Multi-stage cooperative intelligent control and equipment operation and maintenance data management platform for chemical production

Abstract

The invention relates to the technical field of industrial automation control, and discloses a chemical production-oriented multi-stage cooperative intelligent control and equipment operation and maintenance data management platform, which comprises a multi-source data acquisition module for acquiring multi-source heterogeneous data and performing time alignment processing and quality verification; the working condition identification and feature extraction module is used for obtaining equipment operation state data by adopting weighted fusion and principal component analysis; the equipment health assessment module is used for calculating an equipment health degree score and predicting a degradation trend through exponential smoothing; the self-adaptive control module is used for designing a multivariable coordination controller and dynamically generating operation constraints of the multivariable coordination controller according to the equipment health degree score; the layered optimization control module is used for establishing an interlayer feedback mechanism; the collaborative decision-making module generates a collaborative decision-making trigger signal according to the equipment health degree information and generates a collaborative decision-making scheme; according to the invention, by establishing a collaborative decision-making mechanism, a decision-making scheme considering the production task completion degree, the equipment health degree and the maintenance cost is generated.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control technology, and more specifically, to a multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production. Background Technology

[0002] Chemical production processes are highly complex, strongly coupled, and subject to stringent safety and environmental protection requirements, posing significant challenges to production control and equipment operation and maintenance management. In core units such as continuous catalytic cracking in petrochemical enterprises, multiple operating parameters of the reaction regeneration system exhibit complex coupling relationships. Frequent changes in operating conditions necessitate the coordinated adjustment of multiple parameters, which traditional single-loop control methods struggle to handle. Simultaneously, critical rotating equipment such as main fans and flue gas fans operate under prolonged high-temperature and high-load conditions, leading to frequent equipment vibration and bearing temperature issues. Unplanned shutdowns of any critical piece of equipment can cause the entire unit to cease production, resulting in substantial economic losses.

[0003] Currently, chemical enterprises generally deploy multiple information systems, such as DCS control systems, equipment vibration monitoring systems, video surveillance systems, and equipment management systems. However, these systems operate independently, creating significant data silos. When performing optimization control, production control systems only consider process and product quality constraints, neglecting equipment health status as an optimization constraint. This can lead to optimization control requiring equipment to operate under extreme conditions for extended periods, accelerating equipment degradation. Furthermore, equipment maintenance departments lack access to production schedules when developing maintenance plans, potentially disrupting production continuity. Production control decisions and equipment maintenance decisions are independent, resulting in decision conflicts.

[0004] Existing equipment monitoring systems rely on fixed threshold alarms, which cannot adapt to the frequent changes in operating conditions of chemical equipment. For example, a high vibration level in a main fan during full-load operation is normal, but the same vibration level at low load may indicate a malfunction. Fixed thresholds cannot distinguish this difference, leading to numerous false alarms under high-load conditions and potential missed faults under low-load conditions. Equipment health is influenced by various factors, including process operating parameters, vibration monitoring data, temperature data, lubricating oil analysis data, and historical maintenance records. Existing systems store this data in a scattered manner, lacking in-depth fusion and analysis, making it difficult to comprehensively and accurately assess equipment health. Therefore, there is an urgent need for a technical solution that enables deep integration of control and maintenance systems, multi-level collaborative control, and condition-adaptive equipment health management to comprehensively improve the level of intelligence in chemical production. Summary of the Invention

[0005] This invention provides a multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production, which solves the technical problems of the separation between the control system and the operation and maintenance system, poor coordination of multi-level control, and inaccurate equipment health assessment in related technologies in chemical production.

[0006] This invention provides a multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production, including: The multi-source data acquisition module establishes a unified time base based on edge computing nodes, collects heterogeneous data from multiple sources, performs time alignment processing and quality verification, and obtains a time-synchronized data stream. The working condition identification and feature extraction module uses a sliding window to identify working conditions based on time-synchronized data streams, extracts equipment features through frequency domain analysis and temperature trend analysis, and obtains equipment operating status data through weighted fusion and principal component analysis. The equipment health assessment module calculates the equipment health score based on equipment operating status data and predicts the deterioration trend through exponential smoothing to obtain equipment health information; The adaptive control module designs a multivariable coordinated controller based on equipment health information, and dynamically generates the operating constraints of the multivariable coordinated controller according to the equipment health score to obtain the set values ​​of the operating parameters; The hierarchical optimization control module executes optimization control, coordinated control, and PID control respectively based on the set values ​​of operating parameters, establishes an inter-level feedback mechanism, and derives multi-level coordinated control commands. The collaborative decision-making module, based on equipment health information and multi-level coordinated control commands, generates collaborative decision-making trigger signals according to equipment health information, and uses a rule matching method to generate the final collaborative decision-making scheme.

[0007] In a preferred embodiment, the time alignment process includes: Acquire multi-source heterogeneous data collected by edge computing nodes, establish a data buffer for the multi-source heterogeneous data based on a unified time base, aggregate all data with the same timestamp within the same time unit, and eliminate time deviations between data sources through data aggregation processing to obtain a time-unified multi-source data set.

[0008] In a preferred embodiment, the acquisition of multi-source heterogeneous data includes: The system acquires high-frequency vibration data, process operating parameter data, and equipment temperature data from chemical production equipment. The process operating parameter data and equipment temperature data are directly assigned to the corresponding time units according to the timestamp. The high-frequency vibration data is processed using a sliding window statistical method to calculate statistical characteristic values. The statistical characteristic values ​​are then aligned with the time base to form a multi-source data stream with a unified time format.

[0009] In a preferred embodiment, the operating condition identification includes: The process operation parameter data in the time-synchronized data stream is acquired. The difference between the process operation parameter data and different operating conditions is calculated based on historical operating data. The difference is measured and calculated using the variance analysis method. The parameter with the largest difference is selected as the operating condition sensitive parameter based on the difference measurement result for subsequent operating condition status judgment.

[0010] In a preferred embodiment, the use of a sliding window for operating condition identification includes: The system acquires real-time values ​​of operating condition sensitive parameters, performs time-series analysis on these parameters using a sliding window method, calculates the changing trend of the operating condition sensitive parameters within the window, and calculates the slope using a linear regression method to characterize the changing trend. When the absolute value of the slope of the changing trend is greater than a preset threshold, it is determined that the operating condition switch has started, and an operating condition status indicator is output.

[0011] In a preferred embodiment, the weighted fusion includes: The equipment features extracted from frequency domain analysis and temperature trend analysis are obtained. Based on the operating condition identification results, an operating condition feature weight mapping table is established. According to the current operating condition type, the corresponding feature weight coefficient is queried from the operating condition feature weight mapping table. Each component of the equipment feature is multiplied by the corresponding weight coefficient, and a weighted comprehensive state feature vector is obtained through weight fusion processing.

[0012] In a preferred embodiment, the generation of the device health information includes: Obtain normal operating data of equipment under different operating conditions from the historical operating database. Calculate the statistical characteristic parameters of the equipment status feature vector sample set under each operating condition. Use the statistical characteristic parameters as the feature baseline of the normal state of the equipment under this operating condition. Establish an operating condition baseline mapping library to store the feature baseline for equipment health score calculation.

[0013] In a preferred embodiment, the multivariate coordination controller includes: The equipment health level is obtained from the equipment health score calculation. Based on the equipment health level, the equipment load rate constraint range is determined. When the health level is high, a wider load rate constraint range is set to improve production efficiency. When the health level is low, a stricter load rate constraint range is set to protect equipment safety, thus obtaining dynamic equipment status constraint parameters.

[0014] In a preferred embodiment, it further includes: The cloud-edge-device collaborative data management module manages cloud-edge-device collaborative data based on the final collaborative decision-making scheme. It achieves data management and model updates through data collection, edge processing, cloud storage, offline model training, and model deployment.

[0015] In a preferred embodiment, a computer-readable storage medium is provided for storing computer-readable instructions that, when read by a computer, enable the operation of the multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production as described above.

