Design of Expert System for Trend Diagnosis of Process Parameters Based on Data Mining

1) Wang Zhu, China University of Petroleum (Beijing), School of Information Science and Engineering/School of Artificial Intelligence, Department of Automation, Teacher

2) Wang Shaoxian, China University of Petroleum (Beijing), School of Information Science and Engineering/School of Artificial Intelligence, Control Engineering, Master’s Student 21

3) Zhang Shaokang, Manager, Electrical Instrument Department, Sinopec Shijiazhuang Refinery

4) Zhan Jiale, China University of Petroleum (Beijing), School of Information Science and Engineering/School of Artificial Intelligence, Control Science and Engineering, Master’s Student 22

To diagnose abnormal trends in process parameters during industrial processes, this paper proposes an expert system based on rolling data KPCA and SVDD (ES-KPCA and ES-SVDD). This expert system utilizes data mining techniques to identify significant trend anomalies and fluctuations in process parameters, triggering alarms in a timely manner. The system includes rule-based stability assessment of process parameters to evaluate whether they are stable. When parameters are unstable, the rolling data KPCA and SVDD methods are used to diagnose abnormal trends in process parameters. Finally, the expert system is experimentally validated using UniSim simulations and data from chemical plants. Results show that this expert system has good diagnostic performance for abnormal trends in process parameters.

In petrochemical DCS, alarms related to process parameters are often about setting trigger alarm lines. However, when significant anomalies or fluctuations in process parameters occur before reaching the alarm line, there are no alarm prompts. Therefore, diagnosing and alarming the volatility of process parameters addresses this gap, and volatility alarms before triggering alarm lines can serve as early warnings for processes. This plays a critical role in timely emergency handling and maintenance inspections on-site.

Step 1: Obtain process parameter data through interfaces.

Step 2: Construct a forward time series judgment data matrix and diagnosis vector.

Step 3: If the forward time series judgment data matrix is successfully normalized, assess the stability of the process parameters and utilize rolling data KPCA and SVDD for anomaly trend detection.

Step 4: If the forward time series judgment data matrix normalization fails, assess the stability of the process parameters.

Step 5: Offline parameter adjustment phase: Adjust threshold parameters and check alarm locations to ensure their rationality.

Step 6: Online diagnosis phase: Obtain and apply threshold parameters adjusted in the offline phase.

Figure 1 Algorithm Flow Chart

1. Rolling Data Matrix

Based on different real-time moments i, construct the most recent parameter column vector as the diagnosis vector to be identified, and then construct the forward time series data matrix for similarity comparison, as shown below.

Forward Time Series Judgment Data Matrix Diagnosis Vector

2. Rule-Based Stability Assessment of Process Parameters

Obtain historical data of process parameters, construct a rolling data matrix, and calculate the mean and standard deviation by columns.

Calculate the mean and standard deviation of all data in the rolling data matrix. If the current data satisfies Equation (2), set outliers equal to the mean.

Perform statistical analysis on the current process parameters.

Where, if Equation (3) is satisfied, the current process parameters exhibit abnormal volatility.

3. Parameter Trend Diagnosis Based on Rolling Data KPCA

Rolling data KPCA is performed by transforming the original data into the forward time series judgment data matrix and diagnosis vector, utilizing KPCA for feature extraction of the forward time series data judgment matrix, and then conducting diagnosis through similarity comparison, as shown in the diagnostic flow in Figure 2.

Figure 2 Rolling Data KPCA Process Parameter Abnormal Trend Diagnosis Flow

4. Parameter Trend Diagnosis Based on Rolling Data SVDD

By transforming the original data into the forward time series judgment data matrix and diagnosis vector, utilizing SVDD to train the forward time series judgment data matrix into a hypersphere, and recording the center and radius of this hypersphere. Then, using the SVDD similarity comparison method for fault alarming. The specific flow chart is shown in Figure 3.

Figure 3 Rolling Data SVDD Process Parameter Abnormal Trend Diagnosis Flow

Data from the offline debugging parameter phase of ES-KPCA and ES-SVDD totaled 8726 sets, with 687 diagnoses performed.

(a)

(b)

Figure 4 Offline Parameter Adjustment Results

Table 1 Related Parameter Adjustment Results

During online diagnosis, three process parameters were diagnosed online a total of 444 times. Comparative experiments were set with ES-PCA and ES-OCSVM.

The online diagnosis results are as follows:

(a)

(c)

Figure 5 Online Diagnosis Results

In Figure 5a, the ES-PCA method detected 9 significant pulses and 2 significant operational adjustments, with 3 false alarms for minor operational adjustments. In Figure 5b, the ES-OCSVM method detected 9 significant pulses and 2 significant operational adjustments. Figures 5c and d show that the proposed method detected all 10 significant pulses and 2 significant operational adjustments without any false alarms.

This paper describes in detail an expert system for diagnosing trends in process parameters based on data mining. The expert system combines assessments of process parameter stability and data cleaning. When process parameters are unstable, it further evaluates abnormal trends using rolling data KPCA and SVDD methods. Additionally, effectiveness validation was conducted using UniSim and data from domestic refineries, and comparative experiments were carried out using ES-PCA and ES-OCSVM. Results indicate that the proposed expert system effectively diagnoses significant fluctuations and abnormal trends in process parameters, with diagnostic performance superior to ES-PCA and ES-OCSVM methods through ablation studies. Therefore, the expert system can monitor abnormal changes in process parameters in real-time and take timely actions to improve production efficiency and quality levels.

Wang Zhu, Associate Professor

Male, Ph.D., Communist Party member, currently an associate professor and master’s supervisor at China University of Petroleum (Beijing). He has been working in the Department of Automation at China University of Petroleum (Beijing) since 2016 and is currently a council member of the Beijing Artificial Intelligence Society and a youth member of the Information Technology Application Committee of the Chinese Chemical Society.

He has been engaged in research on system identification and intelligent control, petrochemical process fault early warning, and deep learning-based time series forecasting, publishing more than 20 high-level academic papers as the first or corresponding author. He has presided over National Natural Science Foundation projects and several corporate projects.