Dynamic permission floating control method and system suitable for DCS system of thermal power plant
By using multimodal data acquisition and fusion model prediction of cognitive status, the problems of inaccurate trust assessment and insecure permission adjustment in the DCS system of thermal power plants are solved. It realizes the proactive calibration of operator trust status and secure adjustment of permissions, thereby improving the stability and security of human-machine collaboration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI THERMOELECTRIC GRP CO LTD
- Filing Date
- 2026-04-27
- Publication Date
- 2026-07-07
AI Technical Summary
The existing human-machine collaboration mode of the DCS system in thermal power plants cannot accurately assess the operator's trust in the intelligent agent, lacks an active calibration mechanism, and lacks security constraints on permission adjustment, leading to problems of excessive or insufficient trust.
A cognitive state prediction model based on multimodal human factors data collection is constructed. The model predicts cognitive state through physiological signals, operation logs and voice input data, performs trust quantification and active calibration, calculates permission coefficients, adjusts permission levels within the security envelope boundary, and optimizes the model using a self-supervised learning mechanism.
It enables accurate assessment of operator cognitive status, avoids over-trust and under-trust, ensures that permission adjustments are within security constraints, and improves the stability and security of human-machine collaboration performance.
Smart Images
Figure CN122346103A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of human-machine interaction technology in industrial control systems, specifically a dynamic permission floating control method and system applicable to DCS systems in thermal power plants. Background Technology
[0002] With the development of artificial intelligence technology, the main control room of thermal power plants is gradually evolving from manual operation to a human-machine collaborative model. The intelligent agent in a thermal power plant's DCS system refers to an intelligent software module embedded in the DCS. Its core function is to provide auxiliary decision support around the DCS. The intelligent agent can provide operational suggestions, fault diagnosis, and auxiliary decision-making, significantly improving operational efficiency. However, human-machine collaboration faces a core challenge: operators may exhibit either over-trust (leading to monitoring negligence) or under-trust (leading to untimely manual intervention), resulting in two suboptimal states.
[0003] Existing human-machine trust management technologies mainly suffer from the following problems: (1) Limited application scenarios: Existing human-machine collaborative trust calibration technology is mostly applied to dynamic environments such as autonomous driving and robot remote operation, while the main control room of thermal power plants has the characteristics of long-term steady-state monitoring, high consequences and risks, and team collaboration, making it difficult to directly transfer existing technologies.
[0004] (2) Insufficient dimensions of trust modeling: At present, the human-machine collaboration mode of DCS system in thermal power plants is mostly based on operational performance indicators or single physiological signals to build trust models, ignoring the deep human factors such as the operator's cognitive load and situational awareness, making it difficult to accurately distinguish between excessive trust and cognitive inertia.
[0005] (3) Lack of proactive calibration mechanism: Existing technologies are mostly passive detection response modes, that is, intervention is carried out after a trust anomaly occurs, and there is a lack of strategies for proactively calibrating the operator's trust level.
[0006] (4) Lack of security constraints in permission adjustment: The existing dynamic permission adjustment method does not fully consider the security envelope constraints under the critical operating conditions of the main control room of the thermal power plant, and there is a possibility of security risks due to improper permission adjustment. Summary of the Invention
[0007] The technical problem to be solved by the present invention is to provide a dynamic permission floating control method and system applicable to the DCS system of thermal power plants, which solves the technical problems in the prior art that cannot accurately assess the operator's trust status of the intelligent agent, cannot actively calibrate trust deviation, and lack security constraints for permission adjustment.
[0008] The technical solution of this invention is as follows: A dynamic permission floating control method applicable to DCS systems in thermal power plants, characterized by comprising the following steps: (1) Collect multimodal human factors data, which includes the operator's physiological signals, the operation log of the power plant DCS system, and the operator's voice input data to the power plant DCS system. (2) Construct a cognitive state prediction model based on multimodal fusion. Input multimodal human factors data into the cognitive state prediction model to predict the cognitive state. The cognitive state includes the operator's cognitive load index, situational awareness index and emotional arousal. (3) Based on the predicted cognitive state and comprehensive risk level, trust is quantified and actively calibrated to obtain the calibrated trust level; (4) Calculate the permission coefficient based on the comprehensive risk level, the calibrated trust level, and the situational awareness index; (5) The DCS system of thermal power plant maps the permission coefficient to the permission level, and performs the corresponding permission operation according to the different permission levels.
[0009] The operator's physiological signals include electrocardiogram (ECG), electrodermal conductance (EDA), photoplethysmography (PPG), respiratory signal (RESP), and dual-channel eye movement signals. The operator's physiological signals are acquired through a wearable physiological signal acquisition device. The operation log of the DCS system of the thermal power plant includes the operation type, operation object, operation value, response time, and cancellation and confirmation records recorded by the DCS system of the thermal power plant. The voice input data includes the voice input signal and the corresponding transcribed text data.