[0016] The beneficial effects of this invention are as follows: By establishing a unified data acquisition and fusion platform, multi-source heterogeneous data such as process operation data from the DCS system, equipment status data from the vibration monitoring system, image data from the video monitoring system, and maintenance records from the equipment management system are time-aligned and deeply fused, breaking down data silos; in multivariate coordinated control, equipment health information is used as a constraint condition, and the allowable range of equipment load rate is dynamically adjusted according to the equipment health level to ensure that control optimization does not exacerbate equipment deterioration; by establishing a control and maintenance collaborative decision-making mechanism, when the equipment health deteriorates, a collaborative decision-making program is automatically triggered to generate a decision scheme that takes into account the completion of production tasks, equipment health, and maintenance costs, and the control strategy adjustment and maintenance plan arrangement are uniformly optimized, realizing the unification of production optimization and equipment protection, and fundamentally solving the decision conflict problem caused by the independence of control decisions and maintenance decisions in the traditional mode; This invention addresses the variable operating conditions of chemical equipment by proposing an adaptive equipment health assessment method based on operating condition identification. It rapidly identifies operating condition switching states using a sliding window method, accurately identifies specific operating condition types using cluster analysis, establishes a characteristic baseline for the normal equipment state for each operating condition type, and calculates the deviation between the current state and the corresponding operating condition baseline using Mahalanobis distance, converting the deviation into a health score. This method accurately distinguishes between operating condition changes and equipment degradation, avoiding false alarms and missed alarms caused by operating condition changes in traditional fixed threshold methods. By fusing vibration, temperature, and operational characteristics, a weighted fusion method is used to construct a comprehensive equipment status feature, fully reflecting the equipment's health condition. Exponential smoothing and linear extrapolation methods are used to predict future equipment degradation trends, enabling early identification of equipment health risks, providing sufficient preparation time for preventative maintenance, and effectively reducing the risk of unplanned equipment shutdowns. Attached Figure Description

[0017] Figure 1 This is a module diagram of the multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production in this invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production, such as... Figure 1 As shown, it includes: The multi-source data acquisition module establishes a unified time base based on edge computing nodes, collects heterogeneous data from multiple sources, performs time alignment processing and quality verification, and obtains a time-synchronized data stream. The purpose of this module is to collect heterogeneous data from multiple systems in the chemical production site and solve problems such as data time synchronization and data quality through edge preprocessing, so as to provide a reliable data foundation for subsequent analysis.

[0020] Based on edge computing nodes deployed at the chemical production site, a unified time reference server is established using the Network Time Protocol (NTP) to obtain a standard time reference at the plant level. Specifically, an edge server is deployed as the time reference server in the plant's central control room. This server acquires satellite time signals via a GPS antenna and synchronizes its own clock with UTC time, achieving a time accuracy down to the microsecond level. Based on this time reference server, time synchronization signals are sent to various data acquisition systems, including the DCS system, vibration monitoring system, video surveillance system, and equipment management system, using the NTP. Each system adjusts its own clock upon receiving the synchronization signal, ensuring that the time error of all data acquisition systems is less than 10 milliseconds, resulting in a time-unified data acquisition network.

[0021] Based on a time-unified data acquisition network, the DCS system's data acquisition interface collects process operating parameters from the reaction regeneration system, including reactor temperature, reactor pressure, catalyst circulation rate, regenerator temperature, regenerator oxygen content, and regeneration air volume. Simultaneously, it collects process parameters from the fractionation system, such as fractionation column top temperature, reflux ratio, and side stream extraction rate. The sampling frequency is set to 1 second, resulting in a process operating parameter data stream. At each sampling moment, the DCS system reads the process variable values ​​at each measuring point and obtains the current timestamp from the unified time reference server. The process variable values ​​and timestamps are packaged into a data frame. The data frame format is: measuring point identifier, value, unit, timestamp, and quality flag. The data frame is then sent to the edge computing node via industrial Ethernet.

[0022] Based on a time-unified data acquisition network, vibration signals from equipment such as the main fan bearing, flue gas fan bearing, and reactor agitator are collected through the data interface of the vibration monitoring system at a sampling frequency of 100 milliseconds, resulting in a high-frequency vibration raw data stream. Simultaneously, temperature signals from the main fan bearing, flue gas fan bearing, and main fan motor are collected through the temperature monitoring system at a sampling frequency of 1 second, resulting in an equipment temperature data stream. Based on the high-frequency vibration raw data stream, a sliding window statistical processing method is employed, with a window length of 1 second and a window sliding step of 1 second. Statistical measures are calculated for 10 sampling points within each window. Specifically, the mean, standard deviation, maximum value, and minimum value of the vibration signal within the window are calculated to obtain vibration statistical characteristics. These vibration statistical characteristics are correlated with the timestamp of the center moment of the window to obtain a 1-second frequency vibration characteristic data stream aligned with the sampling frequency of the process parameters. The equipment temperature data stream and the vibration characteristic data stream are merged to obtain the equipment status data stream.

[0023] Based on a time-unified data acquisition network, video image data from the main fan operating area, flue gas fan operating area, and reactor area are collected through an industrial video monitoring system. The video resolution is 1920×1080, and the frame rate is 25 frames per second, resulting in a raw video data stream. Based on this raw video data stream, intelligent image analysis methods are used for real-time processing. Specifically, video frames are preprocessed, including noise reduction and enhancement, to obtain enhanced video frames. Then, target detection methods are used to identify key areas in the video, including equipment flange connections, valve seals, and equipment exhaust ports, to determine if there are any abnormal phenomena such as leaks, smoke, or discoloration. When an anomaly is detected, feature descriptions of the abnormal area are extracted, including the anomaly type, anomaly location coordinates, and anomaly severity score. These feature descriptions are then correlated with the current timestamp to obtain video anomaly event data. For normally operating videos, only keyframes are extracted (one frame every 10 seconds) for storage, without retaining the complete video stream, resulting in a compressed video feature data stream.

[0024] Based on the database interface of the equipment management system, historical maintenance records for each piece of equipment are extracted. Specifically, this includes maintenance records such as the time of each maintenance, maintenance type, maintenance content, list of replaced parts, and post-maintenance trial operation records. Simultaneously, lubricating oil analysis records are extracted, including physicochemical indicators such as lubricating oil sampling time, viscosity, moisture content, and metal particle content. This structured data is then organized according to equipment number and time sequence to obtain the equipment historical maintenance dataset.

[0025] Based on process operation parameter data streams, equipment status data streams, and video feature data streams, a time alignment processing method is used for data fusion. Specifically, a data buffer is established, using 1 second as the base time unit, and all data within the same time unit (with the same integer second timestamp value) are aggregated. For process operation parameter data and equipment temperature data, since the sampling frequency is 1 second, they are directly assigned to the corresponding time unit according to the timestamp. For equipment vibration feature data, since it has already undergone frequency downsampling and is aligned with the 1-second time base, it is also directly assigned to the corresponding time unit according to the timestamp. For video feature data, detected abnormal events are assigned to the corresponding time unit according to the event occurrence timestamp. If a certain type of data is missing within the same time unit, a forward padding method is used to fill it with data from the previous time unit. After time alignment processing, a time-synchronized multi-source fused data stream is obtained, where each time unit contains all types of data at that moment.

[0026] Based on the obtained time-synchronized multi-source fused data stream, a data quality verification method is used to identify and process outliers. The specific processing procedure is as follows: Data abrupt changes are detected. For each process variable, the difference between the current value and the previous value is calculated. Simultaneously, the mean and standard deviation of the normal rate of change for that variable over the past hour are calculated. When the difference exceeds the mean of the normal rate of change plus three times the standard deviation, it is marked as a suspected abrupt change. For marked suspected abrupt changes, secondary confirmation is performed to check whether other variables related to that variable have also undergone synchronous abrupt changes. For example, when a suspected abrupt change occurs in the reactor temperature, related variables such as catalyst circulation rate and feed rate are checked. If related variables also undergo synchronous abrupt changes, it is determined to be a genuine change in operating conditions, and the abrupt change data is retained. If related variables do not undergo abrupt changes, it is determined to be an outlier caused by measurement anomalies or communication failures. The outlier is corrected using a linear interpolation method, that is, the linear interpolation result of normal data points before and after the suspected abrupt change point is used to replace the outlier. After data quality verification and outlier processing, a time-synchronized data stream is obtained.

[0027] Furthermore, deep learning-based data quality assessment methods can replace traditional statistical methods. Specifically, a data quality assessment model is constructed using a Long Short-Term Memory (LSTM) network. This model takes the sequence of process variable values ​​from the past 10 time units as input and predicts the normal range of values ​​at the current moment. When the actual collected value falls outside the predicted range, it is judged as an outlier. This method can learn the normal fluctuation patterns of variables under different operating conditions, improve the accuracy of outlier identification, and reduce false positives and false negatives. The advantage of this alternative embodiment is that it can adapt to the nonlinear and time-varying characteristics of chemical processes and has better adaptability to scenarios with frequently changing operating conditions.

[0028] The multi-source data acquisition module outputs a time-synchronized data stream containing information such as process parameters, equipment status, and video features. The data timestamps are unified, ensuring reliable quality and allowing for direct use in subsequent analysis.

[0029] The working condition identification and feature extraction module uses a sliding window to identify working conditions based on time-synchronized data streams, extracts equipment features through frequency domain analysis and temperature trend analysis, and obtains equipment operating status data through weighted fusion and principal component analysis. The purpose of this module is to identify the current operating conditions of the device, extract the status characteristics of the device under these conditions, and provide accurate feature input for device health assessment.

[0030] Based on time-synchronized data streams, a rapid operating condition identification method is employed to determine the unit's operating status. Sensitive operating parameters are selected by analyzing historical operating data to identify the most sensitive characteristic parameters to changes in operating conditions. Specifically, the method involves calculating the degree of difference between various process parameters under different operating conditions, measuring this difference using analysis of variance, and selecting the parameter with the largest difference as the sensitive operating condition parameter. For the catalytic cracking unit, the selected sensitive operating conditions parameters include reactor temperature, feed rate, and catalyst circulation rate, which exhibit certain differences under different load conditions. Based on these sensitive operating parameters, a sliding window method is used for rapid operating condition identification. The sliding window length is set to 3 minutes, and the sliding step size is 1 minute. For each window, the changing trend of the sensitive operating parameters within the window is calculated. The changing trend is characterized by the slope of linear regression. When the absolute value of the slope of the feed rate change trend exceeds a set threshold (the slope corresponding to a feed rate change exceeding 5% within 3 minutes), it is determined that an operating condition switch has begun, and the current operating condition state is marked as a transition state, resulting in an operating condition transition state identifier.

