A flour production line safety intelligent supervision system based on a cloud platform

Abstract

The present application relates to flour production line safety supervision technical field, specifically to a kind of flour production line safety intelligent supervision system based on cloud platform;It contains data acquisition module, risk assessment module, decision module and execution and repair module;System obtains production line sensor data and HMI operation log, calculates instantaneous physical risk;Its core is that the trust deficit of operator is quantified based on HMI operation log, and bypass probability of safety interlock is predicted accordingly;When instantaneous physical risk exceeds preset threshold, the system will execute final action according to whether the trust deficit exceeds its threshold;The present application includes the trust state of operator into assessment, solves the systemic failure problem caused by neglecting human factor consideration and being bypassed in traditional system.

Description

Technical Field

[0001] This invention relates to the field of flour production line safety supervision technology, specifically a cloud-based intelligent safety supervision system for flour production lines. Background Technology

[0002] In high-risk human-machine interaction environments such as flour production lines, dust explosions are a persistent and significant threat. Traditional AI safety monitoring systems tend to operate as black boxes, focusing on physical risk assessment but generally lacking consideration of the operator's trust level. When the AI ​​system issues an alarm or performs an action, the operator may choose to bypass the safety interlock due to distrust, misunderstanding, or alarm fatigue, thus deactivating it manually. This bypass behavior exposes the safety system to the risk of systemic failure. Therefore, how to solve the monitoring failure problem caused by the lack of human factors and the ease with which operators can bypass traditional black-box AI safety systems has become a major technical challenge that urgently needs to be addressed in this field. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a cloud-based intelligent safety monitoring system for flour production lines. Specifically, the technical solution of this invention includes:

[0004] The data acquisition module is used to acquire multi-dimensional sensor data vectors from production line sensors in real time, as well as HMI operation logs from the human-machine interface.

[0005] The risk assessment module is used to determine the instantaneous ignition source probability based on multi-dimensional sensor data vectors, and calculate the instantaneous physical risk by combining the preset probability of combustible material concentration exceeding the standard.

[0006] The risk assessment module is also used to quantify the operator's trust deficit based on the HMI operation log, and determine the bypass probability of the safety interlock based on the trust deficit.

[0007] The decision-making module is used to determine the final action based on the comparison results of instantaneous physical risk and preset physical risk threshold, as well as trust deficit and preset trust deficit threshold.

[0008] The execution and repair module is used to perform the final execution action. When the final execution action is a preset resilience repair action, it will simultaneously push an interpretability warning message and provide feedback to correct the trust deficit.

[0009] Preferably, the multidimensional sensor data vector includes: ambient relative humidity, material flow rate, electrostatic sensor readings, and key bearing vibration.

[0010] Preferably, the risk assessment module determines the instantaneous ignition source probability, including:

[0011] By using a pre-trained federated graph neural network, a nonlinear mapping is performed on multidimensional sensor data vectors using neural network functions to output the instantaneous ignition source probability.

[0012] The pre-trained model weights used in the neural network function are obtained by federated learning on a cloud platform using historical sensor data and simulated failure events.

[0013] Preferably, the risk assessment module calculates instantaneous physical risk, including:

[0014] Based on the event concurrency probability principle of classic failure tree analysis, the instantaneous ignition source probability is multiplied by the preset probability of combustible material concentration exceeding the standard.

[0015] Among them, the preset probability of exceeding the standard of combustible concentration is a high-value constant, which represents the high-risk steady state in which the production line operates with the dust concentration maintained in the lower explosive limit region.

[0016] Preferably, the risk assessment module quantifies the operator's trust deficit, including:

[0017] Monitor HMI operation logs to identify instances where operators repeatedly ignore AI alerts triggered by the probability of instantaneous ignition sources and label them as trust breach events;

[0018] Based on trust breach events, a pre-defined exponential decay and impulse accumulation model is applied to calculate the trust deficit.

[0019] Preferably, the risk assessment module determines the bypass probability of the safety interlock, including:

[0020] The trust deficit is mapped to a bypass probability using a pre-defined standard logistic function.

[0021] Preferably, the decision module determines the final action to be performed, including:

[0022] When the instantaneous physical risk is greater than the preset physical risk threshold and the trust deficit is not greater than the preset trust deficit threshold, the preset default optimal action will be determined as the final action to be executed.

[0023] When the instantaneous physical risk is greater than the preset physical risk threshold and the trust deficit is greater than the preset trust deficit threshold, the total explosion probability of the default optimal action is calculated through the preset resilience utility function. The total explosion probability is determined based on the expected bypass probability, which is equal to the bypass probability.

[0024] The decision module is also used to calculate the total probability of a resilient repair action, where the total probability of a repair action is determined based on an expected bypass probability that is expected to be extremely low.

[0025] The decision module is also used to compare the total explosion probability of the default optimal action with the total explosion probability of the resilience repair action, and select the action with the lower total explosion probability as the final action to be executed.