[0010] The processing steps of the cognitive state prediction model based on multimodal fusion specifically include the following steps: S21. Preprocess the multimodal human factors data, including time synchronization, missing value imputation, noise suppression, and normalization, to obtain preprocessed physiological signal input sequences, operational input sequences, and speech input sequences. Then, encode the physiological signal input sequences, operational input sequences, and speech input sequences to obtain the physiological feature vector at time t. , Operational feature vector and speech feature vectors ; S22. According to the following formula (1), the cross-attention mechanism is used to fuse multimodal features; (1); In equation (1), , and These are linear mapping matrices for query, key, and value, respectively. Represents transposition; The dimension representing the key; Represents the softmax function; This represents the query at time t; This represents the key at time t; This represents the value at time t. Represents the cross-attention fusion feature matrix; S23. The multimodal features and the cross-attention fusion feature matrix are concatenated using the following formula (2), and a unified fusion feature vector is obtained through fully connected mapping and normalization. ; (2); In equation (2), This represents a vector concatenation operation; Represents the fusion weight matrix; Represents fusion bias; Representative layer normalization operation; This represents the unified fused feature vector at time t; S24. Perform temporal modeling on the unified fused feature vectors at consecutive time points to obtain the temporal state vector of the fused features. , see the following formula (3) for details: (3); In equation (3), Represents the time series modeling module. Represents the timing window length; S25, Temporal state vector of fused features Three independent output heads are input respectively, which are used to predict the cognitive load index, situational awareness index and emotional arousal. The prediction process of the three output heads is shown in the following formulas (4)-(6): (4); In equation (4), The weight matrix representing the cognitive load output head; The bias term representing the cognitive load output head; Represents the softmax function; This represents the probability distribution of cognitive load into three categories: low, medium, and high. This represents the probability of low cognitive load. The probability of cognitive load in the middle; This represents the probability of high cognitive load; This represents the cognitive load index, with a value ranging from 0 to 1. (5); In equation (5), The weight matrix representing the context-aware output head; The bias term representing the context-aware output head; Represents the softmax function; The probability distribution represents the low, medium, and high levels of situational awareness; This represents a low probability of situational awareness. Represents the probability of contextual awareness; This represents a high probability of situational awareness; This represents the situational awareness index, with a value ranging from 0 to 1; (6); In equation (6), The weight matrix representing the emotional arousal output head; The bias term representing the output head of emotional arousal; Represents the Sigmoid activation function; Represents emotional arousal, with a value ranging from 0 to 1; Cognitive Load Index Situational awareness index and emotional arousal Composition of cognitive state .
[0011] The process of quantifying and actively calibrating trust based on the predicted cognitive state and comprehensive risk level to obtain the calibrated trust level is as follows: S31. The adoption rate of the recommendations is calculated using the following formulas (7) and (8) respectively. With human intervention rate ; (7); In equation (7), This represents the number of times the agent in the DCS system of the thermal power plant is adopted by the operator within the sliding time window corresponding to time t; This represents the total number of agent suggestions made by the DCS system of the thermal power plant within the sliding time window corresponding to time t. This represents a small constant to prevent the denominator from being zero; (8); In equation (8), This represents the number of times manual intervention, cancellation, or correction is performed within the sliding time window corresponding to time t. This represents the total number of operations performed by the agent and the human within the sliding time window corresponding to time t. This represents a small constant to prevent the denominator from being zero; S32. Calculate the real-time trust level , see the following formula (9) for details: (9); In equation (9), , , and They are respectively , , and The weighting coefficients of the four items; among which, and Adaptive adjustment is performed using the following formula (10); (10); In equation (10), and All are initial weights; Represents the cognitive load adjustment coefficient; S33, Calculate the comprehensive risk level , see the following formula (11) for details: (11); In equation (11), The risk of equipment status is calculated by the following formula (12); The risk of the representative process parameter is calculated by the following formula (13); The risk of the operating environment is calculated by the following formula (14); (12); In equation (12), This represents the number of key equipment items included in the assessment; Representing the The weight of each device satisfies ; Representing the The state risk value of a device at time t; Normalized value representing the severity of equipment failure; Normalized value representing the degree of equipment health degradation; A normalized value representing the degree of interlock protection or backup limitation; , , The weight coefficients representing the three normalized values mentioned above satisfy the following conditions: ; (13); In equation (13), To represent the total number of key process parameters included in the assessment, Representing the The weights of the key process parameters satisfy... ; Representing the The real-time values of key process parameters at time t; Representing the Safety reference values for key process parameters; Representing the The maximum allowable deviation of each key process parameter; all key process parameters are obtained in real time from the DCS system of the thermal power plant. (14); In equation (14), The total number of environmental risk factors. Representing the The weights of each environmental risk factor satisfy... ; Representing the The normalized values of each environmental risk factor at time t; S34. Calculate the optimal trust level , see the following formula (15) for details: (15); In equation (15), , , They are respectively , and The weight coefficients of the three terms, and satisfying ; S35. Calculate the trust deviation index , see the following formula (16) for details: (16); S36. First, set the symmetrical dead zone threshold. ,in, This represents the threshold for determining excessive trust. This represents the threshold for determining insufficient trust, within a certain range. Trust Bias Index The normal fluctuation range; Calculate the strength of over-trust calibration And insufficient trust calibration strength See equations (17) and (18) below for details: (17); (18); Then calculate the calibration action. , see the following formula (19) for details: (19); In equation (19), and These represent over-trust calibration gain and under-trust calibration gain, respectively; when the trust bias index... Greater than This indicates that the operator has excessive trust in the agent, and at this time, the action quantity should be calibrated. Taking a negative value is used to reduce the post-calibration trust level; when Less than When this occurs, it indicates that the operator lacks sufficient trust in the agent, and the calibration action should be performed. A positive value is used to improve the trust level after calibration; when ≤ ≤ hour, =0; Finally, calculate the post-calibration trust level. , see the following formula (20) for details: (20); In equation (20), The representative will calibrate the trust level A saturated function restricted to the interval between 0 and 1.
