CN112904810B - Process industry nonlinear process monitoring method based on effective feature selection - Google Patents
Process industry nonlinear process monitoring method based on effective feature selection Download PDFInfo
- Publication number
- CN112904810B CN112904810B CN202110044130.5A CN202110044130A CN112904810B CN 112904810 B CN112904810 B CN 112904810B CN 202110044130 A CN202110044130 A CN 202110044130A CN 112904810 B CN112904810 B CN 112904810B
- Authority
- CN
- China
- Prior art keywords
- self
- hidden layer
- encoder
- data set
- online detection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 179
- 230000008569 process Effects 0.000 title claims abstract description 132
- 238000012544 monitoring process Methods 0.000 title claims abstract description 44
- 238000012549 training Methods 0.000 claims abstract description 17
- 238000007781 pre-processing Methods 0.000 claims abstract description 14
- 238000001514 detection method Methods 0.000 claims description 135
- 239000000523 sample Substances 0.000 claims description 86
- 230000004913 activation Effects 0.000 claims description 11
- 239000012468 concentrated sample Substances 0.000 claims description 9
- 238000000605 extraction Methods 0.000 claims description 8
- 238000012360 testing method Methods 0.000 claims description 8
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000005315 distribution function Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 4
- 230000006870 function Effects 0.000 description 17
- 238000004519 manufacturing process Methods 0.000 description 11
- 238000009776 industrial production Methods 0.000 description 5
- 238000011161 development Methods 0.000 description 4
- 238000004445 quantitative analysis Methods 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000013178 mathematical model Methods 0.000 description 2
- 238000000513 principal component analysis Methods 0.000 description 2
- 238000004451 qualitative analysis Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000010485 coping Effects 0.000 description 1
- 238000004134 energy conservation Methods 0.000 description 1
- 238000011031 large-scale manufacturing process Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000002156 mixing Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B19/00—Programme-control systems
- G05B19/02—Programme-control systems electric
- G05B19/418—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM]
- G05B19/41865—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM] characterised by job scheduling, process planning, material flow
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B2219/00—Program-control systems
- G05B2219/30—Nc systems
- G05B2219/32—Operator till task planning
- G05B2219/32252—Scheduling production, machining, job shop
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]
Landscapes
- Engineering & Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Manufacturing & Machinery (AREA)
- Quality & Reliability (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Automation & Control Theory (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention provides a process industry nonlinear process monitoring method based on effective feature selection, which comprises the following steps: step 1, acquiring process variables during normal operation of a process industry, constructing a normal data set according to the acquired process variables, and performing data preprocessing on the normal data set; and 2, inputting the preprocessed normal data sample into a preset stacking self-encoder model, extracting hidden layer characteristics of the preprocessed normal data sample, pre-training a plurality of self-encoders, sequentially connecting and stacking input layers and hidden layers of the pre-trained self-encoders, constructing the trained stacking self-encoder model, and finely adjusting parameters of the trained stacking self-encoder model to obtain the hyper-parameters of the trained stacking self-encoder model. According to the invention, the extracted hidden layer characteristics are selected through an effective characteristic selection strategy, and the characteristics more effective for fault monitoring are selected, so that the fault monitoring accuracy is greatly improved, and the fault monitoring effect is greatly improved.
Description
Technical Field
The invention relates to the technical field of process industry, in particular to a process industry nonlinear process monitoring method based on effective characteristic selection.
Background
The process industry refers to modern manufacturing industry that produces process-based products, primarily processing products based on continuous and batch material flow, energy flow. The processing method of the product mainly comprises chemical and physical reactions, separation, mixing and the like, and the process industry is always the prop industry for national economy and social development in China in recent decades. With the development of production capacity and science and technology, the Chinese process industry has been greatly improved and advanced. Nowadays, the industrial production process faces new situations and new problems, and automation, informatization and intellectualization become development directions of transformation and upgrading, safe production, energy conservation and emission reduction in the industrial production process. Therefore, the capacity and the modernization level of the system are gradually improved, the industrial scale is rapidly expanded, and for a large-scale production system, the occurrence of system abnormity can cause production loss, operation unit damage and even casualties. In order to ensure the normal operation and product quality of an industrial system and ensure the safety of personnel and property, the existence of abnormal working conditions needs to be identified in real time and the operation and production recovery conditions need to be given in time, namely, the industrial production process is monitored on line.
Computer technology and distributed control systems are widely used in the process industry today to allow large amounts of process variable data to be collected and utilized, however, many product manufacturing process failure events are still handled manually by operators. Due to the lack of experience and capability of operators, when the abnormal conditions of production are faced, the adoption of wrong coping modes is inevitable, so that the industrial production cannot run stably. Therefore, the key content of the research is to introduce a computer control system to monitor the process industrial production process accurately in real time and to guide and assist field operators when dealing with production faults.
