjusted.

[0014] Preferably, the specific process for determining the related factors is as follows:

[0015] Based on the risk node topology constructed using the dynamic knowledge graph, and combined with the effective associated risk data screened in the preprocessing stage, candidate associated factors are initially determined. For the candidate associated factors, the co-occurrence probability of associated events is calculated using historical data from the same period, and factors with a co-occurrence probability higher than a basic threshold are selected. Combining the risk feature correlation degree, if the risk feature correlation degree between the associated factor feature and the target risk feature exceeds a preset threshold, a strong correlation is confirmed between the two features, and the factor is retained; otherwise, it is removed. For the finally retained associated factors, the weight coefficient of each associated factor is calculated through historical data statistics and expert experience correction, and the weight coefficient is ensured to meet the normalization constraint.

[0016] Preferably, the specific process for determining the risk level is as follows:

[0017] The preprocessed multidimensional data is input into a Bayesian network model. Combined with a pre-defined risk prior probability and conditional probability matrix, the initial posterior probability of a single-dimensional risk is obtained. A dynamic knowledge graph is invoked, and by incorporating related factors and weight coefficients, the initial posterior probability is adjusted using a total weight enhancement formula to obtain an enhanced posterior probability. The enhanced posterior probability is compared with a threshold to determine the single-dimensional risk level. For multi-risk events with strong correlations, independent risks and associated risks are distinguished, and the comprehensive risk probability of independent and associated risks is calculated separately using different methods. If the comprehensive risk probability exceeds the escalation threshold, a cross-dimensional risk level escalation warning is triggered. The overall risk level is determined by combining the single-dimensional level and the comprehensive risk probability.

[0018] Preferably, the specific process for obtaining the optimal handling strategy is as follows:

[0019] The risk level, risk management strategy knowledge base, and grain warehouse resource status are input into the Actor network. The Actor network outputs a probability distribution of candidate strategies through a multilayer perceptron and initially matches management strategies that are suitable for the current risk and resource status from the knowledge base based on the probabilities. The Critic network calculates the strategy value based on the current status and the initially matched management strategy. Combining the immediate reward, discount factor, and expected value of the next state, it updates the value function through a temporal difference learning algorithm to obtain the value of each candidate strategy. The Actor network adjusts its network parameters based on the value of each candidate strategy fed back by the Critic network through a strategy gradient algorithm.

[0020] Preferably, after processing according to the optimal processing strategy, the model is optimized based on the processing effect. The specific process is as follows:

[0021] The system obtains the execution results of the disposal strategy and changes in risk data. It calculates the disposal efficiency index for each risk disposal using the disposal efficiency index formula, combined with the disposal time and resource input coefficients. Based on the disposal efficiency index, it updates the association strength coefficient of the risk association links in the knowledge graph and corrects the prior probability and conditional probability of the risk assessment model. It adjusts the Actor-Critic algorithm according to the execution results of the disposal strategy and updates the risk disposal strategy knowledge base according to the disposal efficiency index.

[0022] The present invention also provides an artificial intelligence-based digital collaborative management method for grain warehouses, wherein the system is used to execute the aforementioned artificial intelligence-based digital collaborative management method for grain warehouses.

[0023] The present invention also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the aforementioned artificial intelligence-based digital collaborative management method for grain warehouses.

[0024] The beneficial effects of this invention are:

[0025] 1. By mining causal links such as "flood control water level - abnormal grain temperature" and "personnel violations - open flame risk" through dynamic knowledge graphs, and combining Bayesian networks and weighted augmentation models to dynamically adjust risk probabilities, this approach breaks down dimensional barriers in risk assessment, accurately identifies potential related risks, avoids missed or misjudged risks due to single-dimensional judgments, significantly improves the comprehensiveness and accuracy of risk level assessment, provides a scientific basis for subsequent targeted measures, and reduces the probability of food security incidents caused by the superposition of related risks.

[0026] 2. Adopting an Actor-Critic reinforcement learning architecture, strategies are dynamically generated by combining real-time grain warehouse resources (personnel qualifications, equipment availability) with risk characteristics. The value of the strategies is continuously evaluated through the Critic network, realizing the transformation of disposal strategies from "passive template matching" to "active dynamic optimization". Operation steps can be flexibly adjusted according to the actual situation of resources, avoiding resource waste or insufficient disposal, greatly improving the efficiency and adaptability of strategy execution, and ensuring efficient risk management even in scenarios with limited personnel and equipment.

