A method and system for intelligent detection of small faults based on weighted divergence

Abstract

The invention relates to the technical field of minor fault detection, and discloses a method and system for intelligent minor fault detection based on weighted divergence. The detection method comprises: collecting offline production data at various moments of an industrial operation process; constructing a principal component analysis model and extracting principal components; calculating a data score matrix; performing a JS divergence test under a sliding window strategy; performing local outlier outlier detection on a sequence of JS divergence statistics by a local outlier factor weighting method, and introducing the obtained local outlier weight factor into the JS divergence statistics; fitting the distribution of data by a smooth kernel function; calculating the weighted JS divergence statistics of online production data, and judging whether a fault occurs. The invention constructs an efficient and stable minor fault detection model by introducing the JS divergence calculation weighted by the local outlier factor and combining the principal component analysis technology, thereby improving the detection capability of minor faults under specific conditions or in a local area.

Description

Technical Field

[0001] The present invention relates to the technical field of minor fault detection, and in particular to a method and system for intelligent minor fault detection based on weighted divergence. Background Art

[0002] As industrial processes become more modernized and the scale of industrial processes becomes larger and more complex, the number of sensors in industrial processes increases rapidly, leading to a sharp increase in minor faults in the operation process. How to accurately and effectively detect minor faults in industrial processes has become one of the hot issues in the field.

[0003] According to the fault characteristics and their development mode, minor faults can be roughly divided into the following three categories: slow-changing minor faults, sudden minor faults and intermittent minor faults. The fault characteristics of slow-changing minor faults are not obvious, the amplitude is low and the development is slow, but they will gradually accumulate over time. For example, bearings gradually wear out during long-term friction and load use, and the bearing performance declines, which may eventually have a significant impact on system performance or equipment safety; while sudden minor faults have a certain amplitude at the beginning of the fault, and may suddenly deteriorate in a short period of time. Although it has not yet reached the level of significantly affecting system performance, its greater uncertainty and suddenness may cause minor faults to expand rapidly and cause more serious faults; for intermittent minor faults that appear and disappear from time to time and are unstable, intermittent abnormal readings or signal interruptions may occur when the external environment interferes or the internal components age.

[0004] The root causes of minor faults are diverse, including natural aging of equipment, negligence and errors in operation, negligence of human factors, and insufficient maintenance. If minor faults are ignored in the early stage and are not detected, identified and properly handled in time through detection algorithms or systems, over time, minor faults may eventually evolve into serious system failures or even mechanical disasters, endangering personnel safety, causing significant economic losses such as a surge in equipment repair and replacement costs, a sharp drop in production efficiency, and unplanned downtime losses.

[0005] Therefore, the detection of minor faults has become the lifeline for maintaining the efficient operation of industrial systems. It is not only a guarantee for the smooth operation of equipment, but also the cornerstone of the reliability, availability and maintainability of the industrial system, which is directly related to the safety of operators and the economic health of enterprises. Therefore, the industry is increasing its investment in the research and development of fault warning and health management technologies, using advanced monitoring and analysis methods to intervene in the early stages of faults to ensure the continuous stability and safety of production processes.

[0006] According to the nature and development trend of different faults, data-driven methods can be used for targeted detection and early warning. In practical applications, various detection technologies and methods are used in combination with the characteristics of specific equipment and systems to timely discover and handle minor faults and ensure the safe, stable and efficient operation of industrial systems. The core advantage of data-driven minor fault detection methods is that they do not rely on detailed mathematical models or complex physical mechanisms of the system, but directly use process data collected in real time by a large number of sensors throughout the production line. Data-driven methods achieve industrial process monitoring by modeling sensor data.

[0007] Principal component analysis (PCA) in data-driven analysis can effectively reduce the dimensionality of high-dimensional and complex raw data, while retaining the main elements of the raw data and simplifying the complexity of data analysis; however, due to the low value of minor faults, weak signals, and easy to be masked by environmental noise, their feature extraction is extremely challenging, resulting in the difficulty of detecting minor faults in time by traditional fault detection methods based on principal component analysis and canonical variable analysis. This type of method relies too much on the changes in the mean and variance of the original data, but minor faults change weakly, and the variance of the original data does not change significantly and is difficult to capture, and is easily ignored, making it impossible for traditional fault detection methods to accurately and timely identify minor faults.

[0008] Some technical solutions in the prior art overcome this limitation by combining algorithms such as nonlinear processing, machine learning, and new data-driven methods with advanced technologies such as time series analysis. Among them, divergence-based micro-fault detection has good results, especially KL divergence (Kullback-Leibler divergence), which shows good application effects because it can measure the difference between two probability distributions. KL divergence is an asymmetric measure of the difference between two probability distributions P and Q. It represents the amount of information lost when distribution Q is used to approximate distribution P. KL divergence can measure the difference in probability distribution between normal state and potential fault state in industrial processes. By comparing the KL divergence values ​​of real-time data and historical normal data, it can be determined whether a micro-fault has occurred in the system. However, due to the asymmetry of KL divergence, using KL divergence alone may not fully reflect the complexity of the fault state. Summary of the invention

[0009] In order to solve the above technical problems, the present invention provides a method and system for intelligent detection of minor faults based on weighted divergence, aiming to improve the sensitivity, accuracy and stability of fault detection by optimizing the divergence calculation method, and realize intelligent monitoring and early warning of industrial processes.

