Carbon capture process intelligent diagnosis control method and system based on AI decision

Abstract

The invention discloses an AI decision-based carbon capture process intelligent diagnosis control method and system, and particularly relates to the technical field of industrial process intelligent control, and the method comprises the following steps: obtaining control intentions of an edge calculation side and a cloud side, and aligning the control intentions to generate a synchronous control sequence; calculating six types of conflict factors to form a standardized conflict factor vector; generating a limited adaptive weight vector in combination with the data freshness, the source credibility and the conflict continuity; calculating an uncalibrated risk value through bounded gain aggregation; determining a coupling risk coefficient according to the uncalibrated risk-coupling risk comparison table and outputting a grading processing suggestion; according to the method, six types of conflict factors are calculated through time alignment and unit unification, so that the conflict identification precision in the carbon capture process is improved; by generating a limited adaptive weight, the risk assessment dynamic adaptability is realized; an uncalibrated risk value is generated based on bounded gain polymerization, the performability of risk quantification and regulation is enhanced, and safe and stable operation of the carbon capture process is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for industrial processes, and more specifically, to an intelligent diagnostic control method and system for carbon capture processes based on AI decision-making. Background Technology

[0002] The carbon capture process involves multiple core units, including absorption towers, desorption towers, regeneration heat exchangers, solvent circulation pumps, compressors, and steam supply systems. These units interact with multi-dimensional process variables such as temperature, pressure, flow rate, and concentration, exhibiting significant nonlinearity and strong coupling characteristics. In practical industrial scenarios, to reduce energy consumption, ensure process stability, and improve carbon dioxide capture rates, enterprises are increasingly adopting edge computing and cloud-based optimized collaborative operation architectures. The edge computing side handles local rapid response control, while the cloud side executes decision-making instructions based on historical data and global optimization strategies. Theoretically, this collaborative operation can balance local stability and global economic efficiency; however, it also introduces more complex interactions and conflict risks at the control level, demanding higher levels of intelligence in diagnosis and regulation.

[0003] Existing carbon capture process control methods primarily rely on local controllers or single optimization platforms for decision-making, lacking a synchronous comparison mechanism between the edge and cloud, and failing to establish a unified time benchmark and unit of measurement system. Due to the lack of data alignment and metadata supplementation, data collected at the edge often suffers from communication delays, time drift, and a high proportion of missing data, making it difficult to directly use for cloud-based decision-making, resulting in biases in the basis of control decisions. Furthermore, traditional conflict identification methods typically only judge one dimension, such as whether the adjustment direction is consistent, failing to comprehensively quantify multiple factors such as direction reversal, amplitude asymmetry, phase mismatch, energy impact, sensitivity amplification, and boundary approach, thus failing to fully reflect the essence of control conflicts. Further, existing risk assessment mechanisms lack the ability to dynamically weight data freshness, source reliability, and conflict persistence, making it difficult to provide effective early warnings and grading suggestions in the early stages of control conflicts. Ultimately, intervention relies on human experience, leading to delayed control responses and potentially causing energy waste, equipment oscillations, or even safety accidents. Therefore, this paper proposes an AI-based intelligent diagnostic control method and system for carbon capture processes to address the above problems. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent diagnostic control method for carbon capture processes based on AI decision-making, comprising the following steps: Step 1: Acquire control intent data from the edge computing side and the cloud side within the same time period, align the data time reference of the control intent and unify the units of measurement, and generate a synchronous control sequence for comparative analysis; Step 2: Calculate six types of conflict factors based on the synchronous control sequence. The six types of conflict factors include: direction reversal, amplitude asymmetry, phase mismatch, energy impact, sensitivity amplification, and boundary approximation. Direction reversal is used to measure the angle between command directions and the proportion of opposite signs; amplitude asymmetry is used to measure the relative difference in adjustment amplitudes on both sides; phase mismatch is used to measure time center offset and period offset; and energy impact is used to measure the strength of expected power step or steam load step. Sensitivity amplification is used to measure the amplification effect on key process quantities under the current operating conditions, while boundary approximation is used to measure the inverse value of the remaining margin of the preset constraints. Then, the data is summarized and standardized to obtain a standardized conflict factor vector. Step three: Based on data freshness, source reliability, and conflict persistence, an adaptive weight vector is calculated, and a weight upper limit is set according to the device safety level to obtain a restricted adaptive weight vector. Step four: Bounded gain aggregation is used to weight and superimpose the standardized conflict factor vector and the restricted adaptive weight vector to obtain an uncalibrated risk value. Step five: The uncalibrated risk value is looked up according to a preset uncalibrated risk-coupled risk lookup table to obtain the corresponding coupled risk coefficient. Based on the preset coupled risk coefficient and the corresponding graded interval, disposal suggestions are output to guide the carbon capture process control system in implementing dynamic regulation.

[0005] In a preferred embodiment, the synchronous comparison sequence is obtained through the following steps: First, valve step, pump speed transition, setpoint change, and alarm confirmation are identified on the edge computing side and the cloud side using preset event judgment rules. Valid anchor points are then selected using two types of invariant constraints: carbon conservation and energy balance. A piecewise linear time mapping with global non-decreasing and upper and lower bounds of slope is solved on the set of anchor points that pass the constraints, thereby obtaining an initial estimate of clock drift and communication delay. Constrained dynamic time warping and phase correction are implemented on the initial time mapping to maintain the monotonicity of the mapping function and limit the local slope. At the same time, a small phase term is added within a limited lag range based on the cross-correlation peak value. The aligned time base and the aligned residual sequence containing the mean, variance, and maximum deviation are output uniformly. Based on the process property relationship, the state dependence conversion between volumetric flow rate and mass flow rate, and between steam flow rate and heat load and different concentration expressions is completed. The unit of measurement is unified, and source credibility, alignment residual confidence band, version summary identifier, and integrity verification summary are added to each moment. Finally, a synchronous comparison sequence carrying complete metadata is formed.

[0006] In a preferred embodiment, during the formation of a synchronous control sequence carrying complete metadata, density and enthalpy are calculated based on temperature, pressure, and solvent composition; volumetric flow rate is converted to mass flow rate; vapor flow rate is converted to heat load; mass fraction and volume fraction are converted to mole fraction; and pressure, temperature, and concentration are unified to a preset unit of measurement system. A source confidence level is added for each moment, calculated by weighting the instrument's most recent calibration date, online self-test results, and the proportion of missing measurements. An alignment residual confidence band is added for each moment, with upper and lower boundaries given by interval estimation of the alignment residual sequence. A version summary identifier and an integrity verification summary are added for each moment. The version summary identifier consists of the model name, model version, and training time; the integrity verification summary consists of a deterministic summary string of the configuration parameter list. These are then combined to generate a synchronous control sequence carrying metadata.

[0007] In a preferred embodiment, step two includes the following steps: extracting the control intent increments of the edge computing side and the cloud side within the same control cycle based on the synchronous comparison sequence, pairing them one by one according to the control channels to obtain paired datasets of the two sides with different preset quantization increments, and recording the rated range, rated energy capacity and constraint limit of each channel.

[0008] In a preferred embodiment, six types of conflict factors are calculated on paired datasets, specifically: Directional reversal is calculated by first calculating the directional angle between the increments on both sides of the same channel, with the directional angle ranging from 0 to 180 degrees; then calculating the proportion of channels with opposite signs; the directional reversal is equal to the weighted average of the directional angle divided by 180 degrees and this proportion; amplitude asymmetry is calculated by taking the absolute value of the difference between the absolute values ​​of the increments on both sides of the same channel, then dividing by the sum of the absolute values ​​of the increments on both sides plus a pre-set minimum positive lower limit as the denominator, and finally taking the arithmetic mean of all channels; phase mismatch is calculated by calculating the cross-correlation function of the control intentions on both sides within a specified maximum time lag range, taking the absolute value of the time lag that produces the maximum correlation and normalizing it with the maximum time lag, and simultaneously calculating the time center difference of the control intentions on both sides and normalizing it with the channel control period; the phase mismatch is equal to the weighted average of the two normalization results; energy impact... The calculation of the degree is as follows: the control intention increment of each channel is converted into the expected power step and expected heat load step using the energy conversion coefficient under the current operating condition. The absolute value of the sum of the two is obtained and normalized by the sum of the rated power and rated heat load of the corresponding unit. The calculation of the sensitivity amplification degree is as follows: the local first-order sensitivity coefficient of the key process quantity to each input channel is obtained at the current operating point. The control intention increment of each channel is multiplied by the corresponding sensitivity coefficient and the sum is obtained to obtain the expected change amplitude of each key process quantity. The maximum value among all key process quantities is taken and normalized by the allowable change bandwidth of the process quantity. The calculation of the boundary approximation degree is as follows: for each controlled constraint, the remaining margin of the constraint after executing the control intentions on both sides is calculated. If the constraint is the upper limit, the upper limit is subtracted from the expected value. If the constraint is the lower limit, the expected value is subtracted from the lower limit. Negative values ​​are counted as zero. Then, the constraint is normalized by the safe bandwidth of the corresponding constraint and the normalization result is subtracted. Finally, the maximum value among all constraints is taken as the boundary approximation degree.

[0009] In a preferred embodiment, the standardized conflict factor vector refers to: performing preset standard processing on the six types of conflict factors calculated based on the synchronous control sequence, mapping each conflict factor to a standardized range of zero to one, and then summarizing them according to a preset order to obtain the standardized conflict factor vector.

[0010] In a preferred embodiment, the constrained adaptive weight vector is obtained through the following steps: Calculating the data freshness score, specifically: on a unified time base, for each data channel used for the six conflict factors, calculate the time lag value, communication delay value, and missing sample ratio. The time lag value is the current unified time minus the latest valid timestamp of the channel; the communication delay value is the communication delay estimate obtained from the aforementioned time mapping; and the missing sample ratio is the proportion of invalid samples to all samples within a fixed evaluation window. The time lag value is linearly back-mapped according to a preset maximum acceptable time difference to obtain the timeliness component; the communication delay value is linearly back-mapped according to a preset maximum acceptable delay to obtain the channel delay component; and the missing sample ratio is calculated using a one-to-one relationship to obtain the integrity component. The timeliness component, channel delay component, and integrity component are then geometrically averaged to obtain a data freshness score ranging from zero to one. Calculating the conflict persistence, specifically: within the same evaluation window, construct an immediate conflict indicator. The immediate conflict indicator is equal to the product of the structural conflict intensity and the amplification intensity, multiplied by the boundary intensity, and the maximum value is taken. The structural conflict intensity is the direction reversal degree and amplitude asymmetry. The arithmetic mean of degree and phase mismatch is used, the amplification intensity is the arithmetic mean of energy impact and sensitivity amplification, and the boundary intensity is the boundary approximation. A conflict threshold is set, and time periods when the indicator value continuously exceeds the threshold are identified. The proportion of time exceeding the threshold and the average value of the indicator value when exceeding the threshold are calculated. The product of the time proportion and the average value is used as the unweighted conflict duration. Then, a linearly increasing weight is applied according to the time from near to far to obtain a conflict duration in the range of zero to one. A restricted adaptive weight vector is formed, specifically: for each component of the six conflict factors, the source credibility weight is multiplied by the data freshness score, and the result is multiplied by the exponential function of the conflict duration. The exponent is given by the sensitivity coefficient corresponding to the device safety level. The sensitivity coefficient takes corresponding fixed values ​​for device safety levels one, two, three, and four. The obtained current weight and the previous period weight are convexly combined with a fixed smoothing coefficient to obtain a smoothed weight. The weight is truncated according to the upper limit set by the device safety level. The upper limit of the weight for levels one, two, three, and four is taken in ascending order. Finally, the restricted adaptive weight vector is output.