[0031] Based on rapid identification results and time-synchronized data streams, a precise operating condition classification method is employed to identify specific operating condition types. A precise identification window length of 10 minutes is set; once a transitional state is determined, data is continuously collected for a 10-minute window. Based on the 10-minute window data, operating condition feature vectors are extracted. These feature vectors include statistical values ​​of process parameters such as average reactor temperature, average feed rate, average catalyst circulation rate, average regenerator temperature, and average reaction pressure. Cluster analysis is used to classify operating conditions. An operating condition classification model is established based on historical operating data, dividing the unit's operating conditions into six typical types: start-up, low-load operation, rated-load operation, high-load operation, shutdown, and feedstock switching. Each operating condition type is represented by the central value of its feature vector. For the feature vector extracted in the current 10-minute window, its Euclidean distance to the central values ​​of each operating condition type is calculated. The operating condition type with the smallest distance is the current precise operating condition type, thus obtaining the precise operating condition classification identifier. Once the precise operating condition is identified, it is determined whether the precise operating condition is consistent with the operating condition type before entering the transition state. If they are consistent, the operating condition state is restored from the transition state to the stable state. If they are inconsistent, it is confirmed that the operating condition has been switched to the new operating condition type, and the operating condition state is updated to the stable state of the new operating condition.

[0032] Furthermore, a support vector machine (SVM)-based work condition classification method can be used instead of clustering analysis. The SVM method, by constructing classification decision boundaries, can handle the nonlinear boundary problems between work conditions. Specifically, a multi-class SVM model is trained using historically labeled work condition data. The model input is the work condition feature vector, and the output is the work condition type label. During online recognition, the feature vector of the current window is input into the trained SVM model, and the work condition type is directly output without calculating distance. The purpose of this alternative embodiment is to improve the accuracy of work condition classification, especially for work conditions with unclear boundaries.

[0033] Based on precise operating condition classification and time-synchronized data streams, frequency domain analysis is employed to extract equipment vibration characteristics. For the main fan bearing vibration signal, the obtained 1-second window vibration statistical characteristics are used as time-domain features, including vibration mean, vibration standard deviation, and vibration peak value. Further frequency domain feature extraction is performed by applying a Fast Fourier Transform to the raw high-frequency data (100ms sampling) of the main fan bearing vibration, converting the time-domain vibration signal into a frequency-domain spectrum to obtain the frequency component distribution of the vibration signal. Based on the vibration frequency-domain spectrum, the dominant frequency and dominant frequency amplitude are extracted; the dominant frequency refers to the frequency component with the largest amplitude in the frequency-domain spectrum. Harmonic features are also extracted, with harmonic frequencies being integer multiples of the dominant frequency, and the amplitudes of the second and third harmonics are extracted. Combining the time-domain and frequency-domain features yields the vibration feature vector of the main fan, with the format being vibration mean, vibration standard deviation, vibration peak value, dominant frequency, dominant frequency amplitude, second harmonic amplitude, and third harmonic amplitude. The same method is used to extract vibration feature vectors for other rotating equipment such as flue gas fans and reactor agitators.

[0034] Based on equipment temperature data from a data stream with precise operating condition classification and time synchronization, a temperature trend analysis method is used to extract temperature features. For the main fan bearing temperature signal, the absolute temperature value at the current moment is extracted as one of the temperature features. The temperature rise rate feature is further calculated; the temperature rise rate is defined as the difference between the current temperature and the temperature 10 minutes ago divided by the 10-minute time interval, reflecting the speed of temperature change. Temperature fluctuation features are calculated, extracting the standard deviation of the temperature over the past 30 minutes, reflecting temperature stability. The absolute temperature value, temperature rise rate, and temperature fluctuation features are combined to obtain the temperature feature vector of the main fan. The same method is used to extract temperature feature vectors for other equipment such as the smoke hood.

[0035] Based on the process operation parameter data stream from the time-synchronized data stream, equipment operating characteristics are extracted. For the main fan, its operating characteristics include fan speed, fan current, and fan outlet pressure. The current values ​​of these parameters are collected from the DCS system, and the equipment load rate is calculated. The load rate is defined as the ratio of the current speed to the rated speed, reflecting the intensity of equipment operation. The load change rate is calculated, defined as the difference between the current load rate and the load rate 5 minutes ago, divided by the 5-minute time interval, reflecting the severity of load adjustments. The operating characteristic vector of the main fan is obtained by combining the speed, current, outlet pressure, load rate, and load change rate.

[0036] Based on vibration feature vectors, temperature feature vectors, and operational feature vectors, a weighted fusion method is used to construct the comprehensive state characteristics of the equipment. The weights for weighted fusion are determined according to the current operating condition type and the reliability of each feature. Specifically, an operating condition feature weight mapping table is established, which is obtained through expert knowledge and historical data analysis. For example, under startup conditions, the weight of the temperature rise rate feature is set to 0.5, the weight of the vibration feature is set to 0.3, and the weight of the operational feature is set to 0.2, because temperature change is the most critical monitoring indicator during startup; under rated load operating conditions, the weight of the vibration feature is set to 0.5, the weight of the temperature feature is set to 0.3, and the weight of the operational feature is set to 0.2, because vibration is the main indicator reflecting equipment health during stable operation. The corresponding feature weights are retrieved from the mapping table according to the current operating condition type. Each component of the vibration feature vector, temperature feature vector, and operational feature vector is multiplied by its corresponding weight, and then concatenated into a comprehensive feature vector to obtain the weighted comprehensive state characteristics of the main fan.

[0037] Based on the weighted comprehensive state characteristics, normalization is used to eliminate the influence of different physical quantities' dimensions. The normalization method employs minimum and maximum normalization. The specific calculation process is as follows: For each component in the feature vector, the numerical distribution of that component is statistically analyzed from the historical database to determine its minimum and maximum values ​​in historical data; the actual measured value of the feature component at the current moment is obtained; the difference between the current value and the historical minimum value is calculated to obtain the offset of the current value relative to the minimum value; the difference between the historical maximum value and the historical minimum value is calculated to obtain the historical variation range of the feature component; the offset is divided by the variation range to obtain the normalized feature component value; the normalization result is verified to be within the range of 0 to 1, and if it exceeds the range, boundary processing is performed.

[0038] Through the above processing, the original value of each feature component is converted into a dimensionless value in the range of 0 to 1, eliminating the dimensional differences between different physical quantities, so that each feature component has the same weight basis in subsequent calculations.

[0039] All components of the eigenvector are normalized to obtain a normalized comprehensive state eigenvector. Based on this normalized comprehensive state eigenvector, principal component analysis (PCA) is used for dimensionality reduction. PCA projects the high-dimensional feature space onto a low-dimensional principal component space through linear transformation, retaining the principal components with the largest variance contribution rate. Specifically, the feature covariance matrix is ​​calculated based on historical equipment state data. The covariance matrix is ​​then decomposed into eigenvalues, and the eigenvalues ​​are sorted from largest to smallest. The top few principal components with a cumulative variance contribution rate of 95% are selected. The normalized eigenvector at the current moment is projected onto the selected principal component space to obtain the dimensionality-reduced equipment state eigenvector, decreasing the dimension from the original dozen or so dimensions to 3 to 5 dimensions.

[0040] Output of the working condition identification and feature extraction module: Equipment operating status data including accurate working condition identifiers, transition state identifiers, and dimensionality-reduced comprehensive state feature vectors. The equipment operating status data accurately reflects the operating status of the equipment under the current working conditions and can be used for health assessment.

[0041] The equipment health assessment module calculates the equipment health score based on equipment operating status data and predicts the deterioration trend through exponential smoothing to obtain equipment health information; The purpose of this module is to assess the current health of the equipment based on its operating status data and working condition information, and to predict its future deterioration trend, providing a basis for control and maintenance decisions.

[0042] Based on the types and characteristics of equipment in catalytic cracking units, an equipment classification system is established using an equipment classification method. The equipment within the unit is divided into three main categories according to its structural characteristics and failure modes: rotating equipment, static equipment, and electrical equipment. Rotating equipment includes main fans, flue gas fans, feed pumps, and circulating pumps. Typical failure modes for this type of equipment include bearing wear, impeller wear, and rotor imbalance. Key health assessment indicators are vibration and temperature. Static equipment includes reactors, regenerators, fractionation towers, heat exchangers, and pipelines. Typical failure modes for this type of equipment include corrosion thinning, crack propagation, and seal leakage. Key health assessment indicators are wall thickness, corrosion rate, and operating time. Electrical equipment includes transformers, motors, and switchgear. Typical failure modes for this type of equipment include insulation aging, poor contact, and overheating. Key health assessment indicators are insulation resistance, temperature rise, and current. For each type of equipment, a corresponding health assessment indicator system is established, resulting in a mapping relationship between equipment classification and assessment indicators.

