Abnormality monitoring method and system for multi-working-condition data

Abstract

The invention discloses a multi-working-condition data-oriented anomaly monitoring method and system, which are used for carrying out distinguishing representation learning of common characteristics and characteristic characteristics on the data of a multi-working-condition process on the premise of not carrying out working-condition attribution division of the data, and do not need priori knowledge to determine the working-condition attribution division of the data. The brand new algorithm of common feature analysis is adopted to carry out differentiated representation learning on the change features of the multi-working-condition data, and the change features of the multi-working-condition data can be described more comprehensively. And describing an algorithm which can define a hyper-ellipsoid in a high-dimensional space to quantify a characteristic change range through support vector data so as to deal with the monitoring problem of transient process data between working condition switching. In addition, the invention also discloses a corresponding improved technical scheme for determining the generality and characteristic transformation matrix, dividing the sub-matrix, further updating the characteristic and the like, and aims to better complete the anomaly monitoring task of the multi-working-condition process.

Description

Technical Field

[0001] The present invention belongs to the technical field of data-driven process monitoring, and in particular relates to an abnormality monitoring method and system for multi-operating condition data. Background Art

[0002] With the development of social productivity, modern industry is also moving towards larger-scale and more complex processes. The requirements for production process safety and control and management capabilities are also increasing, which has promoted the research of real-time monitoring of abnormal operating conditions in the manufacturing industry. Data-driven process monitoring methods are widely used and studied due to their low computing resource consumption, convenient deployment, and excellent monitoring results. In practice, industrial production processes often exhibit multi-condition characteristics due to changes in raw materials, setpoints, external environment, and product types. Currently, most data-driven process monitoring methods and technologies are based on the premise that process variables are stationary, which makes it difficult to cope with the non-stationary characteristics of multi-condition production process data. Although statistical feature representation learning, deep learning, reinforcement learning, and other methods have been applied to process monitoring, they generally require the system to operate under a single operating condition.

[0003] The Chinese invention patent with patent number 202010464703.5 discloses a multi-condition process monitoring method with improved PCA with continuous learning capability. It is necessary to combine system principles and prior knowledge to determine the weight matrix to improve the interpretability of the method. Therefore, this method is not purely data-driven and cannot be used without prior knowledge of system principles and multi-condition switching. The Chinese invention patent with patent number 202110581890.X discloses a multi-condition industrial process monitoring method and system that integrates mechanism and data. It requires mechanism knowledge to complete the condition attribution division of multi-condition data, so that the common feature analysis and extraction between multi-condition data can be realized on this basis. It is not a completely data-driven method technology. The Chinese invention patent with application number / patent number 202311287791.6 proposes a multi-condition industrial process intelligent monitoring method and system based on continuous learning of element-aware dictionaries. This method also requires that the data sets corresponding to multiple different conditions are known in advance, and does not take into account the monitoring problem of transient processes of condition switching.

[0004] In addition, an English paper titled "Multirate Mixture Probability Principal Component Analysis for Process Monitoring in Multimode Processes" published in the journal "IEEE Transactions on Automation Science and Engineering" proposed using a mixed probability PCA model to represent the data change distribution of the multimodal process, thereby realizing the monitoring of multi-condition processes. Although this method does not require the working condition attribution of multi-condition data in advance, it fails to learn the feature representation of the transient process data of the working condition switching. It can be seen that these methods either require the working condition attribution of the data to be known, or ignore the monitoring of the transient process of the working condition switching. Therefore, the application scenarios of these methods are limited, and they cannot truly solve the problem of abnormal monitoring of multi-condition processes.

[0005] For multi-condition processes, classifying and attributing data conditions from a data perspective is a technically challenging task, because the transient process data of condition switching exhibits significant non-stationary change characteristics, and its feature representation learning is relatively difficult. However, multi-condition data are all collected from the same production process. In addition to showing multi-cluster distribution characteristics, their data distribution also has certain common change characteristics. Only by learning the common change characteristics and characteristic change characteristics of multiple conditions without the need for precise condition classification and attribution, can we provide a feasible technical solution to the problem of abnormal monitoring in multi-condition processes without being restricted by application scenarios. Summary of the Invention

[0006] The main technical problem to be solved by the present invention is: how to learn the discriminative representation of common features and characteristic features of data of multi-operating condition processes without dividing the data into operating condition categories, and then implement abnormal monitoring of multi-operating condition processes including transient processes of operating condition switching on this basis.

[0007] The technical solution adopted by the present invention to solve the above problems is: a method and system for abnormality monitoring of multi-operating condition data; wherein, the implementation process of the abnormality monitoring method for multi-operating condition data disclosed by the present invention includes steps 1 to 4 shown below.

[0008] Step 1: Obtain historical data of normal operation of multiple working conditions, perform standardization on them to form a training data matrix X, and then divide X into M sub-matrices X1, X2, ..., X with equal number of columns. M ; Each column vector of X corresponds to a set of historical data, and M is an integer not less than 3.

[0009] Step 2: For X1, X2,…, X M Perform common feature analysis to obtain the common transformation matrix C and characteristic transformation matrix D, and then calculate the common feature matrix F = C T X and its corresponding covariance matrix Θ = FF T / (N-1), determine the threshold of common monitoring indicators Where the superscript T denotes the transpose of a matrix or vector, and N is equal to the total number of column vectors in X.