[0011] In a preferred embodiment, the uncalibrated risk value is obtained through the following steps: Gain constraint functions are established for direction reversal, amplitude asymmetry, phase mismatch, energy impact, sensitivity amplification, and boundary approximation, respectively. The gain constraint functions adopt a smoothed saturation form, i.e., the ratio of the gain coefficient multiplied by the index value as the numerator and the sum of the gain coefficient and the index value as the denominator, then the smaller value is taken as the gain upper limit. The growth coefficients and gain upper limits for each factor are preset according to the device safety level. The standardized conflict factor vector is input component-by-component into the gain constraint function to obtain a gain-limited factor vector, and then multiplied component-by-component by the limited adaptive weight vector to form a weighted component set. A truncated arithmetic mean is performed on the weighted component set to obtain a aggregate base value, with the truncated ratio taking a fixed value between a preset minimum and maximum ratio. The aggregate base value obtained from the truncated arithmetic mean is then subjected to double-threshold shaping and convex combination smoothing to obtain the uncalibrated risk value.

[0012] In a preferred embodiment, the steps of performing double-threshold shaping and convex combination smoothing on the aggregate base value obtained from the truncated arithmetic mean include: setting a low threshold and a high threshold, with the low threshold and high threshold taking fixed values ​​and the low threshold being less than the high threshold; outputting a zero value when the aggregate base value does not exceed the low threshold; outputting a proportional value equal to the aggregate base value minus the low threshold, divided by the high threshold minus the low threshold when the aggregate base value is between the low threshold and the high threshold; outputting a value equal to a function minus an exponential function, where the independent variable of the exponential function is the product of the aggregate base value minus the high threshold and a fixed convergence coefficient, the fixed convergence coefficient being preset according to the device safety level; combining the output of the current cycle with the output of the previous evaluation cycle using a fixed smoothing coefficient, the fixed smoothing coefficient taking a definite value within the range of zero to one, and limiting the convex combination result to the range of zero to one to obtain an uncalibrated risk value.

[0013] In a preferred embodiment, the AI-based intelligent diagnostic control system for carbon capture processes specifically includes: an alignment and unification module, which acquires control intent data from the edge computing side and the cloud side within the same time period, aligns the data time reference of the control intent and unifies the units of measurement, and generates a synchronous control sequence for comparative analysis; and a conflict quantification module, which calculates six types of conflict factors based on the synchronous control sequence. These six types of conflict factors include direction reversal, amplitude asymmetry, phase mismatch, energy impact, sensitivity amplification, and boundary approximation. Direction reversal measures the angle between command directions and the proportion of opposite signs; amplitude asymmetry measures the relative difference in adjustment amplitudes on both sides; phase mismatch measures time center offset and period offset; energy impact measures the intensity of expected power step or steam load step; and sensitivity amplification... The boundary approximation is used to measure the amplification effect on key process quantities under the current operating conditions and sensitivity. The boundary approximation is used to measure the inverse value of the residual margin of the preset constraints. Then, the data is summarized and standardized to obtain a standardized conflict factor vector. The weight generation module calculates an adaptive weight vector based on data freshness, source credibility, and conflict persistence, and sets an upper limit on the weight according to the device safety level to obtain a restricted adaptive weight vector. The gain aggregation module uses bounded gain aggregation to weight and superimpose the standardized conflict factor vector and the restricted adaptive weight vector to obtain an uncalibrated risk value. The mapping and handling module looks up the uncalibrated risk value according to the preset uncalibrated risk-coupled risk lookup table to obtain the corresponding coupled risk coefficient, and outputs handling suggestions based on the graded interval corresponding to the preset coupled risk coefficient to guide the carbon capture process control system to implement dynamic control.

[0014] The technical effects and advantages of this invention are as follows: By acquiring control intent data from both the edge computing side and the cloud side within the same time period, this invention aligns the time base and unifies the units of measurement, ensuring that subsequent analysis is completed under a unified standard. This effectively avoids misjudgments caused by time deviations or unit inconsistencies between different data sources. Based on this, the invention calculates six types of conflict factors: direction reversal, amplitude asymmetry, phase mismatch, energy impact, sensitivity amplification, and boundary approximation, and performs standardization processing to form a directly quantifiable standardized conflict factor vector. This method can comprehensively analyze multiple dimensions, including command direction, amplitude difference, time offset, energy impact, sensitivity amplification, and constraint approximation. Compared to traditional methods based on only a single indicator, it can significantly improve the accuracy and reliability of conflict identification in the carbon capture process.

[0015] This invention effectively solves the problem of fixed weight allocation and inability to adapt to changes in operating conditions in traditional risk assessments by calculating an adaptive weight vector based on data freshness, source reliability, and conflict duration, and generating a constrained adaptive weight vector under the constraints of the device's safety level. The invention dynamically adjusts the weights during the calculation process, ensuring that factors with fresher data, more reliable sources, and longer conflict durations have a higher weight in the risk assessment, thereby ensuring that the calculation results reflect the real-time operating status. This mechanism guarantees a high degree of consistency between the risk assessment and the actual operating environment of the device, improving the agility and adaptability of the carbon capture process control strategy.

[0016] This invention employs a bounded gain aggregation method to weight and superimpose a standardized conflict factor vector with a constrained adaptive weight vector to obtain an uncalibrated risk value. This uncalibrated risk value is then compared with a pre-defined uncalibrated risk-coupled risk lookup table to output a coupled risk coefficient and corresponding mitigation recommendations. This method not only transforms complex conflict characteristics into quantifiable and interpretable risk coefficients but also directly provides corresponding control suggestions, offering a clear basis for the dynamic control of carbon capture process control systems. Through this closed-loop mechanism from conflict identification to risk mapping and mitigation output, this invention achieves seamless integration of risk assessment and control, thereby effectively reducing system operational risks and improving process control stability. Attached Figure Description

[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of the intelligent diagnostic control method for carbon capture process based on AI decision-making in this invention.

[0018] Figure 2 This is a schematic diagram of the intelligent diagnostic control system for carbon capture process based on AI decision-making in this invention. Detailed Implementation

[0019] 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.

[0020] Reference Figure 1 - Figure 2 The following examples were obtained: Example 1: An intelligent diagnostic control method for carbon capture processes based on AI decision-making, comprising the following steps: Step 1: Acquire control intent data from the edge computing side and the cloud side within the same time period. Align the time reference of the control intent data and unify the units of measurement to generate a synchronous comparison sequence for comparative analysis. This step ensures that data from different computing sides can be accurately compared and analyzed under the same time series and the same measurement system. Time reference alignment eliminates time errors caused by clock drift, network latency, or differences in data acquisition cycles between edge computing nodes and the cloud platform, ensuring the comparability of control intent behavior at the same point in time. Unifying the units of measurement avoids numerical deviations caused by different dimensional systems used by different computing sides, such as volumetric flow rate versus mass flow rate, steam flow rate versus heat load, and concentration representation methods. The synchronous comparison sequence generated in this step not only includes aligned control intent data but also incorporates metadata such as source credibility and time residuals, ensuring that subsequent conflict analysis is based on traceability, verifiability, and unambiguity.

[0021] Step 2: Calculate six types of conflict factors based on the synchronous comparison sequence. These six types of conflict factors include: direction reversal, amplitude asymmetry, phase mismatch, energy impact, sensitivity amplification, and boundary approximation. Direction reversal measures the proportion of opposite angles and signs of the command directions; amplitude asymmetry measures the relative difference in adjustment amplitudes on both sides; phase mismatch measures time center offset and period offset; energy impact measures the intensity of expected power step or steam load step; sensitivity amplification measures the amplification effect on key process quantities under current operating conditions; and boundary approximation measures the reverse value of the remaining margin of the preset constraints. Then, the data is summarized and standardized to obtain a standardized conflict factor vector. This step quantifies the differences in control behavior between the edge computing side and the cloud side into comparable indicators. Direction reversal reveals whether there is a completely opposite trend in the adjustment directions on both sides. If such conflicts occur frequently, it will cause the equipment to perform reciprocating adjustments in a short period of time. Amplitude asymmetry reflects a significant difference in the adjustment amplitudes of the commands on both sides, and can identify over-adjustment or under-adjustment caused by model parameter drift. Phase mismatch, by analyzing the temporal offset of commands on both sides, can identify coordination imbalances caused by delayed response or premature execution. Energy impact directly quantifies the impact of control actions on power and heat load, avoiding the impact of short-term energy fluctuations on system operation. Sensitivity amplification considers the amplification effect of control actions on critical process quantities under different operating conditions and can be used to identify risk amplification effects. Boundary approach measures whether control behavior pushes the system towards constraint limits, serving as an important early warning factor to prevent boundary breaches. After standardizing the six types of conflict factors, a standardized conflict factor vector is formed, providing a unified input basis for subsequent risk aggregation and adaptive weight allocation.

[0022] Step 3: Calculate the adaptive weight vector based on data freshness, source reliability, and conflict persistence, and set a weight cap according to the device safety level to obtain a restricted adaptive weight vector. This step is used to dynamically adjust the influence weight of each conflict factor in the risk aggregation process. Data freshness reflects the timeliness and completeness of the input data; high freshness means that the data can more accurately represent the current state. Source reliability measures the reliability of the data source, preventing the amplification of distorted results caused by abnormal sensors or model version errors. Conflict persistence measures the stability of the conflict phenomenon over time. If the conflict is only short-term noise, its influence weight should be reduced; if the conflict lasts for a long time and its intensity is stable, its weight should be increased. Finally, by setting a weight cap in conjunction with the device safety level, it is possible to prevent a single conflict factor from being infinitely amplified due to short-term anomalies, ensuring the stability and controllability of the risk assessment results.

[0023] Step four involves using bounded gain aggregation to weight and superimpose the standardized conflict factor vector and the constrained adaptive weight vector to obtain the uncalibrated risk value. This step fuses multiple conflict factors according to their weights through a gain aggregation mechanism and avoids abnormal amplification of the risk index caused by extreme values ​​through a gain constraint function. The core of bounded gain aggregation is to map conflict factors and their weights to a controlled gain response, ensuring that the contribution of each factor to the final risk value is within a predictable range. The aggregation process can also incorporate truncated averaging or smoothing mechanisms to suppress the impact of short-term outliers, thereby ensuring that the uncalibrated risk value accurately reflects the comprehensiveness of multi-source conflicts without distortion due to local anomalies.