[0012] The calculation of the permission coefficient based on the comprehensive risk level, the calibrated trust level, and the situational awareness index is as follows: S41. Calculate the original permission coefficient using the following formula (21). : (twenty one); In equation (21), , , They are respectively , and The weight coefficients of the three terms, and satisfying ; Represents the original permission coefficient at time t; S42. Calculate the permission coefficient using the following formula (22). : (twenty two); In equation (22), This represents the permission coefficient. The clipping function is restricted to the safe envelope boundary interval, i.e. , The minimum permission boundary, or auxiliary lower bound, represents the boundary of the safety envelope. The maximum permission boundary, i.e., the security limit, represents the boundary of the security envelope. Represents the rate coefficient for adjusting permissions; Represents the permission coefficient at time t-1. This represents the permission coefficient at time t.
[0013] The minimum authority boundary of the security envelope boundary It is calculated from the following formula (23): (twenty three); In equation (23), This represents the minimum auxiliary boundary obtained based on fatigue. = , The fatigue level is represented by the following formula (24), represent Follow A monotonically increasing function that changes; This represents the minimum auxiliary boundary obtained based on cognitive load. , Represents the cognitive load index. represent Follow A monotonically increasing function that changes; Represents the minimum authority boundaries required by the protocol; This represents taking the maximum value; (twenty four); In equation (24), A temporal state vector representing the fused features; The weight matrix representing the fatigue level mapping. The bias term representing the fatigue mapping; Represents the Sigmoid activation function; The value ranges from 0 to 1, with a larger value indicating a higher level of operator fatigue. The maximum authority boundary of the security envelope boundary It is calculated from the following formula (25): (25); In equation (25), The highest authority boundary required by the representative procedure; This represents the maximum limit of basic permissions; The risk correction factor is used to compress the maximum authority limit when the risk increases, and is calculated by the following formula (26); This represents taking the minimum value; (26); In equation (26), Represents the overall risk level; Represents the risk sensitivity coefficient; with Increase This reduces the maximum permission boundary, thus automatically shrinking it.
[0014] The DCS system of the thermal power plant maps permission coefficients to permission levels, and performs corresponding permission operations according to different permission levels, specifically: S51, The continuous permission coefficients Mapped to discrete permission levels , see the following formula (27) for details: (27); In equation (27), To preset the total number of permission levels, Indicates rounding down; S52, When the permission level is 1 to When the permission level is determined to be low, the agent only has the permission to suggest displays; when the permission level is higher... and not higher than When the permission level is determined to be medium, the agent performs restricted parameter writing with human confirmation; when the permission level is greater than... and not higher than When a high-privilege level is determined, the agent performs semi-automatic or fully automatic operations within preset boundaries; among which, This represents rounding down; According to different permission levels, after performing corresponding permission operations, a self-supervised learning mechanism is used to take the changes in operator adoption rate and error rate as feedback signals to continuously optimize the weights in the cognitive state prediction model, the weights and calibration gains in the active calibration, and the parameters in the safety envelope boundary, so as to realize the continuous iterative optimization of the dynamic permission floating control method.
[0015] A dynamic authority floating control system suitable for DCS systems in thermal power plants, comprising a perception layer, an evaluation layer, a decision-making layer, and an execution and evolution layer; The perception layer is used to acquire the operator's physiological signals, the operation logs of the power plant's DCS system, and the operator's voice input data for the power plant's DCS system. The assessment layer uses a cognitive state prediction model to predict the cognitive state, and based on the predicted cognitive state and comprehensive risk level, it performs trust quantification and active calibration to obtain the calibrated trust level. The decision-making body calculates the authority coefficient based on the comprehensive risk level, the calibrated trust level, and the situational awareness index. The execution and evolution layer maps permission coefficients to permission levels, performs corresponding permission operations according to different permission levels, and continuously optimizes the weights in the cognitive state prediction model, the weights and calibration gains in trust quantification and active calibration, and the parameters in the security envelope boundary using a self-supervised learning mechanism.
[0016] Advantages of this invention: (1) The present invention constructs a cognitive state prediction model based on multimodal fusion, using multimodal human factors data as input to achieve accurate assessment of the operator's cognitive state, thereby capturing the dynamic change pattern of the operator's cognitive state over time and improving the stability and accuracy of the permission level assessment results.
[0017] (2) Based on the predicted cognitive state and comprehensive risk level, the present invention performs trust quantification and active calibration, actively calibrates the operator's trust level, avoids the problems of excessive trust and insufficient trust, and facilitates subsequent adjustment of the permission level according to the calibrated trust level.
[0018] (3) The present invention calculates the permission coefficient based on the comprehensive risk level, the calibrated trust level and the situational awareness index, and maps it to the permission level to adjust the control range and coordination strength of the intelligent agent in the DCS system of the thermal power plant. Specifically, it can be applied to control objects such as suggestion display permission, parameter writing permission, automatic execution permission, whether manual confirmation is required, single parameter adjustment range and continuous automatic operation duration.
[0019] (4) In order to ensure that the adjustment of the permission coefficient does not exceed the safety envelope boundary, the present invention sets a minimum permission boundary and a maximum permission boundary. When the operator is fatigued, has a heavy cognitive load, or when the regulations require the agent to provide the minimum assistance support, the minimum permission boundary ensures that the agent's permission is not lower than the necessary lower limit of assistance, so as to avoid insufficient assistance capability due to excessively low permission. In high-risk working conditions or regulatory restriction scenarios, the maximum permission boundary restricts the agent's permission to not exceed the current acceptable safety upper limit, so as to prevent the permission expansion from exceeding the safety boundary. Moreover, the minimum permission boundary and the maximum permission boundary under the current working condition are adaptively adjusted according to the current working condition type, equipment status and operation category, matching the corresponding constraint rules to meet the adjustment requirements of the permission coefficient under the current working condition.