[0010] Step 3: Apply support vector data description algorithm to the characteristic feature matrix S=D T X performs analysis and obtains the radius of the hyperellipsoid R, n support vectors s1, s2, ..., s n and its corresponding n coefficients α1,α2,…,α n .

[0011] Step 4: Repeat the following steps 4-1 to 4-2 according to the sampling time interval of the multi-operating process.

[0012] Step 4-1: Get a set of data at the latest sampling time t, and perform normalization on it to obtain the column vector x t Then calculate the corresponding common monitoring indicators and characteristic monitoring index R t ;

[0013] Step 4-2: Determine whether the conditions are met And R t ≤R; if so, the multi-condition process is running normally; if not, the multi-condition process is running abnormally.

[0014] In the above implementation steps, the core key technology that distinguishes the method of the present invention from the existing technical solutions is that: step 1 is to randomly divide the training data matrix X into M sub-matrices with equal number of columns; the common feature analysis in step 2 includes the following steps 2-1 to step 2-6.

[0015] Step 2-1: After initializing the eigenvector w, calculate the matrix L = XX T .

[0016] Step 2-2: Determine the symbol coefficient β according to formula ① m,k ; Among them, number m = 1, 2,…, M; number k = 1, 2,…, M.

[0017] Step 2-3, calculate the matrix After solving the generalized eigenvalue problem Gw=λLw, we can find the eigenvector w corresponding to the largest eigenvalue λ. Update w.

[0018] Step 2-4: Determine whether w has converged; if not, return to step 2-2; if so, retain the eigenvector and execute step 2-5.

[0019] Step 2-5: X=X-pw T Update X and return to step 2-1 until B eigenvectors are obtained, where B equals the total number of row vectors in X and the loading vector p = Lw.

[0020] Step 2-6: The first b eigenvectors are combined into a common transformation matrix C, and the remaining Bb eigenvectors are combined into a characteristic transformation matrix D.

[0021] In addition, the above step 4-1 is based on the formula Calculate with the following formula ② and R t :

[0022] In the above formula ②, K(s t ,s i ) indicates the calculation of s t =D T x t With the i-th support vector s i The kernel function between i With α j denote the i-th and j-th coefficients in step 3 respectively.

[0023] The implementation process of the above steps 2-2 and 2-3 is to solve the maximization problem with constraints as shown below:

[0024] In the above formula ②, Represents calculation The absolute value of the absolute value; considering that the calculation of the absolute value only involves positive and negative without changing the numerical value, the symbol coefficient β defined in formula ① can be introduced m,k , and then the above formula ③ can be transformed into the following maximization problem with constraints:

[0025] The solution to equation ④ above can be achieved through the Lagrange multiplier method, that is, constructing the Lagrange function J:

[0026] Then, setting the differential of J with respect to w equal to 0, we can obtain: Gw-λLw=0; this is the generalized eigenvalue problem in steps 2-3; in addition, if we multiply both sides of the equality of Gw=λLw by w on the leftT And consider the constraint w T Lw=1, we can infer that: w T Gw=λw T Lw = λ; therefore, λ is equivalent to the objective function that needs to be maximized in equation ④ above; this is why only the largest eigenvalue needs to be solved in steps 2-3.

[0027] It should be further explained that the above formulas ③ and ④ are both used to maximize the correlation between different sub-matrices after transformation by the eigenvector w, that is, to analyze and extract the common change characteristics between multiple working condition data; in order to obtain multiple eigenvectors for transformation, in steps 2-5, X=X-pw T X updates X, thus avoiding the redundancy of information extracted by common feature analysis.

[0028] The advantages of the method described above are as follows: First, it eliminates the need for prior knowledge to determine the data's operating condition classification. Instead, it randomly partitions X into several submatrices with the same number of columns, making it very easy to use. Second, it uses a novel algorithm, common feature analysis, to analyze and extract the common variation characteristics between data from multiple operating conditions. This allows for discriminative representation learning of these variation characteristics, providing a more comprehensive description of these variation characteristics. Finally, using Support Vector Data Description (SVDD), an algorithm that defines hyperellipsoids in high-dimensional space to quantify the range of characteristic variation, it effectively addresses the complex variations in transient process data between operating condition switching.

[0029] As an improvement to the abnormality monitoring method of the present invention, it is preferred to determine the common transformation matrix C and the characteristic transformation matrix D in step 2-6 according to the following steps A-1 to A-3.

[0030] Step A-1, after setting X0 equal to the training data matrix X in step 1, Calculate the score vectors u1,u2,…,u respectively B ; Among them, w1,w2,…,w B represents the B feature vectors obtained in steps 2-5, w c and u c They represent the cth eigenvector and score vector respectively, numbered c=1,2,…,B.

[0031] Step A-2: Generate a 1×N-dimensional data vector v randomly according to a normal distribution with a mean of 0 and a standard deviation of 1, and then use J c =[E{g(u c )}-E{g(v)}] 2Calculate the commonality index J1, J2, ..., J corresponding to B eigenvectors B ; Wherein, g(v)=ln(cosh(v)), E{} represents the calculated average value.

[0032] Step A-3, for J1, J2, ..., J B After arranging in ascending order, the eigenvectors corresponding to the first b smallest commonality degree indices are combined into a commonality transformation matrix C, and the remaining Bb eigenvectors are combined into a characteristic transformation matrix D.