[0010] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0011] A method for intelligent detection of minor faults based on weighted divergence, comprising:

[0012] Collect offline production data at each moment of the industrial operation process and divide the offline production data into training data sets And the test dataset ;

[0013] based on Construct the principal component analysis model and obtain the load matrix , determine the number of principal components of the principal component analysis model ;based on , and Calculate the training data score matrix in the principal component subspace and the test data score matrix ;

[0014] Given the width W of the sliding window, calculate The i-th test score in Variance of the first W elements ,calculate The i-th training score in Variance of the first W elements , ;based on and Perform JS divergence test; obtain a sequence composed of JS divergence statistics;

[0015] The local outlier weight factor is introduced into the JS divergence statistic through the local outlier factor weighting method to obtain the weighted JS divergence statistic ; Based on a given impact factor ,set up Control limits ;

[0016] Collect online production data of current industrial operation processes ,calculate The weighted JS divergence statistic , by comparison and The size of the fault is used to determine whether a fault has occurred.

[0017] Furthermore, the Construct the principal component analysis model and obtain the load matrix , determine the number of principal components of the principal component analysis model , specifically including:

[0018] Based on the training data set Construct a principal component analysis model and obtain a training data set The corresponding eigenvalues ​​of the covariance matrix and the corresponding unit eigenvector, and the load matrix ; is the mth eigenvalue of the covariance matrix, and m is the total number of observed variables in the offline production data; the cumulative variance contribution rate method is used to determine the number of principal components of the principal component analysis model :

[0019] Starting from the largest eigenvalue, the variance contribution rates corresponding to the eigenvalues ​​are accumulated in sequence until the cumulative variance contribution rate reaches the preset threshold. :

[0020] ;

[0021] represents the i-th eigenvalue of the covariance matrix, represents the jth eigenvalue of the covariance matrix, Before The cumulative contribution rate of variance corresponding to the principal component.

[0022] Furthermore, the , and Calculate the training data score matrix in the principal component subspace and the test data score matrix , specifically including:

[0023] Through the load matrix and training dataset The corresponding matrix , calculate the training data score matrix in the principal component subspace ;

[0024] ;

[0025] Through the load matrix And the test dataset The corresponding matrix , calculate the test data score matrix in the principal component subspace ;

[0026] ;

[0027] for The training scores, for The Test scores.

[0028] Furthermore, the online production data of the current industrial operation process is collected ,calculate The weighted JS divergence statistic , by comparison and The size of the fault is used to determine whether a fault has occurred, which specifically includes the following steps:

[0029] S71, continuously collects online production data of the current industrial operation process , forming an online dataset ; for Middle Online production data, for The total number of online production data;

[0030] S72, using the principal component analysis model and utilizing the load matrix Calculate the score matrix for online production data ; For online datasets The corresponding matrix, Scoring online production data;

[0031] S73, given sliding window width , through the JS divergence test under the sliding window strategy, calculate the mean of the online production data score in each sliding window and variance , and combined with the training data score to obtain the online JS divergence statistics of the online production data ;

[0032] S74, online JS divergence statistics by local outlier factor weighting method Perform local outlier detection and introduce the obtained local outlier weight factor into In the example, we get the weighted JS divergence statistic of the online production data. ;

[0033] S75, by judgment and The size of the fault is used to determine whether a fault has occurred: , then the current industrial operation process is a fault process; if , then the current industrial process is a normal process;

[0034] S76, collect online production data of the next industrial process, and perform steps S72 to S75 to continuously perform online fault detection.

[0035] A small fault intelligent detection system based on weighted divergence, comprising:

[0036] The data acquisition module collects offline production data at each moment of the industrial operation process and divides the offline production data into training data sets And the test dataset ;

[0037] Model building module, based on Construct the principal component analysis model and obtain the load matrix , determine the number of principal components of the principal component analysis model ;based on , and Calculate the training data score matrix in the principal component subspace and the test data score matrix ;

[0038] Divergence test module, given the width W of the sliding window, calculates The i-th test score in Variance of the first W elements ,calculate The i-th training score in Variance of the first W elements , ;based on and Perform JS divergence test; obtain a sequence composed of JS divergence statistics;

[0039] The control limit setting module introduces the local outlier weight factor into the JS divergence statistic through the local outlier factor weighting method to obtain the weighted JS divergence statistic ; Based on a given impact factor ,set up Control limits ;

[0040] Fault diagnosis module, collects online production data of the current industrial operation process ,calculate The weighted JS divergence statistic , by comparison and The size of the fault is used to determine whether a fault has occurred.

[0041] Compared with the prior art, the beneficial technical effects of the present invention are:

[0042] Through refined data preprocessing and principal component analysis, the present invention can accurately extract key information in the industrial process, remove redundant data, and provide a high-quality data foundation for subsequent fault detection; secondly, the introduction of sliding window strategy and JS divergence test, combined with the local outlier factor weighting method, enables the present invention to keenly capture subtle changes in the data, especially those abnormal signals that may be ignored in traditional detection methods. This differentiated weight distribution mechanism allows the detection system to focus more on areas that may indicate the occurrence of faults, thereby achieving early warning of potential faults.