[0020] (5) The present invention adopts a self-supervised learning mechanism for continuous optimization, so that the intelligent agent of the DCS system of thermal power plant can adapt to the individual differences and collaborative styles of different operators, and achieve long-term stable improvement of human-machine collaborative performance. Attached Figure Description
[0021] Figure 1 This is a flowchart of the dynamic permission floating control method of the present invention.
[0022] Figure 2 This is a block diagram illustrating the principle of the dynamic permission floating control system of this invention. Detailed Implementation
[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] See Figure 1 A dynamic permission floating control method applicable to DCS systems in thermal power plants, specifically including the following steps: (1) Collect multimodal human factors data, which includes the operator's physiological signals, the operation log of the power plant DCS system, and the operator's voice input data to the power plant DCS system. The operator's physiological signals include electrocardiogram (ECG), electrical activity of the skin (EDA), photoplethysmography (PPG), respiratory signal (RESP), and dual-channel eye movement signals. These signals are acquired using wearable physiological signal acquisition devices. Specifically: ECG uses a chest patch or lead-type ECG acquisition module to acquire one channel of ECG signal; EDA uses an electrical activity of the skin sensor module to acquire one channel of electrical activity of the skin; PPG uses a clip-on or wristband-type photoplethysmography sensor module to acquire one channel of pulse wave signal; RESP uses a breathing band or chest band-type breathing sensor module to acquire one channel of respiratory signal; and eye movement signals are acquired using a head-mounted or screen-based eye-tracking device to acquire two channels: horizontal and vertical eye movement. All six channels are synchronously acquired at a preset sampling frequency and aligned with the timestamps of the operation log and voice input data. The operation log of the DCS system of the thermal power plant includes the operation type, operation object, operation value, response time, as well as cancellation and confirmation records recorded by the DCS system of the thermal power plant; Voice input data includes voice input signals and corresponding transcribed text data; (2) Construct a cognitive state prediction model based on multimodal fusion, and input multimodal human factors data into the cognitive state prediction model to predict the cognitive state. The specific steps include: S21. Preprocess the multimodal human factors data, including time synchronization, missing value imputation, noise suppression, and normalization, to obtain preprocessed physiological signal input sequences, operational input sequences, and speech input sequences. Then, encode the physiological signal input sequences, operational input sequences, and speech input sequences to obtain the physiological feature vector at time t. , Operational feature vector and speech feature vectors ; S22. According to the following formula (1), the cross-attention mechanism is used to fuse multimodal features, and the physiological state is used to guide attention to operational behavior and speech information, thereby enhancing important features related to cognitive state. (1); In equation (1), , and These are linear mapping matrices for query, key, and value, respectively. Represents transposition; The dimension representing the key; Represents the softmax function; This represents the query at time t; This represents the key at time t; This represents the value at time t. Represents the cross-attention fusion feature matrix; S23. The multimodal features and the cross-attention fusion feature matrix are concatenated using the following formula (2), and a unified fusion feature vector is obtained through fully connected mapping and normalization. ; (2); In equation (2), This represents a vector concatenation operation; Represents the fusion weight matrix; Represents fusion bias; Representative layer normalization operation; This represents the unified fusion feature vector at time t, which has 512 dimensions. S24. To capture the dynamic changes in the operator's cognitive state over time and improve the stability and accuracy of the evaluation results, time-series modeling is performed on the unified fusion feature vectors at continuous time points to obtain the time-series state vector of the fusion features. , see the following formula (3) for details: (3); In equation (3), This represents the temporal modeling module, specifically any one of the following: Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU), or Temporal Transformer. Represents the timing window length; S25, Temporal state vector of fused features Three independent output heads are input respectively, which are used to predict the cognitive load index, situational awareness index and emotional arousal. The prediction process of the three output heads is shown in the following formulas (4)-(6): (4); In equation (4), The weight matrix representing the cognitive load output head; The bias term representing the cognitive load output head; Represents the softmax function; This represents the probability distribution of cognitive load into three categories: low, medium, and high. This represents the probability of low cognitive load. The probability of cognitive load in the middle; This represents the probability of high cognitive load; This represents the cognitive load index, with a value ranging from 0 to 1. (5); In equation (5), The weight matrix representing the context-aware output head; The bias term representing the context-aware output head; Represents the softmax function; The probability distribution represents the low, medium, and high levels of situational awareness; This represents a low probability of situational awareness. Represents the probability of contextual awareness; This represents a high probability of situational awareness; This represents the situational awareness index, with a value ranging from 0 to 1; (6); In equation (6), The weight matrix representing the emotional arousal output head; The bias term representing the output head of emotional arousal; Represents the Sigmoid activation function; Represents emotional arousal, with a value ranging from 0 to 1; Cognitive Load Index Situational awareness index and emotional arousal Composition of cognitive state ; When the cognitive state prediction model is pre-trained, The true label is obtained by having operators fill out a cognitive load assessment form after completing typical DCS monitoring and operation tasks, scoring their psychological burden, operational pressure, and attentional resource consumption during the task process; at the same time, objective performance indicators such as task completion time, number of misoperations, alarm response delay, and operation accuracy are used for auxiliary evaluation; and then the cognitive load level is reviewed and labeled in combination with the task difficulty level, working condition complexity, and operational performance, and finally the cognitive load is divided into three categories: low, medium, and high, which are used as the true labels of the cognitive load index CL. The true labels are set during the training sample collection process, including status identification questions, trend judgment questions, anomaly diagnosis questions, or situation freeze test questions related to the unit's operating conditions, to assess the operator's perception, understanding, and prediction ability of the current operating conditions; at the same time, a comprehensive evaluation is carried out by combining indicators