[0027] 3. Construct a closed loop of "data collection - risk assessment - strategy execution - effect feedback - model optimization". Quantify the effect through the disposal efficiency index, and feed back into the knowledge graph, risk model and strategy knowledge base for updating, forming a virtuous cycle of continuous iteration. As disposal cases accumulate, model parameters (such as risk association weights and algorithm learning rate) and strategy base are continuously optimized, enabling the risk management system to maintain high adaptability and accuracy in the long term, gradually reducing the cost of manual intervention, and promoting the upgrade of grain warehouse management to full automation and intelligence. Attached Figure Description

[0028] Figure 1 This is a flowchart of an artificial intelligence-based digital collaborative management method for grain warehouses according to the present invention;

[0029] Figure 2 This is a flowchart illustrating the management process of a digital collaborative management method for grain warehouses based on artificial intelligence, as described in this invention. Detailed Implementation

[0030] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0031] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0032] like Figure 1 and Figure 2 As shown, the intelligent sensor node performs preliminary cleaning on the image data collected by the camera and the temperature data collected by the temperature sensor. For the image data, image quality assessment algorithms are used, such as calculating the peak signal-to-noise ratio (PSNR) of the image, using the formula: ,in ( For the original image, For reference image, , (for image size), when At that time, images deemed blurry, distorted, or otherwise invalid are discarded; for temperature sensor data, based on the safe temperature range for different grain varieties (e.g., the safe storage temperature for wheat is generally -5℃ to 25℃), if data is collected that exceeds this range (e.g., ... -10℃ or Data at 50℃ was identified as incorrect and discarded.

[0033] A lightweight federated learning algorithm is used to perform feature extraction locally on each node. For image data, specifically targeting the work scenario inside a grain silo, a convolutional neural network (CNN) is used to extract the "personnel work behavior - violation probability" feature (the extraction process is existing technology and will not be elaborated upon). The constructed feature mapping relationship is expressed by the formula... Calculation, where Indicates the work behavior as The probability of violations (such as not following procedures, entering dangerous areas without authorization, etc.) It is a work behavior The number of violation records, It is a work behavior Total number of records The behavior weighting coefficient (set according to the degree of danger of the behavior, such as unauthorized entry into the fumigation area) A value of 1.5 indicates a common violation. (Value 1.0), and at the same time, it is necessary to... Normalization is performed to ensure that the sum of all behavior weights is 1, guaranteeing the comparability of different behavior weights. For temperature sensor data, the "grain temperature gradient - anomaly probability" feature is extracted, calculated using the following formula: ,here The grain temperature gradient is The probability of an anomaly at (unit: ℃ / h) The grain temperature gradient is The number of abnormal records, The grain temperature gradient is Total number of records This is the normal grain temperature gradient value (usually taken as 0.5℃ / h). The gradient influence coefficient (usually set to 2℃ / h) is used to constrain the degree of influence of the grain temperature gradient deviating from the normal level on the probability of anomalies.

[0034] After feature extraction is complete, cross-node data feature fusion is achieved through a secure and encrypted parameter exchange mechanism. Each node will exchange its extracted image feature parameters (as calculated above). ) and temperature characteristic parameters (such as The parameters are encrypted using an asymmetric encryption algorithm, with the public key used to encrypt the parameters before sending them to other relevant nodes. The receiving nodes decrypt the data using their private keys and then perform a fusion calculation on the feature parameters from different nodes. For example, for the features "abnormal grain temperature" and "personnel violation of operating procedures," the correlation between them is calculated. ,in and They are the first The data includes the probability of personnel violating regulations and the probability of abnormal grain temperature, where n is the number of data groups. When the correlation coefficient C is greater than a set threshold (e.g., 0.3), it is determined that these two types of risk data are correlated, thus initially screening out potentially correlated risk data.

[0035] The data processed above is organized to generate corresponding tables. The cleaned personnel operation image data table standardizes valid personnel operation image records by fields such as storage location name and operation date, and marks whether the personnel's operation behavior in the images violates regulations; the personnel positioning trajectory data combines the personnel position in the images with positioning device data, performs noise reduction and interpolation on the personnel positioning information, and records the personnel's position changes in the grain warehouse in time series form; the grain temperature anomaly data marking table marks temperature sensor data where the grain temperature exceeds the normal range (e.g., rice grain temperature exceeding 30℃ is marked as abnormal), and records information such as the warehouse where the anomaly occurred and the detection time; the grain condition anomaly type coding table encodes abnormal grain temperature and abnormal grain surface movements (…). Different types of grain condition anomalies are coded based on image recognition (e.g., grain temperature anomaly is coded as 01, grain surface anomaly is coded as 02); the flood control water level deviation table (if relevant images or sensor data are available) calculates the difference between the actual flood control water level and the safe water level threshold, and records the reservoir name, river name, water level value, etc.; the open flame verification data table is generated by performing flame feature algorithm recognition on open flame monitoring image data (e.g., extracting the shape, temperature, and other features of the red high-brightness area in the image, and excluding false alarms such as light), recording information such as the location and duration of the open flame, and finally outputting these processed and organized tabular data.

[0036] The preprocessing stage outputs a multi-dimensional monitoring data set, including cleaned personnel operation image data (fields such as employee ID, operation time, whether safety helmet was worn, and whether there were any violations), personnel location trajectory data (fields such as personnel location coordinates, timestamp, and warehouse area), grain temperature anomaly data marking table (fields such as warehouse number, detection time, grain temperature gradient, and anomaly type), flood control water level deviation table (fields such as warehouse name, detection time, actual water level, safe water level threshold, and water level deviation value), and open flame verification data table (fields such as open flame location coordinates, detection time, open flame duration, and flame intensity). These data, along with preset risk assessment model parameters (such as prior probabilities and conditional probability matrices for various risk events), are imported into a Bayesian network model to calculate the posterior probability of single-dimensional risk events.