[0033] The above-mentioned preferred technical solution for determining the common transformation matrix C and the characteristic transformation matrix D utilizes a non-Gaussian degree test; since the common characteristics of multi-operating condition data will not show multi-cluster distribution characteristics, its non-Gaussian degree should be low, that is, the value of the common degree index in step A-2 should be relatively small; in this way, each eigenvector is objectively evaluated, and the best b eigenvectors are selected to form the common transformation matrix C, which can effectively avoid the problem of errors in common feature analysis and extraction.

[0034] As another improvement of the abnormality monitoring method of the present invention, it is preferred to divide X into M sub-matrices X1, X2, ..., X with equal number of columns according to the following steps B-1 to B-3. M .

[0035] Step B-1: Randomly select a column vector from the training data matrix X and record it as z1. Then, determine a column vector from X that has the largest distance to z1 and record it as z2.

[0036] Step B-2: Determine a column vector from X that has the largest sum of distances from z1 and z2, and record it as z3; and so on until M column vectors z1, z2, ..., z3 are determined. M .

[0037] Step B-3: Find the K column vectors with the smallest distance to z1 from X to form the first sub-matrix X1, then find the K column vectors with the smallest distance to z2 from the remaining NK column vectors in X to form the second sub-matrix X2, then find the K column vectors with the smallest distance to z3 from the remaining N-2K column vectors in X to form the third sub-matrix X3; and so on until M sub-matrices X1, X2, ..., X are obtained. M ; Wherein, K is the largest integer not greater than N / M.

[0038] Since the common feature analysis algorithm involved in the method of the present invention is intended to represent the learning X1, X2, ..., X M The common change characteristics between them, so X1,X2,…,X MThe greater the spatial distribution difference between them, the better; and the M sub-matrices obtained by random division are convenient for operation and implementation, but in extreme cases, M sub-matrices with similar spatial distribution may be obtained; this preferred technical solution for dividing the sub-matrices first determines the M column vectors with the largest distance between each other, and then sequentially forms the corresponding sub-matrices with the K column vectors with the smallest distance between these column vectors, which can better distinguish X1, X2, ..., X M The distribution differences between them ensure the effectiveness of the subsequent common feature analysis algorithm.

[0039] As a further improvement of the present invention, when applying the support vector data description algorithm to analyze the characteristic matrix S in step 3, the autoencoder model of S is first established and trained to obtain the estimated value of S Then pass After updating S, the support vector data description algorithm is applied to S to obtain the hyperellipsoid radius R, n support vectors and their corresponding n coefficients; at the same time, the characteristic monitoring index R is calculated. t When, first pass s t =D T x t Calculate s t , and then s t As input, pass the autoencoder model to obtain the corresponding estimated value Afterwards, follow Updates t , and then use formula ② to calculate R t .

[0040] The above improved technical solution uses the autoencoder to perform the feature matrix S and the feature score vector s t After treatment, S and s can be significantly reduced t The negative impact of non-stationary changes in the characteristic components of multiple working conditions, which makes the subsequent application of the support vector data description algorithm to achieve the updated S or s t The accurate description of the range of changes in the working condition can improve the sensitivity of abnormal data detection during the working condition switching process.

[0041] As a further improvement of the present invention, when establishing and training the autoencoder model of S, it is preferred to use the loss function The autoencoder model is trained; the parameter set of the autoencoder model is V and σ represent the weight matrix and bias vector of the coding layer respectively, and Represent the weight matrix and bias vector of the decoding layer, y d represents the dth column vector in S, the range of the regularization coefficient η is 0<η≤0.1, and the encoding layer output vector h d =f(yd ), the decoding layer output vector f() and ζ() represent the activation functions of the encoding layer and decoding layer respectively. and Represents S and y respectively d The e column vectors with the smallest distance between them.

[0042] When establishing and training an autoencoder model, those skilled in the art usually assume that the loss function to be minimized is That is, the first half of the loss function preferably used in the above-mentioned improved technical solution. In contrast, the regularization term added in this improved technical solution further considers the neighbor distribution relationship between data points; since data belonging to the same working condition are usually close in spatial distribution, the introduction of this regularization term will enable the autoencoder model to further maintain the neighbor relationship characteristics of multi-working condition data during training, and can represent the learning of more complex change characteristics of multiple working conditions, which is beneficial to the subsequent R-based t Abnormal monitoring.

[0043] Similarly, as another preferred solution for training the autoencoder model, The calculation formula involves You can choose S and y d The e column vectors with the closest corresponding sampling times; the improved technical solution is detailed as follows.

[0044] Based on the loss function The autoencoder model is trained; the parameter set of the autoencoder model is V and σ represent the weight matrix and bias vector of the coding layer respectively, and Represent the weight matrix and bias vector of the decoding layer, y d represents the dth column vector in S, the range of the regularization coefficient η is 0<η≤0.1, and the encoding layer output vector h d =f(y d ), the decoding layer output vector f() and ζ() represent the activation functions of the encoding layer and decoding layer respectively. and Represents S and y respectively d The e column vectors with the closest corresponding sampling times.