[0043] Based on equipment operating status data and the established equipment classification system, a health assessment baseline based on operating conditions is established for rotating equipment. The specific method is as follows: Normal operating data of the main fan under different operating conditions is extracted from the historical operating database. Normal operating data refers to data from periods when the equipment has not experienced failures and performs well under that operating condition. For each operating condition type, a sample set of equipment state feature vectors is extracted, and the statistical characteristics of the sample set are calculated, including the mean, standard deviation, and quantiles of each component of the feature vector. These statistical characteristics are used as the characteristic baseline of the equipment's normal state under that operating condition and stored in the operating condition baseline mapping library. For example, for rated load operating conditions, the baseline for the main fan's mean vibration is 2.5 mm / s, the baseline for the standard deviation is 0.3 mm / s, the baseline for the main frequency amplitude is 1.8 mm / s, and the baseline for the bearing temperature is 75 degrees Celsius. The same method is used to establish operating condition baseline mapping libraries for other rotating equipment such as flue gas fans and feed pumps.

[0044] Based on the established operating condition-baseline mapping library and equipment operating status data, the health score of the rotating equipment is calculated. According to the current operating condition identifier, the corresponding feature baseline is queried from the operating condition-baseline mapping library. The current status feature vector of the equipment is extracted, and the deviation between the current feature vector and the baseline feature is calculated. The deviation is calculated using the Mahalanobis distance metric. The specific calculation process is as follows: Obtain the current state feature vector of the equipment, denoted as the current feature vector; extract the characteristic baseline mean vector of the normal state under this operating condition from the operating condition-baseline mapping library, denoted as the baseline mean vector; extract the characteristic covariance matrix of the normal state under this operating condition from the operating condition-baseline mapping library, denoted as the baseline covariance matrix; calculate the difference vector between the current feature vector and the baseline mean vector to obtain the feature deviation vector; calculate the inverse matrix of the baseline covariance matrix to obtain the inverse covariance matrix; multiply the transpose of the feature deviation vector by the inverse covariance matrix to obtain an intermediate calculation result; multiply the intermediate calculation result by the feature deviation vector to obtain a quadratic calculation result; take the square root of the quadratic calculation result to obtain the Mahalanobis distance value, which reflects the degree of deviation between the current state and the normal state.

[0045] Based on the calculated Mahalanobis distance, a scoring mapping function is used to convert the distance into a health score. The scoring mapping function is designed as follows: when the Mahalanobis distance is less than threshold 1, the health score is 100 points minus the Mahalanobis distance multiplied by 10, indicating the equipment is in an excellent state; when the Mahalanobis distance is between threshold 1 and threshold 2, the health score linearly maps to 75 to 90 points, indicating the equipment is in a good state; when the Mahalanobis distance is between threshold 2 and threshold 3, the health score linearly maps to 60 to 75 points, indicating the equipment is in a normal state; when the Mahalanobis distance is between threshold 3 and threshold 4, the health score linearly maps to 45 to 60 points, indicating the equipment is in a watchful state; when the Mahalanobis distance is greater than threshold 4, the health score linearly maps to 0 to 45 points, indicating the equipment is in an abnormal state. The health score of the main fan is calculated based on its current Mahalanobis distance, resulting in the main fan's health score value and corresponding health level.

[0046] Furthermore, an anomaly detection method based on isolation forests can be used instead of the Mahalanobis distance method. Isolation forests are an unsupervised anomaly detection method that isolates anomalous points by constructing random trees. Specifically, an isolation forest model is trained using historical normal operation data. The model input is a device state feature vector, and the output is an anomaly score. A higher anomaly score indicates a more abnormal current state and lower health. The anomaly score from the isolation forest is then normalized and converted into a health score. The purpose of this alternative embodiment is to improve the ability to identify complex anomaly patterns, especially to more accurately identify outliers in a multi-dimensional feature space.

[0047] Based on the equipment classification system, a time- and operating parameter-based health assessment method is adopted for static equipment. For reactors, the health assessment considers the following factors: operating time since the last inspection, the ratio of cumulative operating time to design life, the proximity of operating temperature to design temperature, pressure fluctuation amplitude, and exposure time to corrosive media. Based on these factors, a weighted scoring method is used to calculate the reactor's health. Specifically, the weight of the runtime factor is 0.3, calculated by dividing the time since the last inspection by the inspection cycle, multiplying the result by 100, and then subtracting from 100 to obtain the factor's score; the weight of the cumulative runtime factor is 0.3, calculated by dividing the cumulative runtime by the design life, multiplying the result by 100, and then subtracting from 100 to obtain the factor's score; the weight of the temperature factor is 0.2, calculated as the ratio of the current operating temperature to the design temperature. When the ratio is less than 0.9, the score is 100; when the ratio is between 0.9 and 1.0, the score linearly decreases to 80; when the ratio is greater than 1.0, the score decreases rapidly; the weight of the pressure fluctuation factor is 0.1, calculated as the standard deviation of pressure over the past 30 days, with a higher standard deviation resulting in a lower score; the weight of the corrosion factor is 0.1, calculated as the ratio of exposure time to corrosive media to the total exposure time. The overall health score of the reactor is obtained by multiplying each factor's score by its weight and summing the results.

[0048] Based on the equipment health score, a time-series analysis method is used to predict the future degradation trend of the equipment. The specific method is as follows: Extract the time-series data of the equipment's health score for the past 30 days from the historical database, and arrange the time-series data in chronological order to form a health score sequence. Apply exponential smoothing to this sequence for trend extraction. The calculation process of exponential smoothing is as follows: Obtain the actual health score value at the current time (day t); obtain the smoothed value at the previous time (day t-1), and if it is the first day, initialize the smoothed value to the actual health score of the first day; set the smoothing coefficient, which is 0.3 in this embodiment, and this coefficient controls the weight ratio between new data and historical data; multiply the current actual health score by the smoothing coefficient to obtain the weighted value of the current data; multiply the smoothed value at the previous time by (1 minus the smoothing coefficient) to obtain the weighted value of the historical data; add the weighted value of the current data and the weighted value of the historical data to obtain the smoothed value at the current time.

[0049] Exponential smoothing is used to obtain a smoothed trend curve for the health score. Based on the smoothed trend curve, the rate of decline in health score is calculated, defined as the average slope of the smoothed values ​​over the past 7 days. A linear extrapolation method is used to predict the health score for the next 7 days. The prediction calculation process is as follows: obtain the smoothed health score at the current moment; calculate the rate of decline in health score, i.e., the average rate of change of the smoothed values ​​over the past 7 days; for the k-th day in the future (k takes values ​​from 1 to 7), multiply the rate of decline by the number of prediction days k to obtain the total change during the prediction period; add the total change to the smoothed health score at the current moment to obtain the predicted health score for the k-th day in the future.

[0050] Based on the predicted health score for the next 7 days, it is determined whether the equipment is in a degradation phase: if the predicted health score is below 60 on any day, the equipment is considered to be in a degradation phase and maintenance needs to be scheduled in advance. The maintenance priority of the equipment is determined by combining the current health score, health level, and degradation trend assessment results. The priority rules are as follows: equipment with a current health level of "Attention" or "Abnormal" and in a degradation phase has high priority; equipment with a current health level of "General" and in a degradation phase has medium priority; other equipment has low priority.

[0051] The equipment health assessment module outputs equipment health information, including health scores, health levels, degradation trend prediction results, and maintenance priorities for each piece of equipment. This output will serve as a constraint for subsequent control and operation and maintenance decisions.

[0052] The adaptive control module designs a multivariable coordinated controller based on equipment health information, and dynamically generates the operating constraints of the multivariable coordinated controller according to the equipment health score to obtain the set values ​​of the operating parameters; The purpose of this module is to incorporate equipment state constraints into the control strategy when performing multivariate control of a reaction regeneration system based on equipment health information, so as to ensure that control actions do not exacerbate equipment degradation.

[0053] Based on the process mechanism of the reaction-regeneration system, a simplified mechanistic model is used to describe the coupling relationship between reactor temperature, regenerator temperature, and catalyst circulation rate. The heat balance of the reactor is as follows: the heat entering the reactor equals the heat carried in by the catalyst plus the heat of reaction; the heat leaving the reactor equals the heat carried away by the catalyst plus the heat carried away by the product. The heat balance of the regenerator is as follows: the heat entering the regenerator equals the heat carried in by the catalyst; the heat leaving the regenerator equals the heat carried away by the catalyst plus the heat carried away by the flue gas plus the heat released by coke combustion. The catalyst circulation rate connects the heat balance of the reactor and the regenerator. Based on these balance relationships, a simplified mathematical model is established, and the model parameters are identified through historical operating data. The data identification method involves collecting steady-state operating data of variables such as reactor temperature, regenerator temperature, catalyst circulation rate, feed rate, and regeneration air volume under different operating conditions. The least squares method is used to fit the model parameters, minimizing the sum of squared errors between the model predictions and actual measurements. This yields the coupling relationship model and its parameters for the reaction-regeneration system.