[0037] According to Bayes' theorem ,in: In the evidence When it occurs, the event The posterior probability of occurrence; It is an event When it happens, evidence The conditional probability of occurrence; For the event The prior probability of occurrence; As evidence The marginal probability of occurrence is determined by the law of total probability. calculate( (For all possible events).

[0038] Taking the calculation of personnel violation risk as an example, let "personnel violation of work regulations" be the event. "Image recognition of violations" as evidence Prior probability Based on historical violation data, such as the frequency of personnel violations in the past year; conditional probability. When it is confirmed that personnel are indeed violating regulations, the probability that the image recognition system correctly detects the violation is determined by the historical accuracy rate of the image recognition algorithm. Marginal probability is the probability of a person being mistakenly judged as violating regulations when they are not, based on image recognition false alarm rate statistics. ,in Posterior probability This value reflects the probability that a person actually violated regulations when evidence of image violations was detected.

[0039] In the calculation of the risk of abnormal grain temperature, "abnormal grain temperature" is defined as an event. Evidence that "grain temperature gradient exceeds normal range" Prior probability It is based on the frequency of occurrence of historical grain temperature anomalies. The probability of grain temperature gradient exceeding the standard when grain temperature is abnormal is derived from the relationship between historical grain temperature anomalies and gradient data. This represents the probability of the gradient exceeding the limit when the grain temperature is normal. It is calculated using the law of total probability. Then, the posterior probability is obtained using Bayes' theorem. .

[0040] The risk calculation for higher-than-normal flood control water levels in the reservoir area is based on the assumption that "higher-than-normal flood control water levels" is an event. "Water level deviation exceeding the safety threshold" as evidence Prior probability Based on the frequency of historically high flood season water levels; This represents the probability that the deviation exceeds the safety threshold when the water level is high. This represents the probability of exceeding the deviation limit when the water level is normal. Calculate the marginal probability. posterior probability .

[0041] The risk calculation for open flames in the reservoir area is based on the scenario of "open flame causing a fire". Evidence that "the duration of open flame exceeded the danger threshold" Prior probability Determined based on the frequency of fires caused by open flames in the historical reservoir area; This represents the probability that the duration of open flame exceeds a dangerous threshold during a fire. This represents the probability of continuously exceeding a threshold without the open flame igniting a fire. It is calculated using the law of total probability. Bayes' theorem yields the posterior probability. .

[0042] After calculating the posterior probability of a single-dimensional risk event, a dynamic knowledge graph is invoked to mine cross-dimensional risk associations. The dynamic knowledge graph has constructed a topological structure containing nodes such as "personnel violations," "abnormal grain temperature," "high flood control water level," and "open flame risk," along with the relationships between these nodes. Examples of causal links include "high flood control water level → increased warehouse humidity → abnormal grain temperature" and "personnel violations → equipment failure → open flame risk." Through association mining algorithms within the graph (such as path similarity-based association analysis), potential connections between different risk dimensions can be identified, and the effectiveness of causal links can be verified based on historical data. For instance, "abnormal grain temperature" is not only related to "high flood control water level" but may also be associated with "personnel violating regulations by opening warehouse doors, leading to changes in temperature and humidity."

[0043] After determining the cross-dimensional association, use the total weight enhancement formula. The posterior probabilities of correlated risk events are weighted and enhanced. For the enhanced posterior probability of risk; This is the initial posterior probability calculated from a single dimension; For the first The weight coefficients of each related factor are calculated based on the frequency ratio of related events occurring simultaneously in historical data, and must meet the following requirements: The normalization constraint reflects the degree of influence of related factors on the target risk; Related factors The influence function is calculated based on the specific related factors, such as the water level deviation value. Influence function ( (This represents the highest historical water level deviation) and the duration of open flame. Influence function ( (This is the danger threshold).

[0044] Based on the enhanced posterior probabilities of each risk event, combined with a preset risk level classification threshold (low risk: 0.3, Medium Risk: High risk: Determine the risk level. For multiple risk events with strong correlation (such as high flood control water levels and abnormal grain temperatures occurring simultaneously with a certain probability), it is necessary to distinguish between independent risks and related risks, and use different risk superposition formulas. If the risk events are independent of each other (such as open flame risk and personnel violation risk), the risk superposition formula can be used. Calculate the overall risk probability, where To determine the number of associated risks, Enhance the posterior probability for each risk. If there is a causal relationship between risk events (such as higher than normal flood control water levels)... (Abnormal grain temperature) is assessed using a conditional probability model, with the following formula: ,in, Risk at the start of the causal chain For risk The conditional probability (extracted from a Bayesian network) is used. If the overall risk probability exceeds a set threshold (e.g., 0.8), a cross-dimensional risk level upgrade warning is triggered (e.g., raising the overall risk level of the storage area from medium to high). Finally, a risk level assessment result is generated, encompassing personnel, grain, and the storage area, clearly indicating the specific level of each risk category and its associated influencing factors (e.g., "Abnormal grain temperature - medium risk, associated factors: high flood control water level, personnel violating regulations regarding warehouse door operation"), ensuring precise alignment with subsequent response strategy generation.