[0024] Step 5: The uncalibrated risk value is looked up according to a pre-defined uncalibrated risk-coupled risk lookup table to obtain the corresponding coupled risk coefficient. Based on the pre-defined tiered intervals corresponding to the coupled risk coefficients, disposal recommendations are output to guide the carbon capture process control system in implementing dynamic regulation. This step maps uncalibrated risk values ​​to coupled risk coefficients, achieving a direct correlation between risk outcomes and executable control strategies. Through the pre-defined correspondence between uncalibrated risk and coupled risk, complex multidimensional conflict information can be transformed into a risk quantification result within a range of zero to one. The coupled risk coefficient is divided into multiple tiered intervals, each corresponding to different disposal recommendations, such as parameter fine-tuning, command freezing, model shadow running, manual intervention, or emergency de-rating. Through this mapping relationship, the system can automatically trigger dynamic regulation strategies matching the current state based on the risk level, achieving closed-loop management from risk assessment to decision execution, thereby improving the safety, stability, and operational efficiency of the carbon capture process.

[0025] The synchronous comparison sequence is obtained through the following steps: First, valve step, pump speed transition, setpoint change, and alarm confirmation are identified on both the edge computing side and the cloud side using preset event judgment rules. The valve step judgment rule is that the change in valve opening exceeds one percent of the valve range within a single control cycle and lasts for at least two seconds. For example, if the valve opening changes from 20% to 35% and remains there for three seconds, it is considered a valid valve step event. The pump speed transition judgment rule is that the circulating pump speed changes by more than three percent of the rated speed within five seconds and in the same direction. For example, if the circulating pump speed increases from 1000 rpm to 1100 rpm without reverse fluctuation, it is considered a valid pump speed transition event. The setpoint change judgment rule is that if a change in the target setpoint is detected in the control loop exceeding one percent of the control range and is a unidirectional adjustment, for example, if the target value of the solvent circulation volume is increased from 120 tons per hour to 125 tons per hour, it is considered a valid setpoint change event. The rule for determining alarm confirmation is to record the change in alarm status from unconfirmed to confirmed in the control system. For example, if the alarm of excessive outlet pressure of carbon dioxide compressor is cleared after confirmation by the operator, a valid alarm confirmation event can be generated.

[0026] After identifying these events, all events are mapped to time anchors, and valid anchors are selected based on two types of invariant constraints: carbon conservation and energy balance. The carbon conservation constraint requires that the change in carbon dioxide capture at the time anchor point must simultaneously satisfy the material balance deviation between the calculated results of process quantities such as flue gas flow rate, flue gas carbon dioxide concentration, solvent circulation rate, and reboiler heat load not exceeding 2%. For example, if the difference between the calculated capture amount and the measured value is less than 1%, the anchor point is considered valid. The energy balance constraint requires that the difference between the steam heating supply and the stripping tower heat load must be within the allowable deviation range. For example, if the stripping tower heat load is 10 MW and the steam heating supply is 10.1 MW, the deviation is 1%, meeting the selection criteria for valid anchors. For the set of valid anchors selected through constraints, a piecewise linear time mapping function with globally non-decreasing characteristics and upper and lower bounds on the slope is solved using the least squares fitting method. This time mapping function requires that time reversal is not allowed globally, and the slope range is limited to between 0.9 and 1.1, thereby avoiding excessive alignment errors due to time stretching or compression. This mapping process provides an initial estimate of clock drift and communication latency. For example, calculations show that edge computing nodes have a clock offset of 500 milliseconds and a communication latency of 100 milliseconds relative to the cloud.

[0027] After obtaining the initial time map, constrained dynamic time warping and phase correction operations are performed. Dynamic time warping is used to adjust the time map function within a small range to align the time axes of the edge computing side and the cloud side, while maintaining the monotonicity of the time map function and limiting the local slope to no more than 1.1 or less than 0.9. During the warping process, a cross-correlation function is used to calculate the correlation peak between the edge control intent and the cloud control intent within a maximum allowable lag of two seconds. When a correlation peak is detected to be lagging by 0.5 seconds, the phase offset is set to 0.5 seconds and superimposed on the time map function for correction. The final output time alignment result not only provides a unified time reference but also generates an alignment residual sequence containing the mean, variance, and maximum deviation. For example, after alignment, the residual mean is 0.1 seconds, the residual variance is 0.02 seconds, and the maximum deviation is 0.3 seconds.

[0028] After time alignment, state-dependent conversions are performed based on process property relationships to unify the units of measurement for different data. Volumetric flow rate is converted to mass flow rate using measured temperature, pressure, and solvent composition. For example, if the solvent density at room temperature is 1,000 kg / m³ and the volumetric flow rate is 120 m³ / h, the converted mass flow rate is 120,000 tons / h. Steam flow rate is converted to heat load using steam enthalpy difference. For example, if the steam mass flow rate is 30 tons / h and the enthalpy difference is 2,500 kJ / kg, the heat load is 75 MW. Solvent concentration is converted from mass fraction and volume fraction to mole fraction when necessary. For example, if the solvent mass fraction is 30% and the molar mass is 45 g / mol, the converted mole fraction is 32%. Pressure, temperature, and concentration are also unified to a standard unit system to ensure that subsequent calculations of conflict factors are numerically unambiguous.

[0029] To ensure data traceability and reliability, after unit conversion, a source confidence score, alignment residual confidence band, version summary identifier, and integrity verification summary are appended to each time point. The source confidence score is calculated by weighting the instrument's most recent calibration date, online self-test results, and the percentage of missing tests. For example, if an instrument's most recent calibration was within one week and all online self-tests passed, with a missing test percentage below one percent, the confidence score can reach 0.95%. The alignment residual confidence band is estimated based on the mean and variance of the alignment residual sequence; for example, the residual confidence interval is 0.05 seconds to 0.15 seconds. The version summary identifier includes the model name, model version number, and training time; for example, the model name is "Absorber Regulation Model," the version is 1.0, and the training time is December 2024. The integrity verification summary is calculated based on a hash function of the configuration parameter list to prevent configuration parameters from being tampered with or lost. Finally, the data after time alignment, unit conversion, and metadata annotation are integrated to form a synchronous comparison sequence carrying complete metadata. This sequence can not only accurately represent the control intentions of the edge computing side and the cloud side at the same time, but also provide data-related credibility, error range, and model information, providing a reliable input basis for subsequent conflict factor calculation, weight generation, and risk aggregation.

[0030] In the process of forming a synchronized reference sequence carrying complete metadata, the density and enthalpy are first calculated based on temperature, pressure, and solvent composition. Density is calculated based on the solvent's property equations. For example, when the solvent is a single amine solution, the density can be calculated using the functional relationship between temperature, pressure, and solvent composition. When the temperature is 35 degrees Celsius, the pressure is atmospheric pressure, and the solvent mass fraction is 30%, the density can be calculated as 105 kg / m³. Enthalpy is calculated using the thermodynamic equation of state. For example, under the same conditions, the enthalpy is 150 kJ / kg, used for subsequent heat conversion. After completing the density and enthalpy calculations, the volumetric flow rate is converted to the mass flow rate. The unit of volumetric flow rate is cubic meters per hour. The mass flow rate is obtained by multiplying it by the density. For example, if the solvent flow rate is 120 cubic meters per hour and the density is 105 kg / m³, the converted mass flow rate is 120,600 kg / h. Subsequently, the steam flow rate is converted into heat load. Heat load is calculated by multiplying the steam mass flow rate by the enthalpy difference. For example, if the steam mass flow rate is 30 tons per hour and the enthalpy difference is 2,500 kilojoules per kilogram, the converted heat load is 75 megawatts. Next, the mass fraction and volume fraction are converted into mole fraction. Mole fraction ensures that the concentrations of different components can be expressed consistently. For example, if the solvent mass fraction is 30%, the molar mass is 45 grams per mole, and the density of the solvent solution is 1,000 kilograms per cubic meter, the converted mole fraction of the solvent is 32%. For gaseous components, conversion can be performed using volume fraction combined with molar volume under standard conditions to ensure consistent concentration data across different phases. After the above conversions, pressure, temperature, and concentration are unified to a preset unit system. For example, pressure is unified to kilopascals, temperature to degrees Celsius, and concentration to mole fraction to ensure consistency and comparability of data in subsequent calculations.

[0031] After unit conversion, a source reliability score is applied to each time point. Source reliability is a crucial indicator of data reliability at that moment, calculated by weighting the instrument's most recent calibration date, online self-test results, and the percentage of missing data. For example, if a pressure sensor has undergone traceability calibration within the past seven days, online self-tests show stable and unbiased output, and the percentage of missing data is less than one percent, the sensor's source reliability score can reach 0.95. If a temperature sensor's calibration cycle exceeds thirty days or the percentage of missing data exceeds five percent, the source reliability score will drop below 0.8. The source reliability score can be calculated using a weighted average method, with the calibration date weighted at 0.4, the online self-test result weighted at 0.4, and the missing data percentage weighted at 0.2, ensuring that the reliability reflects both the sensor's health status and the availability of real-time data. Subsequently, an aligned residual confidence band is applied to each time point. The aligned residual confidence band is obtained based on interval estimation of the time-aligned residual sequence, for example, by statistically analyzing the mean, variance, and percentiles of the residual sequence to provide upper and lower confidence boundaries. If the mean residual is 0.1 seconds and the variance is 0.02 seconds for a given time period, the confidence interval can be calculated to be from 0.05 seconds to 0.15 seconds. The confidence band provides error boundaries for subsequent risk analysis, used to determine the reliability of time alignment. A version summary identifier and an integrity verification summary are appended to each time point. The version summary identifier includes the model name, model version, and training time; for example, the name of the absorber control model is "Absorber Regulation Model," the version is 1.0, and the training time is December 2024. The integrity verification summary generates a unique string based on the model and configuration parameter list using a hash algorithm. For example, the SHA-256 algorithm is used to summarize the configuration file, which includes model hyperparameters, input variable definitions, boundary conditions, etc., ensuring that any changes can be quickly detected. This summary string effectively prevents data inconsistencies caused by model version confusion or configuration file tampering.

[0032] Finally, time alignment, unit conversion, source confidence, alignment residual confidence band, version summary identifier, and integrity check summary are summarized to form a synchronous comparison sequence carrying complete metadata. For example, a record at a certain moment in the synchronous comparison sequence includes the following: time 10:30, solvent mass flow rate 120,600 kg / h, steam heat load 75 MW, solvent mole fraction 32%, source confidence 0.95, alignment residual confidence band from 0.05 to 0.15 seconds, model version 1.0 (absorbing tower regulation model), and integrity check summary (a 64-bit hexadecimal string). This synchronous comparison sequence provides a traceable, verifiable, and highly consistent data foundation for subsequent conflict factor calculation, adaptive weight generation, and risk aggregation, ensuring high reliability and interpretability of intelligent diagnosis and control of the carbon capture process.