[0045] Since the multi-operating condition data with continuous sampling time can best reflect the state change of the operating condition switching of the multi-operating condition process, introducing a regular term with similar sampling time to train the autoencoder model can better represent the temporal relationship characteristics between the multi-operating condition data, which can also be beneficial to the subsequent R-based t Abnormal monitoring.

[0046] As another improvement to the anomaly monitoring method of the present invention, it is preferred that when applying the support vector data description algorithm to analyze the characteristic matrix S in step 3, a stacked autoencoder model of S is first established and trained to obtain an estimated value of S Then pass After updating S, the support vector data description algorithm is applied to S to obtain the hyperellipsoid radius R, n support vectors and their corresponding n coefficients; at the same time, the characteristic monitoring index R is calculated. t When, first pass s t =D T x t Calculate s t , and then s t As input, pass the stacked autoencoder model to obtain the corresponding estimated value Afterwards, follow Updates t , and then use formula ② to calculate R t .

[0047] Compared with the autoencoder model composed of three layers of neural network structure, namely input layer, encoding layer, and decoding layer, the stacked autoencoder model adopts more encoding layers and decoding layers. Through the greedy training method of layer-by-layer pre-training and fine-tuning of full network parameters, it takes advantage of the technical advantages of deep learning to learn more abstract features, and should obtain better performance on the feature matrix S=D T The feature representation learning effect of X can improve the feature-based monitoring index R t Abnormal detection sensitivity.

[0048] As another improvement of the present invention, when the support vector data description algorithm is applied and the characteristic monitoring index R is calculated by formula ② t When the kernel function K(s t ,s i ) preferably uses the radial basis kernel function, namely: K(s t ,s i )=exp(-||s t -s i || 2 / ε); where exp() represents an exponential function with a natural constant as the base, ||s t -s i || 2 Indicates calculation s t With s i The parameter ε is determined by first calculating the square distances between the column vectors in S, then calculating the average of these square distances, and setting ε equal to the average.

[0049] While many kernel functions can be used in support vector data description algorithms, the radial basis kernel is the most widely used, with excellent generalization performance. However, the selection of the parameter ε in the radial basis kernel is highly subjective. This technical solution provides a relatively objective method for determining the parameter ε by calculating the squared distances between all column vectors and averaging these squared distances to form the parameter ε. Using the average value also largely avoids overfitting and underfitting problems caused by excessively large or small ε.

[0050] Based on the same inventive concept, the present invention also discloses an abnormality monitoring system for multi-operating condition data, including a software subsystem and a hardware subsystem; wherein, the software subsystem stores programs for offline training and online monitoring; the hardware subsystem consists of a data collector, a central controller, a monitoring display and an alarm; the functions of each module are introduced as follows.

[0051] The offline training part of the program executes and implements the above steps 1 to 3; while the online monitoring part of the program executes and implements the above steps 4-1 to 4-2.

[0052] The data collector can obtain a set of data of multiple working conditions at the latest sampling time t in real time according to the sampling time interval and send it to the central controller; after receiving the set of data, the central controller loads and runs the program of the online monitoring part in the software subsystem to obtain common monitoring indicators and characteristic monitoring index R t Afterwards, and R t The specific value is sent to the monitoring display, and at the same time, the And R t The judgment result of ≤R will trigger the alarm.

[0053] The monitoring display is divided into two parts for real-time display and R t The specific value of the change, and at the same time the threshold The radius R of the hyperellipsoid is shown as a horizontal line.

[0054] Most importantly, the offline training part of the program stored in the software subsystem randomly divides the training data matrix X into M sub-matrices with equal number of columns, and performs common feature analysis according to the above steps 2-1 to 2-6; while the online monitoring part of the program stored in the software subsystem is based on the formula Calculate with formula ② and R tIn addition, the central controller is also connected to an initialization button. When the initialization button is pressed, the central controller immediately loads and runs the offline training part of the program stored in the software subsystem, and updates and stores the common transformation matrix C, characteristic transformation matrix D, covariance matrix Θ, threshold value after the execution is completed. The radius of the hyperellipsoid R, n support vectors s1, s2, ..., s n and its corresponding n coefficients α1,α2,…,α n .

[0055] It can be seen that the above-mentioned abnormality monitoring system executes the implementation process of the above-mentioned abnormality monitoring method by combining software and hardware, and is based on the same inventive concept as the above-mentioned abnormality monitoring method.

[0056] As an improvement of the abnormality monitoring system of the present invention, the offline training part of the software subsystem is to divide X into M sub-matrices X1, X2, ..., X with equal number of columns according to the above steps B-1 to B-3. M , and the common transformation matrix C and the characteristic transformation matrix D are determined according to the aforementioned steps A-1 to A-3.

[0057] This improved anomaly monitoring system uses two improvement strategies at the same time. Compared with the original anomaly monitoring system, its corresponding anomaly monitoring can simultaneously utilize the technical advantages of these two improvement strategies, thereby ensuring a better anomaly monitoring effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of the process flow of the coal pulverizing system.

[0059] Figure 2 The present invention is an implementation flow chart of the abnormality monitoring method.

[0060] Figure 3 This is an abnormality monitoring curve diagram corresponding to the first embodiment of the present invention.

[0061] Figure 4 In the second embodiment of the present invention, R t Abnormal monitoring curve diagram.

[0062] Figure 5 R in the fourth embodiment of the present invention t Abnormal monitoring curve diagram.