such as fault identification accuracy, decision-making timeliness, and operational rationality during the task process; and then the operator's overall grasp of the operating conditions is judged into a level, and finally the situational awareness level is divided into three categories: low, medium, and high, which are used as the true labels of the situational awareness index SA. The true label is generated by the operator rating their tension, alertness and emotional activation after the task is completed. At the same time, a normalized arousal score is constructed by combining physiological indicators such as heart rate changes, skin conductance fluctuation amplitude, respiratory rhythm changes and eye movement activity. The above results are then fused to obtain a continuous emotional arousal level in the range of 0 to 1, which serves as the true label for emotional arousal AR. (3) Based on the predicted cognitive state and comprehensive risk level, trust is quantified and actively calibrated to obtain the calibrated trust level, specifically: S31. The adoption rate of the recommendations is calculated using the following formulas (7) and (8) respectively. With human intervention rate ; (7); In equation (7), This represents the number of times the agent in the DCS system of the thermal power plant is adopted by the operator within the sliding time window corresponding to time t; This represents the total number of agent suggestions made by the DCS system of the thermal power plant within the sliding time window corresponding to time t. This represents a small constant to prevent the denominator from being zero; Used to characterize the degree to which an operator relies on the agent's recommendations; (8); In equation (8), This represents the number of times manual intervention, cancellation, or correction is performed within the sliding time window corresponding to time t. This represents the total number of operations performed by the agent and the human within the sliding time window corresponding to time t. This represents a small constant to prevent the denominator from being zero; Used to characterize the frequency with which an operator actively intervenes or corrects the behavior of an agent; S32. Calculate the real-time trust level , see the following formula (9) for details: (9); In equation (9), , , and They are respectively , , and The weighting coefficients of the four items; among which, and Adaptive adjustment is performed using the following formula (10); (10); In equation (10), and All are initial weights; Represents the cognitive load adjustment coefficient; When the operator's cognitive load index When it is too high, the adoption rate is recommended. weight Decrease, while the rate of human intervention weight This enhances trust assessment by demonstrating the importance of human control behavior under high cognitive load. S33, Calculate the comprehensive risk level , see the following formula (11) for details: (11); In equation (11), The risk of equipment status is used to characterize the current equipment's failure, degradation, and protection limitations, and is calculated by the following formula (12); The risk of the process parameter is used to characterize the degree to which the key process parameter deviates from the safety reference range, and is calculated by the following formula (13); Represents the risk of the operating environment, used to characterize the impact of the current human-computer interaction environment and work context on operational safety, and is calculated by the following formula (14); (12); In equation (12), This represents the number of key equipment items included in the assessment; Representing the The weight of each device satisfies ; Representing the The state risk value of a device at time t; Normalized value representing the severity of equipment failure; Normalized value representing the degree of equipment health degradation; A normalized value representing the degree of interlock protection or backup limitation; , , The weight coefficients representing the three normalized values mentioned above satisfy the following conditions: ; (13); In equation (13), To represent the total number of key process parameters included in the assessment, Representing the The weights of the key process parameters satisfy... ; Representing the The real-time values of key process parameters at time t; Representing the Safety reference values for key process parameters; Representing the The maximum allowable deviation of key process parameters; key process parameters include, but are not limited to, unit load, main steam pressure, main steam temperature, feedwater flow rate, furnace negative pressure and steam drum water level, all of which are obtained in real time by the DCS system of the thermal power plant. (14); In equation (14), The total number of environmental risk factors. Representing the The weights of each environmental risk factor satisfy... ; Representing the The normalized values of environmental risk factors at time t; environmental risk factors include, but are not limited to, alarm flood intensity, information congestion level, shift handover complexity, on-site noise interference level, and current task switching frequency, which are obtained by combining HMI logs, alarm system, shift operation records and environmental data collection information; S34. Calculate the optimal trust level , see the following formula (15) for details: (15); In equation (15), , , They are respectively , and The weight coefficients of the three terms, and satisfying ; S35. Calculate the trust deviation index , see the following formula (16) for details: (16); S36. First, set the symmetrical dead zone threshold. ,in, This represents the threshold for determining excessive trust. This represents the threshold for determining insufficient trust, within a certain range. Trust Bias Index The normal fluctuation range; Calculate the strength of over-trust calibration And insufficient trust calibration strength See equations (17) and (18) below for details: (17); (18); Then calculate the calibration action. , see the following formula (19) for details: (19); In equation (19), and These represent over-trust calibration gain and under-trust calibration gain, respectively; when the trust bias index... Greater than This indicates that the operator has excessive trust in the agent, and at this time, the action quantity should be calibrated. Taking a negative value is used to reduce the post-calibration trust level; when Less than When this occurs, it indicates that the operator lacks sufficient trust in the agent, and the calibration action should be performed. A positive value is used to improve the trust level after calibration; when ≤ ≤ hour, =0; Finally, calculate the post-calibration trust level. , specifically see the following formula (20: (20); In equation (20), The representative will calibrate the trust level A saturated function restricted to the interval between 0 and 1; (4) Calculate the permission coefficient based on the comprehensive risk level, the calibrated trust level, and the situational awareness index, specifically as follows: S41. Calculate the original permission coefficient using the following formula (21). : (twenty one); In equation (21), , , They are respectively , and The weight coefficients of the three terms, and satisfying ; Represents the original permission coefficient at time t; S42. Calculate the permission coefficient using the following formula (22). : (twenty two); In equation (22), This represents the permission coefficient. The clipping function is restricted to the safe envelope boundary interval, i.e. , The minimum permission boundary, or auxiliary lower bound, represents the boundary of the safety envelope. The maximum permission boundary, i.e., the