[0026] Preferably, the calculation of the total explosion probability is defined as the sum of the probabilities of two mutually exclusive cases, including: the probability that the action is performed but a physical explosion still occurs, and the probability that the action is bypassed and the bypass causes an explosion.

[0027] The conditional probability of the bypass causing the explosion is preset to 1.0.

[0028] Preferably, the toughness repair action is a predefined combination of physical actions, including:

[0029] Send a command to the ventilation and dust removal system to instantly increase the power to the preset peak value;

[0030] The command is sent to the production line control system to reduce the speed of the grinding and screening units by a preset percentage.

[0031] Preferably, the execution and repair module provides feedback to correct the trust deficit, including:

[0032] Monitor the human-machine interface to identify operator actions that do not perform bypass operations within a preset decision window, and mark them as trust restoration events;

[0033] The trust repair event is fed back to the risk assessment module, and by introducing a preset negative weight, the cumulative value of the trust deficit is proactively reduced.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. This system incorporates the operator's trust status into the security assessment model. By dynamically quantifying the trust deficit and predicting the probability of security interlocks being bypassed, it solves the problem of systemic failure caused by traditional security systems neglecting human factors and being easily bypassed by operators.

[0036] 2. Under the conflict state of high physical risk and high trust deficit, this system can make intelligent decisions through the resilience utility function, actively abandon the theoretically optimal but likely to be bypassed black box action, and instead execute the suboptimal resilience repair action with a lower total explosion probability.

[0037] 3. This system can proactively repair operator trust by executing highly perceptible resilience repair actions and simultaneously pushing interpretable warning messages; at the same time, the system monitors the operator's acceptance behavior of the action and provides feedback to correct the trust deficit, thus constructing a dynamic closed-loop management of trust destruction-resilience repair-trust recovery.

[0038] 4. This system adopts a federated graph neural network model, which can deeply explore the complex nonlinearity and spatial correlation between multi-dimensional sensor data, significantly improving the accuracy and reliability of instantaneous ignition source probability assessment, and providing a solid data foundation for subsequent trust assessment and safety decision-making. Attached Figure Description

[0039] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0040] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0042] Example 1:

[0043] Please see Figure 1 A cloud-based intelligent safety monitoring system for flour production lines includes:

[0044] The data acquisition module is used to acquire multi-dimensional sensor data vectors from production line sensors in real time, as well as HMI operation logs from the human-machine interface.

[0045] The risk assessment module is used to determine the instantaneous ignition source probability based on multi-dimensional sensor data vectors, and calculate the instantaneous physical risk by combining the preset probability of combustible material concentration exceeding the standard.

[0046] The risk assessment module is also used to quantify the operator's trust deficit based on the HMI operation log, and determine the bypass probability of the safety interlock based on the trust deficit.

[0047] The decision-making module is used to determine the final action based on the comparison results of instantaneous physical risk and preset physical risk threshold, as well as trust deficit and preset trust deficit threshold.

[0048] The execution and repair module is used to perform the final execution action. When the final execution action is a preset resilience repair action, it will simultaneously push an interpretability warning message and provide feedback to correct the trust deficit.

[0049] This embodiment provides a cloud-based intelligent safety monitoring system for flour production lines, including a data acquisition module, a risk assessment module, a decision-making module, and an execution and repair module. These modules work together to form a safety monitoring system with dynamic assessment, resilient decision-making, and closed-loop repair capabilities.

[0050] The data acquisition module aims to provide real-time, multi-dimensional on-site data input for subsequent risk assessment; in this embodiment, the module is configured to perform two main tasks.

[0051] Real-time acquisition of multi-dimensional sensor data vectors from production line sensors; in the specific technical context of this invention, this vector refers to a vector combination containing key physical parameters. The source of these sensors is physical sensors installed at key locations in flour production lines, such as grinding and screening units.

[0052] Obtain the operation logs of the human-machine interface (HMI); the purpose is to capture the interaction behavior between the operator and the AI ​​system, especially the alarm system, as an objective basis for assessing the human-machine trust relationship.

[0053] The core purpose of the risk assessment module is to quantify two coupled risks: the immediate physical explosion risk and the human-caused bypass risk due to operator distrust. In this embodiment, the function of this module is decomposed into two parallel paths.

[0054] Physical risk path; this module is based on multi-dimensional sensor data vectors acquired from the data acquisition module. Determine the instantaneous ignition source probability Furthermore, the module combines the preset probability of combustible material concentration exceeding the standard. Calculate instantaneous physical risk ;

[0055] Human factors risk path; this module is also used to quantify the operator's trust deficit based on HMI operation logs. Furthermore, the module calculates the trust deficit. Determine the bypass probability of the safety interlock. Here, safety interlock refers to hardware or software protection mechanisms used to automatically stop the production line under dangerous conditions, while bypass probability refers to the probability that the operator ignores the interlock and forcibly continues production.