[0063] Figure 6 The figure is a schematic diagram of the structure of the abnormality monitoring system involved in the present invention. DETAILED DESCRIPTION

[0064] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] In order to clearly and specifically describe the specific implementation of the abnormality monitoring method of the present invention, it is applied to Figure 1 In the abnormal monitoring task of the coal pulverizing system shown in . Figure 1 The coal pulverizing system shown in the figure consists of a coal feeder, a coal mill, a rotary separator, a raw coal hopper, and a stone coal hopper. In the actual production process, the type of coal and the unit load will change frequently, resulting in a series of different operating modes. Figure 1 The coal pulverizing system shown is a typical multi-condition process.

[0066] Figure 1 The coal pulverizing system shown can collect data on nine key variables, including product discharge temperature, pressure, flow rate, and temperature of the air-powder mixture, and main heating or cooling air feed flow rate, temperature, and pressure, at fixed sampling time intervals. Figure 1 The multi-condition process object shown in the figure applies the abnormality monitoring method disclosed in the present invention, and its implementation process is as follows: Figure 2 As shown, it includes steps 1 to 4 shown below.

[0067] Step 1: Obtain 2000 sets of historical data when the multi-condition process is operating normally, perform standardization on them, and form a training data matrix X. Then, randomly divide the training data matrix X into M = 3 sub-matrices X1, X2, and X3 with equal numbers of columns; where the total number of column vectors of X is equal to 2000, and the total number of row vectors is equal to 9, which is the number of key variables.

[0068] Step 2: Perform common feature analysis on X1, X2, and X3 according to the following steps 2-1 to 2-6 to obtain the common transformation matrix C and the characteristic transformation matrix D, and then calculate the common feature matrix F = C T X and its corresponding covariance matrix Θ = FF T / (N-1), determine the threshold of common monitoring indicators

[0069] Step 2-1: After initializing the eigenvector w, calculate the matrix L = XX T .

[0070] Step 2-2: Determine the symbol coefficient β according to the above formula ① mk ; Among them, number m = 1, 2, 3; number k = 1, 2, 3.

[0071] Step 2-3, calculate the matrix After solving the generalized eigenvalue problem Gw=λLw, we can find the eigenvector w corresponding to the maximum eigenvalue λ. Update w.

[0072] Step 2-4: Determine whether w has converged; if not, return to step 2-2; if so, retain the eigenvector and execute step 2-5.

[0073] Step 2-5: X=X-pw T Update X and return to step 2-1 until B eigenvectors are obtained, where B = 9 and the load vector p = Lw.

[0074] Step 2-6: The first b eigenvectors are combined into a common transformation matrix C, and the remaining Bb eigenvectors are combined into a characteristic transformation matrix D.

[0075] In this technical field, the threshold value of common monitoring indicators is determined There are many ways, including but not limited to: using the maximum value of the common monitoring indicators corresponding to each column vector in the common feature matrix F; assuming that the common monitoring indicators conform to the chi-square distribution, using the chi-square distribution to determine their values under the 99% confidence limit; performing kernel density estimation on the common monitoring indicators corresponding to each column vector in the common feature matrix F, and then determining their values under the 99% confidence limit; in this embodiment, the kernel density estimation method is selected.

[0076] Step 3: Use support vector data description algorithm to describe the characteristic matrix S = D T X performs analysis and obtains the radius of the hyperellipsoid R, n = 50 support vectors s1, s2, ..., s n and its corresponding n=50 coefficients α1,α2,…,α n .

[0077] Step 4: Repeat the following steps 4-1 to 4-2 according to the sampling time interval of the multi-operating process.

[0078] Step 4-1: Get a set of data at the latest sampling time t and perform normalization on it to obtain the column vector x t Then, according to the formula Calculate common monitoring indicators using the above formula ② and characteristic monitoring index R t .

[0079] Step 4-2: Determine whether the conditions are met And R t ≤R; if so, the multi-condition process is running normally; if not, the multi-condition process is running abnormally.

[0080] Monitor the data of the 2800 latest sampling moments through step 4 and and R t These two indicators are plotted on Figure 3 In the middle; the horizontal line represents the threshold and the hyperellipsoid radius R, while the vertical line indicates the sampling moment when the anomaly begins.

[0081] from Figure 3 It can be found that from the 900th latest sampling moment, the characteristic monitoring indicators have basically exceeded the radius R of the hyperellipsoid, and the abnormality monitoring sensitivity is relatively high; in contrast, the abnormality monitoring effect of the common monitoring indicators is relatively poor, as only a small part exceeds its threshold; and R t Any exceeding of the limit can reflect an abnormality. Figure 3 The monitoring curves in FIG1 illustrate the effectiveness of the method of the present invention in monitoring processes under multiple operating conditions.

[0082] The second embodiment of the present invention adopts the following steps A-1 to A-3 to determine the common transformation matrix C and the characteristic transformation matrix D in step 2-6.

[0083] Step A-1, after setting X0 equal to the training data matrix X in step 1, Calculate the score vectors u1,u2,…,u9 respectively.

[0084] Step A-2: Generate a 1×N-dimensional data vector v randomly according to a normal distribution with a mean of 0 and a standard deviation of 1, and then use J c =[E{g(u c )}-E{g(v)}] 2 Calculate the commonality indexes J1, J2,…, J9 corresponding to the B eigenvectors.

