Robust fault isolation method for industrial dynamic processes based on variational bayesian inference

Abstract

The application discloses a kind of industrial dynamic process robust fault isolation methods based on variational bayesian inference.The data in process industrial production process under normal working condition and fault condition are collected by sensor as the entire data set;Robust fault isolation model is established for the data set under normal working condition and fault condition;Suitable prior distribution is selected for each variable in the established robust fault isolation model, then the method of variational bayesian inference is used to solve, the posterior distribution of each variable is calculated, and fault indication matrix in the solution result is used for fault judgment and isolation.The method of the application can adapt to industrial dynamic process data with abnormal value or larger noise, and can realize accurate fault isolation, to provide effective support for industrial production control behavior.

Description

Technical Field

[0001] This invention relates to an isolation method in the field of process monitoring and fault diagnosis in industrial control systems, specifically a robust fault isolation method for industrial dynamic processes based on variational Bayesian inference. Background Technology

[0002] With the increasing demands for reliability and safety in industrial production processes, data-driven process monitoring has become a simple and effective monitoring method widely used. Among the many methods, fault detection based on Principal Component Analysis (PCA) is one of the fundamental methods. PCA projects data onto the principal component space and residual space, constructing statistics to monitor industrial processes. Subsequently, traditional contribution plot methods are used to locate and isolate fault variables. Reconstruction methods based on the least absolute shrinkage and selection operator (Lasso) are also used for fault variable location and isolation. These methods exhibit high location and isolation performance when data fluctuations are small and interference is minimal. However, when the data contains many outliers or significant noise, traditional fault isolation models become ineffective.

[0003] In recent years, probabilistic fault location and isolation models have also received widespread attention. Compared with general methods, probabilistic methods can effectively combine prior knowledge using appropriate probability distributions and can handle uncertainties and missing data in industrial processes, such as Probabilistic Principal Component Analysis (PPCA) and Hidden Markov Models (HMMs). Meanwhile, to address the impact of outliers on results, researchers have proposed using heavy-tailed distributions or mixture distributions to improve the robustness of the models, such as Student's t-distribution and Gaussian mixture distribution. Researchers have also found that utilizing the invariance of sparse variables in the MMV over a short period can effectively capture the temporal characteristics of process data.

[0004] Therefore, in order to ensure the stable operation of industrial production processes and improve fault isolation performance, a robust fault isolation method for industrial dynamic processes based on variational Bayesian inference was developed to realize the location and identification of industrial fault variables. Summary of the Invention

[0005] To address the technical problems existing in the background art, this invention provides a robust fault isolation method for industrial dynamic processes based on variational Bayesian inference. This method is applicable to fault location problems in complex, large-scale industrial production systems and has far-reaching significance for promoting the development of industrial automation and big data technologies.

[0006] This invention addresses the characteristics of process data containing outliers or significant noise, aiming to meet the robustness requirements of industrial processes. This invention provides effective technical support for robust fault isolation methods in dynamic industrial processes.

[0007] The technical solution adopted in this invention is:

[0008] Step 1: Collect data from the process industry production process under normal and fault conditions using sensors to create the entire dataset.

[0009] Step 2: Establish robust fault isolation models with a window width of d for the datasets under normal operating conditions and fault operating conditions.

[0010] Step 3: Select appropriate prior distributions for each variable in the established robust fault isolation model, and then use variational Bayesian inference to solve the model, calculate the posterior distributions of each variable, and perform fault judgment and isolation based on the fault indication matrix in the solution results.

[0011] In step 1, data collected by the sensor under normal operating conditions and fault conditions in the process industry are used as training data and test data, respectively. The training data and test data are then merged and normalized to obtain a dataset with a mean of zero and a unit variance of one. The training data and test data serve as the datasets for normal operating conditions and fault conditions, respectively.

[0012] In the dataset, both training and testing data consist of data collected sequentially by different sensors. Data collected by all different sensors at the same time point forms a data vector, and a single data vector within the time interval td to t is represented as... Where m represents the total number of sensors in the dataset, τ represents the delay steps, d represents the time delay, t represents the current time, and x r (t-τ) represents the data vector collected at time t-τ. Represents a vector set of size m*1;

[0013] Step 2 specifically involves:

[0014] For training data under normal operating conditions, the following robust fault isolation model is established:

[0015] X={x r}

[0016] x r ={x i}

[0017] D = {D i}

[0018]