security limit, represents the boundary of the security envelope. Represents the rate coefficient for adjusting permissions; Represents the permission coefficient at time t-1. This represents the permission coefficient at time t; Among them, the minimum authority boundary of the security envelope boundary. It is calculated from the following formula (23): (twenty three); In equation (23), This represents the minimum auxiliary boundary obtained based on fatigue. = , The fatigue level is represented by the following formula (24), represent Follow A monotonically increasing function that changes; This represents the minimum auxiliary boundary obtained based on cognitive load. , Represents the cognitive load index. represent Follow A monotonically increasing function that changes; Represents the minimum authority boundaries required by the protocol; This represents taking the maximum value; (twenty four); In equation (24), A temporal state vector representing the fused features; The weight matrix representing the fatigue level mapping. The bias term representing the fatigue mapping; Represents the Sigmoid activation function; The value ranges from 0 to 1, with a larger value indicating a higher level of operator fatigue. Maximum authority boundary of the security envelope boundary It is calculated from the following formula (25): (25); In equation (25), The highest authority boundary required by the representative procedure; This represents the maximum limit of basic permissions; The risk correction factor is used to compress the maximum authority limit when the risk increases, and is calculated by the following formula (26); This represents taking the minimum value; (26); In equation (26), Represents the overall risk level; Represents the risk sensitivity coefficient; with Increase Reduce this, thereby automatically shrinking the maximum permission boundary; and The regulations and constraints are determined by the constructed regulations and constraints library, which includes thermal power unit operation regulations, DCS operation authority specifications, key operating condition dual-person review requirements, and items prohibited from automatic execution in accident operating conditions. (5) The DCS system of the thermal power plant maps the permission coefficient to the permission level, and performs the corresponding permission operation according to different permission levels, specifically: S51, The continuous permission coefficients Mapped to discrete permission levels , see the following formula (27) for details: (27) Equation (27), To preset the total number of permission levels, Indicates rounding down; S52, When the permission level is 1 to When the permission level is determined to be low, the agent only has the permission to suggest displays; when the permission level is higher... and not higher than When the permission level is determined to be medium, the agent performs restricted parameter writing with human confirmation; when the permission level is greater than... and not higher than When the privilege level is determined to be high, the intelligent agent will perform semi-automatic or fully automatic operations within the preset boundaries. (6) The self-supervised learning mechanism is adopted to use the changes in operator adoption rate and error rate as feedback signals to continuously optimize the weights in the cognitive state prediction model, the weights and calibration gains in trust quantification and active calibration, and the parameters in the safety envelope boundary, so as to realize the continuous iterative optimization of the dynamic permission floating control method.
[0025] The self-supervised learning mechanism specifically involves statistically analyzing changes in the operator's adoption rate of suggestions from the agent and the corresponding error rate within a preset sliding time window. An increase in adoption rate and a decrease in error rate are considered positive feedback, while a decrease in adoption rate and an increase in error rate are considered negative feedback. Based on these positive or negative feedbacks, a collaborative effect feedback index is constructed to characterize the quality of human-machine collaboration under the current parameter configuration. When the collaborative effect feedback index indicates an improvement in the current collaborative effect, it suggests that the current cognitive state prediction model, trust quantification and active calibration mechanism, and safety envelope boundary parameter settings are relatively reasonable, and the system maintains the current parameter configuration or enhances effective parameters. When the collaborative effect feedback index indicates a decrease in the current collaborative effect, it indicates that the current parameter configuration does not adequately match the operator's actual state or current working conditions. The system updates the weight parameters in the cognitive state prediction model, the weight parameters in trust quantification, the calibration gain in active calibration, and the parameters in the safety envelope boundary to ensure that the subsequently output permission coefficients and permission levels better match the operator's actual cognitive state, trust state, and risk conditions.
[0026] See Figure 2 A dynamic authority floating control system suitable for DCS systems in thermal power plants includes a perception layer 1, an evaluation layer 2, a decision-making layer 3, and an execution and evolution layer 4. The perception layer 1 is used to acquire the operator's physiological signals, the operation log of the power plant's DCS system, and the operator's voice input data for the power plant's DCS system. The assessment layer 2 uses a cognitive state prediction model to predict the cognitive state, and based on the predicted cognitive state and comprehensive risk level, it performs trust quantification and active calibration to obtain the calibrated trust level. The decision-making level 3 calculates the authority coefficient based on the comprehensive risk level, the calibrated trust level, and the situational awareness index. The execution and evolution layer 4 maps the permission coefficients to permission levels. Based on different permission levels, it performs corresponding permission operations and uses a self-supervised learning mechanism to continuously optimize the weights in the cognitive state prediction model, the weights and calibration gains in trust quantification and active calibration, and the parameters in the safety envelope boundary.
[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A dynamic permission floating control method applicable to DCS systems in thermal power plants, characterized in that: Specifically, it includes the following steps: (1) Collect multimodal human factors data, which includes the operator's physiological signals, the operation log of the power plant DCS system, and the operator's voice input data to the power plant DCS system. (2) Construct a cognitive state prediction model based on multimodal fusion. Input multimodal human factors data into the cognitive state prediction model to predict the cognitive state. The cognitive state includes the operator's cognitive load index, situational awareness index and emotional arousal. (3) Based on the predicted cognitive state and comprehensive risk level, trust is quantified and actively calibrated to obtain the calibrated trust level; (4) Calculate the permission coefficient based on the comprehensive risk level, the calibrated trust level, and the situational awareness index; (5) The DCS system of thermal power plant maps the permission coefficient to the permission level, and performs the corresponding permission operation according to the different permission levels.