[0085] Step A-3, for J1, J2, ..., J B After being arranged in ascending order, the eigenvectors corresponding to the b=3 smallest commonality degree indices are combined into a commonality transformation matrix C, and the remaining Bb=6 eigenvectors are combined into a characteristic transformation matrix D.

[0086] Monitor the same 2800 latest sampling data. Figure 4 The characteristic monitoring index R in the second embodiment of the present invention is given in t The curve graph of Figure 3 Compared with the characteristic monitoring indicators in the second embodiment, the second embodiment can more accurately divide the common change characteristics and characteristic change characteristics, thereby improving the sensitivity of abnormality monitoring.

[0087] The third embodiment of the present invention is based on the first embodiment and randomly divides the sub-matrices according to the following steps B-1 to B-3.

[0088] Step B-1: Randomly select a column vector from the training data matrix X and record it as z1. Then, determine a column vector from X that has the largest distance to z1 and record it as z2.

[0089] Step B-2: Determine from X a column vector whose sum of distances from z1 and z2 is the largest, and record it as z3.

[0090] Step B-3. Find the K column vectors with the shortest distance to z1 from X to form the first sub-matrix X1. Then, find the K column vectors with the shortest distance to z2 from the remaining NK column vectors in X to form the second sub-matrix X2. Then, find the K column vectors with the shortest distance to z3 from the remaining N-2K column vectors in X to form the third sub-matrix X3.

[0091] Since only M=3 sub-matrices X1, X2, X3 are divided in the first embodiment, there is no need to determine z4 and X4 in steps B-2 and B-3; this embodiment only avoids the small probability problem of small differences among X1, X2, X3 due to randomness, and its corresponding abnormal monitoring results are the same as Figure 3 The curves shown are basically the same and will not be repeated here.

[0092] The fourth embodiment of the present invention can be improved on the basis of any of the first three embodiments. Specifically, when the support vector data description algorithm is used to analyze the characteristic matrix S in step 3 of the first embodiment, an autoencoder model of S is first established and trained to obtain an estimated value of S. Then pass After updating S, the support vector data description algorithm is applied to S to obtain the hyperellipsoid radius R, n support vectors and their corresponding n coefficients; at the same time, the characteristic monitoring index R is calculated. t When, first pass s t =D T x t Calculate s t , and then s t As input, it is passed through the autoencoder model to obtain the corresponding estimated value Afterwards, follow Updates t , and then use formula ② to calculate R t .

[0093] Figure 5 The abnormal monitoring curve corresponding to the fourth embodiment of the present invention is shown. Figure 3 Therefore, Figure 5 Only the curves of characteristic monitoring indicators are shown in Figure 3 , Figure 5The characteristic monitoring indicators in are more sensitive to abnormal data monitoring, which is mainly because the autoencoder model eliminates the negative impact of non-stationary changes in multiple working conditions in the characteristic features.

[0094] The next two embodiments of the present invention are similar to the fourth embodiment. They are all about updating the characteristic matrix S when applying the support vector data description algorithm. The difference is that different loss functions are used to train the autoencoder model. The abnormality monitoring results obtained are different from those obtained in the fourth embodiment. Figure 5 Similar, no repetition is shown here.

[0095] In addition, the present invention can also use a deep learning framework such as a stacked autoencoder to further explore the deep abstract features of the characteristic matrix S to better eliminate the negative impact of non-stationary changes in multiple working conditions.

[0096] In all the above embodiments, support vector data description algorithm and characteristic monitoring index calculation by formula ② are involved; the kernel function is selected as K(s t ,s i )=exp(-||s t -s i || 2 / ε); the parameter ε is determined by first calculating the square distances between the column vectors in S, then calculating the average of these square distances, and setting ε equal to the average.

[0097] The present invention also provides an abnormal monitoring system for multi-condition data, the composition diagram of which is as follows: Figure 6 As shown, it includes a software subsystem that stores two parts of programs: offline training and online monitoring, and a hardware subsystem composed of a data collector, a central controller, a monitoring display and an alarm.

[0098] Figure 6 The program of the offline training part shown in the embodiment executes the aforementioned steps 1 to 3; wherein step 1 is to randomly divide the training data matrix X into M sub-matrices with equal number of columns, and step 2 is to perform common feature analysis according to steps 2-1 to 2-6 in claim 1.

[0099] Figure 6 The program of the online monitoring part shown in the embodiment implements steps 4-1 to 4-2 in claim 1; wherein step 4-1 is performed according to the formula Calculate with formula ② and R t .

[0100] Figure 6The functions of the various components of the hardware subsystem are as follows: the data collector can obtain a set of data of the multi-condition process at the latest sampling time t in real time according to the sampling time interval and send it to the central controller; after receiving the set of data, the central controller loads and runs the program of the online monitoring part of the software subsystem to obtain the common monitoring indicators and characteristic monitoring index R t Afterwards, and R t The specific value is sent to the monitoring display, and at the same time, the And R t The judgment result of ≤R triggers the alarm; the monitoring display is divided into two parts for real-time display and R t The specific value of the change, and at the same time the threshold The radius R of the hyperellipsoid is shown as a horizontal line.