[0054] Based on a coupling relationship model, a multivariable coordinated controller (MCC) is designed using predictive control. The basic idea of ​​predictive control is to predict the system's output response over a future period using a model in each control cycle, and then determine the control sequence for that period through optimization calculations, ensuring the predicted output is as close as possible to the desired target and that the changes in control actions are as smooth as possible. Specific settings are as follows: the prediction time domain is set to 30 control cycles, the control time domain to 10 control cycles, and the control cycle to 1 minute. The controlled variables of the controller are reactor temperature and regenerator temperature, with setpoints derived from the optimization layer's instructions. The manipulated variables of the MCC are catalyst circulation rate and regeneration airflow. An objective function for predictive control is established, comprising three terms: the first term is the weighted sum of squares of the deviations between the predicted output and the setpoint, with the weight matrix set according to the importance of temperature control, where reactor temperature deviation has a greater weight than regenerator temperature deviation; the second term is the weighted sum of squares of the changes in the manipulated variables, used to suppress drastic changes in control actions; the third term is the energy consumption term, represented by the square of the regeneration airflow, encouraging the MCC to reduce energy consumption while meeting temperature control requirements. The objective function is calculated as follows: The deviation between the predicted output and the setpoint is calculated. For each moment in the prediction time domain, the deviation between the predicted reactor temperature and the setpoint, and the deviation between the predicted regenerator temperature and the setpoint are calculated separately. The weighted sum of squares of the predicted output deviation is calculated by squaring the temperature deviation at each moment, multiplying it by the corresponding weighting coefficient, and finally summing the weighted squared deviations over all moments in the prediction time domain. The squared term of the manipulated variable change is calculated. For each moment in the control time domain, the change in catalyst circulation rate and the change in regeneration air rate are calculated separately, squared, and multiplied by the corresponding weighting coefficient. The weighted sum of squares of the manipulated variable change is calculated by summing the weighted squared terms of the manipulated variable change at each moment in the control time domain. The energy consumption term is calculated. For each moment in the control time domain, the square of the regeneration air rate is taken as the energy consumption term, multiplied by the corresponding energy consumption penalty coefficient, and then summed over the control time domain. The predicted output deviation term, the manipulated variable change term, and the energy consumption term are added together to obtain the complete objective function value.

[0055] The prediction time domain length is set to 30, and the control time domain length is set to 10. The output error weight matrix is ​​a diagonal matrix, with the reactor temperature weight set to 1.0 and the regenerator temperature weight set to 0.8. The control increment weight matrix is ​​also a diagonal matrix, with the catalyst circulation rate change weight set to 0.1 and the regeneration air volume change weight set to 0.05. The energy consumption penalty coefficient is set to 0.01.

[0056] Based on equipment health scores and health levels, a constraint generation method is used to convert equipment states into operational constraints for a multivariable coordinated controller. Specific rules are as follows: For the main fan, when its health level is excellent, the upper limit constraint for the main fan load rate is set to 100%, and the lower limit constraint is set to 30%; when the health level is good, the upper limit constraint is set to 85%, and the lower limit remains at 30%; when the health level is average, the upper limit constraint is set to 70%, and the lower limit remains at 30%; when the health level is warning, the upper limit constraint is set to 50%, and the lower limit remains at 30%, while triggering a maintenance warning; when the health level is abnormal, the equipment is prohibited from operation, and the upper limit constraint is set to 0. Since the main fan load rate is directly proportional to the regenerated air volume, the main fan load rate constraint is converted into a regenerated air volume constraint. The upper limit of the regenerated air volume is the main fan rated air volume multiplied by the upper limit of the load rate, and the lower limit is the rated air volume multiplied by the lower limit of the load rate. Constraints are generated for the flue gas fan using the same method; the flue gas fan constraints affect the allowable range of catalyst circulation. The equipment health constraints are expressed as inequalities, resulting in a set of equipment constraints.

[0057] Based on the predictive control objective function and equipment health constraints, an optimization method is used to calculate the optimal control sequence. The predictive control problem is formulated as a constrained optimization problem: minimizing the objective function while satisfying equipment constraints, manipulated variable constraints, and manipulated variable rate of change constraints. Constraints include: the regenerated air volume must be within defined upper and lower limits, the catalyst circulation volume must be within its corresponding upper and lower limits, and process safety constraints such as reactor temperature and regenerator temperature not exceeding their safety limits must also be met. A quadratic programming solver is used to solve this constrained optimization problem, obtaining the optimal control sequence in the future control time domain. Following the rolling optimization strategy of predictive control, the first element of the control sequence is taken as the control output for the current control cycle, thus obtaining the current catalyst circulation setpoint and regenerated air volume setpoint. In some cases, equipment health constraints may lead to an infeasible optimization problem, for example, when all critical equipment health is low but the production task must be completed. To address this, a constraint softening mechanism is introduced: hard constraints are converted into soft constraints, allowing appropriate violations of equipment constraints, but imposing a high penalty weight on the amount of constraint violation. The calculation process for the soft constraint penalty term is as follows: For each moment in the control time domain, check whether the control quantity exceeds the upper limit of the soft constraint. If the control quantity exceeds the upper limit, calculate the excess (control quantity minus the soft upper limit); otherwise, the excess is zero. Square the excess and multiply it by the penalty weight coefficient to obtain the upper limit violation penalty term. For each moment in the control time domain, check whether the control quantity is lower than the lower limit of the soft constraint. If the control quantity is lower than the lower limit, calculate the deficiency (soft lower limit minus the control quantity); otherwise, the deficiency is zero. Square the deficiency and multiply it by the penalty weight coefficient to obtain the lower limit violation penalty term. Summate the upper limit violation penalty term and the lower limit violation penalty term in the control time domain to obtain the total constraint violation penalty term.

[0058] The upper and lower limits of the softened constraints are appropriately relaxed based on the health of the equipment, and the penalty weight coefficient is set to a very large value, such as 1000, to ensure that the constraints are violated only when necessary.

[0059] When the optimization solution results in a constraint violation, an alarm message for constraint violation is output to indicate to the operator that the equipment is in an abnormal state but is still running.

[0060] The adaptive control module outputs operating parameter settings, including catalyst circulation rate settings and regeneration air rate settings, as well as possible constraint violation alarm information.

[0061] The hierarchical optimization control module executes optimization control, coordinated control, and PID control respectively based on the set values ​​of operating parameters, and establishes an inter-layer feedback mechanism to obtain multi-level coordinated control commands; The purpose of this module is to establish a hierarchical control architecture from the device layer to the loop layer, achieve coordinated optimization of each level, and adaptively adjust the control strategy according to changes in operating conditions.

[0062] Based on operating condition indicators and equipment health information, steady-state optimization calculations are performed at the unit level to determine the optimal operating target values ​​for the entire unit. The goal of unit-level optimization is to maximize the economic benefits of the unit while meeting production tasks, ensuring product quality, complying with safety and environmental constraints, and considering equipment health constraints. Economic benefits are represented by subtracting raw material and energy costs from product value. Specifically, a unit-level optimization model is established, with decision variables including key operating parameters such as target reactor temperature, target regenerator temperature, and target fractionation column reflux ratio. Constraints include: product yield must meet contractual requirements, product quality indicators must be within acceptable ranges, reactor temperature must be within safe ranges, regenerator oxygen content must be within safe ranges, and the load of each piece of equipment must be within the defined health allowable range.

[0063] Because economic benefit calculations involve parameters with different dimensions (product price in monetary units / units of mass, output in units of mass, raw material cost in monetary units / units of mass, raw material consumption in units of mass, energy cost in monetary units / units of energy, and energy consumption in units of energy), it is necessary to ensure the consistency of dimensions in all calculations. The specific handling method is as follows: The product value item is calculated by multiplying the product price by the product output. The raw material cost item is calculated by multiplying the unit cost of raw materials by the amount of raw materials consumed. The energy cost item is calculated by multiplying the unit cost of energy by the amount of energy consumed.

[0064] All cost items are standardized in monetary units, ensuring dimensional consistency in the calculation of economic benefits. The calculation process for the economic benefits of the plant is as follows: Calculate the value of each product by multiplying its price by its output; calculate the total product value by adding the values ​​of all products; calculate the raw material cost by multiplying the unit cost of raw materials by the amount of raw materials consumed; calculate the energy cost by multiplying the unit cost of energy by the amount of energy consumed; calculate the plant's economic benefits by subtracting the raw material cost and then the energy cost from the total product value.

[0065] The optimization objective is to maximize the economic benefits of the device calculated in step five.

[0066] Based on the current operating condition, the corresponding optimization parameter configuration is invoked, including parameters such as product price, energy cost, and constraint limits. A sequential quadratic programming method is used to solve the unit-level optimization problem, obtaining optimal operational objective values, including optimal reactor temperature, optimal regenerator temperature, and optimal fractionation column reflux ratio.