[0045] For example, in a certain actual monitoring scenario, preprocessed data showed that during the work period from 9:00 to 10:00, personnel Wang Wu was detected by image recognition not to be wearing a safety helmet as required. Historical statistics showed the prior probability of personnel in this warehouse violating the rule of not wearing safety helmets. The probability that image recognition correctly detects violations. False positive probability The probability characteristic value of violation corresponding to the behavior of "not wearing a safety helmet" Calculate the marginal probability. posterior probability .

[0046] Warehouse 6 detected a grain temperature gradient at 11:00. =1.8℃ / h, the normal grain temperature gradient threshold is 1℃ / h; historical grain temperature anomaly prior probability Probability of gradient exceeding standard when grain temperature is abnormal Probability of exceeding the standard under normal circumstances ; Corresponding to the "anomaly probability" feature value Calculate the marginal probability. posterior probability .

[0047] At 13:00, the actual water level at reservoir point C was 11.5m, the safe water level threshold was 11.0m, and the water level deviation was [value missing]. m; Prior probability of historically higher flood season water levels The probability of deviation exceeding the standard when the water level is high. Probability of exceeding the standard under normal circumstances Calculate the marginal probability. posterior probability .

[0048] An open flame was detected near warehouse 3 at 14:00, and the duration was... Danger threshold Prior probability of fire caused by historical open flame Probability of open flame exceeding the threshold during a fire The probability of continuously exceeding the threshold without causing a fire Calculate the marginal probability. posterior probability .

[0049] By utilizing a dynamic knowledge graph, the correlation strength coefficient between "high flood control water level" and "abnormal grain temperature" was found. The warehouse's historical highest water level deviation Influence function The correlation strength coefficient between "personnel illegally opening 6 ventilation vents in the warehouse" and "abnormal grain temperature" Influence function (Calculated based on opening duration and ventilation impact). Substituting into the total weight enhancement formula, the enhanced posterior probability of abnormal grain temperature risk is... .

[0050] Based on the risk level thresholds: personnel violation risk (0.667) is medium risk, abnormal grain temperature risk (0.6545) is medium risk, high flood control water level risk (0.789) is high risk, and open flame risk (0.277) is low risk. Since there is a strong correlation between high flood control water level and abnormal grain temperature, a risk aggregation formula is used to calculate the overall risk probability. This triggered an upgrade of the overall risk level of the reservoir area to high risk, and the final overall risk level of the reservoir area was determined to be "high risk." The cause of the risk was that "the flood control water level at reservoir point C was too high (deviation of 0.4m), causing an increase in humidity in the warehouse, which led to abnormal grain temperature."

[0051] The risk level assessment results (including the level and cause of risks such as personnel violations, abnormal grain temperature, flood control in the storage area, and open flames), the risk management strategy knowledge base (which stores basic management strategies corresponding to different risk types and levels, such as conventional ventilation strategies for low-risk abnormal grain temperature and personnel and material allocation strategies for high-risk flood control warnings, etc., with strategies stored in a structured form, including fields such as strategy number, operation steps, required resources, and expected effects), and the grain storage resource status (real-time updated personnel on-duty status, including the number of personnel and their skills and qualifications in each position; equipment availability status, such as ventilation fan operating parameters, water pump integrity rate, and fire-fighting equipment reserves, etc., which are collected in real time through IoT devices and stored in the database) are used as input parameters and imported into the management strategy generation system based on the deep reinforcement learning Actor-Critic algorithm.

[0052] The Actor network employs a multilayer perceptron (MLP) structure, consisting of an input layer, multiple hidden layers, and an output layer. The input layer receives encoded risk information, encoding the risk level (high / medium / low) as a numerical value of 3 / 2 / 1, and concatenating the risk type (e.g., personnel violations encoded as [1,0,0,0], grain temperature abnormalities encoded as [0,1,0,0], reservoir flood control encoded as [0,0,1,0], and open flame risk encoded as [0,0,0,1]) to form the input feature vector. Assuming the hidden layer contains two layers with 16 and 8 neurons respectively, and uses the ReLU activation function (…),… Add nonlinearity to the network. Taking the risk of abnormal grain temperature (encoded as [2,0,1,0,0]) as an example, the input layer first processes this vector using the formula... Passed to the first hidden layer, where For the input feature vector, It is the weight matrix from the input layer to the first hidden layer (dimension (5,16)). It is the bias vector of the first hidden layer (dimension (16)). The output of the first hidden layer (dimension (16)); the output of the first hidden layer is then processed by the formula Passed to the second hidden layer, It is the weight matrix from the first hidden layer to the second hidden layer (dimension (16,8)). It is the bias vector of the second hidden layer (dimension (8)). The output of the second hidden layer (dimension (8)) is used; finally, the output layer is passed through the Softmax function. The output of the second hidden layer is transformed into a probability distribution of the candidate policies, where It is the unnormalized output of the output layer (dimension is ( ), (the number of candidate strategies), by Calculation yields ( It is the weight matrix from the second hidden layer to the output layer. (This is the bias vector of the output layer). Assuming there are 3 candidate strategies for the risk of abnormal grain temperature, the output layer outputs the selection probabilities of the 3 strategies as [0.3, 0.5, 0.2], and based on this, the corresponding strategies are initially matched from the strategy knowledge base.