[0019] In the formula, X represents the training data under normal operating conditions, x r x represents a data vector under normal operating conditions. i The data vector x represents the data under normal operating conditions. r The i-th measurement value, where D represents the projection matrix, D i It is the i-th row of the projection matrix D, g represents the dimension-reduced vector after the data vector under normal operating conditions is projected onto the dimension-reduced space, and s i The vector a represents the abnormal value indication under normal operating conditions. i The amplitude, v i This represents the result of multiplying the projection matrix by the dimension-reduced vector under normal operating conditions and the i-th measurement value x of the data vector. i The difference between them; a i It is the i-th measurement value x of the data vector under normal operating conditions. i The vector indicating outliers, a i =1 represents x i Includes outliers, otherwise if a i =0 represents x i It does not contain outliers;

[0020] For training data under fault conditions, the following robust fault isolation model is established:

[0021]

[0022]

[0023] D = {D i}

[0024] ΔD={ΔD i}

[0025]

[0026]

[0027] In the formula, This represents training data under normal operating conditions. This represents a data vector under normal operating conditions. Data vector representing normal operating conditions The i-th measurement value, ΔD represents the fault indication matrix, ΔD i This represents the i-th row of the fault indication matrix ΔD. This represents the dimension-reduced vector after projecting the data vector under fault conditions onto the dimension-reduced space. Represents the vector of indicated abnormal values ​​under fault conditions. amplitude, This represents the result of multiplying the projection matrix by the dimension-reduced vector under fault conditions and the i-th measurement value x of the data vector. i The difference between them; It is the i-th measurement value of the data vector under normal operating conditions. The vector of outliers, represent Includes outliers, and vice versa. represent It does not contain outliers; W is the weight matrix; Z is the label parameter matrix of the sparse representation, which controls the sparsity of the fault indication matrix ΔD; symbols This represents the Hadamard product, which is the multiplication of corresponding elements in two matrices.

[0028] This invention includes the addition of s i and a i The product of these factors serves as the robustness factor.

[0029] The robust fault isolation model of the present invention is obtained by linearly reducing the dimension of the multi-measurement vector model (MMV) and then introducing a matrix indicating the fault variable and a robust factor.

[0030] In constructing a robust fault isolation model for fault conditions, this invention establishes a fault indication matrix ΔD, which quantifies the contribution of each variable to the fault. The variable with the highest contribution is the one with the greatest fault. The L2 norm of each element in the fault indication matrix ΔD represents the fault contribution value of each variable. That is, if the L2 norm of a row element is 0, it means that the corresponding variable is not faulty; conversely, if the L2 norm of a row element is not 0, it means that the corresponding variable has a fault.

[0031] In constructing a robust fault isolation model, this invention incorporates robustness factors into both the robust fault isolation model for normal operating conditions and the robust fault isolation model for fault operating conditions. This ingeniously enables accurate identification of outlier location and magnitude, significantly improving the robustness of the method. When data containing outliers is input into the reconstructed SMV model, the robustness factor can correctly identify the location and magnitude of the outliers.

[0032] In step 3, the fault indication matrix ΔD is extracted from the solution results. The L2 norm of each row vector in the fault indication matrix ΔD is calculated as the contribution value of the fault variable. Each row vector represents a sensor. The larger the L2 norm, the greater the fault. Then, a judgment is made.

[0033] If the contribution value of the fault variable is greater than 0, then the sensor location corresponding to the row vector of the fault variable contribution value has a fault. Then the operator adjusts the corresponding sensor according to the specific working condition location of the fault variable to resolve the fault.

[0034] If the contribution value of the fault variable is not greater than 0, then the sensor location corresponding to the row vector of the fault variable contribution value has not experienced a fault and no action is taken.

[0035] In step 3, only the abnormal value vector a under normal operating conditions and fault conditions is selected. i and The label parameter matrix Z is set to a Bernoulli-Gaussian prior distribution, and the dimension-reduced vectors g and g under normal and fault conditions are... The vector of abnormal values ​​a under normal and fault conditions i Amplitude s i and indicator outlier vector amplitude The difference v between normal operating conditions and fault conditions i Sum and Difference The projection matrix D and the weight matrix W are set to Gaussian distribution. This prior distribution setting promotes the sparsity of the fault indication matrix and outlier indication vector, thereby improving the accuracy of subsequent fault isolation judgment.