[0056] The decision-making module aims to dynamically select an action that minimizes the total system risk, i.e., the total probability of explosion, based on a comprehensive assessment of physical and human-caused risks. In this embodiment, the module receives... and As input, based on instantaneous physical risk Compared with the preset physical risk threshold and the trust deficit Compared with the preset trust deficit threshold The comparison results determine the final action to be taken. ;in, This refers to the unacceptable level of physical risk; This refers to a trust deficit that is so high as to potentially lead to bypassing. Specific methods for determining this include, for example, analyzing historical safety data and defining it as events that historically led to minor safety incidents or serious near misses. The 95th percentile of the value; For example, the specific determination method is to set it as calibrated in Example 6. Value, or set to Historically, this has led to a higher probability of bypassing. Exceeding the predetermined safety value value;

[0057] The execution and repair module aims to faithfully execute the instructions of the decision-making module and proactively close the loop to repair human-machine trust when executing specific instructions, namely, resilience repair actions. In this embodiment, this module is used to execute the final execution action determined by the decision-making module. ;

[0058] To further clarify, when the final action is a preset resilience repair action... At that time, this situation occurred and When triggered, this module will execute three coupled actions;

[0059] implement Defined physical actions, such as forced ventilation and deceleration;

[0060] Simultaneously push interpretable warning messages to the HMI, such as clearly stating that the risk of static electricity has increased dramatically and that forced ventilation has been activated. The purpose is to proactively restore trust by providing clear reasons and actions.

[0061] Feedback will correct the trust deficit; the module will monitor whether the operator has accepted it. The action, i.e., without bypassing the process, will be fed back to the risk assessment module as a trust repair event to proactively reduce [risk]. The cumulative value;

[0062] This invention, through the collaborative work of the aforementioned modules, achieves an innovative closed-loop safety supervision system. It not only assesses instantaneous physical risks, but more importantly, it incorporates the operator's trust deficit as a core variable into the safety decision-making model. By dynamically quantifying the trust deficit and predicting bypass probabilities, and in high-risk, high-distrust conflict states, it utilizes a resilience utility function to abandon theoretically optimal but potentially bypassable actions, instead executing suboptimal but acceptable resilience repair actions. Simultaneously, through interpretable warnings and a trust repair closed loop, it proactively manages and repairs the trust deficit. This solves the major technical problem of traditional black-box AI safety systems failing systematically due to a lack of human factor considerations and susceptibility to operator bypass, significantly improving the robustness and practical effectiveness of safety systems in high-risk human-machine interaction environments, such as flour production line safety systems.

[0063] Example 2:

[0064] The multidimensional sensor data vector includes: ambient relative humidity, material flow rate, electrostatic sensor readings, and vibration of key bearings.

[0065] This embodiment, based on embodiment 1, specifies the specific composition of the multidimensional sensor data vector;

[0066] The multidimensional sensor data vector acquired by the data acquisition module The purpose is to comprehensively characterize the physical conditions that may lead to ignition sources, especially electrostatic discharge; in this embodiment, the vector specifically includes ambient relative humidity, material flow rate, electrostatic sensor readings, and critical bearing vibration.

[0067] The selection of this set of vectors has clear technical considerations; relative humidity is a key factor affecting static electricity accumulation, that is, dry environments are prone to generating static electricity; material flow rate is positively correlated with the rate of triboelectric charging; static electricity sensor readings are a direct measurement of static electricity accumulation; and vibration of key bearings may indicate mechanical friction or loosening, which is another potential ignition source, namely mechanical sparks.

[0068] By specifically defining these four key physical parameters, this embodiment ensures the accuracy of the data vectors input to the subsequent intelligent model. It exhibits high physical relevance; compared to broad or poorly defined sensor data, this solution can more accurately and robustly capture the core driving factors leading to electrostatic ignition sources, thereby significantly improving the probability of instantaneous ignition sources output by the risk assessment module. The accuracy and reliability of [the system / mechanism].

[0069] Example 3:

[0070] The risk assessment module determines the probability of an instantaneous ignition source, including:

[0071] By using a pre-trained federated graph neural network, a nonlinear mapping is performed on multidimensional sensor data vectors using neural network functions to output the instantaneous ignition source probability.

[0072] The pre-trained model weights used in the neural network function are obtained by federated learning on a cloud platform using historical sensor data and simulated failure events.

[0073] Based on Example 1, this embodiment limits the specific algorithm implementation for the risk assessment module to determine the instantaneous ignition source probability;

[0074] To obtain multidimensional sensor data vectors The instantaneous ignition source probability is calculated in the middle. This embodiment innovatively employs a pre-trained Federated Graph Neural Network (FedGNN) to achieve this.