[0101] In addition, the central controller is also connected to an initialization button. When the initialization button is pressed, the central controller immediately loads and runs the program of the offline training part stored in the software subsystem, and updates and stores the common transformation matrix C, characteristic transformation matrix D, covariance matrix Θ, threshold value after the program is completed. The radius of the hyperellipsoid R, n support vectors s1, s2, ..., s n and its corresponding n coefficients α1,α2,…,α n .

[0102] Finally, as an improvement to the anomaly monitoring system, Figure 6 The procedure of the offline training part shown in is to divide X into M sub-matrices with equal number of columns according to the aforementioned steps B-1 to B-3, and to determine the common transformation matrix C and the characteristic transformation matrix D according to the aforementioned steps A-1 to A-3.

[0103] The above explanation of the technical solution disclosed in the present invention in combination with the accompanying drawings is illustrative rather than restrictive. The technical solution disclosed in the present invention is not limited to the above embodiments. Any modifications made without departing from the purpose of the technical solution disclosed in the present invention and the scope of protection of the claims shall fall within the scope of protection of the present invention.

Claims

1. A method for abnormality monitoring of multi-operating condition data, comprising steps 1 to 4: Step 1: Obtain historical data of normal operation of multiple working conditions, perform standardization on them to form a training data matrix X, and then divide X into M sub-matrices X1, X2, ..., X with equal number of columns. M ;in, Each column vector of X corresponds to a set of historical data, and M is an integer not less than 3; Step 2: For X1, X2,…, X M Perform common feature analysis to obtain the common transformation matrix C and characteristic transformation matrix D, and then calculate the common feature matrix F = C T X and its corresponding covariance matrix Θ = FF T / (N-1), determine the threshold of common monitoring indicators Where the superscript T represents the transpose of a matrix or vector, and N is equal to the total number of column vectors in X; Step 3: Apply support vector data description algorithm to the characteristic feature matrix S=D T X performs analysis and obtains the radius of the hyperellipsoid R, n support vectors s1, s2, ..., s n and its corresponding n coefficients α1,α2,…,α n ; Step 4: Repeat the steps 4-1 to 4-2 according to the sampling time interval of the multi-operating process; Step 4-1: Get a set of data at the latest sampling time t, and perform normalization on it to obtain the column vector x t back, Then calculate the corresponding common monitoring indicators and characteristic monitoring index R t ; Step 4-2: Determine whether the conditions are met And R t ≤R; if yes, the multi-condition process is running normally; if not, Then the multi-condition process operation is abnormal; The method is characterized in that: the step 1 is to randomly divide the training data matrix X into M sub-matrices with equal number of columns; the common feature analysis includes the following steps 2-1 to 2-6: Step 2-1: After initializing the eigenvector w, calculate the matrix L = XX T ; Step 2-2: Determine the symbol coefficient β according to formula ① m,k ; Among them, number m=1,2,…,M; number k=1,2,…,M; Step 2-3, calculate the matrix After solving the generalized eigenvalue problem Gw=λLw, we can find the eigenvector w corresponding to the largest eigenvalue λ. Update w; Step 2-4, determine whether w converges; if not, return to step 2-2; if so, retain the eigenvector and execute step 2-5; step 2-5, through X = X-pw T Update X and return to step 2-1 until B eigenvectors are obtained; where B equals the total number of row vectors in X and the loading vector p = Lw; Step 2-6: The first b eigenvectors form a common transformation matrix C, and the remaining Bb eigenvectors form a characteristic transformation matrix D; Step 4-1 is based on the formula Calculate with the following formula ② and R t : Among them, K(s t ,s i ) indicates the calculation of s t =D T x t With the i-th support vector s i The kernel function between .

2. The abnormality monitoring method according to claim 1, characterized in that: The process of determining the common transformation matrix C and the characteristic transformation matrix D in step 2-6 includes the following steps A-1 to A-3: Step A-1, after setting X0 equal to the training data matrix X in step 1, Calculate the score vectors u1,u2,…,u respectively B ; Among them, w1,w2,…,w B represents the B feature vectors obtained in steps 2-5, w c and u c Represent the cth eigenvector and score vector, numbered c = 1, 2, ..., B; Step A-2: Generate a 1×N-dimensional data vector v randomly according to a normal distribution with a mean of 0 and a standard deviation of 1, and then use J c =[E{g(u c )}-E{g(v)}] 2 Calculate the commonality index J1, J2, ..., J corresponding to B eigenvectors B ; Where g(v) = ln(cosh(v)), E{} represents the calculated average value; Step A-3, for J1, J2, ..., J B After arranging in ascending order, the eigenvectors corresponding to the first b smallest commonality degree indices are combined into a commonality transformation matrix C, and the remaining Bb eigenvectors are combined into a characteristic transformation matrix D.

3. The abnormality monitoring method according to claim 1, characterized in that: The step 1 is to divide X into M sub-matrices X1, X2, ..., X with equal number of columns according to the following steps B-1 to B-3. M : Step B-1: Randomly select a column vector from the training data matrix X, record it as z1, and then determine a column vector from X that has the largest distance to z1, and record it as z2; Step B-2: Determine a column vector from X that has the largest sum of distances from z1 and z2, and record it as z3; and so on until M column vectors z1, z2, ..., z3 are determined. M ; Step B-3: Find the K column vectors with the smallest distance to z1 from X to form the first sub-matrix X1, then find the K column vectors with the smallest distance to z2 from the remaining NK column vectors in X to form the second sub-matrix X2, then find the K column vectors with the smallest distance to z3 from the remaining N-2K column vectors in X to form the third sub-matrix X3; and so on until M sub-matrices X1, X2, ..., X are obtained. M ; Wherein, K is the largest integer not greater than N / M.