[0033] Step two includes the following steps: First, based on the synchronous comparison sequence, extract the control intent increments of the edge computing side and the cloud side within the same control cycle. Then, pair them one by one according to the control channels to obtain paired datasets of different preset quantization increments on both sides. Record the rated range, rated energy capacity, and constraint limits for each channel. The control intent increment refers to the change in the value of the regulating variable between two adjacent sampling times. For example, in the circulating pump speed control channel, if the speed in the previous sampling cycle was 1000 rpm and the speed in this cycle is 1050 rpm, then the control intent increment is 50 rpm. When pairing channels, the synchronous comparison sequence ensures time consistency between the edge computing side and the cloud side, and matching is done using channel identifiers to ensure that identical devices and variables can be matched one-to-one. For each channel, record the rated range, such as the valve opening range of zero to 100%; record the rated energy capacity, such as the reboiler's heat load capacity of 100 MW; and record the constraint limits, such as the top temperature of the analytical tower not exceeding 120 degrees Celsius. These data are used for subsequent normalization calculations of conflict factors.

[0034] After obtaining the paired datasets, six types of conflict factors are calculated, specifically: First, the direction reversal degree is calculated by first calculating the directional angle between the two increments on both sides of the same channel. The directional angle is taken as an angle value from 0 degrees to 180 degrees. The directional angle is equal to the dot product of the two increment vectors divided by the product of the magnitudes of the two increment vectors and then taking the inverse cosine. For example, in the solvent circulation pump channel, if the increment on the edge calculation side is positive 100 kg / h and the increment on the cloud side is negative 50 kg / h, then the angle between the two is close to 180 degrees. Then, the proportion of channels with opposite signs to all channels is calculated. For example, if four out of ten channels have opposite directions, then the proportion is 0.4. Finally, the direction reversal degree is equal to the weighted average of the directional angle divided by 180 and the proportion. If the weights are 0.6 and 0.4 respectively, then the direction reversal degree is 0.6 multiplied by the normalized value of the angle plus 0.4 multiplied by the proportion of opposite signs.

[0035] Second, the amplitude asymmetry is calculated by taking the absolute value of the difference between the absolute values ​​of the increments on both sides of the same channel, and then dividing by the sum of the absolute values ​​of the increments on both sides plus a pre-set minimum positive lower limit value as the denominator. For example, in the solvent flow control channel, the increment on the edge computing side is 20 tons per hour, and the increment on the cloud side is 30 tons per hour, with a difference of 10 tons per hour. The sum of the absolute values ​​of the two is 50 tons per hour. If the lower limit value is 0.1 tons per hour, then the amplitude asymmetry is 10 divided by 50.1, and the result is approximately 0.199. Finally, the arithmetic mean of all channels is taken as the global amplitude asymmetry.

[0036] Third, the phase mismatch is calculated by calculating the cross-correlation function of the control intentions on both sides within the specified maximum time lag range, taking the absolute value of the time lag that produces the maximum correlation, and normalizing it with the maximum time lag. For example, in the power control channel of the flue gas compressor, there is a 0.5-second lag between the control commands on the edge computing side and the cloud side, and the maximum allowable lag is two seconds. The normalized lag value is 0.25. At the same time, the time center difference between the control intentions on both sides is calculated and normalized with the control cycle of the channel. For example, if the time center difference is 0.2 seconds and the control cycle is one second, the result is 0.2. The phase mismatch is equal to the weighted average of the two normalized results. The weights can be set according to the dynamic characteristics of the actual controlled object. For example, if the lag weight is 0.7 and the time center difference weight is 0.3, the phase mismatch is 0.7 multiplied by 0.25 plus 0.3 multiplied by 0.2, which is approximately 0.235.

[0037] Fourth, the energy impact is calculated by converting the control intention increment of each channel into the expected power step and expected heat load step using the energy conversion factor under the current operating conditions. For example, in the reboiler channel, if the steam flow rate increment is two tons per hour and the enthalpy difference is 2,500 kJ per kilogram, then the heat load step is 1,389 kilowatts. In the circulating pump channel, if the speed increases by 50 revolutions per minute, the corresponding power step is 10 kilowatts. After summing the energy steps of all channels, normalize them by the sum of the rated power and rated heat load of the corresponding unit. For example, if the total energy step is 1,400 kilowatts and the rated capacity is 10,000 kilowatts, then the energy impact is 0.14.

[0038] Fifth, the sensitivity amplification is calculated by obtaining the local first-order sensitivity coefficient of the key process quantity to each input channel at the current operating point. For example, the sensitivity coefficient of the analytical column top temperature to the solvent flow rate is 0.02 degrees Celsius per ton per hour, and the sensitivity coefficient to the reboiler steam flow rate is 0.05 degrees Celsius per ton per hour. Multiplying the solvent flow rate increment of 20 tons per hour by the sensitivity coefficient yields 0.4 degrees Celsius, and multiplying the reboiler steam flow rate increment of 5 tons per hour by the sensitivity coefficient yields 0.25 degrees Celsius. The sum of the two is 0.65 degrees Celsius. Then, taking the allowable variation bandwidth of 1 degree Celsius, the sensitivity amplification is 0.65.

[0039] Sixth, the boundary approximation degree is calculated by determining the remaining margin of each controlled constraint after executing the control intentions on both sides. If the constraint is an upper limit, the upper limit is subtracted from the expected value; if the constraint is a lower limit, the expected value is subtracted from the lower limit. Negative values ​​are counted as zero. For example, if the upper limit of the analytical tower pressure is 200 kPa and the expected pressure is 190 kPa, then the remaining margin is 10 kPa, the safety bandwidth is 50 kPa, and the normalized value is 0.2. Subtracting the normalized result from 1 gives 0.8. If there are multiple constraints, the maximum value among all constraints is taken as the boundary approximation degree. Finally, the calculated six types of conflict factors are standardized and mapped so that the values ​​of each conflict factor are within the range of zero to one, and a standardized conflict factor vector is generated as the basic input for subsequent weight calculation and risk aggregation, ensuring that the conflict diagnosis in the carbon capture process is comparable, traceable, and stable.

[0040] Step two includes the following steps: First, based on the synchronous comparison sequence, extract the control intent increments of the edge computing side and the cloud side within the same control cycle. Then, pair them one by one according to the control channels to obtain paired datasets of different preset quantization increments on both sides. Record the rated range, rated energy capacity, and constraint limits for each channel. The control intent increment refers to the change in the value of the regulating variable between two adjacent sampling times. For example, in the circulating pump speed control channel, if the speed in the previous sampling cycle was 1000 rpm and the speed in this cycle is 1050 rpm, then the control intent increment is 50 rpm. When pairing channels, the synchronous comparison sequence ensures time consistency between the edge computing side and the cloud side, and matching is done using channel identifiers to ensure that identical devices and variables can be matched one-to-one. For each channel, record the rated range, such as the valve opening range of zero to 100%; record the rated energy capacity, such as the reboiler's heat load capacity of 100 MW; and record the constraint limits, such as the top temperature of the analytical tower not exceeding 120 degrees Celsius. These data are used for subsequent normalization calculations of conflict factors.

[0041] After obtaining the paired datasets, six types of conflict factors are calculated, specifically: First, the direction reversal degree is calculated by first calculating the directional angle between the two increments on both sides of the same channel. The directional angle is taken as an angle value from 0 degrees to 180 degrees. The directional angle is equal to the dot product of the two increment vectors divided by the product of the magnitudes of the two increment vectors and then taking the inverse cosine. For example, in the solvent circulation pump channel, if the increment on the edge calculation side is positive 100 kg / h and the increment on the cloud side is negative 50 kg / h, then the angle between the two is close to 180 degrees. Then, the proportion of channels with opposite signs to all channels is calculated. For example, if four out of ten channels have opposite directions, then the proportion is 0.4. Finally, the direction reversal degree is equal to the weighted average of the directional angle divided by 180 and the proportion. If the weights are 0.6 and 0.4 respectively, then the direction reversal degree is 0.6 multiplied by the normalized value of the angle plus 0.4 multiplied by the proportion of opposite signs.

[0042] Second, the amplitude asymmetry is calculated by taking the absolute value of the difference between the absolute values ​​of the increments on both sides of the same channel, and then dividing by the sum of the absolute values ​​of the increments on both sides plus a pre-set minimum positive lower limit value as the denominator. For example, in the solvent flow control channel, the increment on the edge computing side is 20 tons per hour, and the increment on the cloud side is 30 tons per hour, with a difference of 10 tons per hour. The sum of the absolute values ​​of the two is 50 tons per hour. If the lower limit value is 0.1 tons per hour, then the amplitude asymmetry is 10 divided by 50.1, and the result is approximately 0.199. Finally, the arithmetic mean of all channels is taken as the global amplitude asymmetry.

[0043] Third, the phase mismatch is calculated by calculating the cross-correlation function of the control intentions on both sides within the specified maximum time lag range, taking the absolute value of the time lag that produces the maximum correlation, and normalizing it with the maximum time lag. For example, in the power control channel of the flue gas compressor, there is a 0.5-second lag between the control commands on the edge computing side and the cloud side, and the maximum allowable lag is two seconds. The normalized lag value is 0.25. At the same time, the time center difference between the control intentions on both sides is calculated and normalized with the control cycle of the channel. For example, if the time center difference is 0.2 seconds and the control cycle is one second, the result is 0.2. The phase mismatch is equal to the weighted average of the two normalized results. The weights can be set according to the dynamic characteristics of the actual controlled object. For example, if the lag weight is 0.7 and the time center difference weight is 0.3, the phase mismatch is 0.7 multiplied by 0.25 plus 0.3 multiplied by 0.2, which is approximately 0.235.

[0044] Fourth, the energy impact is calculated by converting the control intention increment of each channel into the expected power step and expected heat load step using the energy conversion factor under the current operating conditions. For example, in the reboiler channel, if the steam flow rate increment is two tons per hour and the enthalpy difference is 2,500 kJ per kilogram, then the heat load step is 1,389 kilowatts. In the circulating pump channel, if the speed increases by 50 revolutions per minute, the corresponding power step is 10 kilowatts. After summing the energy steps of all channels, normalize them by the sum of the rated power and rated heat load of the corresponding unit. For example, if the total energy step is 1,400 kilowatts and the rated capacity is 10,000 kilowatts, then the energy impact is 0.14.

[0045] Fifth, the sensitivity amplification is calculated by obtaining the local first-order sensitivity coefficient of the key process quantity to each input channel at the current operating point. For example, the sensitivity coefficient of the analytical column top temperature to the solvent flow rate is 0.02 degrees Celsius per ton per hour, and the sensitivity coefficient to the reboiler steam flow rate is 0.05 degrees Celsius per ton per hour. Multiplying the solvent flow rate increment of 20 tons per hour by the sensitivity coefficient yields 0.4 degrees Celsius, and multiplying the reboiler steam flow rate increment of 5 tons per hour by the sensitivity coefficient yields 0.25 degrees Celsius. The sum of the two is 0.65 degrees Celsius. Then, taking the allowable variation bandwidth of 1 degree Celsius, the sensitivity amplification is 0.65.

[0046] Sixth, the boundary approximation is calculated by calculating the remaining margin of each controlled constraint after executing the control intentions on both sides. If the constraint is an upper limit, the upper limit is subtracted from the expected value; if the constraint is a lower limit, the expected value is subtracted from the lower limit. Negative values ​​are counted as zero. For example, if the upper limit of the analytical tower pressure is 200 kPa and the expected pressure is 190 kPa, then the remaining margin is 10 kPa, the safety bandwidth is 50 kPa, and the normalized value is 0.2. Then, take one and subtract the normalized result to get 0.8. If there are multiple constraints, the maximum value among all constraints is taken as the boundary approximation.