[0075] Should The core function is to utilize neural network functions A nonlinear mapping is performed on the multidimensional sensor data vector to output probability values. The underlying logic of using FedGNN lies in the fact that the physical relationships between sensors on the production line, such as the flow velocity at point A affecting the electrostatic charge at point B, can be modeled as a graph structure. FedGNN can effectively capture this complex spatial dependency. In a specific implementation of this embodiment, this graph structure... Defined as: For the set of sensor nodes, i.e. The vectors include ambient relative humidity, material flow rate, electrostatic sensor readings, and vibration of key bearings. The set of edges is predefined based on the physical proximity of sensors on the production line or known physical causal relationships; The architecture employs a two-layer graph convolutional network, followed by a fully connected layer and a sigmoid activation function to output probability values ​​between 0 and 1. ;

[0076] The calculation method is as follows;

[0077] ;

[0078] in, : Represents the instantaneous ignition source probability, which is a dimensionless probability value calculated by this module;

[0079] : The sensor data vector at time t is a vector provided by the data acquisition module, and its structure is as defined in Example 2;

[0080] The weights of the pre-trained model are represented by a set of parameters, the sources of which are as follows;

[0081] : Represents the inference function of the neural network, which is the algorithm model and is preset in the module;

[0082] Should The function uses pre-trained model weights. Its acquisition method is also specifically designed: it is obtained by training historical sensor data and simulated failure events on a cloud platform using federated learning; the use of federated learning, rather than traditional centralized training, aims to ensure the model's generalization ability and data privacy. Data from individual facilities does not need to be uploaded to the cloud, thus building a more robust model that aggregates experience from multiple scenarios. The predictive model; specifically, the federated learning training uses the FedAvg aggregation algorithm; the cloud platform acts as the central server, and each brownfield facility acts as a client; each client trains its local system using its historical sensor data and simulated failure events. The model is then uploaded to the cloud platform; subsequently, the client only uploads the model weights; the cloud platform performs a weighted average of the collected client weights to update the global model weights. And will update The data is distributed to each client, and this process is repeated until the model converges; simulated failure events can be generated by injecting perturbations that conform to physical laws into historical normal data.

[0083] By adopting This solution can deeply mine the complex nonlinearities and spatial correlations between sensor data, thereby more accurately predicting the probability of ignition sources; it obtains data through federated learning. This solution significantly improves the model's generalization ability and its ability to capture rare failure events without compromising the data privacy of each production line, thus solving the technical problems of sparse training data and poor generalization of traditional single models.

[0084] Example 4:

[0085] The risk assessment module calculates instantaneous physical risk, including:

[0086] Based on the event concurrency probability principle of classic failure tree analysis, the instantaneous ignition source probability is multiplied by the preset probability of combustible material concentration exceeding the standard.

[0087] Among them, the preset probability of exceeding the standard of combustible concentration is a high-value constant, which represents the high-risk steady state in which the production line operates with the dust concentration maintained in the lower explosive limit region.

[0088] This embodiment, based on embodiment 1, calculates instantaneous physical risk for the risk assessment module. The specific mathematical principles and parameter settings were limited;

[0089] The calculation is based on the event concurrency probability principle of classic failure tree analysis (FTA), that is, the explosion risk, i.e. the top event, is simplified into the concurrence of two bottom events: the presence of an ignition source and the excessive concentration of combustibles.

[0090] In this embodiment, this principle is implemented by using the instantaneous ignition source probability calculated by the method of Embodiment 3. The probability of exceeding the preset flammable material concentration standard Perform product operations;

[0091] The calculation method is as follows;

[0092] ;

[0093] in, : Represents instantaneous physical risk, a dimensionless probability value, calculated by this module;

[0094] : Represents the instantaneous ignition source probability, which is a dimensionless probability value, calculated by the previous steps of this module, namely Example 3;

[0095] : Represents the probability of exceeding the preset combustible concentration standard, and is a high-value constant, the source of which is as follows;

[0096] The preset probability of combustible material concentration exceeding the standard here. It has a specific technical meaning: it is set as a high-value constant, for example... The value is close to 1.0; this parameter originates from the specific constraints of the application scenario of this invention, namely, characterizing the production line operating in a high-risk steady state where the dust concentration is maintained in the lower explosive limit (LEL) region; under this high-risk steady state, the concentration of combustibles exceeds the standard. This is considered a background condition that is constantly satisfied;

[0097] This embodiment will... The default setting of a high-value constant greatly simplifies the calculation model of physical risk; the key is that this setting acknowledges that the system is always in a high-risk steady state, making instantaneous physical risk... The change depends entirely on the instantaneous ignition source probability. The dynamic changes allow the entire safety supervision system to focus on the accurate assessment and management of dynamic ignition sources, which is consistent with the actual working conditions of flour production lines where dust concentration is difficult to measure in real time and accurately, and is highly feasible.