[0043] Furthermore, the present invention uses the kernel density estimation (KDE) method to set reasonable control limits, which not only considers the overall distribution characteristics of the data, but also takes into account the local changes of the data, making the setting of the control limits more scientific and reasonable, greatly improving the reliability and stability of fault judgment, and reducing the situation of false alarms and missed alarms. Finally, the real-time data processing and dynamic analysis capabilities of the online fault detection part enable the detection system to quickly respond to any abnormal changes in the industrial process and send out an alarm signal at the first time. This rapid response mechanism provides strong support for the timely adjustment of the production process and the rapid repair of faults, ensuring the continuity and stability of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 The figure is a schematic diagram of a method for intelligently detecting small faults based on weighted divergence in one embodiment of the present invention.

[0046] Figure 2 This is a process diagram of a continuous kettle stirring and heating process in an embodiment of the present invention.

[0047] Figures 3(a) to 3(d) are respectively four effect diagrams of the detection methods for the small precision drop fault in the continuous kettle stirring heating process in one embodiment of the present invention. Among them, Figure 3(a) is the effect diagram of the detection method based on principal component analysis, Figure 3(b) is the effect diagram of the KL divergence detection method based on the PCA framework, Figure 3(c) is the effect diagram of the JS divergence detection method based on the PCA framework, and Figure 3(d) is the effect diagram of the small fault intelligent detection method based on weighted divergence of the present invention.

[0048] Figures 4(a) to 4(d) are respectively four effect diagrams of detecting small deviation faults in a continuous kettle stirring heating process in one embodiment of the present invention. Among them, Figure 4(a) is an effect diagram of a detection method based on principal component analysis (PCA), Figure 4(b) is an effect diagram of a KL divergence detection method based on a PCA framework, Figure 4(c) is an effect diagram of a JS divergence detection method based on a PCA framework, and Figure 4(d) is an effect diagram of a small fault intelligent detection method based on weighted divergence.

[0049] Figures 5(a) to 5(d) are respectively four effect diagrams of detecting small drift faults in a continuous kettle stirring heating process in an embodiment of the present invention. Among them, Figure 5(a) is an effect diagram of a detection method based on principal component analysis, Figure 5(b) is an effect diagram of a KL divergence detection method based on a PCA framework, Figure 5(c) is an effect diagram of a JS divergence detection method based on a PCA framework, and Figure 5(d) is an effect diagram of a small fault intelligent detection method based on weighted divergence of the present invention.

[0050] Figure 6 This is a comparison diagram of the fault detection rate (FDR) when performing minor fault detection on a continuous kettle stirring heating process with three sensor types using four detection methods in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to clearly and completely describe the method and system for intelligent detection of minor faults of the present invention, the following will be described in detail in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] The present invention provides a method for intelligently detecting small faults based on weighted divergence, which specifically includes the following steps:

[0053] Step S1, in the normal industrial operation process, production data is collected through sensors, instruments and other equipment. The production data may include temperature, pressure, flow, speed and other observation variables to construct training data sets and test data sets.

[0054] In order to ensure the consistency and accuracy of data analysis and eliminate the influence of significant variables in the dimensionality reduction process, the collected production data needs to be preprocessed:

[0055] The matrix composed of production data The mean and standard deviation of each column in and , the matrix composed of normalized production data is expressed as :

[0056] ; (1)

[0057] The preprocessed production data is divided into a training data set and a test data set. The former is used to build the principal component analysis model, and the latter is used to verify the performance of the principal component analysis model.

[0058] Training Dataset , the corresponding standardized matrix In the present invention, the principal component analysis (PCA) technique is used to model the training data set, and the training data is projected onto the principal component subspace to obtain its score in the principal component direction. These scores reflect the position of the training data in the principal component subspace, laying the foundation for subsequent fault detection based on JS divergence. Next, the covariance matrix is ​​subjected to eigenvalue decomposition to obtain a series of eigenvalues ​​and their corresponding unit eigenvectors, and the load matrix is ​​obtained at the same time. :

[0059] ; (2)

[0060] in, contains all eigenvalues ​​of S, represents a diagonal matrix, S represents the covariance matrix, represents the operation of constructing a diagonal matrix, represents the feature vector matrix, n is the training data set The total number of training data in , m is The total number of observed variables in . And the eigenvalues ​​are required to satisfy: , and the unit eigenvector matrix corresponding to the corresponding eigenvalue is recorded as , that is, the loading matrix is ​​obtained. The eigenvalue and eigenvector respectively represent the variability size and variability direction of the training data in different directions.

[0061] This method aims to propose an intelligent detection method for minor faults in industrial processes by combining the advantages of principal component analysis (PCA) and Jensen-Shannon divergence (JS divergence). This method uses the PCA dimensionality reduction strategy to retain the most important data feature elements from a large amount of complex original data to reduce computational complexity and improve detection efficiency. At the same time, it combines the JS divergence that measures the probability density distribution to detect minor faults. Through the principal component analysis method, the eigenvectors with a cumulative variance contribution rate of 90% are selected as the principal components, and the space spanned by these eigenvectors is the principal component subspace of PCA. The principal component subspace is the feature space that retains the most important data in the original data. The number of eigenvectors selected by this method is less than the total number of observed variables in the original data, achieving the purpose of dimensionality reduction. The JS divergence detection model is constructed in the PCA principal component subspace, and the intelligent detection of minor faults is achieved by comparing the JS divergence values ​​of real-time data and historical normal data. This preprocessing method can ensure the extraction of minor fault signal features in a complex industrial environment, creating advantages for the difference measurement of JS divergence detection.