[0053] Critic network usage state-action value function Evaluation strategy Representing the current state, it is a multi-dimensional vector that not only includes the risk level and the cause of the risk, but also integrates information on the status of grain storage resources. For example, it represents the risk status of abnormal grain temperature. [2, High flood control water level leads to increased humidity in warehouse, coupled with the illegal opening of 2 ventilation openings in warehouse, 1 maintenance personnel, 60% of the spare dehumidification equipment is in good working order, 4 ventilation fans are operating normally, and 2 are awaiting repair]; 'a' indicates the disposal strategy to be evaluated; To perform the action The immediate rewards are determined by a finely set reward mechanism based on the strategy execution results. For example, successfully starting the equipment earns +10 points, failing to start the equipment due to malfunction earns -5 points, timely personnel deployment earns +8 points, and deployment delays earn -6 points. This is a discount factor with a value of 0.9, used to measure the importance of future rewards; The next state The expected value is predicted using the Temporal Difference Learning (TD) algorithm, which performs rolling predictions based on historical data and the current state. The specific calculations employ TD (…). The algorithm, through formula renew Value, of which This is the learning rate (set to 0.1). It is time difference error. It is the eligibility trace, used to accelerate the learning process and reflect the degree to which past states and actions contribute to the current learning.

[0054] The Actor network is based on feedback from the Critic network. Values, using the policy gradient algorithm Using stochastic gradient descent (SGD), according to the formula ( (This is the learning rate, set to 0.001) Adjust the parameters along the gradient direction. Through multiple rounds of iterative optimization, the probability of selecting high-value strategies is increased, and the optimal handling strategy for each risk is finally output.

[0055] For example, based on monitoring data for a certain day, the assessment results for storage point C are as follows: the risk of abnormal grain temperature is medium, because the flood control water level at storage point C is higher than normal (deviation of 0.4m), leading to increased humidity in the storage room, coupled with the unauthorized opening of two ventilation vents in the storage room; the risk of personnel violations is medium, because safety helmets were not worn during operations; the risk of higher than normal flood control water level in the storage area is high, because the actual water level exceeds the safety threshold by 0.4m, posing a risk of overflow; the risk of open flame in the storage area is low, because an open flame was detected for 30 seconds, posing a fire hazard; the overall risk level of the storage area is high, because the higher than normal flood control water level at storage point C caused increased humidity in the storage room, leading to abnormal grain temperature, and an open flame was also detected, resulting in multiple major risks. The grain depot's resources are as follows: In terms of personnel, there is 1 person with equipment maintenance skills, 2 managers with safety training qualifications, 1 on-site safety officer, and 8 general workers; in terms of equipment, there is 1 spare dehumidifier (60% operational, with a fault warning), 6 ventilation fans (4 operating normally, 2 awaiting maintenance), 3 fire pumps (all normal), and sufficient fire extinguisher reserves.

[0056] For the risk of abnormal grain temperature, the Actor network input feature vector [2,0,1,0,0] is processed by a multilayer perceptron to obtain the candidate strategy probability distribution [0.25,0.6,0.15]. Initial matching is performed for "increase ventilation frequency," "start dehumidification equipment," and "turn the grain over," and corresponding basic strategies are extracted from the strategy knowledge base. During the Critic network evaluation, taking the strategy of "starting backup dehumidification equipment" as an example, the current state s is [2, high flood control water level leading to increased warehouse humidity, plus unauthorized opening of 2 warehouse ventilation openings, 1 maintenance worker, 60% availability of backup dehumidification equipment, 4 ventilators operating normally, 2 awaiting repair]. Assuming that the strategy fails to start due to equipment failure after execution, an immediate reward is given. Points, through TD ( Algorithm predicts the expected value of the next state =20 points, then the strategy The strategy of "increasing ventilation frequency" is expected to effectively reduce grain temperature after implementation due to the availability of ventilation fans, and will be rewarded immediately. 8 points, expected value of the next state =25 points point.

[0057] After optimization using the Actor network and policy gradient algorithm, the optimal strategy is finally output. For the medium-risk situation of abnormal grain temperature, the strategy is to prioritize contacting an external maintenance team to repair the backup dehumidifier (emergency contact: Mr. Wang, phone: 138XXXX1234, expected arrival in 1 hour), while increasing ventilation frequency to 3 times per hour, and assigning Zhang Xiaosan and Li Si to manually check and record the grain temperature in warehouse 2 every 30 minutes. For the medium-risk situation of personnel violations, the strategy is to suspend the work of those violating regulations, have Chen Liu, a qualified safety training administrator, conduct on-site training, and remotely supervise other workers through the monitoring system. For the high-risk situation of flood control in the reservoir area, the strategy is to activate the flood control emergency plan, deploy all 3 fire pumps to the dike at reservoir point C, organize 5 ordinary workers to carry sandbags to reinforce the dike, and have the on-site safety officer in charge of safety command. For the low-risk situation of open flames in the reservoir area, the strategy is to assign the two nearest workers to carry fire extinguishers to the vicinity of warehouse 5 to extinguish the open flames, and maintain real-time communication with walkie-talkies.