[0098] Example 5:

[0099] The risk assessment module quantifies the operator's trust deficit, including:

[0100] Monitor HMI operation logs to identify instances where operators continuously ignore AI alerts triggered by the probability of instantaneous ignition sources, and label these as trust breach events; here, continuous ignoring is specifically defined as: within a preset timeframe... Within the time window, by The cumulative number of similar AI alerts triggered Each time, the operator failed to perform a confirmation or response operation within that window;

[0101] Based on trust breach events, a pre-defined exponential decay and impulse accumulation model is applied to calculate the trust deficit.

[0102] This embodiment, based on embodiment 1, quantifies the trust deficit of operators in the risk assessment module. The specific data sources and model principles were limited;

[0103] To obtain calculation The risk assessment module monitors the HMI operation logs to identify instances where operators continuously ignore the probability of instantaneous ignition sources. The behavior of triggering AI alerts; this behavior of ignoring alerts is defined as a trust breach event in this invention. ;

[0104] After identifying Following the event, the module calculates the trust deficit based on a pre-defined exponential decay and impulse accumulation model, using the trust breach event as a basis. The technical principle of this model is as follows: That is, the trust deficit will decrease exponentially over time, simulating the natural fading of operator forgetfulness or apathy; while each event of trust breach... The occurrence of these events will lead to A pulse accumulation is generated, which simulates the instantaneous collapse of trust or the accumulation of deficit;

[0105] The calculation method is as follows;

[0106] ;

[0107] in, : Represents the trust deficit at time t, which is a cumulative value calculated by this module;

[0108] : Represents the deficit at the previous moment, is a cumulative value, and is derived from historical calculations;

[0109] The forgetting decay coefficient is a preset parameter derived from HMI log analysis to reflect the natural fading of numbness. Specific calibration methods include analyzing historical HMI logs to determine the time required for operators to return to normal response behavior after experiencing a trust breach. , Can be set to The reciprocal, for example ;

[0110] : Represents the time step, which is a unit of time and is derived from the system sampling period;

[0111] : Represents the weight of type i events, which is a preset parameter. Its source is the calibration through HMI logs. For example, continuous ignoring events have high weights. Specific calibration methods include: assigning higher weight values ​​to different types of ignoring behaviors through expert scoring or historical data regression analysis.

[0112] : Indicates whether a type i trust breach event has occurred, and is a Boolean value of 0 or 1. Its source is obtained by monitoring HMI logs;

[0113] and Specific calibration methods include, for example, analyzing historical HMI logs to determine the length of the time window during which operators begin to exhibit continuous ignoring behavior due to alarm fatigue and the number of alarms, and then setting the statistical median or 80th percentile as follows: and The value is used to ensure the objectivity of the calibration;

[0114] By monitoring HMI logs to identify trust breach events, this solution achieves objective and automated quantification of trust as a subjective psychological state. Through the use of an exponential decay and impulse accumulation model, this solution establishes a dynamic system that conforms to the laws of human memory and trust accumulation / decay. Quantization model; this enables subsequent models based on The bypass probability prediction and decision-making have a solid and dynamic input foundation.

[0115] Example 6:

[0116] The risk assessment module determines the bypass probability of safety interlocks, including:

[0117] The trust deficit is mapped to a bypass probability using a pre-defined standard logistic function.

[0118] Based on Example 1, this embodiment determines the bypass probability of safety interlocking in the risk assessment module. The specific mathematical model was defined;

[0119] To address the trust deficit in internal state variables Transformed into external behavior prediction bypass probabilities that can be used for decision-making This embodiment uses a preset standard logistic function, namely the Sigmoid function, to address the trust deficit. Mapped to bypass probability ;

[0120] The technical motivation for using the logistic function is that it can standardly describe how the probability of a binary event, i.e., bypass / non-bypass, changes with a continuous input, i.e., the trust deficit. The function exhibits S-shaped curve characteristics, consistent with the real-world logic that the lower the trust deficit, the lower the bypass probability; and the higher the trust deficit, the more the bypass probability rises sharply and approaches 1.0.

[0121] The calculation method is as follows;

[0122] ;

[0123] in, : Represents the bypass probability, which is a dimensionless probability value calculated by this module;

[0124] : Represents the trust deficit, which is a cumulative value, and its source is the calculation obtained from the previous steps of this module, namely Example 5;

[0125] The slope parameter is a preset parameter; it is derived by fitting and calibrating the trust deficit and corresponding bypass event data points in the historical HMI operation log to reflect the change in bypass probability. The intensity of the growth;

[0126] : Represents the half-probability point deficit threshold, which is a preset parameter; its source is determined by analyzing historical HMI operation logs, specifically representing the threshold corresponding to 50% of bypass operations in history. Cumulative value; and The specific calibration method is as follows: extract the dataset from the historical HMI operation logs, where each data entry contains... , Trust deficit for high-risk moments, To determine whether a bypass event occurred at that moment, a standard logistic regression algorithm is applied to fit the dataset to obtain the result. and The calibration value;

[0127] By using the standard logistic function as arrive The conversion model presented in this scheme provides a mature, interpretable, and easily calibrated approach, requiring only... and Two parameters, mathematical tools; it enables the system to be based on The quantified value is used to probabilistically predict the operator's bypass behavior, which is a key prerequisite for the subsequent decision-making module to activate the resilience utility function.