[0062] Further calculate the training data score matrix :

[0063] ; (3)

[0064] for The mth training score in .

[0065] Step S2, using the cumulative variance contribution method (CPV) to determine the number of principal components. Starting from the largest eigenvalue, the variance contribution rates corresponding to the eigenvalues ​​are accumulated in sequence until the cumulative contribution rate reaches a preset threshold value, which is 90% in the present invention. The formula is as follows:

[0066] ; (4)

[0067] The load matrix is ​​written as , the cumulative variance contribution rate determines the previous The principal component subspace consists of , and the remaining The residual subspace composed of components .

[0068] Established in the principal subspace The statistical test is performed on the statistics, and the fluctuation of the principal vector modulus of the principal subspace is used to reflect the changes of multiple variables and monitor the process. The statistic can be expressed as:

[0069] ; (5)

[0070] Is an integral symbol, indicating Statistics; In addition, SPE represents the squared prediction error in the residual subspace and can be defined as:

[0071] ; (6)

[0072] in, yes The identity matrix of .

[0073] Step S3, given the impact factor In the case of The upper control limit of the SPE statistic can be calculated from its approximate distribution:

[0074] ; (7)

[0075] in, The degrees of freedom are and On the F distribution critical point, , .

[0076] is a normal distribution at the test level In industrial processes, the following logical conditions are used to determine whether a fault has occurred:

[0077] ; (8)

[0078] Secondly, for the residual subspace, if the SPE statistic is greater than If the control limit is greater than 0.05, the process is a fault process; otherwise, the industrial process is a normal process.

[0079] Step S4, JS divergence, i.e. Jensen-Shannon divergence, is a similarity measure between two probability distributions. It is improved based on Kullback-Leibler divergence (KL divergence), solves the problem of asymmetric KL divergence, and has some unique properties. JS divergence can more effectively detect minor faults by calculating the JS divergence of the probability distribution of real-time data and historical normal data. In addition, the smoothness of JS divergence helps to reduce the impact of noise on the detection results. This method can accurately monitor the state of industrial processes, thereby ensuring the safety and stability of industrial process production.

[0080] definition and is a continuous variable JS divergence can quantify the difference between two probability distributions and the similarity between two probability distributions. is defined as:

[0081] ; (9)

[0082] in, For two different probability distributions and The average distribution of .

[0083] Represents the KL divergence of two probability distributions, calculated as:

[0084] . (10)

[0085] KL divergence is asymmetric, so when using the opposite distribution formula to calculate the distance of data with the same probability distribution, KL divergence will produce two different asymmetric quantities. Compared with KL divergence, JS divergence is symmetric, and the statistical quantity detection is more stable, that is, .

[0086] Then, the JS divergence statistic is calculated. According to the load matrix obtained in step S1 of the present invention, , calculate the score matrix of the training data in the principal subspace and the score matrix of the test data in the principal component subspace .

[0087] Given a sliding window width W, calculate the i-th training score and the i-th test score The variance of the first W elements of and . and Satisfies Gaussian distribution and ,in and are the mean and standard deviation of the normal industrial operation process, and are the mean and standard deviation of the faulty industrial operation process. For the convenience of description, the training score and test scores The variances of the first W elements of are recorded as and , for the training score and test scores JS Divergence Perform JS divergence test:

[0088] ; (11)

[0089] Among them, the intermediate variable , is the variance of random noise in the industrial operation process.

[0090] Finally, the JS divergence detection formula can be simplified to:

[0091] ; (12)

[0092] Analysis shows that the JS divergence increases with the increase in the number of faults. Taking the sensor accuracy degradation fault as an example, when the variance of the fault data is greater than the variance of the standard data, the JS divergence increases with the increase in the variance of the fault data.

[0093] Step S5, the implementation of the local outlier factor weighting method specifically includes:

[0094] The local outlier factor (LOF) algorithm is a density-based unsupervised outlier detection algorithm. The core idea of ​​the algorithm is to identify outliers by comparing the local density difference between a data point and other data points in its neighborhood. In the local outlier factor algorithm, the higher the local density of a data point, the lower the possibility of it being an outlier; conversely, the lower the local density, the higher the possibility of it being an outlier. The local outlier factor algorithm is effective in detecting faults that only occur under specific conditions or in a local area. By quantifying the outliers in the data set and assigning the local outlier factor as a weight to the JS divergence statistic, the fault detection rate can be improved to a certain extent. This method can fully mine the probability distribution information implicit in the industrial process data, re-weight the JS divergence according to the degree of local faults, emphasize those faults that contain more important information, make the weight assignment in fault detection more reasonable, and overcome the disadvantage of a single JS divergence that treats all fault components equally.

[0095] The present invention provides a method for weighting outliers based on local outlier factors to emphasize JS divergence components that are more closely related to the degree of faults, so that the re-weighted improved JS divergence has better fault detection performance.

[0096] In the local outlier factor algorithm, the Euclidean distance is used to calculate the distance between data points. The JS divergence detection is integrated, and the weights of the JS divergence distance and the Euclidean distance are weighted to obtain a similarity measure that considers both spatial location and probability distribution. Next, the local outlier factor outlier point detection is performed, which specifically includes the following steps:

[0097] 1. Calculate the Euclidean distance between each JS divergence statistic of the sequence of JS divergence statistics and other JS divergence statistics ,in yes The zth neighbor of , for The total number of neighbors.