[0047] The key parameter definitions and operating condition association rules in the calculation formula of the "six types of conflict factors" are supplemented and disclosed: The method of obtaining the weighting coefficient in the directional reversal degree is as follows: the first weighting term is the normalized value of the directional angle, and the directional angle θ ranges from [0, 180] degrees; the second weighting term is the proportion of channels with opposite signs, defined as the number of channels with opposite signs divided by the total number of channels; the weighted average coefficient can be set at 0.5:0.5, or it can be adjusted according to the sensitivity of directional change based on the response characteristic curve of the control system; this characteristic curve can be established through standard step response test; if there are specific operating conditions that are more sensitive to changes in directional polarity (such as steam regulating valves), the directional angle weight can be increased to 0.7 to reflect the importance of directional conflict. The maximum hysteresis range in phase mismatch: The maximum hysteresis time window Δt_max is 3 to 5 control cycles, with a default value of 3×T_c, where T_c is the duration of a single control cycle (e.g., 5s). If the system response has an inertial amplification effect (e.g., high-temperature heat exchangers), Δt_max can be set to 5×T_c. This value is set with reference to the definition of the hysteresis identification time window in the "Process Control System Engineering Design Standard" (GB / T38898-2020). The truncation ratio in the truncated arithmetic mean: The upper and lower limits are fixed at 15% and 30% respectively, with a default of 20%, derived from the strategy for excluding extreme values ​​during multi-factor aggregation in the "Statistical Quality Control Engineering Handbook". This can be adjusted based on industry experience and historical conflict data distribution curves. Explanation of operating condition-related parameters and data sources: Energy conversion coefficients in energy impact: Control intention increments (such as opening degree Δu) are converted into equivalent electrical power P_e and heat load Q_h, using a linear approximation conversion model: P_e = α_e * Δu, Q_h = α_h * Δu, where α_e and α_h are the electrical and heat energy conversion coefficients corresponding to a unit opening degree. Coefficients α_e and α_h can be extracted from device energy efficiency assessment data or gain values ​​in the control model, such as the flow-load conversion module in the DCS system or P&ID design parameters; if not directly recorded, they can be obtained by fitting a linear regression model to the operating data (refer to IEC61512-3 batch energy mapping specification). Methods for obtaining the sensitivity coefficient in sensitivity amplification: The sensitivity coefficient is defined as the local partial derivative (∂y / ∂u) of the key process quantity with respect to the input variable. It is calculated from historical operation data using the operation point linearization method. It can be obtained by fitting a local first-order linear model using a rolling window (10-30 points recommended). Key process quantities include carbon dioxide absorption rate, solvent temperature, tower pressure, etc., and their selection should meet the principles of "process safety relevance" and "controllability". If a neural network or fuzzy system is used as the control proxy model, the sensitivity coefficient can be directly extracted through backpropagation or gradient tracking.To avoid relying on trial and error adjustments, the default settings and adjustment ranges of all key parameters in this invention should be derived and verified based on the following standards or documents: National standards and industry specifications: GB / T38898-2020 "Process Control System Engineering Design Standard", HG / T20571-2015 "Chemical Control System Design Specification", NB / T20384-2021 "Carbon Dioxide Capture Engineering Design Guidelines", IEC61512-3 "Batch Processing Control Systems - Energy Efficiency Data Interface". Finally, the six types of conflict factors calculated will be standardized and mapped so that the values ​​of each conflict factor are within the range of zero to one, and a standardized conflict factor vector will be generated as the basic input for subsequent weight calculation and risk aggregation, ensuring that the conflict diagnosis in the carbon capture process is comparable, traceable, and stable.

[0048] The standardized conflict factor vector refers to the process of standardizing six types of conflict factors calculated based on synchronous control sequences, mapping each conflict factor to a standardized range of zero to one, and then summarizing them according to a preset order to obtain the standardized conflict factor vector. Standardization of direction reversal: Direction reversal represents the degree of difference in the control intention direction between the edge computing side and the cloud side within the same channel. Its value is derived from the weighted result of the ratio of the angle between the directions of the commands and the opposite signs of the commands. To avoid the influence of dimensional differences between different channels on risk calculation, direction reversal is mapped to a range of zero to one. For example, under a certain operating condition, the angle between the control commands from the edge computing side and the cloud side is 90 degrees, corresponding to a normalized value of 0.5. If three out of ten channels have opposite signs, the proportion of opposite signs is 0.3, and after weighting, the direction reversal is 0.48. This result is standardized using a linear mapping method between the minimum value of zero and the maximum value of one, resulting in a final standardized direction reversal of 0.48. Standardization of amplitude asymmetry: Amplitude asymmetry reflects the relative proportion of the difference in adjustment amplitude between the two sides within the same channel. The calculated amplitude asymmetry is directly mapped to the preset minimum value of zero and maximum value of one. For example, in the solvent circulation pump channel, the edge computing side increment is 20 tons per hour, the cloud side increment is 30 tons per hour, the difference between the two is 10 tons per hour, the sum of the absolute values ​​is 50 tons per hour, and the normalized result is 0.2, which directly enters the normalized range of zero to one.

[0049] Phase mismatch is standardized: Phase mismatch measures the time offset between control commands on the edge computing side and the cloud side, and includes a weighted average of the maximum correlation lag and the time center difference. Taking a certain control channel as an example, if the maximum correlation lag is 0.5 seconds and the allowable lag upper limit is 2 seconds, then the lag part is normalized to 0.25; the time center difference is 0.1 seconds and the control cycle is 1 second, then it is normalized to 0.1. The weighted average of the two is 0.17, which is mapped to a standardized range of 0 to 1 and used as the standardized phase mismatch input. Energy impact is standardized: Energy impact measures the magnitude of the impact of control command adjustments on system power and thermal load. Under a certain operating condition, if the increase in reboiler steam flow rate is converted to 1,200 kW and the increase in circulating pump power is converted to 20 kW, the total is 1,220 kW, and the rated energy capacity is 10,000 kW, then the normalized energy impact is 0.122. After mapping this value to the standardized range of 0 to 1, it can be directly used for subsequent calculations.

[0050] Sensitivity amplification is standardized: Sensitivity amplification measures the intensity of the impact of control commands on critical process quantities at the current operating point. For example, the sensitivity coefficient for the temperature at the top of the analytical column is 0.05 degrees Celsius per ton per hour. When the solvent flow rate increases by 20 tons per hour, the expected temperature change is 1 degree Celsius, and the allowable temperature fluctuation bandwidth is 2 degrees Celsius. The original value of sensitivity amplification is 0.5. This value is mapped to a standardized range of 0 to 1 and then used for risk calculation. Boundary approximation is standardized: Boundary approximation measures whether control commands push process quantities towards constraint limits. For example, the upper limit of the pressure at the top of the analytical column is 200 kPa, the expected pressure is 180 kPa, the margin is 20 kPa, and the safety bandwidth is 50 kPa. The normalized result is 0.6. Subtracting this value from 1 yields a boundary approximation of 0.4. This value is directly mapped to a range of 0 to 1 and then used in the calculation. Finally, the standardized direction reversal, amplitude asymmetry, phase mismatch, energy impact, sensitivity amplification, and boundary approximation are summarized in a fixed order to form a standardized conflict factor vector. For example, if the standardized results of the six conflict factors at a certain moment are 0.48, 0.2, 0.17, 0.122, 0.5, and 0.4, the final standardized conflict factor vector will be [0.48, 0.2, 0.17, 0.122, 0.5, 0.4]. This vector will serve as the core input for subsequent adaptive weight calculation and bounded gain aggregation, ensuring that the basic data for risk calculation has a unified dimension, strong comparability, and is suitable for dynamic analysis under multiple operating conditions.

[0051] The constrained adaptive weight vector is obtained through the following steps: First, a data freshness score is calculated. The data freshness score measures the overall quality of data across the six conflict factors in terms of time validity, communication latency, and completeness. Under a unified time reference, the time lag, communication latency, and missing sample ratio are calculated for each data channel. The time lag is the time difference between the current unified time and the latest valid timestamp for that channel. For example, if the current unified time is 10:00 and the latest valid timestamp is 9:59:50, the time lag is 10 seconds. The communication latency is derived from the communication latency estimate during the time mapping phase. For example, if the data transmission latency from the edge computing side to the cloud side is 0.3 seconds, the communication latency is 0.3 seconds. The missing sample ratio is the proportion of invalid samples to the total number of samples within a fixed evaluation window. For example, in a five-minute evaluation window with 300 sampling points, if 15 sampling points are invalid, the missing sample ratio is 15 divided by 300, which equals 0.05.

[0052] After obtaining the time lag value, communication delay value, and missing test ratio, standardization is performed on each. The time lag value is linearly inversely mapped based on the preset maximum acceptable time difference. For example, if the maximum acceptable time difference is 30 seconds, a 10-second time lag value is mapped to 1 minus 10 divided by 30, resulting in a timeliness component of 0.67. The communication delay value is linearly inversely mapped based on the maximum acceptable delay. For example, if the maximum acceptable delay is 1 second, 0.3 seconds is mapped to 1 minus 0.3 divided by 1, resulting in a channel delay component of 0.7. The missing test ratio is converted to an integrity component through a subtraction mapping relationship. For example, if the missing test ratio is 0.05, the integrity component is 1 minus 0.05, resulting in 0.95. The geometric mean of the timeliness component, channel delay component, and integrity component is taken to obtain the data freshness score, which is the cube root of (0.67 x 0.7 x 0.95), approximately equal to 0.77. This value is used for subsequent weight calculations. Next, the conflict persistence is calculated, which measures the continuity and intensity of conflicts over time. Within the same evaluation window, an instantaneous conflict indicator is constructed. This indicator is equal to the maximum value of the product of the structural conflict intensity and the amplification intensity, divided by the boundary intensity. The structural conflict intensity is the arithmetic mean of direction reversal, amplitude asymmetry, and phase mismatch. For example, if these are 0.4, 0.3, and 0.2 respectively, then the structural conflict intensity is (0.4 + 0.3 + 0.2) divided by 3, which equals 0.33. The amplification intensity is the arithmetic mean of energy impact and sensitivity amplification. For example, if these are 0.5 and 0.4 respectively, then the amplification intensity is (0.5 + 0.4) divided by 2, which equals 0.45. The boundary intensity is the boundary approximation, for example, 0.6. Therefore, the instantaneous conflict indicator is the maximum value between the product of the structural conflict intensity and the amplification intensity (0.15) and the boundary intensity (0.6), resulting in an instantaneous conflict indicator of 0.6.