[0128] Example 7:

[0129] The decision-making module determines the final action to be taken, including:

[0130] When the instantaneous physical risk is greater than the preset physical risk threshold and the trust deficit is not greater than the preset trust deficit threshold, the preset default optimal action will be determined as the final action to be executed.

[0131] When the instantaneous physical risk is greater than the preset physical risk threshold and the trust deficit is greater than the preset trust deficit threshold, the total explosion probability of the default optimal action is calculated through the preset resilience utility function. The total explosion probability is determined based on the expected bypass probability, which is equal to the bypass probability.

[0132] The decision module is also used to calculate the total probability of a resilient repair action, where the total probability of a repair action is determined based on an expected bypass probability that is expected to be extremely low.

[0133] The decision module is also used to compare the total explosion probability of the default optimal action with the total explosion probability of the resilience repair action, and select the action with the lower total explosion probability as the final action to be executed.

[0134] This embodiment, based on embodiment 1, determines the final action to be executed by the decision-making module. The detailed decision-making logic and core algorithm, namely the resilience utility function, were defined;

[0135] The decision-making logic of the decision-making module is divided into two cases;

[0136] The first scenario, namely regular high risk: when there is instantaneous physical risk. Greater than the preset physical risk threshold And a trust deficit Not greater than the preset trust deficit threshold At that time, the system determines that the human-machine interaction is trustworthy and will perform the preset default optimal action. For example, immediately stopping the machine is determined as the final action to be taken. ;

[0137] The second scenario is a high-risk, high-distrust conflict state: when there is instantaneous physical risk. Greater than And a trust deficit Greater than When the system activates resilience decision-making, the module uses a preset resilience utility function, the core of which is to minimize the total probability of explosion. , come in and That is, choosing among resilience repair actions;

[0138] This process specifically includes:

[0139] Calculate the default optimal action Total probability of explosion The key to this calculation is that it is based on an extremely high expected bypass probability. Certainly; in this invention, It is set to be equal to the high bypass probability currently calculated by the risk assessment module, i.e., Example 6. ,because It is the black-box instructions that lead to the high trust deficit;

[0140] Calculate toughness repair action Total probability of explosion The key to this calculation is that it is based on an extremely low expected bypass probability. Determined, for example It is close to 0; this is because Designed to be explainable and intended to restore trust, it is therefore not expected to be bypassed.

[0141] The decision module is also used for comparison. and In a state of high conflict, because Extremely high and close to 1.0, while because Extremely low and close to That is, a lower residual physical risk; due to Much less than 1.0, the module selects actions with a lower total probability of explosion. As the final execution action ;

[0142] This embodiment defines... and Two thresholds clearly distinguish between two modes: conventional decision-making (i.e., good trust) and resilient decision-making (i.e., trust collapse). Its core innovation lies in the design of the resilience utility function: it does not rigidly select the physically optimal solution. Instead, through the That is, the differentiated setting of the expected bypass probability, for example extremely high Extremely low; mathematically proven, under a high trust deficit state, execution... Total risk On the contrary, it is better than performing the second-best action. Total risk This provides a quantitative, non-obvious basis for the system to abandon the optimal solution and instead execute a resilient solution, thus solving the problem of safety decision-making under human-machine conflict.

[0143] Example 8:

[0144] The total explosion probability is calculated as the sum of the probabilities of two mutually exclusive scenarios: the probability that the action is performed but a physical explosion still occurs, and the probability that the action is bypassed and the bypass results in an explosion.

[0145] The conditional probability of the bypass causing the explosion is preset to 1.0.

[0146] This embodiment, based on embodiment 7, adjusts the total explosion probability. The mathematical definition of has been further limited and clarified;

[0147] This is the core of the resilience utility function; in this embodiment, the calculation of the total explosion probability is defined as the sum of the probabilities of two mutually exclusive cases;

[0148] The calculation method is as follows;

[0149] ;

[0150] Among them, the first item For the first scenario: the action is executed, with a probability of... However, the probability of a physical explosion still occurring is that the residual physical risk is... ;

[0151] Second item For the second scenario: the action is bypassed, the probability is... Furthermore, the conditional probability of the bypass causing an explosion is: ;

[0152] : Represents the residual physical explosion probability after performing action A. It is an expected probability value, and its source is a preset value, for example... Close to 0, and Lower but not zero; The specific determination method is based on safety engineering assessment: The residual risk is set to a minimum value close to 0, representing its high physical reliability; The residual risk is assessed and determined through physical simulation or failure mode and effects analysis, and its value is below [value missing]. But significantly higher than This reflects its physical characteristics as a suboptimal solution;

[0153] : Represents the expected bypass probability after executing action A, which is the expected probability value, and its source is as shown in Example 7, i.e. Extremely high, and Extremely low;

[0154] : Represents the conditional probability of the bypass causing an explosion, which is a preset value, and its source is as follows;

[0155] To enable the model to operate under extremely high-risk conditions, i.e. The timing is crucial, and this embodiment makes a key setting for this conditional probability: the conditional probability that bypassing leads to an explosion. It is preset to 1.0; this setting The technical meaning is: in high physical risk Once activated, any bypass operation that bypasses the safety action will be considered by the system as a 100% failure that will result in an explosion.