[0098] 2. Calculate the JS divergence statistic The reachable distance to k neighbors :

[0099] ; (13)

[0100] in, for The Euclidean distance to the kth adjacent JS divergence statistic.

[0101] 3. Calculate the divergence statistics for each JS The local reachable density :

[0102] ; (14)

[0103] 4. Calculate all JS divergence statistics The local outlier factor ,Right now

[0104] ; (15)

[0105] In the present invention, whether the JS divergence statistic is an outlier factor is determined by judging whether the local outlier factor is close to 1. If the local outlier factor is much larger than 1, the JS divergence statistic is considered to be an outlier factor, which indirectly indicates that a fault occurs at the time point corresponding to the JS divergence statistic; however, if the local outlier factor is close to 1, the time point corresponding to the JS divergence statistic is considered to be normal. Therefore, the local outlier factor can indirectly describe the degree of fault at that moment in the entire industrial process. The JS divergence detection method based on the weighting of the local outlier factor can well give a larger weight to places with large faults. The use of the local outlier factor can better emphasize those detection statistic components rich in more important fault information.

[0106] Because the local reachable density of the JS divergence component is the inverse of the average reachable distance of the nearest neighbor based on the KL divergence, the larger the relative distance at this moment, the smaller the density, and the larger the amplitude of the corresponding local outlier factor; otherwise, the smaller the relative distance at this moment, the larger the density, and the smaller the amplitude of the corresponding local outlier factor.

[0107] Therefore, in the present invention, the JS divergence statistic is weighted in the form of exponential operation based on the local outlier factor, and the weighted JS divergence statistic is expressed as :

[0108] ; (16)

[0109] The local outlier factor is the weight factor , because the value of JS divergence is less than 1, and It is consistent with the weighting logic in fault detection. Therefore, the local outlier factor weighting method follows the rule of focusing on key faults to assign weights, making the weight assignment in fault detection more reasonable.

[0110] Step S6, the present invention adopts the kernel density estimation (KDE) method to determine the control limit of the statistic.

[0111] Kernel density estimation is a non-parametric method for estimating the probability density function of a random variable. This method does not require any prior assumptions about the distribution of the data, but directly starts from the data sample itself and fits the distribution of the data through a smooth kernel function. Kernel density estimation methods are widely used in statistics, machine learning, and data science, especially in data visualization, anomaly detection, density estimation, and other fields.

[0112] The kernel density estimation in the present invention can handle distributions of any shape, and the kernel density estimation method can establish reasonable control limits for the weighted improved JS divergence. The specific steps are as follows:

[0113] Setting the Impact Factor , using the kernel density estimation method to estimate the weighted JS divergence statistic Control limits:

[0114] ; (17)

[0115] in, is the probability density function. The core idea of ​​this formula is to regard each data point as a small probability distribution, and then superimpose these small probability distributions to form an overall density estimate.

[0116] In the present invention, the control limit setting methods of KL divergence detection, JS divergence detection and JS divergence detection weighted by local outlier factors all adopt kernel density estimation.

[0117] In actual industrial processes, With control limits , the following logic is used for fault detection:

[0118] ; (18)

[0119] in, is the query value of the observed variable. If the weighted JS divergence statistic Below control limit , it indicates that the current industrial operation process is normal; otherwise, if the weighted JS divergence statistic Beyond the limits of control , it indicates that the current industrial operation process is a fault process, and the system will issue an alarm to remind operators to solve minor faults in time.

[0120] The overall process of the weighted divergence-based intelligent detection method for small faults in the present invention is as follows: Figure 1 In order to fully describe the weighted divergence-based intelligent detection method for small faults in the embodiment of the present invention, the proposed method is applied to the continuous kettle stirring heating process (CSTH) for detailed description.

[0121] The continuous kettle stirring and heating process (CSTH) is an experimental platform developed by the Department of Chemical Engineering at the University of Alberta. The process mainly consists of a reactor body, a stirrer, a heating device (such as a jacket), a cooling device (such as a cooling water cycle), a feed port, a discharge port, a temperature sensor, a pressure sensor, etc. The heating device and the cooling device are used to increase and decrease the temperature in the kettle respectively to meet the reaction requirements; the process mixes the reactants in a certain proportion and puts them into the reactor together with the catalyst, stirs the reactants to make them evenly mixed in the reactor, and controls the reaction temperature through the heating device to achieve a specific chemical reaction or physical change; the process flow chart of the continuous kettle stirring and heating process is shown in the figure Figure 2 shown.

[0122] Figure 2 In the figure, T represents temperature, L represents tank liquid level, and F represents cold water flow rate; T m represents the transmitter, C represents the PI controller. TC, LC, and FC represent the PI controller for temperature, the PI controller for tank level, and the PI controller for cold water flow, respectively. TT m ,LT m , FT m The transmitters respectively represent the temperature, the tank level, and the cold water flow. The continuous kettle stirring and heating process in the present invention includes three types of sensor data: tank level L, cold water flow F, and outlet water temperature T. Production data ,in, They represent the PI controller measurement variables of the tank level L, the PI controller measurement variables of the cold water flow F, and the PI controller measurement variables of the outlet water temperature T, respectively. The PI controllers are closely related to the two closed-loop control loops. For example, the output of the PI controller for the tank level is used as the input of the flow controller, which is used as a closed-loop control loop. Therefore, there are six observation variables, including three sensor data plus the measurement variables of the three PI controllers that monitor the sensor data.