[0053] A conflict threshold is set, for example, 0.5. If the instantaneous conflict indication exceeds 0.5 for 50 consecutive seconds, its proportion within the total 100-second evaluation window is 0.5, and the average instantaneous conflict indication exceeding the threshold is 0.6. Then, the unweighted conflict duration is 0.5 multiplied by 0.6, which equals 0.3. A linearly increasing weight is then applied to this result from most recent to furthest in time. For example, the weight is 1 for the most recent 30 seconds, 0.8 for the middle 30 seconds, and 0.6 for the furthest 40 seconds. After weighted summation and normalization, the conflict duration is 0.35, mapped to the range of 0 to 1.

[0054] In the construction of the constrained adaptive weight vector, the maximum time lag value, which is the upper limit of the time lag parameter used to measure the synchronization of control data, is set to a default range of one to five seconds. This value is derived from the actual sampling period and execution feedback response time of the carbon capture process control system. In industrial coal-fired power plants or natural gas capture scenarios, the typical feedback control period is two to three seconds, so it is recommended to set the maximum acceptable lag time to three seconds. If the control system has a high-speed feedback mechanism (such as operating at a millisecond-level sampling frequency on the edge side), the maximum lag time threshold can be calculated by multiplying the sampling period by ten to twenty times. Secondly, the calibration period for weight updates, i.e., the time window length for updating the constrained adaptive weight vector, is set to a fixed period within the range of ten to thirty minutes by default. This period setting aims to balance the response sensitivity to changes in conflict trends with the stability of system disturbances. Its specific value is recommended to be determined based on the following three types of indicators: the frequency of conflict factor fluctuations, the frequency of device switching, and the response inertia of the control execution structure. If the conflict factor fluctuates frequently (e.g., the conflict lasts for more than three minutes within every five minutes), the calibration cycle should be shortened; if the actuator of the control system has a large inertia (e.g., the temperature control of the solvent regeneration furnace), the calibration cycle can be appropriately extended to avoid repeated adjustments.

[0055] For various weighting coefficients involved in the weighting composition, such as the weighting allocation of timeliness, delay, and completeness in data freshness scoring, and the weighting factors of time proportion and intensity in conflict persistence, it is recommended to use the normalized three-factor equal weighting principle, or to perform weighted optimization based on the key objectives of device operation. For example, in a steam-driven capture system, it is recommended to increase the sensitivity weight of the conflict persistence factor corresponding to energy impact intensity to enhance the response capability to sudden changes in heat load. Regarding the setting of the upper limit of weights and the exponential sensitivity coefficient for device safety level, it is recommended to map parameters according to the definition of safety level in actual industry standards. For example, referring to the constraint intensity of safety levels one to four in the process safety management guidelines, the corresponding upper limits of weights can be set to 20%, 35%, 50%, and 70%, respectively, and the exponential sensitivity coefficients can be set to 1, 1.5, 2, and 3, respectively. This setting method ensures that in high safety level scenarios, changes in conflict factors have a stronger driving effect on the final weighting composition, thereby triggering adjustment recommendations earlier. Furthermore, if uncertainties exist in the actual industrial environment that necessitate dynamic adjustments to the aforementioned parameters, a rolling self-tuning mechanism can be constructed based on process modeling results, historical data distribution, and deviation evaluation indicators to periodically correct these empirical parameters. This self-tuning mechanism can reference parameter update strategies from statistical process control methods, fuzzy weighted optimization methods, or Kalman filter models, enabling the entire weight calculation process to have adaptive adjustment capabilities while maintaining defined boundaries, further enhancing the usability and robustness of the method under complex operating conditions.

[0056] Source reliability measures the reliability of a data source in terms of calibration status, equipment self-testing, and data continuity, and is a crucial factor in the calculation of constrained adaptive weighted vectors. The calculation of source reliability consists of the following components: The instrument's most recent calibration date reflects the reliability of the data source's metrological accuracy. First, the date of the most recent metrological calibration is recorded for each measuring point. The time interval is calculated by subtracting this date from the current unified time. For example, if a pressure transmitter was calibrated 120 days before the current date, the time interval is 120 days. Then, a linear decay mapping is performed based on this time interval and the preset maximum effective calibration period. If the maximum effective calibration period is 180 days, the calibration component equals one minus the time interval divided by 180. For example, the calibration component for 120 days is one minus 120 divided by 180, which equals 0.33. If the time interval exceeds 180 days, the calibration component is zero, indicating that the data lacks calibration reliability. In practical applications, for example, if a flow meter was last calibrated only 30 days ago, the calibration component is one minus 30 divided by 180, which equals 0.83, indicating that the data from this flow meter is highly reliable in terms of calibration.

[0057] Online self-test results reflect the self-diagnostic status of the data acquisition equipment. The equipment generates a self-test score through its self-test function, ranging from zero to one. Zero indicates complete equipment malfunction, and one indicates complete normal operation. The self-test score is calculated based on the following weighted indicators: Signal stability: For example, analyzing the standard deviation of a flow signal over one minute; if the standard deviation is less than one percent of the rated range, a full score of one is awarded; otherwise, the score is reduced proportionally. Hardware status: The equipment self-test function can return health status indicators; for example, a normal power supply voltage scores one, while a low power supply voltage scores less based on the degree of deviation. Historical failure rate: For example, the number of self-test anomalies recorded in the past seven days; zero anomalies score one, while any anomalies result in a proportional deduction. For example, if a temperature sensor has a signal stability score of 0.95, a hardware status score of one, and a historical failure rate of 0.85, then the self-test score, calculated by averaging the three equally weighted indicators, is (0.95 + 1 + 0.85) divided by three, which equals 0.93.

[0058] The missing sample ratio reflects the data continuity within a fixed evaluation window. The missing sample ratio equals the number of invalid samples divided by the total number of samples. For example, if 300 samples are collected within a five-minute evaluation window, and six samples are invalid due to packet loss, the missing sample ratio is six divided by 300, which equals 0.02. The missing sample ratio is mapped to one minus the missing sample ratio; for example, one minus 0.02 equals 0.98, indicating extremely high data continuity. If the missing sample ratio exceeds 20%, this component is set to zero, indicating a loss of data continuity and lack of reliability. The calibration component, online self-test score, and missing sample ratio component are combined using a weighted geometric mean. The weights can be set according to process safety requirements; for example, each of the three weights can be one-third. For example, if a flowmeter's calibration component is 0.83, its self-test score is 0.93, and its missing sample ratio component is 0.98, then the source reliability is equal to the cube root of (0.83 x 0.93 x 0.98), which is approximately 0.91.

[0059] A constrained adaptive weight vector is formed. For each component of the six conflict factors, the source credibility weight is first multiplied by the data freshness score. For example, if the source credibility weight is 0.9 and the data freshness score is 0.77, the result is 0.693. Then, the result is multiplied by the exponential function of the conflict persistence. The exponent is determined by the sensitivity coefficient corresponding to the device's security level. For example, if the device security level is level two and the sensitivity coefficient is 1, then the exponential function of the conflict persistence of 0.35 is 0.35 raised to the power of 1, resulting in 0.35. Multiplying this by 0.693 yields 0.243. The current weight is then convexly combined with the weight of the previous period using a fixed smoothing coefficient. For example, if the previous period weight was 0.2 and the smoothing coefficient was 0.7, the weighted result is 0.7 multiplied by 0.243 plus 0.3 multiplied by 0.2, which equals 0.2341. Finally, the smoothing weights are truncated according to the upper limit of the weights corresponding to the device safety level. For example, the upper limit of the second-level weight is 0.5. The current weight is 0.2341, which is lower than the upper limit, so there is no need to truncate it. The final restricted adaptive weight vector is the weighted result set of the six conflict factors, such as [0.2341, 0.20, 0.21, 0.22, 0.25, 0.23].

[0060] This constrained adaptive weight vector ensures that data quality, conflict intensity, and temporal continuity are taken into account in the conflict risk assessment of the carbon capture process. At the same time, combined with the differentiated weight upper limit of the device safety level, the risk weight can be dynamically adjusted, so that the assessment results are both real-time and safe. This provides a highly reliable basic weight input for subsequent bounded gain aggregation and risk value calculation.

[0061] The uncalibrated risk value is used for basic risk quantification before the calculation of the coupled risk coefficient. Its generation process consists of three parts: gain constraint, weighted aggregation, and double-threshold shaping, and the final result is formed through convex combination smoothing. The specific calculation steps are as follows: Construction and application of the gain constraint function: Gain constraint functions are established for direction reversal, amplitude asymmetry, phase mismatch, energy impact, sensitivity amplification, and boundary approximation. The gain constraint function adopts a smooth saturation form, that is, the index value of the conflict factor is multiplied by the corresponding gain coefficient as the numerator, and the result of multiplying the index value by the gain coefficient is used as the denominator, forming a fractional expression. The smaller value between the result and the corresponding upper limit of gain is taken, thereby avoiding the nonlinear amplification of the overall risk by a single conflict factor in extreme cases. For example, when the direction reversal is 0.6 and the growth coefficient is 2, the calculation result is 2 multiplied by 0.6 divided by 1 plus 2 multiplied by 0.6, resulting in 0.54. If the upper limit of gain is 0.7, then 0.54 is taken as the direction reversal after gain constraint.

[0062] The values ​​of the growth coefficient and upper limit of gain are determined based on the safety level of the device. For example, at safety level 1, the growth coefficient and upper limit of gain are set to lower values ​​to ensure more conservative risk quantification; at safety level 4, the growth coefficient and upper limit of gain are set to higher values ​​to improve sensitivity to risk signals. The six components of the standardized conflict factor vector are input into the gain limiting function to obtain the gain-limited factor vector, thus providing a basis for subsequent weighted calculations.

[0063] The formation and truncated arithmetic mean of the weighted component set: The gain-limited factor vector is multiplied component by component by the limited adaptive weight vector to obtain the weighted component set. This set reflects the actual contribution of each conflict factor in the comprehensive risk quantification. For example, when the energy impact is 0.5 and the corresponding weight is 0.8, the weighted result is 0.4. A truncated arithmetic mean is performed on the weighted component set, with the truncation ratio taking a fixed value between a preset minimum and maximum ratio, for example, the minimum ratio is 10% and the maximum ratio is 20%. The truncated arithmetic mean is used to eliminate the influence of outliers. For example, when the six weighted components are 0.4, 0.38, 0.42, 0.7, 0.69, and 0.2, if the truncation ratio is 10%, the maximum value of 0.7 and the minimum value of 0.2 are removed, and the average of the remaining four values ​​is taken to obtain the aggregate base value of 0.47.