[0058] The generated grain temperature anomaly handling instruction (medium risk) is as follows: Task No. LW-001, Risk Type: Grain Temperature Anomaly (Medium Risk). The handling measures include contacting an external maintenance team to repair the backup dehumidifier (emergency contact: Mr. Wang, phone: 138XXXX1234), increasing the ventilation frequency to 3 times per hour, and assigning Zhang Xiaosan and Li Si to manually check and record the grain temperature in warehouse 2 every 30 minutes. The personnel / equipment to execute this instruction are the external maintenance team, the ventilation fan operator, Zhang Xiaosan, and Li Si. The time requirement is that the maintenance team arrives within 1 hour, the ventilation frequency adjustment is implemented immediately, and the grain temperature is checked every 30 minutes. Precautions include maintaining communication with the team during maintenance, and ensuring standardized operation and accurate recording during checks. The generated handling instruction is sent to relevant personnel through appropriate methods.

[0059] During the data acquisition process for tracking the execution status of the disposal strategy, when executing the task of "increasing the ventilation frequency to 3 times per hour," the PLC control system collects the fan speed in real time. If the actual speed is stable within the parameter range required for the corresponding frequency, the system updates the task status to "in execution (normal)" every 10 minutes and records the current speed and timestamp. If the speed is abnormal, the system immediately marks the task status as "in execution (abnormal)," triggers an equipment fault warning, and records the abnormal speed, time, and warning information in detail. For the task of "arranging for Zhang Xiaosan and Li Si to manually detect and record the grain temperature in warehouse 2 every 30 minutes," the personnel positioning system first confirms that Zhang Xiaosan and Li Si are in warehouse 2, and then waits for them to upload grain temperature data via handheld terminals. If valid data is received within 30 minutes, the system records the detection time, the detection personnel, and the detection data, and marks the task status as "completed (normal)." If no data is received within the time limit, the system marks the task status as "incomplete," sends a reminder message to Zhang Xiaosan, Li Si, and relevant personnel, and records the reason for incompleteness. The system ultimately integrates the execution status information of all tasks, generates complete tracking data on the execution status of disposal strategies, and achieves closed-loop management of the entire process from strategy generation to execution feedback.

[0060] The results of the disposal strategy execution (including the completion status of each risk disposal task, abnormal records during the execution process, etc.) and changes in risk data (changes in risk level and risk causes before and after disposal, such as the grain temperature gradient decreasing from 1.6℃ / h to 0.8℃ / h, and the flood control water level deviation decreasing from 0.4m to 0.2m, etc.) are used as input parameters and imported into the disposal effect feedback and model optimization system.

[0061] A quantitative evaluation index system is constructed. Taking the risk management of abnormal grain temperature as an example, the management efficiency index is calculated using the following formula: ,in, For disposal efficiency index; The risk level before disposal is determined (using a coding system of high risk = 3, medium risk = 2, and low risk = 1). Risk level after treatment; Processing time (unit: hours); This is the resource input coefficient (the coefficient increases by 0.1 for each person employed, and by 0.1 for each piece of equipment employed). For example, if the pre-treatment grain temperature gradient is 1.6℃ / h (medium risk), Two hours after treatment, the temperature dropped to 0.8℃ / h (low risk). Processing time If 2 testing personnel and 1 piece of equipment are deployed per hour, then the resource input coefficient is... Disposal efficiency index .

[0062] Based on the quantification results, the dynamic knowledge graph is invoked to update the relationships. If a relationship link "A→B" exists, its relationship strength coefficient is determined. The updated formula is: in, The updated correlation strength coefficient; The original correlation strength coefficient; This is the processing efficiency index corresponding to this related link; As an adjustment factor, its value is typically between 0.1 and 0.3. Assuming the original correlation strength coefficient for "high flood control water level → abnormal grain temperature" is... =0.55, Disposal Efficiency Index =2.22, adjustment factor =0.2, then .

[0063] Based on feedback regarding the treatment results, adjust the key parameters of the Actor-Critic algorithm. Discount factor The adjustment is based on the long-term risk stability; if the risk remains stable after the intervention... If the temperature remains stable over a period of time (e.g., the grain temperature maintains a low gradient over 24 hours), then: ,in, The updated discount factor; This is the original discount factor; The actual duration of risk stability; The base duration (e.g., 24 hours) To adjust the step size, it is generally between 0.01 and 0.05. For example, the original... =0.9, duration of risk stability =36 hours =24 hours, adjust step size =0.03, then .

[0064] Learning rate (Critic Network) and The adjustment of the (Actor network) is based on policy optimization to improve convergence speed. If in If the expected convergence is not achieved within the next iteration (e.g., no effective strategy is generated), then: , ,in, and These are the updated learning rates for the Critic network and the Actor network, respectively. and The original learning rate; The target number of iterations; This represents the actual number of iterations. To adjust the ratio, it is generally between 0.1 and 0.3. Assuming the original... =0.1, =0.001, target number of iterations =20, actual number of iterations =30, adjust the ratio =0.2, then , .