[0036] This invention preprocesses the training and test data collected by sensors, then uses a multi-measurement vector model to perform time-series modeling of the collected data and capture the dynamics of industrial processes. It also constructs a robust vector to indicate outliers and employs an indicator matrix following a Bernoulli Gaussian distribution to determine fault variables, achieving sparse fault isolation results. Finally, it uses variational Bayesian inference methods for parameter estimation to process the data, adapting to industrial dynamic process data with outliers or significant noise, enabling accurate fault isolation and providing effective support for industrial production control.

[0037] The beneficial effects of this invention are:

[0038] This invention addresses the impact of time-varying characteristics and outliers on fault isolation in industrial production process data by introducing a multi-measurement vector model and an outlier indicator vector based on a Bernoulli Gaussian distribution. This improves the dynamic performance and robustness of fault isolation methods in industrial production processes, enabling data-driven fault isolation models to more accurately locate faulty sensors and more precisely quantify their contribution to the fault, facilitating adjustments or repairs by operators. Attached Figure Description

[0039] Figure 1 This invention relates to the probabilistic graphical model structure;

[0040] Figure 2 The fault scores for each fault sample point involved in this invention;

[0041] Figure 3This is a contribution diagram to the fault isolation involved in the present invention.

[0042] Figure 4 The fault contribution value diagram for this invention. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] The embodiments of the present invention and their implementation process are as follows:

[0045] The following section uses an overheating fault in the water-cooled wall pipes of a coal-fired power plant boiler as an example to describe in detail a robust fault isolation method for industrial dynamic processes based on real data recorded during operation.

[0046] Water-cooled walls are a crucial component of boiler systems. They are neatly arranged metal tubes installed close to the inside of the furnace. They absorb radiant heat from the high-temperature flames or flue gas, generating steam or hot water within the tubes to lower the furnace wall temperature and protect it. To ensure stable operation of coal-fired power plants, the temperature of the water-cooled wall tubes is artificially set to a threshold temperature. In this example, the limit temperature is 440°C. If the temperature of the water-cooled wall tubes exceeds the threshold for an extended period, the metal pipes risk burning out. Therefore, maintaining the pipe temperature below the limit temperature is crucial for safety. The key is monitoring the variables affecting the temperature of the water-cooled wall tubes. A detailed explanation of this process is provided below.

[0047] Table 1. Variables and descriptions of glass melting process and boiler water-cooled wall overheating fault process.

[0048]

[0049] A total of 16 variables were considered in this process, listed in Table 1. Data samples were collected for each variable every 15 seconds, for a total of 1000 samples. Abnormal changes in the combustion swashplate angle and the position of the secondary air damper were the causes of overheating in the boiler water-cooled wall tubes.

[0050] To apply a robust fault isolation method for industrial dynamic processes based on variational inference to diagnose over-temperature faults in boiler water-cooled walls, the following steps were established:

[0051] Step 1: Collect data from the process industry production process under normal and fault conditions using sensors to form the entire dataset; collect process variable data under normal conditions using sensors to form training data; and collect over-temperature fault data using sensors to form test data.

[0052] Step 2: Establish robust fault isolation models with a window width of d for the datasets under normal operating conditions and fault operating conditions.

[0053] Step 3: Select appropriate prior distributions for each variable in the established robust fault isolation model, and then use variational Bayesian inference to solve the model, calculate the posterior distributions of each variable, and perform fault judgment and isolation based on the fault indication matrix in the solution results.

[0054] Determine the prior distribution of each parameter and establish, as follows: Figure 1 The probability diagram shown uses variational Bayesian inference to calculate the posterior distribution of each parameter. The location of the fault and the contribution value of the fault variable are determined by calculating the L2 norm of each row vector in the fault indication matrix.

[0055] Diagnostic results of robust fault isolation in industrial dynamic processes based on variational Bayesian inference, such as... Figure 2 and Figure 3 As shown. Figure 2 The display shows the scores of the fault variables for each sample point. The darker the color, the higher the score, meaning the greater the contribution of the fault. Figure 3 The fault contribution values ​​based on L1 norm reconstruction are shown. Figure 4 The fault contribution value obtained by this invention is shown. From... Figure 3 As can be seen, the method based on the L1 norm reconstruction incorrectly identifies furnace pressure (variable 8) as a fault variable, and the contribution values ​​of other variables to the fault are not zero. In actual operation, the only variables that cause faults are the main steam temperature (variable 4), the sway angle position of burner #1 (variable 11), and the position of secondary air damper #1 (variable 13). Therefore, the proposed method can accurately locate the fault variables and quantify their contribution values.