In recent years, industrial data is collected more conveniently and rapidly, and in many process monitoring research methods, a qualitative analysis method and a quantitative analysis method are mostly used for implementing real-time monitoring of an industrial process. Qualitative analysis methods describe the state of a system by analyzing the relationships between system variables or parameters, and common methods include: an expert system method, a graph theory method and a qualitative simulation method. Quantitative analysis is a method of analyzing the quantitative characteristics, quantitative relationships and quantitative variations of industrial data using a mathematical module to reveal and describe the correlation and development trends among process variables. Generally, quantitative analysis methods are classified into analytical model-based methods and data-driven methods.
Analytical model based process monitoring methods were studied earlier. The method comprises hidden layer knowledge of the inside of the system, but the method is only suitable for the monitored object to have an accurate mathematical model, the modeling process is complex, and in practice, the accurate mathematical model of the industrial process is often difficult to obtain, and the method cannot obtain a satisfactory monitoring effect. Currently, in modern industrial plants, different operating units can measure a large number of variables that reflect the operating state of an industrial process, thereby providing opportunities for data-driven based methods. Data-driven based process monitoring is an effective technique that can extract information from process data that is readily available in modern process systems, thereby providing a better understanding of the process without the need to know the exact analytical model of the system. Due to the nature of this data-driven approach, the data-driven monitoring method is well suited for complex and large-scale flow systems. The data driving method uses historical data to establish a process monitoring model, and then obtains a monitoring index of fault detection.
However, the data-driven based process monitoring also has the following problems. The effective characteristic information extracted by the model is insufficient, how to remove redundant information, and extracting key information is also an important aspect for improving the monitoring performance of the system. However, the existing feature extraction method for process monitoring cannot extract key effective features from the highly nonlinear and dynamically changing data, and the monitoring performance is affected. The stack self-encoder is used as a method for extracting features, a plurality of self-encoder networks are trained through nonlinear mapping, high-dimensional data are converted into low-dimensional features, abstract key features of input data can be obtained, and nonlinear relations of variables can be extracted.
Disclosure of Invention
The invention provides a process industry nonlinear process monitoring method based on effective feature selection, and aims to solve the problems that a traditional self-encoder model only focuses on feature extraction for reconstructing original input as much as possible, and feature information for increasing the difference between fault data and normal data is extracted from the fault detection angle.
In order to achieve the above object, an embodiment of the present invention provides a process industry nonlinear process monitoring method based on effective feature selection, including:
Wherein, the step 1 specifically comprises:
collecting process variables according to a certain time sequence based on normal process industry, and constructing a normal data set according to the collected process variables, wherein the normal data set is represented as X1＝[x1,x2,...,xn]T∈Rn×mWherein n represents the number of normal data sets, and m represents the number of process variables;
to a normal data set X1＝[x1,x2,...,xn]T∈Rn×mData preprocessing was performed as follows:
where i denotes a normal data sample, i 1,2., n,
the mth process variable representing the ith normal data sample,
represents the maximum value of the mth process variable in the normal data sample,
representing the minimum of the mth process variable in the normal data sample.
Wherein, the step 2 specifically comprises:
inputting the preprocessed normal data set into a first self-encoder of a preset stacked self-encoder model for model training, wherein the structures of the self-encoder and a decoder are as follows:
hi＝f(xi；θ1)＝f(W1xi+b1),θ1＝{W1,b1} (2)
ri＝g(hi；θ2)＝g(W2hi+b2),θ2＝{W2,b2} (3)
wherein, h ═ f (x) denotes an activation function of the hidden layer, and r ═ g (h) denotes an activation function of the output layer; theta1Set of parameters, θ, representing the hidden layer2A parameter set representing the output layer, k representing the number of nodes of the hidden layer, xi∈RmRepresents the ith normal data sample, hi∈RkRepresenting the hidden layer features extracted by the self-encoder, g (-) represents the decoding function, ri∈RmRepresenting the output vector, W1Representing the weight of the encoder, W1∈Rk×m，W2Representing the weight of the decoder, W2∈Rm×k，b1Representing the bias parameters of the encoder, b1∈Rk，b2Representing the offset parameters of the decoder, b2∈Rm。
Wherein, the step 2 further comprises:
inputting the normal data set into a first self-encoder of a preset stacked self-encoder model, and extracting hidden layer characteristics through the first self-encoder;
inputting the extracted hidden layer characteristics into a subsequent self-encoder, and extracting the hidden layer characteristics through the subsequent self-encoder;
repeatedly inputting the hidden layer characteristic data extracted by the previous self-encoder into the next self-encoder, extracting the hidden layer characteristic through the next self-encoder until no next self-encoder stops, and obtaining the weight of each self-encoder, the bias parameter of each self-encoder and the hidden layer characteristic of each self-encoder;
sequentially connecting the self-coders after pre-training according to an input layer and a hidden layer to construct a trained stacked self-coder model;
and carrying out global fine adjustment on the parameters of the trained stacked self-encoder model through a back propagation algorithm.