[0067] Based on the optimal operating target value at the device level, multivariate coordinated control is executed at the system level. The objects controlled at the system level are various subsystems such as the reaction regeneration system and the fractionation system. For the reaction regeneration system, the optimal target values ​​for reactor temperature and regenerator temperature are used as the control setpoints for the system, and the designed predictive controller is invoked for control calculations. The parameters of the predictive controller are adaptively adjusted according to the current operating condition. Controller parameters include prediction time domain, control time domain, and weighting coefficients; these parameters require different settings under different operating conditions to achieve optimal control performance. An operating condition controller parameter mapping table is established, obtained through simulation experiments and field debugging. For example, in the transitional state during operating condition switching, the control time domain is shortened and the control action weights are increased to accelerate system response; in steady-state operation, the control time domain is extended and the control action weights are decreased to maintain smooth operation. The corresponding controller parameters are retrieved from the mapping table according to the current operating condition, and the predictive controller configuration is updated. Control optimization is performed to obtain the current catalyst circulation rate setpoint and regeneration air rate setpoint. A similar method was used to control the fractionation system. Based on the target of the optimal reflux ratio of the fractionation tower, a conventional PID controller was used to adjust the reflux flow rate.

[0068] Based on the system-level control setpoint, basic control is executed at the loop level. The loop level contains hundreds of single-loop PID controllers, each responsible for a specific regulation task. For catalyst circulation control, the actuator is a catalyst slide valve; the PID controller calculates the valve opening adjustment based on the deviation between the measured and setpoint catalyst circulation values. For regeneration airflow control, the actuator is the main fan inlet guide vane; the PID controller calculates the guide vane opening adjustment based on the deviation between the measured and setpoint airflow values. The parameters of each PID controller are tuned according to the dynamic characteristics of the controlled object, using either the critical proportional gain method or the attenuation curve method. The PID controller output, after being limited, is sent to the field actuator to drive the regulating valve or frequency converter, achieving closed-loop control of the process variables.

[0069] Based on the execution results of the loop layer, an inter-layer feedback mechanism is established to achieve coordination among different levels. The loop layer feeds back the actual execution results to the system layer, including whether the control setpoint has been reached, whether the actuator is saturated, and the magnitude of the control deviation. The system layer evaluates the control effect based on the feedback information. When a large deviation is found between the actual output and the expected output, the cause of the deviation is analyzed: if it is due to model mismatch, an online model calibration program is initiated to update the model parameters using the latest input and output data; if it is due to disturbances exceeding expectations, the robustness parameters of the controller are adjusted to increase the control action to suppress the disturbances. The system layer feeds back the actual operating results to the device layer, including the actual operating parameter values ​​of each subsystem, the actual product yield, and the actual energy consumption. The device layer evaluates the optimization effect based on the feedback information and calculates the gap between the actual economic benefits and the expected economic benefits. When the actual benefits are consistently lower than expected, an optimization model calibration program is initiated to re-identify the model parameters using accumulated operating data, or to adjust the weight coefficients of the optimization objective to make the optimization scheme more consistent with the actual situation.

[0070] Based on operating condition identification, when a switch from one operating condition type to another is detected, a gradual parameter adjustment strategy is adopted to avoid control system oscillations. The specific method is as follows: When a switch from rated load operation to high load operation is detected, the controller parameters are not immediately switched from rated load parameters to high load parameters. Instead, a parameter switching trajectory is designed, and the parameters are gradually adjusted over 5 to 10 minutes. The parameter switching trajectory uses a linear transition or an S-curve transition. For the weighting coefficients of the predictive controller, a linear interpolation method is used to calculate the weight value at each moment during the transition. The weight value changes linearly from the parameter value of the old operating condition to the parameter value of the new operating condition, with a change time of 10 minutes. During the parameter switching process, the fluctuation of the system output is continuously monitored. If the fluctuation exceeds the allowable range, the parameter adjustment speed is slowed down or the adjustment is paused. Once the system output stabilizes, parameter adjustment continues until the complete switch from the old operating condition parameters to the new operating condition parameters is completed.

[0071] The hierarchical optimization control module outputs multi-level coordinated control commands, including the optimization target value at the device level, the multivariable control output at the system level, the PID control action at the loop level, and the execution feedback information and coordinated adjustment results at each level.

[0072] The collaborative decision-making module, based on equipment health information and multi-level coordinated control commands, generates collaborative decision-making trigger signals according to equipment health information and uses a rule matching method to generate the final collaborative decision-making scheme; The purpose of this module is to quickly generate collaborative decision-making solutions that balance production tasks and equipment protection when equipment health deteriorates, and to adjust control strategies and maintenance plans.

[0073] Based on equipment health scores, health levels, and degradation trend predictions, a trigger condition judgment method is used to determine whether to initiate a collaborative decision-making process. A set of collaborative decision-making trigger condition rules is established, including: triggering when the health level of any critical equipment drops to the attention level; triggering immediately when the health level of any critical equipment drops to the abnormal level; triggering when the rate of equipment health decline exceeds 5 points per day; and triggering when the equipment degradation trend prediction indicates that the health level will drop below 60 points within the next 7 days. All critical equipment is traversed, and its health status is checked to see if it meets any of the trigger conditions. If the main fan's current health level is attention, meeting the trigger conditions, the collaborative decision-making process is initiated, and a collaborative decision-making trigger signal is obtained.

[0074] Based on collaborative decision-making trigger signals, equipment health information, and operational status information, a scenario feature extraction method is used to describe the current decision-making scenario. Scenario features include: the equipment identifier that triggered the decision, the current health level of that equipment, the current load rate of that equipment, the health levels of other critical equipment, the availability status of backup equipment, the urgency of the current production task, and the availability of maintenance resources. For a decision-making scenario triggered by the main fan, the extracted features are: main fan identifier is FM101, health level is "Caution," current load rate is 85%, flue gas fan health level is "Good," no backup main fan, production task urgency is "High," and maintenance team is on duty. These features are organized into a scenario feature vector to obtain a description of the current decision-making scenario.

[0075] Based on the current decision-making scenario description, a rule matching method is used to search for applicable decision rules from the decision rule base. The decision rule base is a pre-built knowledge base containing a large number of typical scenarios and their corresponding decision schemes. The rule format is: when the device identifier is X, the health level is Y, the current load is Z, and the standby device status is W, take decision scheme M. Each rule is accompanied by a confidence score. The rule base is built by offline simulation of a large number of device failure scenarios, running multi-objective optimization to solve for the optimal decision scheme for each scenario, storing the scenario features and decision schemes as rules, and evaluating the confidence level based on the effectiveness of the simulation schemes. For the current decision scenario, the similarity between its feature vector and the feature vectors of each rule in the rule base is calculated. The similarity is measured using a weighted Euclidean distance; the smaller the distance, the higher the similarity. The top 3 rules with the highest similarity are selected, and their confidence levels are checked to see if they exceed a 90% threshold. If the confidence level of the rule with the highest similarity is 95%, exceeding the threshold, the decision scheme given by that rule is directly adopted, resulting in a fast decision result with a decision time of less than 5 seconds.

[0076] Based on rule matching results, when no high-confidence rule matches the current scenario, an online optimization method is used to solve for the decision scheme. A multi-objective optimization model is established, with decision variables including: target load rate of each key piece of equipment, equipment maintenance time window, and production plan adjustment scheme. The optimization objectives include three items: the first is to maximize the completion rate of production tasks, defined as the ratio of actual output to planned output; the second is to minimize the degree of equipment health deterioration, defined as the weighted sum of the decrease in equipment health after the decision is implemented; and the third is to minimize maintenance costs, including maintenance labor costs and downtime losses.

[0077] Since the three objective functions involve parameters with different dimensions (production task completion rate is a dimensionless ratio, equipment health is a 0-100 score, and maintenance cost is in monetary units), normalization is required to ensure the rationality of multi-objective optimization. The specific normalization method is as follows: As for the completion rate of production tasks, since it is a ratio and is a dimensionless parameter with a value range of 0 to 1, no additional normalization is required.

[0078] To determine the degree of equipment health deterioration, a maximum and minimum normalization method is used. The maximum possible value for the health decrease is set to 100 (equipment deteriorating from excellent to abnormal), and the minimum value is 0 (no health decrease). The normalization calculation process is as follows: For each piece of equipment, multiply its health decrease by its importance weight to obtain the weighted health decrease; sum the weighted health decreases of all equipment to obtain the total weighted health decrease; calculate the sum of the importance weights of all equipment; divide the total weighted health decrease by (100 multiplied by the sum of weights) to obtain the normalized degree of equipment health deterioration.

[0079] For maintenance costs, a normalization method based on historical data is adopted. Based on maintenance cost statistics from the past year, the maximum and minimum maintenance costs are determined. The normalization calculation process is as follows: Calculate the current total maintenance cost by adding maintenance labor costs and downtime losses; calculate the historical range of maintenance costs by subtracting the minimum from the maximum historical maintenance cost; calculate the offset of the current maintenance cost relative to the historical minimum by subtracting the historical minimum from the current total maintenance cost; calculate the normalized maintenance cost by dividing the offset by the historical range.

[0080] The normalized multi-objective optimization problem contains three objective functions: Objective function 1: Maximize the production task completion rate. The calculation steps are as follows: sum the actual output of each product, divide by the sum of the planned output of each product, and obtain the production task completion rate.