2. The dynamic permission floating control method for DCS systems in thermal power plants according to claim 1, characterized in that: The operator's physiological signals include electrocardiogram (ECG), electrodermal conductance (EDA), photoplethysmography (PPG), respiratory signal (RESP), and dual-channel eye movement signals. The operator's physiological signals are acquired through a wearable physiological signal acquisition device. The operation log of the DCS system of the thermal power plant includes the operation type, operation object, operation value, response time, and cancellation and confirmation records recorded by the DCS system of the thermal power plant. The voice input data includes the voice input signal and the corresponding transcribed text data.
3. The dynamic permission floating control method for DCS systems in thermal power plants according to claim 1, characterized in that: The processing steps of the cognitive state prediction model based on multimodal fusion specifically include the following steps: S21. Preprocess the multimodal human factors data, including time synchronization, missing value imputation, noise suppression, and normalization, to obtain preprocessed physiological signal input sequences, operational input sequences, and speech input sequences. Then, encode the physiological signal input sequences, operational input sequences, and speech input sequences to obtain the physiological feature vector at time t. , Operational feature vector and speech feature vectors ; S22. According to the following formula (1), the cross-attention mechanism is used to fuse multimodal features; （1）； In equation (1), , and These are linear mapping matrices for query, key, and value, respectively. Represents transposition; The dimension representing the key; Represents the softmax function; This represents the query at time t; This represents the key at time t; This represents the value at time t. Represents the cross-attention fusion feature matrix; S23. The multimodal features and the cross-attention fusion feature matrix are concatenated using the following formula (2), and a unified fusion feature vector is obtained through fully connected mapping and normalization. ; （2）； In equation (2), This represents a vector concatenation operation; Represents the fusion weight matrix; Represents fusion bias; Representative layer normalization operation; This represents the unified fused feature vector at time t; S24. Perform temporal modeling on the unified fused feature vectors at consecutive time points to obtain the temporal state vector of the fused features. , see the following formula (3) for details: （3）； In equation (3), Represents the time series modeling module. Represents the timing window length; S25, Temporal state vector of fused features Three independent output heads are input respectively, which are used to predict the cognitive load index, situational awareness index and emotional arousal. The prediction process of the three output heads is shown in the following formulas (4)-(6): （4）； In equation (4), The weight matrix representing the cognitive load output head; The bias term representing the cognitive load output head; Represents the softmax function; This represents the probability distribution of cognitive load into three categories: low, medium, and high. This represents the probability of low cognitive load. The probability of cognitive load in the middle; This represents the probability of high cognitive load; This represents the cognitive load index, with a value ranging from 0 to 1. （5）； In equation (5), The weight matrix representing the context-aware output head; The bias term representing the context-aware output head; Represents the softmax function; The probability distribution represents the low, medium, and high levels of situational awareness; This represents a low probability of situational awareness; Represents the probability of contextual awareness; This represents a high probability of situational awareness; This represents the situational awareness index, with a value ranging from 0 to 1. （6）； In equation (6), The weight matrix representing the emotional arousal output head; The bias term representing the output head of emotional arousal; Represents the Sigmoid activation function; Represents emotional arousal, with a value ranging from 0 to 1; Cognitive Load Index Situational awareness index and emotional arousal Composition of cognitive state .
4. The dynamic permission floating control method for DCS systems in thermal power plants according to claim 3, characterized in that: The process of quantifying and actively calibrating trust based on the predicted cognitive state and comprehensive risk level to obtain the calibrated trust level is as follows: S31. The adoption rate of the recommendations is calculated using the following formulas (7) and (8) respectively. With human intervention rate ; （7）； In equation (7), This represents the number of times the agent in the DCS system of the thermal power plant is adopted by the operator within the sliding time window corresponding to time t; This represents the total number of agent suggestions made by the DCS system of the thermal power plant within the sliding time window corresponding to time t. This represents a small constant to prevent the denominator from being zero; （8）； In equation (8), This represents the number of times manual intervention, reversal, or correction occurs within the sliding time window corresponding to time t. This represents the total number of operations performed by the agent and the human within the sliding time window corresponding to time t. This represents a small constant to prevent the denominator from being zero; S32. Calculate the real-time trust level , see the following formula (9) for details: （9）； In equation (9), , , and They are respectively , , and The weighting coefficients of the four items; among which, and Adaptive adjustment is performed using the following formula (10); （10）； In equation (10), and All are initial weights; Represents the cognitive load adjustment coefficient; S33, Calculate the comprehensive risk level , see the following formula (11) for details: （11）； In equation (11), The risk of equipment status is calculated by the following formula (12); The risk of the representative process parameter is calculated by the following formula (13); The risk of the operating environment is calculated by the following formula (14); （12）； In equation (12), This represents the number of key equipment items included in the assessment; Representing the The weight of each device satisfies ; Representing the The state risk value of a device at time t; Normalized value representing the severity of equipment failure; Normalized value representing the degree of equipment health degradation; A normalized value representing the degree of interlock protection or backup limitation; , , The weight coefficients representing the three normalized values mentioned above satisfy the following conditions: ; （13）； In equation (13), To represent the total number of key process parameters included in the assessment, Representing the The weights of the key process parameters satisfy... ; Representing the The real-time values of key process parameters at time t; Representing the Safety reference values for key process parameters; Representing the The maximum allowable deviation of each key process parameter; all key process parameters are obtained in real time from the DCS system of the thermal power plant. （14）； In equation (14), The total number of environmental risk factors. Representing the The weights of each environmental risk factor satisfy... ; Representing the The normalized values of each environmental risk factor at time t; S34. Calculate the optimal trust level , see the following formula (15) for details: （15）； In equation (15), , , They are respectively , and The weight coefficients of the three terms, and satisfying ; S35. Calculate the trust deviation index , see the following formula (16) for details: （16）； S36. First, set the symmetrical dead zone threshold. ,in, This represents the threshold for determining excessive trust. This represents the threshold for determining insufficient trust, within a certain range. Trust Bias Index The normal fluctuation range; Calculate the strength of over-trust calibration And insufficient trust calibration strength See equations (17) and (18) below for details: （17）； （18）； Then calculate the calibration action. , see the following formula (19) for details: （19）； In equation (19), and These represent over-trust calibration gain and under-trust calibration gain, respectively; when the trust bias index... Greater than This indicates that the operator has excessive trust in the agent, and at this time, the action quantity should be calibrated. Taking a negative value is used to reduce the post-calibration trust level; when Less than When this occurs, it indicates that the operator lacks sufficient trust in the agent, and the calibration action should be performed. A positive value is used to improve the trust level after calibration; when ≤ ≤ hour, =0; Finally, calculate the calibrated trust level. , see the following formula (20) for details: （20）； In equation (20), The representative will calibrate the trust level A saturated function restricted to the interval between 0 and 1.