[0156] By By explicitly defining it as the sum of two mutually exclusive probabilities, this scheme ensures the mathematical rigor of the utility function; by... With a default setting of 1.0, this scheme significantly strengthens the weight of bypass behavior in the overall risk calculation, thus shifting the focus of the decision-making module's calculations from evaluating a difficult-to-quantify factor. Shift to assessing predictability This greatly enhances the robustness of decision-making and the tendency to avoid detours.

[0157] Example 9:

[0158] The resilience repair action is a predefined combination of physical actions, including:

[0159] Send a command to the ventilation and dust removal system to instantly increase the power to the preset peak value;

[0160] The command is sent to the production line control system to reduce the speed of the grinding and screening units by a preset percentage.

[0161] This embodiment, based on Embodiment 1, focuses on the toughness repair process. The specific physical implementation was defined, and how it achieves interpretability was explained;

[0162] It is the suboptimal action selected by the decision-making module under high conflict conditions; in this embodiment, the resilience repair action is a predefined combination of physical actions. The core idea of ​​its design is to sacrifice certain economic indicators, such as energy consumption and efficiency, in exchange for the operator's perception and trust.

[0163] This combination specifically includes;

[0164] Send a command to the ventilation and dust removal system to instantly increase the power to a preset peak, such as 150% of the rated power; this action sacrifices energy consumption indicators, but its purpose is to rapidly reduce dust concentration, and it can generate a lot of wind noise, providing the operator with a strong tactile and verifiable signal to make them believe that the system is responding, which corresponds to the strong exhaust in the HMI warning.

[0165] The command is sent to the production line control system to reduce the speed of the grinding and screening units by a preset percentage, such as 30%. This action sacrifices efficiency and throughput, but its purpose is to reduce the rate of dust and static electricity generation at the source, which corresponds to the increased risk of static ignition sources in the HMI warning.

[0166] By Defined as a combination of two concrete and perceptible physical actions—forced ventilation and speed reduction—this solution enables the implementation of the abstract concept of resilience repair. This design, sacrificing economic indicators for high perceptibility, along with the interpretable warnings pushed by the execution and repair modules (such as static electricity risk and forced ventilation), forms a strong physical-information corroboration. This allows operators to understand the motivations and behaviors behind the system's decisions, thus dispelling suspicion of "black box" AI and achieving trust restoration and risk reduction. That is, the key physical basis for the expected bypass probability.

[0167] Example 10:

[0168] The execution and repair module provides feedback to correct the trust deficit, including:

[0169] Monitor the human-machine interface to identify operator actions that do not perform bypass operations within a preset decision window, and mark them as trust restoration events;

[0170] The trust repair event is fed back to the risk assessment module, and by introducing a preset negative weight, the cumulative value of the trust deficit is proactively reduced.

[0171] Based on Example 1, this embodiment further defines the closed-loop mechanism for the execution and repair module to correct the trust deficit.

[0172] In execution After the action, as defined in Example 9, is pushed and an interpretability warning is sent, the execution and repair module also needs to implement a trust closed loop.

[0173] This process monitors the human-machine interface (HMI) to identify instances where the operator fails to perform a bypass operation within a preset decision window; this failure to bypass is considered, in this invention, as the operator accepting the bypass. Warnings regarding the interpretability of actions and systems were therefore flagged as trust restoration events. ;

[0174] After identifying Subsequently, the event was reported to the risk assessment module; upon receiving the event, the risk assessment module introduced a preset negative weight. Actively reduce the trust deficit The cumulative value;

[0175] This leads to the revision of the trust deficit model in Example 5;

[0176] ;

[0177] in, : Represents the trust repair weight, which is a preset calibration parameter. Its source is also determined through historical data analysis, and the value is positive. Specific calibration methods include, for example: setting... The value makes it in one After the incident is fixed, it can be offset once. The impact of the sabotage incident, or by observing the operator during the operation. The time required to resume normal operation after an action is used for reverse calibration. Size to ensure It can fall back to within a reasonable time. the following;

[0178] : Represents the event that the repair action is accepted, a Boolean value of 0 or 1, the source of which is monitored and reported by this module;

[0179] The item is the trust repair event and In the mathematical model, the term "trust breach event" plays the opposite role, achieving the effect of... Two-way dynamic correction; among which, adopting The function is to ensure a trust deficit The calculation result is not less than 0, which meets the lower limit constraint in its physical meaning;