[0081] Objective function 2: Minimize the degree of equipment health deterioration. The calculation steps are as follows: For each piece of equipment, multiply its health decline by its importance weight; sum the weighted health decline of all equipment; divide the sum by (100 multiplied by the sum of the importance weights of all equipment) to obtain the normalized degree of equipment health deterioration.

[0082] Objective function three: Minimize maintenance costs. The normalized maintenance cost calculated in step four is used.

[0083] Among them, the production task completion rate is a dimensionless ratio (between 0 and 1), the equipment health deterioration is normalized by dividing by 100 and the sum of weights, and the maintenance cost is normalized by the historical maximum and minimum values, ensuring that the three objective functions are within a similar numerical range, which facilitates multi-objective optimization solutions.

[0084] The constraints include: the sum of the load rates of all equipment must meet the minimum requirements of the production task; the maintenance time window cannot conflict with critical production periods; and the load adjustment range cannot exceed the allowable rate of change of the equipment. A heuristic optimization method is used to solve this multi-objective optimization problem, specifically a genetic algorithm. The initial population of the genetic algorithm is initialized from matching decision schemes for similar scenarios to accelerate the convergence speed. An optimization time limit of 30 seconds is set; if the global optimum is not reached within 30 seconds, the best feasible solution found is output. The resulting optimization decision schemes include reducing the main fan load from 85% to 60%, increasing the flue gas fan load from 80% to 90% to share the load, scheduling maintenance for the main fan during the planned maintenance window after 72 hours, and adjusting the production plan to reduce the output target by 5%.

[0085] Furthermore, particle swarm optimization (PSO) can be used instead of genetic algorithms. PSO is an optimization method based on swarm intelligence, which accelerates the search process through information sharing among particles. Specifically, a swarm of particles is initialized, with each particle representing a decision scheme. Particles update their velocity and position based on their historical best position and the global best position, gradually moving closer to the optimal solution. PSO typically converges faster than genetic algorithms and can search a wider solution space within a 30-second time limit. The purpose of this alternative embodiment is to improve the efficiency and quality of the optimization solution.

[0086] Based on the collaborative decision-making scheme, a scheme decomposition method is used to break down the decision scheme into control adjustment instructions and maintenance plan instructions. Control adjustment instructions include new load rate target values ​​for each piece of equipment, which are then converted into constraints for the unit-level optimization. Specifically, constraints are added to the unit-level optimization model: the main fan load rate does not exceed 60%, and the flue gas fan load rate does not exceed 95%. These modified constraints are sent to the unit-level optimization program, triggering a re-solution of the optimization model to obtain new operational target values. Maintenance plan instructions include information such as maintenance equipment identification, maintenance time window, maintenance type, and estimated maintenance duration. This information is formatted into maintenance work orders and pushed to the operation and maintenance management system via an interface. Simultaneously, the maintenance plan is pushed to the production scheduling system to update the production plan and adjust the production target and raw material procurement plan for the next 72 hours, ensuring the reasonable arrangement of production tasks during maintenance.

[0087] Based on the issued decision plan, an effect tracking method is used to monitor the actual effects of the decision implementation. Tracking indicators include: the actual trend of equipment health changes, the actual completion of production tasks, and the actual execution results of maintenance activities. Relevant data is continuously collected within 7 days of decision implementation. On day 7, effect evaluation indicators are calculated: whether the main fan health has stopped declining or begun to recover, whether the production task completion rate has reached the adjusted target, and whether maintenance activities have been executed as planned and achieved the expected results. The decision scenario, decision plan, and implementation effect triple are recorded in the decision case library. Monthly analysis is performed on the accumulated decision cases to extract common features of successful cases, generate new decision rules, or update the confidence level of existing rules. For cases with poor results, the reasons for failure are analyzed, and the characteristics of failed cases are marked as negative rules for exclusion during rule matching. Through continuous case accumulation and rule updates, the coverage and accuracy of the decision rule library are continuously improved.

[0088] The collaborative decision-making module outputs the final collaborative decision-making scheme, including issuing equipment load adjustment instructions to the control system, pushing maintenance plans to the operation and maintenance system, notifying the scheduling system of production plan adjustments, and the expected evaluation of the decision-making effect.

[0089] In one embodiment of the present invention, in order to achieve hierarchical data storage, reasonable allocation of real-time and offline tasks, and continuous optimization and updating of intelligent models, the multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production further includes: The cloud-edge-device collaborative data management module manages cloud-edge-device collaborative data based on the final collaborative decision-making scheme. It achieves data management and model updates through data collection, edge processing, cloud storage, offline model training, and model deployment. The device deploys smart sensors and industrial gateways to collect process parameters, equipment status, and video features locally, and then uploads the data to edge nodes via the MQTT / OPC UA protocol. To ensure data reliability, the device caches data from the last 24 hours and implements a breakpoint resume mechanism, automatically re-uploading missing data once the network is restored.

[0090] Based on edge computing servers deployed in the device control room, real-time data processing and short-term data storage functions are implemented at the edge. The edge server receives data streams from the device and executes various real-time computing tasks in the first six modules. Specifically, these tasks include: data preprocessing, condition identification and feature extraction, health assessment, multivariate control calculation, hierarchical optimization control, and collaborative decision-making. These tasks have high real-time requirements and must be completed at the edge, with response times required to be in the second range. The edge server is configured with industrial-grade computing resources, including multi-core processors, large-capacity memory, and solid-state drives, to ensure the performance requirements of real-time computing tasks. Short-term data storage is established at the edge, using a time-series database to store high-frequency raw data and real-time calculation results from the past seven days. The time-series database is optimized for time-series data, supporting efficient time-range queries and data compression. The storage strategy uses a cyclic overwrite approach; when storage space reaches its limit, the oldest data is automatically deleted, and the latest seven days of data are retained.

[0091] Based on real-time computation results at the edge, anomaly detection methods are employed to identify abnormal data requiring detailed analysis in the cloud. An anomaly detection program runs at the edge, monitoring device status characteristics and health scores online. Anomaly detection rules are defined: a sudden drop in device health score exceeding 10 points is considered an anomaly; a deviation of the device status feature vector from the normal baseline exceeds a preset threshold; and video surveillance detects safety hazards such as leaks or smoke. Once an anomaly is detected, a selective upload mechanism is triggered: high-frequency raw data, device status characteristics, health scores, and operating information for 30 minutes before and after the anomaly event are packaged and uploaded to the cloud data platform via an encrypted channel for detailed analysis. For normal operating data, high-frequency raw data is not uploaded; only hourly summary statistics, including the mean, maximum, minimum, and standard deviation of each variable, are uploaded to the cloud. This selective upload strategy ensures the integrity of abnormal data while reducing network bandwidth consumption; the amount of data uploaded during normal operation is only about 1% of the total upload volume.

[0092] Based on a cloud computing platform deployed in a data center, long-term storage and data warehouse construction of massive amounts of data are achieved in the cloud. The cloud data platform receives uploaded data, establishes a distributed data storage system, and uses the Hadoop Distributed File System to store massive amounts of historical data. The data retention strategy is as follows: key data such as health scores, operating condition information, control commands, and decision-making schemes for all devices are retained for 3 years; summary statistics of normal operation are retained for 3 years; and detailed data of abnormal events are permanently retained. A data warehouse is established, organizing raw data according to device, time, and operating condition dimensions, creating a multi-dimensional data model to support complex data queries and analysis. A time-series database index is established, indexing timestamp and device identifier fields to accelerate data retrieval. A data query interface is provided, supporting queries for historical data by device, time period, operating condition type, etc., with query response time controlled within seconds.

[0093] Based on massive historical data stored in the cloud, offline model training and optimization tasks are run in the cloud. Offline tasks have high computational resource requirements but low real-time requirements, making them suitable for execution in the cloud. A model retraining task is initiated monthly, specifically including: retraining the operating condition recognition model using labeled operating condition data accumulated in the past month, evaluating model performance using cross-validation, and updating the model if the new model's recognition accuracy is more than 2% higher than the existing model; retraining the health assessment model using equipment operation data and maintenance records from the past month, updating the operating condition-baseline mapping library, and adjusting the threshold parameters for health scoring; and re-identifying the parameters of the predictive control model using recursive regeneration system operation data from the past month, employing recursive least squares to identify model parameters online, making the model more consistent with the current device status. After model training is complete, the performance of the new model is evaluated on an independent validation dataset in the cloud, comparing the prediction error, classification accuracy, and other metrics of the old and new models on the validation set. If the new model outperforms the old model and the improvement reaches a preset threshold, such as 5%, the new model is approved for deployment.