4. An abnormality monitoring method according to any one of claims 1 to 3, characterized in that: When applying the support vector data description algorithm to analyze the characteristic matrix S in step 3, first establish and train the autoencoder model of S to obtain the estimated value of S Then pass After updating S, the support vector data description algorithm is applied to S to obtain the hyperellipsoid radius R, n support vectors and their corresponding n coefficients; At the same time, in the calculation of characteristic monitoring index R t When, first pass s t =D T x t Calculate s t , and then s t As input, it is passed through the autoencoder model to obtain the corresponding estimated value Afterwards, follow Updates t , and then use formula ② to calculate R t .

5. The abnormality monitoring method according to claim 4, characterized in that: When building and training the autoencoder model S, based on the loss function The autoencoder model is trained; wherein the parameter set of the autoencoder model is V and σ represent the weight matrix and bias vector of the coding layer respectively, and Represent the weight matrix and bias vector of the decoding layer, y d represents the dth column vector in S, the range of the regularization coefficient η is 0<η≤0.1, and the encoding layer output vector h d =f(y d ), the decoding layer output vector f() and ζ() represent the activation functions of the encoding layer and decoding layer respectively. and Represents S and y respectively d The e column vectors with the smallest distance between them.

6. The abnormality monitoring method according to claim 4, characterized in that: When building and training the autoencoder model S, based on the loss function The autoencoder model is trained; wherein the parameter set of the autoencoder model is V and σ represent the weight matrix and bias vector of the coding layer respectively, and Represent the weight matrix and bias vector of the decoding layer, y d represents the dth column vector in S, the range of the regularization coefficient η is 0<η≤0.1, and the encoding layer output vector h d =f(y d ), the decoding layer output vector f() and ζ() represent the activation functions of the encoding layer and decoding layer respectively. and Represents S and y respectively d The e column vectors with the closest corresponding sampling times.

7. An abnormality monitoring method according to any one of claims 1 to 3, characterized in that: When the support vector data description algorithm is used to analyze the characteristic matrix S in step 3, a stacked autoencoder model of S is first established and trained to obtain an estimated value of S Then pass After updating S, the support vector data description algorithm is applied to S to obtain the hyperellipsoid radius R, n support vectors and their corresponding n coefficients; At the same time, in the calculation of characteristic monitoring index R t When, first pass s t =D T x t Calculate s t , and then s t As input, it is passed through the stacked autoencoder model to obtain the corresponding estimated value Afterwards, follow Updates t , and then use formula ② to calculate R t .

8. An abnormality monitoring method according to any one of claims 1 to 3, characterized in that: The step 3 implements the support vector data description algorithm and the step 4-1 calculates the characteristic monitoring index R by formula ② t When the kernel function K(s t ,s i )=exp(-||s t -s i || 2 / ε); where exp() represents an exponential function with a natural constant as the base, ||s t -s i || 2 Indicates calculation s t With s i The parameter ε is determined by first calculating the square distances between the column vectors in S, then calculating the average of these square distances, and setting ε equal to the average.

9. An abnormality monitoring system for multi-operating condition data, comprising a software subsystem storing two programs: offline training and online monitoring, and a hardware subsystem consisting of a data collector, a central controller, a monitoring display, and an alarm; wherein, The program of the offline training part executes steps 1 to 3 in claim 1; and the program of the online monitoring part executes steps 4-1 to 4-2 in claim 1; The data collector can obtain a set of data of the multi-condition process at the latest sampling time t in real time according to the sampling time interval and send it to the central controller; After receiving the data, the central controller loads and runs the online monitoring program in the software subsystem to obtain common monitoring indicators. and characteristic monitoring index R t Afterwards, and R t The specific value is sent to the monitoring display, and at the same time, the And R t The judgment result of ≤R is used to trigger the alarm; The monitoring display is divided into two parts for real-time display and R t The specific value of the change, and at the same time the threshold and the hyperellipsoid radius R are displayed as a horizontal line; It is characterized in that the program of the offline training part is to randomly divide the training data matrix X into M sub-matrices with equal number of columns, and perform common feature analysis according to steps 2-1 to 2-6 in claim 1; the program of the online monitoring part is to perform common feature analysis according to the formula Calculate with formula ② and R t The central controller is also connected to an initialization button. When the initialization button is pressed, the central controller immediately loads and runs the offline training part of the program stored in the software subsystem, and updates and stores the common transformation matrix C, the characteristic transformation matrix D, the covariance matrix Θ, and the threshold after the execution is completed. The radius of the hyperellipsoid R, n support vectors s1, s2, ..., s n and its corresponding n coefficients α1,α2,…,α n .

10. The abnormality monitoring system according to claim 9, characterized in that: The procedure of the offline training part is to divide X into M sub-matrices with equal number of columns according to steps B-1 to B-3 in claim 3, and to determine the common transformation matrix C and the characteristic transformation matrix D according to steps A-1 to A-3 in claim 2.

Images