[0180] This embodiment defines a trust repair event. That is, accept Actions were incorporated into the trust deficit model, constructing a complete dynamic evaluation closed loop for trust destruction and trust repair; negative weights were introduced. right This solution enables proactive corrections, ensuring the system performs optimally during execution. After the action, it can enable them to trust the deficit assessment value. It was truly reduced; this avoided The system rigidity caused by values ​​only increasing and never decreasing ensured Able to dynamically fall back to Below the threshold, the system is at physical risk. After being reduced, it can automatically revert to the normal decision-making mode, i.e., trust. This achieves true resilience in human-machine systems;

[0181] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0182] It should be noted that 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 with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cloud-based intelligent safety monitoring system for flour production lines, characterized in that, include: The data acquisition module is used to acquire multi-dimensional sensor data vectors from production line sensors in real time, as well as HMI operation logs from the human-machine interface. The risk assessment module is used to determine the instantaneous ignition source probability based on multi-dimensional sensor data vectors, and calculate the instantaneous physical risk by combining the preset probability of combustible material concentration exceeding the standard. The risk assessment module is also used to quantify the operator's trust deficit based on the HMI operation log, and determine the bypass probability of the safety interlock based on the trust deficit. The decision-making module is used to determine the final action based on the comparison results of instantaneous physical risk and preset physical risk threshold, as well as trust deficit and preset trust deficit threshold. The execution and repair module is used to perform the final execution action, and when the final execution action is a preset resilience repair action, it simultaneously pushes an interpretability warning message and provides feedback to correct the trust deficit. The multidimensional sensor data vector includes: ambient relative humidity, material flow rate, electrostatic sensor readings, and key bearing vibration. The risk assessment module quantifies the operator's trust deficit, including: Monitor HMI operation logs to identify instances where operators repeatedly ignore AI alerts triggered by the probability of instantaneous ignition sources and label them as trust breach events; Based on trust breach events, a pre-defined exponential decay and impulse accumulation model is applied to calculate the trust deficit; The decision-making module determines the final action to be executed, including: When the instantaneous physical risk is greater than the preset physical risk threshold and the trust deficit is not greater than the preset trust deficit threshold, the preset default optimal action will be determined as the final action to be executed. When the instantaneous physical risk is greater than the preset physical risk threshold and the trust deficit is greater than the preset trust deficit threshold, the total explosion probability of the default optimal action is calculated through the preset resilience utility function. The total explosion probability is determined based on the expected bypass probability, which is equal to the bypass probability. The decision module is also used to calculate the total probability of a resilient repair action, where the total probability of a repair action is determined based on an expected bypass probability that is expected to be extremely low. The decision module is also used to compare the total explosion probability of the default optimal action with the total explosion probability of the resilience repair action, and select the action with the lower total explosion probability as the final action to be executed. The risk assessment module determines the instantaneous ignition source probability, including: By using a pre-trained federated graph neural network, a nonlinear mapping is performed on multidimensional sensor data vectors using neural network functions to output the instantaneous ignition source probability. The pre-trained model weights used in the neural network function are obtained by federated learning training on historical sensor data and simulated failure events on a cloud platform. The risk assessment module calculates instantaneous physical risk, including: Based on the event concurrency probability principle of classic failure tree analysis, the instantaneous ignition source probability is multiplied by the preset probability of combustible material concentration exceeding the standard. Among them, the preset probability of combustible concentration exceeding the standard is a high-value constant, which characterizes the production line operating in a high-risk steady state where the dust concentration is maintained in the lower explosive limit region; The risk assessment module determines the bypass probability of safety interlocks, including: The trust deficit is mapped to a bypass probability using a pre-defined standard logistic function.

2. The cloud-based intelligent safety monitoring system for flour production lines according to claim 1, characterized in that, The total explosion probability is calculated as the sum of the probabilities of two mutually exclusive cases, including the probability that the action is performed but a physical explosion still occurs, and the probability that the action is bypassed and the bypass causes an explosion. The conditional probability of the bypass causing the explosion is preset to 1.

0.

3. The cloud-based intelligent safety monitoring system for flour production lines according to claim 1, characterized in that, The toughness repair action is a predefined combination of physical actions, including: Send a command to the ventilation and dust removal system to instantly increase the power to the preset peak value; The command is sent to the production line control system to reduce the speed of the grinding and screening units by a preset percentage.

4. The cloud-based intelligent safety monitoring system for flour production lines according to claim 1, characterized in that, The execution and repair module provides feedback to correct the trust deficit, including: Monitor the human-machine interface to identify operator actions that do not perform bypass operations within a preset decision window, and mark them as trust restoration events; The trust repair event is fed back to the risk assessment module, and by introducing a preset negative weight, the cumulative value of the trust deficit is proactively reduced.

Images