[0064] Double-threshold shaping and convex combination smoothing: Performs double-threshold shaping and convex combination smoothing on the aggregate base value to ensure stable output of uncalibrated risk values. Low and high threshold settings: The low and high thresholds are fixed values, with the low threshold being less than the high threshold. For example, the low threshold is 0.2 and the high threshold is 0.7. Low threshold interval processing: When the aggregate base value does not exceed the low threshold, a zero value is output. For example, when the aggregate base value is 0.15, the output is 0.0. Threshold interval proportional mapping: When the aggregate base value is between the low and high thresholds, the output value equals the aggregate base value minus the low threshold, divided by the high threshold minus the low threshold. For example, when the aggregate base value is 0.5, the result is (0.5 minus 0.2) divided by (0.7 minus 0.2), which equals 0.6. High-threshold interval processing: When the aggregate base value exceeds the high threshold, the output value is equal to 1 minus the exponential function. The independent variable of the exponential function is the product of the aggregate base value minus the high threshold and a fixed convergence coefficient. For example, if the convergence coefficient is 5 and the aggregate base value is 0.8, the exponential term is 5 multiplied by 0.1, which equals 0.5. The exponential function is approximately 0.607, and the output value is 1 minus 0.607, which equals 0.393. The output of the current period and the output of the previous evaluation period are combined into a convex combination using a fixed smoothing coefficient. For example, if the smoothing coefficient is 0.7, the final output is equal to 0.7 multiplied by the result of the current period plus 0.3 multiplied by the result of the previous period. The smoothed result is limited to the range of 0 to 1, which is the uncalibrated risk value. For example, if the uncalibrated risk value of the previous period is 0.5 and the calculated value of the current period is 0.6, then the final uncalibrated risk value is 0.7 multiplied by 0.6 plus 0.3 multiplied by 0.5, which equals 0.57.

[0065] It should be further explained that the growth coefficient is used to control the response slope of the conflict factor, and it is defined as the rate at which the increase in the conflict factor contributes to the gain. In practical applications, the determination of the growth coefficient must simultaneously consider system sensitivity, control rhythm, and safety level: For safety level 1 (high safety sensitivity): the growth coefficient is recommended to be set to a small value, such as 0.5, to avoid triggering a high-risk assessment with a small increase in the conflict factor value; for safety level 2 or 3: the growth coefficient is recommended to be set to 1 to 1.5 to maintain a moderate linear response; for safety level 4 (low safety sensitivity): the growth coefficient can be set to 2 to 3, allowing for rapid amplification of the conflict index. This coefficient can be fitted to the system disturbance caused by historical conflict events, and the optimal response slope can be found using the minimum mean square error principle.

[0066] The gain cap is used to constrain the maximum impact of a single conflicting factor, preventing an outlier from having an overwhelming effect in weighted aggregation. It is recommended to set the gain cap in relation to the device's tolerance bandwidth or load margin: 0.6 to 0.75 for high-safety-level scenarios (such as supercritical capture devices); 0.8 for medium-safety-level scenarios (such as standard amine liquid capture processes); and 0.9 to 1 for low-safety-level scenarios. The gain cap value can be set with reference to the constraint margin indicators in the system process design (such as the maximum steam load safety threshold or the maximum control loop overshoot ratio) to ensure consistency with physical boundaries.

[0067] The truncation ratio controls the impact of the anomaly weighting factor on the aggregate base value. It is recommended to select a ratio within the following ranges: when system disturbances are frequent or data quality is prone to fluctuations, a truncation ratio of 20% to 30% is suggested; if the system has a stable operating environment, it can be set to 10% to 20%; under extreme conditions (such as high-load operation or switching phases), it can be increased to 40% for noise suppression. This ratio can be dynamically set after statistical analysis of the skewness and kurtosis of the aggregate distribution of conflict factors, or the optimal truncation interval can be deduced by back-calculating the proportion of historical anomalies' contribution to index disturbances.

[0068] Low and high thresholds: The low threshold should be set as the lower limit of acceptable fluctuations in the aggregate base value, with a recommended range of 0.1 to 0.2; the high threshold should be set as the actual risk inflection point of the conflict response, with a recommended range of 0.5 to 0.65. Convergence coefficient: Used to control the rate at which the output value converges to one under high-risk conditions; it is recommended to set it to one to two for high security levels to improve response sensitivity; for medium to low security levels, it is recommended to use 0.5 to one. The convergence coefficient can be determined by fitting the exponential function with respect to the upper limit value through numerical simulation based on the functional relationship between conflict impact delay and response cost.

[0069] Regarding the smoothing coefficient and its corresponding mixing weight in the convex combination, it is used for the weighted fusion of current and historical risk values. Recommendations: The default range should be set to 0.6 to 0.8; when operating conditions fluctuate drastically, the smoothing coefficient can be appropriately increased to maintain response stability; during steady-state operation, it can be reduced to 0.5 to enhance sensitivity. This coefficient can be automatically adjusted by constructing a function model using the standard deviation of historical risk value changes and the frequency of process variable changes. All parameter settings have clear engineering correspondences and can be obtained through industry experience, industrial data backtesting analysis, or controller parameter tuning methods.

[0070] The uncalibrated risk value is looked up according to a preset uncalibrated risk-coupled risk lookup table to obtain the corresponding coupled risk coefficient. Based on the preset coupled risk coefficient and the corresponding graded interval, disposal suggestions are output to guide the dynamic control of the carbon capture process control system. The specific process is as follows:

[0071] A mapping table for uncalibrated risk and coupled risk is established: First, based on a large amount of historical operational data, accident records, and intervention effects, a mapping relationship between uncalibrated risk values ​​and coupled risk coefficients is established. Uncalibrated risk values ​​are continuous values ​​within the range of zero to one, and coupled risk coefficients are also continuous values ​​within the range of zero to one. The mapping table is formed by combining piecewise linear interpolation and nonlinear correction. For example, uncalibrated risk values ​​are divided into low-risk intervals (0 to 0.3), medium-risk intervals (0.3 to 0.7), and high-risk intervals (0.7 to 1). Within each interval, historical data is used to fit the response curve of the coupled risk coefficient. In the low-risk interval, an approximate linear mapping is used; for example, when the uncalibrated risk value is 0.2, the coupled risk coefficient is 0.21. In the medium-risk interval, a nonlinear amplification factor is introduced; for example, when the uncalibrated risk value is 0.5, the coupled risk coefficient is 0.6. In the high-risk interval, a saturated convergence curve is used; for example, when the uncalibrated risk value is 0.9, the coupled risk coefficient is 0.95. This lookup table is dynamically updated through a combination of offline training and online verification. When new accident cases or intervention effect data are added, the lookup table can be adaptively adjusted to ensure that the mapping of the coupling risk coefficient is consistent with the current operating characteristics of the device.

[0072] Based on the coupling risk coefficient, the risk level is determined by comparing the obtained coupling risk coefficient with the preset risk level range, and a basis is provided for generating subsequent disposal suggestions.

[0073] Risk grading intervals are divided into ranges from zero to one. For example: Low risk interval: 0 to 0.3, indicating the device is in normal operation and only requires maintaining the current control strategy; Medium risk interval: 0.3 to 0.7, indicating the device has potential fluctuation risks and requires parameter pre-adjustment or load balancing; High risk interval: 0.7 to 1, indicating significant risks to device operation and requiring the triggering of preventative control measures or mandatory intervention strategies. For example, when the uncalibrated risk value is 0.6, the coupling risk coefficient obtained from the lookup table is 0.65, which falls within the medium risk interval. The system will then mark the risk level as medium and invoke the corresponding medium-risk handling strategy.

[0074] Output disposal recommendations and implement dynamic control: Based on the risk level and coupled risk coefficient, generate disposal recommendations for the carbon capture process and transmit the disposal recommendations to the carbon capture process control system to guide it in implementing dynamic control.

[0075] The recommended measures include: when the risk level is low, it is recommended to keep the current operating parameters unchanged and only conduct periodic monitoring, such as maintaining the current absorber temperature at 120 degrees Celsius and the regeneration tower pressure at 0.2 MPa; when the risk level is medium, it is recommended to make slight adjustments, such as reducing the absorber liquid circulation flow rate by 5% or reducing the reboiler heat load of the regeneration tower by 3%, in order to reduce the possibility of subsequent risk escalation. When the risk level is high, it is recommended to take immediate mandatory intervention measures, such as reducing the steam flow rate of the carbon capture unit by 20% or directly triggering the bypass mode, to avoid equipment damage or safety accidents due to excessive risk. For example, when the coupled risk coefficient calculated by the system is 0.8, the automatically generated response suggestion is to reduce the heat load of the regeneration tower by 10% and simultaneously reduce the solvent flow rate by 5%. This adjustment instruction is issued to the field execution unit through the process control system, and the parameter adjustment is completed within two minutes, ensuring that the risk falls back from the high-risk range to the medium-risk range. This step, through the mapping between uncalibrated risk values ​​and coupled risk coefficients, makes the risk quantification process interpretable and executable, ensuring that the carbon capture process control system can dynamically adjust operating parameters according to the actual risk level, effectively improving the safety and stability of the system.

[0076] Example 2: Intelligent diagnostic control system for carbon capture process based on AI decision-making, specifically including: The alignment and unification module acquires control intent data from the edge computing side and the cloud side within the same time period, aligns the data time base of the control intent and unifies the units of measurement, and generates a synchronous control sequence for comparative analysis. The conflict quantification module calculates six types of conflict factors based on synchronous comparison sequences. These six types of conflict factors include direction reversal, amplitude asymmetry, phase mismatch, energy impact, sensitivity amplification, and boundary approximation. Direction reversal measures the proportion of opposite angles and signs of commands; amplitude asymmetry measures the relative difference in adjustment amplitudes on both sides; phase mismatch measures time center offset and period offset; energy impact measures the intensity of expected power step or steam load step; sensitivity amplification measures the amplification effect on key process quantities under current operating conditions; and boundary approximation measures the reverse value of the remaining margin of preset constraints. The results are then summarized and standardized to obtain a standardized conflict factor vector. The weight generation module calculates an adaptive weight vector based on data freshness, source credibility, and conflict persistence, and sets a weight upper limit according to the device security level to obtain a restricted adaptive weight vector. The gain aggregation module uses bounded gain aggregation to weight and superimpose the standardized conflict factor vector and the restricted adaptive weight vector to obtain the uncalibrated risk value. The mapping and handling module looks up the uncalibrated risk value according to the preset uncalibrated risk-coupled risk lookup table, obtains the corresponding coupled risk coefficient, and outputs handling suggestions based on the preset coupled risk coefficient corresponding to the graded interval, guiding the carbon capture process control system to implement dynamic regulation.

[0077] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0078] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0079] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0080] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0081] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for AI decision based intelligent diagnostic control of carbon capture process, characterized in that, The method comprises the following steps: Step one, obtain the control intention data of the edge computing side and the cloud side in the same time period, align the data time base of the control intention and unify the measurement units, and generate a synchronous control sequence for comparative analysis; Step two, calculate six types of conflict factors based on the synchronous control sequence, including direction reversal degree, amplitude asymmetry degree, phase mismatch degree, energy impact degree, sensitive amplification degree and boundary approximation degree, wherein the direction reversal degree is used to measure the proportion of instruction direction angle and sign opposite, the amplitude asymmetry degree is used to measure the relative difference of adjustment amplitude on both sides, the phase mismatch degree is used to measure the time center offset and period offset, the energy impact degree is used to measure the expected power step or steam load step strength, the sensitive amplification degree is used to measure the amplification effect of key process quantity under current working condition sensitivity, and the boundary approximation degree is used to measure the reverse value of the remaining margin of the preset constraint; then, the standardization processing is performed to obtain a standardized conflict factor vector; Step three, calculate an adaptive weight vector according to data freshness, source credibility and conflict duration, and set a weight upper limit according to the device safety level to obtain a limited adaptive weight vector; Step four, perform weighted superposition on the standardized conflict factor vector and the limited adaptive weight vector by using a bounded gain aggregation to obtain an uncalibrated risk value; Step five, look up the uncalibrated risk value according to a preset uncalibrated risk-coupling risk correspondence table to obtain a corresponding coupling risk coefficient, and output a handling suggestion based on a preset grading interval corresponding to the coupling risk coefficient to guide the carbon capture process control system to implement dynamic regulation and control.