[0056] This invention proposes a robust fault isolation method for industrial dynamic processes based on variational Bayesian inference. It can effectively diagnose fault variables causing overheating in boiler water-cooled wall pipes during industrial dynamic processes and processes containing outliers, namely, main steam temperature (variable 4), #1 burner tilt position (variable 11), and #1 secondary air damper position (variable 13). Furthermore, it can accurately quantify their fault contribution values, with the #1 burner tilt position showing the largest fault contribution. In summary, this invention effectively improves the accuracy of fault isolation.

Claims

1. A robust fault isolation method for industrial dynamic processes based on variational Bayes, characterized in that: The method includes the following steps: Step 1: Collect data from the process industry production process under normal and fault conditions using sensors to create the entire dataset. Step 2: Establish robust fault isolation models for datasets under normal operating conditions and fault operating conditions. Step 3: Select appropriate prior distributions for each variable in the established robust fault isolation model, and then use variational Bayesian inference to solve the model, calculate the posterior distributions of each variable, and perform fault judgment and isolation based on the fault indication matrix in the solution results. In step 1, the data collected by the sensor under normal operating conditions and under fault conditions in the process industry production process are used as training data and test data, respectively. The training data and test data are merged and normalized to obtain a dataset with a mean of zero and a unit variance of one. Step 2 specifically involves: For training data under normal operating conditions, the following robust fault isolation model is established: X={x r } x r ={x i } D ={ } In the formula, X represents the training data under normal operating conditions, x r x represents a data vector under normal operating conditions. i The data vector x represents the data under normal operating conditions. r The i-th measurement value, where D represents the projection matrix. It is the i-th row of the projection matrix D, g represents the dimension-reduced vector after the data vector under normal operating conditions is projected onto the dimension-reduced space, and s i Represents an indicator vector of abnormal values ​​under normal operating conditions. The amplitude, v i This represents the result of multiplying the projection matrix by the dimension-reduced vector under normal operating conditions and the i-th measurement value x of the data vector. i The difference between them; It is the i-th measurement value x of the data vector under normal operating conditions. i The vector of outliers, represent Includes outliers, and vice versa. represent It does not contain outliers; For training data under fault conditions, the following robust fault isolation model is established: D ={ } ={ } In the formula, This represents training data under normal operating conditions. This represents a data vector under normal operating conditions. Data vector representing normal operating conditions The i-th measurement value, Represents the fault indication matrix. Fault indication matrix The i-th row, This represents the dimension-reduced vector after projecting the data vector under fault conditions onto the dimension-reduced space. Represents the vector of indicated abnormal values ​​under fault conditions. amplitude, This represents the result of multiplying the projection matrix by the dimension-reduced vector under fault conditions and the i-th measurement value x of the data vector. i The difference between them; It is the i-th measurement value of the data vector under normal operating conditions. The vector of outliers, represent Includes outliers, and vice versa. represent It does not contain outliers; It is a weight matrix; It is a sparse representation of the label parameter matrix; symbols This represents the Hadama product; In step 3, only the abnormal value vectors under normal and fault conditions are included. and The label parameter matrix Z is set to a Bernoulli-Gaussian prior distribution, and the dimension-reduced vectors g and g under normal and fault conditions are... Vector of abnormal values ​​under normal and fault conditions Amplitude and indicator outlier vector amplitude The difference v between normal operating conditions and fault conditions i Sum and Difference The projection matrix D and the weight matrix W are set to Gaussian distribution.

2. The robust fault isolation method for industrial dynamic processes based on variational Bayes as described in claim 1, characterized in that: In the dataset, both training and testing data consist of data collected sequentially by different sensors. Data collected by all different sensors at the same time point forms a data vector, and a single data vector within the time interval td ~ t is represented as... Where m represents the total number of sensors, Indicates the delay steps. This represents the time delay, where t represents the current time. Indicates t- Data vectors collected at any time, This represents a vector set of size m*1.

3. The robust fault isolation method for industrial dynamic processes based on variational Bayes as described in claim 1, characterized in that: In step 3, the fault indication matrix is ​​extracted from the solution results. Calculate the fault indication matrix The L2 norm of each row vector is used as the contribution value of the fault variable, and then a judgment is made: If the contribution value of the fault variable is greater than 0, then the fault occurs at the location of the sensor corresponding to the row vector of the fault variable contribution value. Then, the sensor is adjusted according to the specific working condition location of the fault variable to resolve the fault. If the contribution value of the fault variable is not greater than 0, then the sensor location corresponding to the row vector of the fault variable contribution value has not experienced a fault and no action is taken.

Images