5. The dynamic permission floating control method for DCS systems in thermal power plants according to claim 4, characterized in that: The calculation of the permission coefficient based on the comprehensive risk level, the calibrated trust level, and the situational awareness index is as follows: S41. Calculate the original permission coefficient using the following formula (21). : （21）； In equation (21), , , They are respectively , and The weight coefficients of the three terms, and satisfying ; Represents the original permission coefficient at time t; S42. Calculate the permission coefficient using the following formula (22). : （22）； In equation (22), This represents the permission coefficient. The clipping function is restricted to the safe envelope boundary interval, i.e. , The minimum permission boundary, or auxiliary lower bound, represents the boundary of the safety envelope. The maximum permission boundary, i.e., the security limit, represents the boundary of the security envelope. Represents the rate coefficient for adjusting permissions; Represents the permission coefficient at time t-1. This represents the permission coefficient at time t.
6. The dynamic permission floating control method for DCS systems in thermal power plants according to claim 5, characterized in that: The minimum authority boundary of the security envelope boundary It is calculated from the following formula (23): （23）； In equation (23), This represents the minimum auxiliary boundary obtained based on fatigue. = , The fatigue level is represented by the following formula (24), represent Follow A monotonically increasing function that changes; This represents the minimum auxiliary boundary obtained based on cognitive load. , Represents the cognitive load index. represent Follow A monotonically increasing function that changes; Represents the minimum authority boundaries required by the protocol; This represents taking the maximum value; （24）； In equation (24), A temporal state vector representing the fused features; The weight matrix representing the fatigue level mapping. The bias term representing the fatigue mapping; Represents the Sigmoid activation function; The value ranges from 0 to 1, with a larger value indicating a higher level of operator fatigue. The maximum authority boundary of the security envelope boundary It is calculated from the following formula (25): （25）； In equation (25), The highest authority boundary required by the representative procedure; This represents the maximum limit of basic permissions; The risk correction factor is used to compress the maximum authority limit when the risk increases, and is calculated by the following formula (26); This represents taking the minimum value; （26）； In equation (26), Represents the overall risk level; Represents the risk sensitivity coefficient; along with Increase This reduces the maximum permission boundary, thus automatically shrinking it.
7. The dynamic permission floating control method for DCS systems in thermal power plants according to claim 5, characterized in that: The DCS system of the thermal power plant maps permission coefficients to permission levels, and performs corresponding permission operations according to different permission levels, specifically: S51, The continuous permission coefficients Mapped to discrete permission levels , see the following formula (27) for details: （27）； In equation (27), To preset the total number of permission levels, Indicates rounding down; S52, When the permission level is 1 to When the permission level is determined to be low, the agent only has the permission to suggest displays; when the permission level is higher... and not higher than When the permission level is determined to be medium, the agent performs restricted parameter writing with human confirmation; when the permission level is greater than... and not higher than When the privilege level is determined to be high, the intelligent agent will perform semi-automatic or fully automatic operations within the preset boundaries.
8. The dynamic permission floating control method for DCS systems in thermal power plants according to claim 5, characterized in that: According to different permission levels, after performing corresponding permission operations, a self-supervised learning mechanism is used to take the changes in operator adoption rate and error rate as feedback signals to continuously optimize the weights in the cognitive state prediction model, the weights and calibration gains in the active calibration, and the parameters in the safety envelope boundary, so as to realize the continuous iterative optimization of the dynamic permission floating control method.
9. A system for implementing the dynamic permission floating control method for a DCS system in a thermal power plant as described in claim 8, characterized in that: It includes a perception layer, an evaluation layer, a decision-making layer, and an execution and evolution layer; The perception layer is used to acquire the operator's physiological signals, the operation logs of the power plant's DCS system, and the operator's voice input data for the power plant's DCS system. The assessment layer uses a cognitive state prediction model to predict the cognitive state, and based on the predicted cognitive state and comprehensive risk level, it performs trust quantification and active calibration to obtain the calibrated trust level. The decision-making body calculates the authority coefficient based on the comprehensive risk level, the calibrated trust level, and the situational awareness index. The execution and evolution layer maps permission coefficients to permission levels, performs corresponding permission operations according to different permission levels, and continuously optimizes the weights in the cognitive state prediction model, the weights and calibration gains in trust quantification and active calibration, and the parameters in the security envelope boundary using a self-supervised learning mechanism.