2. The AI decision based intelligent diagnostic control method for carbon capture process according to claim 1, wherein, The synchronous control sequence is obtained by the following steps: First, identify valve steps, pump speed transitions, set value changes and alarm confirmations on the edge computing side and the cloud side through preset event determination rules, and filter effective anchor points by carbon element conservation and energy balance two types of invariants, solve the piecewise linear time mapping with global non-decreasing and slope upper and lower bounds for the anchor point set passing through the constraints, so as to obtain the initial estimation of clock drift and communication delay; Implement limited dynamic time warping and phase correction on the initial time mapping to maintain the monotonicity of the mapping function and limit the local slope, and add a small phase item within a limited lag range according to the cross-correlation peak, thereby unifying the aligned time base and the aligned residual sequence containing mean, variance and maximum deviation; Based on the process property relationship, complete the state-dependent conversion between volume flow and mass flow, steam flow and heat load, and different concentration expressions, unify the measurement units, and add source credibility, version summary identifier and integrity check summary to each time point, and finally form a synchronous control sequence carrying complete metadata.

3. The AI decision based intelligent diagnostic control method for carbon capture process according to claim 2, wherein, In the process of forming the synchronous control sequence carrying complete metadata, the density and enthalpy value are calculated according to the temperature, pressure and solvent composition, the volume flow is converted into mass flow, the steam flow is converted into heat load, the mass fraction and volume fraction are converted into mole fraction, and the pressure, temperature and concentration are unified to the pre-designed measurement unit system; The source credibility is calculated by the latest calibration date of the instrument, the online self-checking result and the missing rate; The version summary identifier is composed of the model name, the model version and the training time; The integrity check summary is composed of the deterministic summary string of the configuration parameter list, and the synchronized control sequence carrying the metadata is generated after summarization.

4. The AI decision based intelligent diagnostic control method for carbon capture process according to claim 3, wherein, The specific steps of step two include: Based on the synchronized control sequence, the control intention increments of the edge computing side and the cloud side in the same control period are extracted, and the paired data sets of the two sides in different preset quantization increments are obtained according to the control channels, and the rated range, the rated energy capacity and the constraint limit value of each channel are recorded; Six types of conflict factors are calculated on the paired data set, which are: The calculation of the direction reverse degree is to calculate the direction angle of the increments of the same channel on the two sides first, and the direction angle takes the angle value from zero to one hundred and eighty degrees, then the proportion of the channels with opposite signs in all channels is calculated, and the direction reverse degree is equal to the weighted average of the direction angle divided by one hundred and eighty and the proportion; The calculation of the amplitude asymmetry degree is to take the absolute value of the difference between the absolute values of the increments on the two sides of the same channel, and then divide by the sum of the absolute values of the increments on the two sides plus the pre-set minimum positive lower limit value as the denominator, and finally take the arithmetic average of all channels; The calculation of the phase mismatch degree is to calculate the cross-correlation function of the control intentions on the two sides within the specified maximum time lag range, take the absolute value of the time lag that produces the maximum correlation and normalize it with the maximum time lag, and at the same time, calculate the time center difference of the control intentions on the two sides and normalize it with the control period of the channel, and the phase mismatch degree is equal to the weighted average of the two normalized results; The calculation of the energy impact degree is to convert the control intention increments of each channel into expected electric power steps and expected heat load steps through the energy conversion coefficient under the current working condition, take the absolute value of the sum of the two, and normalize it with the sum of the rated electric power and the rated heat load of the corresponding unit; The calculation of the sensitive amplification degree is to obtain the local first-order sensitivity coefficients of the key process quantities to each input channel at the current operating point, multiply the control intention increments of each channel by the corresponding sensitivity coefficients, and then sum them up to get the expected change amplitude of each key process quantity, take the maximum value of all key process quantities and normalize it with the allowed change bandwidth of the process quantity; The calculation of the boundary approximation degree is to calculate the remaining margin of each controlled constraint after executing the control intentions on the two sides, if the constraint is an upper limit, subtract the expected value from the upper limit, if the constraint is a lower limit, subtract the lower limit from the expected value, and negative values are counted as zero, then normalize it with the safety bandwidth of the corresponding constraint and take one minus the normalized result, finally take the maximum value of all constraints as the boundary approximation degree.

5. The AI decision based intelligent diagnostic control method for carbon capture process according to claim 3, wherein, The standardized conflict factor vector refers to: The six types of conflict factors calculated based on the synchronized control sequence are subjected to a pre-set standardization process, each conflict factor is mapped to a standardized range of zero to one, and then the standardized conflict factor vector is obtained by summarizing according to the pre-set arrangement order.

6. The AI decision based intelligent diagnostic control method for carbon capture process according to claim 5, wherein, The restricted adaptive weight vector is obtained by the following steps: The data freshness score is calculated as follows: in a unified time reference, the time lag value, the communication delay value and the missing proportion are calculated for each data channel of the six types of conflict factors, the time lag value is the current unified time minus the latest valid timestamp of the channel, the communication delay value is the communication delay estimation obtained by the aforementioned time mapping, and the missing proportion is the proportion of invalid samples in all samples within a fixed evaluation window; the timeliness component is obtained by linearly mapping the time lag value in reverse according to a preset maximum acceptable time difference, the channel delay component is obtained by linearly mapping the communication delay value in reverse according to a preset maximum acceptable delay, the integrity component is obtained by taking the missing proportion according to a minus relationship, and the geometric mean of the timeliness component, the channel delay component and the integrity component is taken to obtain a data freshness score in the range of zero to one; The conflict duration is calculated as follows: in the same evaluation window, the instantaneous conflict indicator is constructed, the instantaneous conflict indicator is equal to the maximum value of the structural conflict strength and the amplification strength multiplied by the boundary strength, the structural conflict strength is the arithmetic mean of the direction reversal degree, the amplitude asymmetry degree and the phase mismatch degree, the amplification strength is the arithmetic mean of the energy impact degree and the sensitive amplification degree, and the boundary strength is the boundary approximation degree; The conflict threshold is set, and the time period during which the indicator continuously exceeds the threshold is identified, the time proportion of the time during which the threshold is exceeded is calculated, and the average value of the indicator when the threshold is exceeded is calculated, the product of the time proportion and the average value is taken as the unweighted conflict duration, and the conflict duration in the range of zero to one is obtained by weighting according to a linearly increasing weight from near to far; The restricted adaptive weight vector is formed as follows: for each component of the six types of conflict factors, the source credibility weight is multiplied by the data freshness score, and the result is multiplied by an exponential function of the conflict duration, the exponential is given by the sensitive coefficient corresponding to the device security level, and the sensitive coefficient is determined according to the device security level, the device security level is one, two, three or four, and the weight upper limit is set according to the device security level, the weight upper limit of one, two, three and four is determined in the order from small to large, and finally the restricted adaptive weight vector is output.

7. The AI decision based intelligent diagnostic control method for carbon capture process according to claim 5, wherein, The uncalibrated risk value is obtained by the following steps: Gain limiting functions are established for the direction reversal degree, the amplitude asymmetry degree, the phase mismatch degree, the energy impact degree, the sensitive amplification degree and the boundary approximation degree, the gain limiting function adopts a smooth saturation form, that is, the ratio of the gain coefficient multiplied by the index value as the numerator and one plus the coefficient multiplied by the index value as the denominator, and the smaller value is taken with the corresponding gain upper limit, the determined values of the growth coefficient and the gain upper limit of each factor are preset according to the device security level; The standardized conflict factor vector is input into the gain limiting function to obtain a gain limited factor vector, and multiplied by the restricted adaptive weight vector to form a weighted component set, and the aggregated base value is obtained by performing a truncated arithmetic mean on the weighted component set, and the truncation proportion is a fixed value between the preset minimum proportion and the maximum proportion; The aggregated base value obtained from the truncated arithmetic average result is subjected to double-threshold shaping and convex combination smoothing to obtain the uncalibrated risk value.

8. The AI decision based intelligent diagnostic control method for carbon capture process according to claim 7, wherein, The step of performing double-threshold shaping and convex combination smoothing on the aggregated base value obtained from the truncated arithmetic average result comprises: setting a low threshold and a high threshold, the low threshold and the high threshold being fixed values and the low threshold being smaller than the high threshold; outputting a zero value when the aggregated base value does not exceed the low threshold; outputting a proportional value equal to the aggregated base value minus the low threshold, divided by the high threshold minus the low threshold when the aggregated base value is between the low threshold and the high threshold; outputting a value equal to one minus an exponential function when the aggregated base value exceeds the high threshold, the argument of the exponential function being the aggregated base value minus the product of the high threshold and a fixed convergence coefficient, the fixed convergence coefficient being preset according to the safety level of the device; performing convex combination on the current period output and the output of the previous evaluation period with a fixed smoothing coefficient, the fixed smoothing coefficient being a determined value in the range of zero to one, and limiting the convex combination result in the range of zero to one to obtain the uncalibrated risk value.

9. The AI decision based intelligent diagnostic control system for carbon capture process according to any one of claims 1-8, characterized in that, Specifically comprising: an alignment and unification module, which acquires control intention data of the edge computing side and the cloud side in the same time period, aligns the data time reference of the control intention, and unifies the units of measurement, to generate a synchronous control sequence for comparative analysis; a conflict quantification module, which calculates six types of conflict factors based on the synchronous control sequence, the six types of conflict factors including direction reversal degree, amplitude asymmetry degree, phase mismatch degree, energy impact degree, sensitive amplification degree, and boundary approximation degree, wherein the direction reversal degree is used to measure the proportion of instruction direction angle and sign reversal, the amplitude asymmetry degree is used to measure the relative difference of adjustment amplitude on both sides, the phase mismatch degree is used to measure the time center offset and period offset, the energy impact degree is used to measure the expected power step or steam load step strength, the sensitive amplification degree is used to measure the amplification effect on key process quantities under the current working condition sensitivity, and the boundary approximation degree is used to measure the reverse value of the remaining margin of the preset constraint; then performing summary and standardization processing to obtain a standardized conflict factor vector; a weight generation module, which calculates an adaptive weight vector according to data freshness, source credibility, and conflict duration, and sets a weight upper limit according to the safety level of the device to obtain a restricted adaptive weight vector; a gain aggregation module, which performs bounded gain aggregation on the standardized conflict factor vector and the restricted adaptive weight vector to obtain the uncalibrated risk value; a mapping and treatment module, which looks up the corresponding coupling risk coefficient according to a preset uncalibrated risk-coupling risk correspondence table based on the uncalibrated risk value, and outputs a treatment suggestion based on the preset classification interval corresponding to the coupling risk coefficient to guide the dynamic regulation and control of the carbon capture process control system.

Images