[0123] When constructing the mathematical model of the continuous tank stirring heating process, the following reasonable assumptions are made to simplify the problem and ensure the accuracy of the model:

[0124] 1. Heat transfer independence assumption: It is assumed that at a given steam valve setting, the heat transferred from the heater to the reactor does not depend on the water level change in the heater. This assumption simplifies the complexity of the heat transfer process, making the heat input a relatively stable parameter that is only affected by the steam valve opening.

[0125] 2. Uniform mixing assumption: It is assumed that the stirring device in the heater can effectively achieve uniform mixing of the materials, so that the temperature of the flowing water at the heater outlet can closely reflect the average temperature inside the heater. This assumption ensures the uniformity of the temperature distribution in the reactor and is the basis for the application of the energy conservation principle in the continuous tank stirring heating process.

[0126] Based on the above assumptions, a mathematical model of the continuous kettle stirring heating process is constructed according to the principle of energy conservation. This mathematical model will describe the change of temperature in the reactor over time, and how this change is affected by various factors such as steam input heat, material inlet and outlet flow, reaction heat release (or endothermicity), etc.

[0127] Specifically, the mathematical model includes one or more differential equations that describe the accumulation and dissipation of energy in the reactor. These equations will take into account the heat input, heat output, and the heat changes that may be generated by the chemical reaction. Based on the principle of conservation of energy, a mathematical model of the continuous stirred tank heating process can be established:

[0128] ; (19)

[0129] ; (20)

[0130] in is the flow rate of hot water into the heater, is the flow rate of cold water into the heater, is the flow rate of water out of the heater, is the heat of steam, is the density of the hot water flowing into the heater, is the density of the cold water flowing into the heater, is the density of the water leaving the heater, is the specific enthalpy of the hot water flowing into the heater, is the specific enthalpy of the cold water flowing into the heater, is the specific enthalpy of the water leaving the heater, V is the volume of water, and H is the total specific enthalpy in the tank.

[0131] Table 1 shows the fault descriptions of three minor faults (sensor deviation, sensor accuracy degradation, and sensor drift) in the continuous tank stirring heating process. To ensure the high reliability of the simulation study, three test data sets were generated in the Simulink software under normal operating mode 1. The startup mode of the simulation model was set to 12 mA for tank level, 10.5 mA for outlet water temperature, and 11.89 mA for cold water flow. The training data set contains 500 training data collected under normal operating conditions. Each test data set has 500 test data, and minor faults are introduced from the 201st sampling moment.

[0132] Table 1 Three minor faults during continuous kettle stirring heating

[0133]

[0134] represents the standard normal distribution, The distribution function of the precision-degraded fault data, represents the distribution function of sensor deviation fault data, Represents the distribution function of sensor drift fault data.

[0135] 1. Sensor accuracy degradation fault

[0136] In the case of a minor precision reduction fault in the continuous kettle stirring and heating process of the present invention, a sensor precision reduction fault is introduced at the outflow water temperature T, and the sampling time is 201 to 500. In order to verify the effectiveness of the intelligent detection method for minor faults based on weighted divergence in the present invention, the detection performance of the present invention for minor precision reduction faults of the sensor is first tested in the continuous kettle stirring and heating process, and then the effects of the detection method based on principal component analysis (PCA), the KL divergence detection method based on the PCA framework (PCA-KLD), and the JS divergence detection method based on the PCA framework (PCA-JSD) are compared with the intelligent detection method for minor faults based on weighted divergence of the present invention.

[0137] Next, the fault detection rate FDR and the fault false alarm rate FAR are used to quantitatively evaluate the performance of all detection methods. The specific formula is as follows:

[0138] ; (twenty one)

[0139] ; (twenty two)

[0140] in, represents the number of normal samples that are misdiagnosed as faults, represents the total number of normal samples, represents the number of samples of detected faults, Indicates the total number of failed samples.

[0141] First, PCA modeling is performed on the precision-decreased sample training data set, and the load matrix and training dataset Calculate the training data score matrix in the principal component subspace These scores represent the projection length of the original training data in the principal component direction, that is, the position of the training data in the principal component subspace. The score matrix of the training data is used for subsequent analysis and modeling, while the score matrix of the test data is used for It is used to verify the performance of the model.

[0142] Next, perform JS divergence detection and calculate the variance of the test data and the training data in each sliding window. and , use the JS divergence metric to evaluate the distribution difference between the test data and the training data. The smaller the JS divergence, the closer the distributions of the two are; conversely, the greater the distribution difference. At this time, the JS divergence test result is obtained.

[0143] Furthermore, the local outlier factor weighting method is applied to the JS divergence statistic sequence for detection. Outliers are identified by comparing the local density difference between a data point and its neighbors. After reasonable weighting, the weighted JS divergence statistic is obtained. .

[0144] Finally, given the impact factor , set reasonable control limits, and use kernel density estimation to fit the weighted The distribution of the statistic and the control limits are obtained.