[0065] reward function The adjustments are based on the specific strategy's effectiveness. Taking the "start dehumidification equipment" strategy as an example, the equipment availability rate is... The adjusted reward rules are as follows: ,in, This is the final reward value; The original success reward; The original penalty for failure; This represents the actual equipment availability rate. The threshold (e.g., 0.8); The increase for rewards and the increase for penalties are generally between 0.2 and 0.5.

[0066] Based on changes in risk data after intervention, the parameters of the risk assessment model are adjusted. Prior probability. The update formula is: ,in, This is the updated prior probability; This is the original prior probability; This represents the actual number of times it occurred. This refers to the number of times something has occurred in history. To adjust the coefficient, it is generally between 0.1 and 0.3. For example, the prior probability of the original "personnel violation risk". =0.12, the actual number of occurrences in the current month =8 times, historical average number of occurrences per year =10 times, adjustment coefficient =0.2, then .

[0067] Conditional probability The update formula is: ,in, This is the updated conditional probability; This is the original conditional probability; and These are the association strength coefficients before and after the update of the associated link "X→Y"; The influence coefficient is generally between 0.1 and 0.3. Assuming the conditional probability of abnormal grain temperature when the water level deviation is >0.3m... =0.85, the correlation strength coefficient of "high flood control water level → abnormal grain temperature" is from =0.55 becomes =0.7942, influence coefficient =0.2, then .

[0068] In the total weight enhancement formula In, weighting coefficient Adjustments are made synchronously based on the updated results of the knowledge graph association strength.

[0069] Based on the successes and failures of this handling, the strategy knowledge base will be expanded or revised. The strategy value scoring update formula is as follows: ,in, Give a new rating; The original rating; For disposal efficiency index; This is a scoring adjustment factor, set according to the importance of the strategy, and is generally between 0.1 and 0.3.

[0070] In the handling of the situation on September 15, 2025, after implementing relevant strategies at storage point C, the grain temperature gradient decreased from 1.6℃ / h to 0.8℃ / h within 2 hours following the handling of the abnormal grain temperature (medium risk). ) downgraded to low risk ( Processing time Hourly rate: 2 testing personnel and 1 piece of equipment; resource input coefficient. Disposal efficiency index After the high-risk flood control measures were implemented, the water level deviation decreased from 0.4m to 0.2m, and the risk level was downgraded from high risk (…). Downgraded to medium risk Processing time Hourly rate, 5 workers and 3 water pumps were deployed, resource input coefficient 5 × 0.1 + 3 × 0.1 = 0.8, Disposal efficiency index =1÷(2×0.8)=0.625.

[0071] The correlation strength coefficient of "high flood control water level → abnormal grain temperature" Updated to Discount factor Because the risk remained stable within 36 hours after grain temperature treatment, the adjustment was made to Critic network learning rate and Actor Network Learning Rate Because the flood control strategy convergence speed is slow, it has been adjusted to , .

[0072] The prior probability of the risk of "abnormal grain temperature" was adjusted due to the frequent occurrence of abnormal events that month. When the water level deviation is >0.3m, the conditional probability of abnormal grain temperature is adjusted as follows: .

[0073] A new combined strategy, LW-007, for "medium-risk abnormal grain temperature and equipment failure," has been added. The strategy is to "contact an external maintenance team (arriving within 1 hour), increase ventilation frequency to 3 times per hour, and conduct manual inspections every 30 minutes." The expected effect is "grain temperature gradient decrease ≥0.5℃ / h within 2 hours." The original strategy of "activating backup dehumidification equipment" has been supplemented with the condition of "equipment integrity rate ≥80%," and its recommended priority when the equipment integrity rate is <80% has been reduced from level 1 to level 3. The basic value score of the original strategy of "increasing ventilation frequency" has been increased from 60 points to 60×(1+2.22×0.2)=60×1.444=86.64, which is rounded to 87 points.

[0074] When operators perform operations, based on the displayed risk level (high, medium, low), they contact the corresponding emergency contact. At the same time, the system sends relevant information and the optimal handling strategy to the emergency contact (the complete strategy is sent, but for the sake of a concise and clear interface, the operator only sees a simplified version of the strategy on the operation interface).