[0094] Based on model training and performance evaluation, a model deployment method is adopted to distribute updated model parameters to the edge. A cloud-edge communication mechanism is established, using message queue technology to achieve asynchronous communication between the cloud and the edge. The cloud packages the new model parameters, new rule base, and new configuration file into a deployment package and sends it to the cloud's instruction queue. The edge queries the cloud instruction queue hourly to check for new deployment packages. When a new deployment package is detected, the edge downloads the deployment package to its local machine, performs integrity verification, and decompresses the deployment package after successful verification. Model parameter updates adopt a hot deployment method to avoid interrupting the running real-time task. Specifically, the new model parameters are first loaded into the memory reserve area, and then, at the end of the control cycle of the real-time task, the model parameter pointer is atomically switched, switching the new model parameters to the active model. After the switch is completed, the memory occupied by the old model parameters is released. The model switching process is logged, including the switching time, the new and old model version numbers, and the switching result. The logs are uploaded to the cloud for auditing.

[0095] Based on a three-tier cloud-edge-device architecture, a cloud-edge collaborative working mechanism and fault tolerance mechanism are established to ensure high system availability. During normal operation, the edge device handles real-time tasks, while the cloud handles offline tasks, communicating and synchronizing asynchronously via message queues. When network connectivity is normal, the edge device periodically uploads data and logs to the cloud, and the cloud periodically sends model and configuration updates to the edge device. When a network failure causes an interruption in cloud-edge communication, the edge device automatically switches to autonomous operation mode, using locally stored models and rule bases to continue executing real-time tasks, ensuring the control system is unaffected by network failures. The edge device caches data during network outages in local storage and uploads it to the cloud in batches after network recovery. When the cloud detects that the edge device has not uploaded data for an extended period, it triggers an alarm to notify operations personnel to check the network status. When edge device hardware fails, a backup edge server is activated. The backup server and the primary server use a dual-machine hot standby mode, with the primary server's status data synchronized to the backup server in real time, and the failover time is less than one minute.

[0096] The cloud-edge-device collaborative data management module outputs a long-term historical database in the cloud, regularly updated intelligent analysis models and decision rule bases, real-time processing results at the edge, cloud-edge-device collaborative operation status information, and high availability guaranteed by fault tolerance mechanisms.

[0097] A computer-readable storage medium for storing computer-readable instructions that, when read by a computer, enable the operation of the aforementioned multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production.

[0098] In one embodiment of the present invention, a specific example is provided: The application was validated on a continuous catalytic cracking unit in a petrochemical company. The unit has an annual processing capacity of 3 million tons and is equipped with 10 key rotating units, including main fans and flue gas fans, and over 300 process control loops. Before the platform implementation, the unit faced problems such as large fluctuations in reaction temperature, frequent unplanned equipment shutdowns, and conflicts between production and maintenance decisions. After the platform implementation, through multi-level collaborative control and collaborative decision-making in control and maintenance, the unit's operational stability improved, and equipment reliability was significantly enhanced.

[0099] The platform's adaptive control capability was verified under operating conditions where the raw material properties switched from light to heavy. Key parameter data for the operating condition switching process are shown in Table 1. Table 1: Key parameter data for the operating condition switching process;

[0100] Table 1 shows that during the operating condition switchover, the feed rate gradually increased from 120 tons per hour to 140 tons per hour, with corresponding adjustments to the reactor temperature, catalyst circulation rate, regenerator temperature, and main blower load rate. Through the platform's rapid operating condition identification function, the operating condition switchover signal was detected at the 10-minute mark, and transitional control parameters were immediately activated. Precise operating condition identification was completed at the 30-minute mark, updating the operating condition type to high-load operation with heavy feedstock, and gradually adjusting the controller parameters. Throughout the switchover process, the maximum reactor temperature fluctuation was 2.1 degrees Celsius, which decreased to 0.9 degrees Celsius after the switchover, indicating stable temperature control. Compared to the reactor temperature fluctuation exceeding 5 degrees Celsius under the same scenario before platform implementation, the control effect was significantly improved.

[0101] The platform's adaptive health assessment capability was verified for monitoring the health status of the main fan. The main fan health assessment data is shown in Table 2. Table 2: Main fan health assessment data;

[0102] Table 2 data shows that, under the same operating conditions and load rate, the average vibration of the main fan gradually increased from 2.8 mm / s to 5.1 mm / s, the bearing temperature increased from 76 degrees Celsius to 90 degrees Celsius, the health score decreased from 88 points to 44 points, and the health level dropped from good to abnormal. The platform, through adaptive assessment of operating conditions, accurately identified this as equipment degradation rather than a deterioration caused by changes in operating conditions. On day 20, the platform's degradation trend prediction function predicted that the health score would drop below 60 points within the next 7 days, triggering a maintenance warning. On day 25, when the health score dropped to the attention level, a collaborative decision was triggered, automatically reducing the main fan load rate from 80% to 60% and scheduling a maintenance operation.

[0103] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production, characterized in that: include: The multi-source data acquisition module establishes a unified time base based on edge computing nodes, collects heterogeneous data from multiple sources, performs time alignment processing and quality verification, and obtains a time-synchronized data stream. The working condition identification and feature extraction module uses a sliding window to identify working conditions based on time-synchronized data streams, extracts equipment features through frequency domain analysis and temperature trend analysis, and obtains equipment operating status data through weighted fusion and principal component analysis. The equipment health assessment module calculates the equipment health score based on equipment operating status data and predicts the deterioration trend through exponential smoothing to obtain equipment health information; The adaptive control module designs a multivariable coordinated controller based on equipment health information, and dynamically generates the operating constraints of the multivariable coordinated controller according to the equipment health score to obtain the set values ​​of the operating parameters; The hierarchical optimization control module executes optimization control, coordinated control, and PID control respectively based on the set values ​​of operating parameters, establishes an inter-level feedback mechanism, and derives multi-level coordinated control commands. The collaborative decision-making module, based on equipment health information and multi-level coordinated control commands, generates collaborative decision-making trigger signals according to equipment health information, and uses a rule matching method to generate the final collaborative decision-making scheme.

2. The multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production according to claim 1, characterized in that, The time alignment process includes: Acquire multi-source heterogeneous data collected by edge computing nodes, establish a data buffer for the multi-source heterogeneous data based on a unified time base, aggregate all data with the same timestamp within the same time unit, and eliminate time deviations between data sources through data aggregation processing to obtain a time-unified multi-source data set.

3. The multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production according to claim 1, characterized in that, The collection of multi-source heterogeneous data includes: The system acquires high-frequency vibration data, process operating parameter data, and equipment temperature data from chemical production equipment. The process operating parameter data and equipment temperature data are directly assigned to the corresponding time units according to the timestamp. The high-frequency vibration data is processed using a sliding window statistical method to calculate statistical characteristic values. The statistical characteristic values ​​are then aligned with the time base to form a multi-source data stream with a unified time format.

4. The multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production according to claim 1, characterized in that, The operating condition identification includes: The process operation parameter data in the time-synchronized data stream is acquired. The difference between the process operation parameter data and different operating conditions is calculated based on historical operating data. The difference is measured and calculated using the variance analysis method. The parameter with the largest difference is selected as the operating condition sensitive parameter based on the difference measurement result for subsequent operating condition status judgment.

5. The multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production according to claim 1, characterized in that, The use of a sliding window for operating condition identification includes: The system acquires real-time values ​​of operating condition sensitive parameters, performs time-series analysis on these parameters using a sliding window method, calculates the changing trend of the operating condition sensitive parameters within the window, and calculates the slope using a linear regression method to characterize the changing trend. When the absolute value of the slope of the changing trend is greater than a preset threshold, it is determined that the operating condition switch has started, and an operating condition status indicator is output.

6. The multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production according to claim 1, characterized in that, The weighted fusion includes: The equipment features extracted from frequency domain analysis and temperature trend analysis are obtained. Based on the operating condition identification results, an operating condition feature weight mapping table is established. According to the current operating condition type, the corresponding feature weight coefficient is queried from the operating condition feature weight mapping table. Each component of the equipment feature is multiplied by the corresponding weight coefficient, and a weighted comprehensive state feature vector is obtained through weight fusion processing.

7. The multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production according to claim 1, characterized in that, The generation of the device health information includes: Obtain normal operating data of equipment under different operating conditions from the historical operating database. Calculate the statistical characteristic parameters of the equipment status feature vector sample set under each operating condition. Use the statistical characteristic parameters as the feature baseline of the normal state of the equipment under this operating condition. Establish an operating condition baseline mapping library to store the feature baseline for equipment health score calculation.

8. The multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production according to claim 1, characterized in that, The multivariable coordination controller includes: The equipment health level is obtained from the equipment health score calculation. Based on the equipment health level, the equipment load rate constraint range is determined. When the health level is high, a wider load rate constraint range is set to improve production efficiency. When the health level is low, a stricter load rate constraint range is set to protect equipment safety, thus obtaining dynamic equipment status constraint parameters.

9. The multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production according to claim 1, characterized in that, Also includes: The cloud-edge-device collaborative data management module manages cloud-edge-device collaborative data based on the final collaborative decision-making scheme. It achieves data management and model updates through data collection, edge processing, cloud storage, offline model training, and model deployment.

10. A computer-readable storage medium, characterized in that, It is used to store computer-readable instructions, which, when read by a computer, enable the operation of the multi-level collaborative intelligent control and equipment operation and maintenance data management platform for chemical production as described in any one of claims 1-9.

Images