[0145] FIG3(a) to FIG3(d) are effect diagrams of four methods for detecting minor faults of sensor accuracy degradation in a continuous kettle stirring heating process according to an embodiment of the present invention. As can be seen from FIG3(a), the conventional detection method based on principal component analysis is Statistics ( ) and SPE statistics ( ) are 0.66% and 0% respectively, and basically no minor faults can be detected. Because the minor faults are assigned low values ​​and the data variance does not change significantly, traditional detection methods represented by PCA cannot detect minor faults in time; Figure 3(b) is a KL divergence detection effect diagram based on the PCA framework. Its highest fault detection rate in the principal component subspace is 52.49%, which is much higher than that of traditional fault detection methods. JS divergence is a symmetric form of KL divergence, and the two have almost the same fault detection effect.

[0146] However, compared with the previous traditional fault detection and single divergence detection methods, the weighted divergence-based intelligent detection method for small faults proposed in this invention has better fault detection capability. As can be seen from Figure 3(d), in the principal element subspace, the weighted JS divergence statistic is With the highest fault detection rate, up to 92.03%, the method of the present invention can make fault weight allocation more effective and have higher adaptability and stability. and Represent the weighted JS divergence statistics in the principal component subspace and the residual subspace, respectively.

[0147] 2. Sensor deviation failure

[0148] In the continuous kettle stirring heating process deviation fault case of the present invention, the cold water flow rate at each moment is A deviation fault is introduced, and the sampling time is from 201 to 500. In order to verify the effectiveness of the intelligent detection method of small faults based on weighted divergence in the present invention, the detection method based on principal component analysis (PCA), the KL divergence detection method based on the PCA framework (PCA-KLD) and the JS divergence detection method based on the PCA framework (PCA-JSD) are compared with the intelligent detection method of small faults based on weighted divergence in the present invention.

[0149] Figures 4(a) to 4(d) show in detail the comparison of the effects of four detection methods in the field of small deviation fault detection in continuous kettle stirring heating process, aiming to better describe the advantages and disadvantages of different strategies in identifying small faults. Figure 4(a) shows the traditional detection method based on principal component analysis, which captures fault signals through T² statistics and SPE statistics, but the statistical fault detection rate of T² statistics is only 1.33%, and SPE statistics is even lower, and small faults cannot be detected at all.

[0150] In order to solve the limitations of PCA in detecting minor faults, Figures 4(b) and (c) introduce the KL divergence detection method based on the PCA framework and the JS divergence detection method based on the PCA framework. These methods have significant improvements in the principal component subspace, and the highest fault detection rate has jumped to about 60%, which is significantly better than the traditional detection method based on principal component analysis. Therefore, the divergence can measure the probability density distribution of normal processes and fault processes, and the fault detection performance is good.

[0151] However, as can be seen from Figure 4(d), the weighted divergence-based intelligent detection method for small faults of the present invention achieves precise focusing and efficient extraction of key fault information by intelligently allocating fault weights. The results show that the fault detection rate of the weighted statistics in the principal component subspace is 92.03%, and the fault detection rate in the residual subspace is 91.36%, far exceeding other methods. The method of the present invention has made a breakthrough, which not only proves the high adaptability and stability of weighted JS divergence detection in complex industrial processes, but also shows its great potential in improving the accuracy and sensitivity of small fault detection.

[0152] 3. Sensor drift failure

[0153] In the drift fault case of the continuous kettle stirring heating process of the present invention, at the 201st to 500th sampling moments, the cold water flow rate at each moment is A small drift fault is introduced. The fault starts at sampling time 201 and is introduced with a slope of 0.001 mA, and lasts until sampling time 500. The effectiveness of the intelligent detection method for small faults based on weighted divergence proposed in the present invention is verified by comparing PCA, PCA-KLD, PCA-JSD with the present invention.

[0154] Figures 5(a) to 5(d) show in detail the effect comparison of four detection methods in the field of drift deviation fault detection in the continuous kettle stirring heating process. Similar to the first two fault detection methods, the weighted divergence-based intelligent detection method for small faults of the present invention has the highest fault detection rate. Figure 5(a) shows that the detection method based on principal component analysis (PCA) is completely unable to detect small faults; but from Figures 5(b) and (c), it can be seen that the KL divergence detection method based on the PCA framework (PCA-KLD) and the JS divergence detection method based on the PCA framework (PCA-JSD) can effectively improve the detection rate of small drift faults. However, a single divergence treats all statistical components equally and cannot effectively emphasize the more critical small fault information. Although a single divergence can detect small faults to a certain extent, its effect is still not ideal.

[0155] As can be seen from Figure 5(d), the weighted divergence-based intelligent detection method for small faults of the present invention has a fault detection rate of more than 90% in the principal component subspace. It is more sensitive to small faults in complex industrial systems, has more accurate detection, and has the highest fault detection rate.

[0156] Figure 6 This is a comparison chart of the fault detection rate (FDR) when detecting minor faults in a continuous kettle stirring heating process using four detection methods. represents the maximum JS divergence statistic in the residual subspace, represents the maximum JS divergence statistic in the principal subspace, represents the maximum KL divergence statistic in the residual subspace, represents the maximum KL divergence statistic in the principal subspace, represents the SPE statistic of the detection method based on principal component analysis, Represents the detection method based on principal component analysis Statistics.

[0157] Therefore, the weighted divergence-based intelligent detection method and system for minor faults of the present invention can detect minor faults in industrial processes in a timely and accurate manner. It not only significantly improves the fault detection rate, but also helps enterprises improve production efficiency. Through accurate and timely minor fault detection, it avoids production interruptions and product quality degradation, greatly reduces maintenance costs, and reduces economic losses caused by faults.