[0075] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0076] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0077] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. An artificial intelligence-based digital warehouse collaborative management method, characterized in that, The method comprises: The method comprises: Through the mining algorithm, the potential connection of different dimensions is found out, the associated risk data is combined, the associated factors are determined, the preprocessed multi-dimensional data is input into the Bayesian network model, the posterior probability of the single-dimensional risk event is obtained, the weight coefficient and the influence function of the associated factors are combined, the enhanced posterior probability is obtained through the total weight enhancement formula, and the risk level is determined according to the enhanced posterior probability; The Actor network preliminarily matches the disposal strategy from the policy knowledge base according to the risk level and type, the Critic network evaluates the policy value based on the real-time resource state and feeds back to the Actor network, and the Actor network adjusts and generates the optimal disposal strategy according to the feedback result; The specific determination process of the associated factors is as follows: Based on the risk node topology structure constructed by the dynamic knowledge graph, in combination with the effective associated risk data screened in the preprocessing stage, the candidate associated factors are preliminarily determined; for the candidate associated factors, in combination with the historical same period data, the co-occurrence probability of the associated events is calculated, and the factors with the co-occurrence probability higher than the basic threshold are screened out; in combination with the risk feature correlation degree, if the risk feature correlation degree between the associated factor features and the target risk features exceeds the preset threshold, it is confirmed that there is a strong correlation between the features of the two, the factor is retained, otherwise it is excluded; for the finally retained associated factors, the weight coefficients of the associated factors are calculated through historical data statistics and combined with expert experience correction, and it is ensured that the weight coefficients satisfy the normalization constraint; The specific determination process of the risk level is as follows: The preprocessed multi-dimensional data is input into the Bayesian network model, the initial posterior probability of the single-dimensional risk is obtained in combination with the preset risk prior probability and conditional probability matrix; the dynamic knowledge graph is called, the associated factors and the weight coefficients are combined, the initial posterior probability is adjusted in weight through the total weight enhancement formula, and the enhanced posterior probability is obtained; the enhanced posterior probability is compared with the threshold value, and the single-dimensional risk level is determined; for the multi-risk events with strong association, the independent risk and the associated risk are distinguished, and the comprehensive risk probability of the independent risk and the associated risk is calculated according to different methods; if the comprehensive risk probability exceeds the upgrade threshold, the cross-dimensional risk level upgrade warning is triggered, the single-dimensional level and the comprehensive risk probability are combined to determine the overall risk level.

2. The method according to claim 1, wherein, The specific process of extracting features from the preprocessed multi-dimensional data is as follows: The personnel operation area is located by a convolutional neural network, the behavior type is identified, the number of behavior violations and the total frequency are counted according to historical records, a behavior weight coefficient is introduced, and a personnel operation behavior violation probability feature is constructed; the grain temperature gradient is calculated according to the grain temperature abnormal data, the abnormal frequency under different gradients is counted according to historical abnormal records, a normal gradient reference value and an influence coefficient are introduced, and a grain temperature gradient abnormal probability feature is constructed; the personnel positioning trajectory data is analyzed in time sequence, the activity area, the stay time and the path are identified, whether the restricted area is violated and the stay time is exceeded is judged according to the grain warehouse area function division, and the personnel positioning trajectory area compliance feature is extracted; the water level data is extracted from the flood prevention water level deviation table, the deviation value is calculated, the historical flood period record is combined, and the water level deviation overflow risk probability feature is extracted; the open fire position and duration are extracted from the open fire verification data table, and the open fire position and duration fire risk feature is extracted according to the distribution of flammable areas; The node feature fusion results are combined, multi-dimensional data correlation is mined, and a risk correlation network is constructed.

3. The method according to claim 2, wherein, The specific process of screening the associated risk data is as follows: Each node processes the single-dimensional features extracted after preprocessing by an asymmetric encryption algorithm, encrypts using a public key and sends to the associated node, and the receiving node decrypts using a private key; for the multi-dimensional feature parameters after decryption, a core risk feature combination is selected, the correlation strength between different risk features is calculated through an association degree calculation formula, combined with multiple sets of historical feature data of the same period, and the association degree is obtained; A preset association degree threshold is compared with the calculated association degree, if the association degree is greater than the threshold, it is determined that the corresponding two sets of risk data are associated, otherwise it is determined that they are not associated; The historical risk association cases stored in the dynamic knowledge graph are called to verify the initially screened associated risk data, if the current associated risk data is consistent with the historical rules, it is determined that the association is valid, if there is a contradiction, the data is further checked or the association threshold is adjusted.

4. The method according to claim 3, wherein, The specific process of obtaining the optimal disposal strategy is as follows: The risk level, risk disposal strategy knowledge base and grain warehouse resource state are input into the Actor network, the Actor network outputs a candidate strategy probability distribution through a multi-layer perceptron, and the disposal strategy suitable for the current risk and resource state is preliminarily matched from the knowledge base according to the probability; the Critic network calculates the strategy value with the current state and the preliminarily matched disposal strategy, combines the immediate reward, the discount factor and the expected value of the next state, updates the value function through the time difference learning algorithm, and obtains the value of each candidate strategy; the Actor network adjusts the network parameters through the strategy gradient algorithm according to the value of each candidate strategy fed back by the Critic network.

5. The method according to claim 4, wherein, After disposal according to the optimal disposal strategy, the model is optimized according to the disposal effect, and the specific process is as follows: The disposal strategy execution result and the risk data change are obtained, the disposal efficiency index formula is used, the disposal time and resource input coefficient are combined, and the disposal efficiency index of each risk disposal is calculated; the association strength coefficient of the risk association link in the knowledge graph is updated based on the disposal efficiency index, and the prior probability and conditional probability of the risk assessment model are corrected; The Actor-Critic algorithm is adjusted according to the treatment policy execution result, and the risk treatment policy knowledge base is updated according to the treatment efficiency index.

6. An artificial intelligence-based digital warehouse collaborative management system, characterized in that, The system is used for executing the grain warehouse digital collaborative management method based on artificial intelligence in any one of claims 1-5.

7. A computer readable storage medium characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the grain warehouse digital collaborative management method based on artificial intelligence in any one of claims 1-5.

Images