[0158] The weighted divergence-based intelligent detection method and system for minor faults proposed in the present invention realizes accurate capture and early warning of minor faults in industrial processes by integrating the two core links of offline training and online fault detection. The system uses the PCA model to extract key principal components, combines the JS divergence test under the sliding window strategy and the local outlier factor weighting method, and effectively improves the sensitivity and accuracy of fault detection. At the same time, reasonable control limits are set by the kernel density estimation method to ensure the reliability of fault judgment. The online fault detection part realizes real-time data processing and dynamic analysis, and can respond quickly and alarm when a fault occurs, providing a solid guarantee for the safe and stable operation of the industrial process.

[0159] The system of the present invention corresponds to the method of the present invention, and the specific technical solution adopted by the method of the present invention is also applicable to the system.

Claims

1. A method for intelligent detection of small faults based on weighted divergence, characterized in that: include: Collect offline production data at each moment of the industrial operation process and divide the offline production data into training data sets And the test dataset ; based on Construct the principal component analysis model and obtain the load matrix , determine the number of principal components of the principal component analysis model ;based on , and Calculate the training data score matrix in the principal component subspace and the test data score matrix ; Given the width W of the sliding window, calculate The i-th test score in Variance of the first W elements ,calculate The i-th training score in Variance of the first W elements , ;based on and Perform JS divergence test; obtain a sequence composed of JS divergence statistics; The local outlier weight factor is introduced into the JS divergence statistic through the local outlier factor weighting method to obtain the weighted JS divergence statistic ; Based on a given impact factor ,set up Control limits ; Collect online production data of current industrial operation processes ,calculate The weighted JS divergence statistic , by comparison and The size of the fault is used to determine whether a fault has occurred.

2. The method for intelligent detection of small faults based on weighted divergence according to claim 1 is characterized in that: The basis Construct the principal component analysis model and obtain the load matrix , determine the number of principal components of the principal component analysis model , specifically including: Based on the training data set Construct a principal component analysis model and obtain a training data set The corresponding eigenvalues ​​of the covariance matrix and the corresponding unit eigenvector, and the load matrix ; is the mth eigenvalue of the covariance matrix, and m is the total number of observed variables in the offline production data; Using the cumulative variance contribution method to determine the number of principal components in the principal component analysis model : Starting from the largest eigenvalue, the variance contribution rates corresponding to the eigenvalues ​​are accumulated in sequence until the cumulative variance contribution rate reaches the preset threshold. : ； represents the i-th eigenvalue of the covariance matrix, represents the jth eigenvalue of the covariance matrix, Before The cumulative contribution rate of variance corresponding to the principal component.

3. The method for intelligent detection of small faults based on weighted divergence according to claim 1 is characterized in that: The basis , and Calculate the training data score matrix in the principal component subspace and the test data score matrix , specifically including: Through the load matrix and training dataset The corresponding matrix , calculate the training data score matrix in the principal component subspace ; ； Through the load matrix And the test dataset The corresponding matrix , calculate the test data score matrix in the principal component subspace ; ； for The training score, for The Test scores.

4. The method for intelligent detection of small faults based on weighted divergence according to claim 1 is characterized in that: The online production data of the current industrial operation process is collected ,calculate The weighted JS divergence statistic , by comparison and The size of the fault is used to determine whether a fault has occurred, which specifically includes the following steps: S71, continuously collects online production data of the current industrial operation process , forming an online dataset ; for Middle Online production data, for The total number of online production data; S72, using the principal component analysis model and utilizing the load matrix Calculate the score matrix for online production data ; For online datasets The corresponding matrix, Scoring online production data; S73, given sliding window width , through the JS divergence test under the sliding window strategy, calculate the mean of the online production data score in each sliding window and variance , and combined with the training data score to obtain the online JS divergence statistics of the online production data ; S74, online JS divergence statistics by local outlier factor weighting method Perform local outlier detection and introduce the obtained local outlier weight factor into In the example, we get the weighted JS divergence statistic of the online production data. ; S75, by judgment and The size of the fault is used to determine whether a fault has occurred: , then the current industrial operation process is a fault process; if , then the current industrial process is a normal process; S76, collect online production data of the next industrial process, and perform steps S72 to S75 to continuously perform online fault detection.

5. An intelligent detection system for small faults based on weighted divergence, characterized in that: include: The data acquisition module collects offline production data at each moment of the industrial operation process and divides the offline production data into training data sets and test dataset ; Model building module, based on Construct the principal component analysis model and obtain the load matrix , determine the number of principal components of the principal component analysis model ;based on , and Calculate the training data score matrix in the principal component subspace and the test data score matrix ; Divergence test module, given the width W of the sliding window, calculates The i-th test score in Variance of the first W elements ,calculate The i-th training score in Variance of the first W elements , ;based on and Perform JS divergence test; obtain a sequence composed of JS divergence statistics; The control limit setting module introduces the local outlier weight factor into the JS divergence statistic through the local outlier factor weighting method to obtain the weighted JS divergence statistic ; Based on a given impact factor ,set up Control limits ; Fault diagnosis module, collects online production data of the current industrial operation process ,calculate The weighted JS divergence statistic , by comparison and The size of the fault is used to determine whether a fault has occurred.

Images