Wherein, the step 3 specifically comprises:
final extraction of stacked autocoder by active feature selection strategyThe feature of the hidden layer is extracted effectively, and statistic T constructed in the feature space2As follows:
defining a calculation formula of the contribution degree of each feature of the normal data set hidden layer according to the formula (4) and calculating the contribution degree of each feature of the normal data set hidden layer as follows:
wherein, cjRepresenting the contribution of the jth feature of the hidden layer of the normal data set, f representing the self-encoder activation function, X1Representing a normal data set, w1jSelf-encoder weights representing the jth feature of the hidden layer of the normal data set,
a bias parameter representing the jth characteristic of the hidden layer of the normal data set,
means, σ, representing the jth feature in the hidden layer of normal data set extraction1j 2Representing the variance of the jth feature in the hidden layer extracted from the normal data set.
Wherein, the step 3 further comprises:
and performing descending order on the calculated contribution degrees of the characteristics of the hidden layer, selecting d characteristics with the largest contribution degree, wherein the hyper-parameter d of the trained stacked self-encoder model meets the following conditions as follows:
C(d)≥C* (6)
wherein, C*Indicates a predetermined threshold value, C*Set between 80% and 90%;
assuming that the hidden layer of the I-th self-encoder has y features, the contribution degrees of the y features are in descending orderIs arranged as c1≥c2≥...≥cyThe minimum trained hyper-parameter d of the stacked self-coder model satisfying the condition of equation (6) is derived as follows:
wherein d represents a hyper-parameter of the trained stacked self-encoder model, d is more than or equal to 1 and less than or equal to y, y represents the number of features, k represents the kth hidden layer feature, ckRepresenting the contribution degree of the characteristic of the kth hidden layer;
statistics of samples in the normal dataset were calculated as follows:
wherein, T1 2Statistics representing samples in the normal data set, H1jRepresenting the value of the jth feature in the hidden layer extracted from the normal data set.
Wherein, the step 3 further comprises:
the control limit of the statistic is calculated by kernel density estimation as follows:
wherein p (X) denotes a normal data set X1Of a probability distribution function of X1iThe ith normal data sample of the normal data set is represented, n represents the number of the normal data samples, h represents a bandwidth parameter, and K (·) represents a kernel function;
the kernel function K (·) satisfies the following condition:
K(x)≥0 (11)。
wherein, the step 4 specifically comprises:
acquiring real-time process variables in the process industry, and constructing an online detection data set according to the acquired real-time process variables, wherein the online detection data set is represented as X2＝[x1,x2,...,xn]T∈Rn×mThe online detection data set comprises normal online detection data samples and fault data samples, and the online detection data set X2＝[x1,x2,...,xn]T∈Rn×mData preprocessing was performed as follows:
where θ represents an online detection data sample, θ ═ 1,2., n,
represents the mth process variable of the theta online test data sample.
Wherein, the step 5 specifically comprises:
inputting the preprocessed online detection data sample into the trained stacked self-encoder model, extracting the characteristics of the hidden layer of the online detection data sample, selecting the effective characteristics of the hidden layer of the online detection data sample through an effective characteristic selection strategy, and calculating the statistic of the online detection data concentrated sample according to the effective characteristics of the selected characteristics of the hidden layer of the online detection data sample, wherein the statistic is as follows:
wherein, T2 2Statistics representing samples in the on-line test dataset, H2jAnd representing the value of the jth feature in the hidden layer extracted by the online detection data set.
Wherein, the step 6 specifically comprises:
the failure detection rate was calculated as follows:
wherein, fdr represents the fault detection rate, a represents the number of online detection data samples judged to be fault in the online detection data set, and s represents the number of fault data samples in the online detection data set.
The scheme of the invention has the following beneficial effects:
according to the method for monitoring the nonlinear process of the process industry based on the effective feature selection, the extracted original features are selected through the effective feature selection strategy, the features more effective for fault monitoring are selected, the fault monitoring accuracy is greatly improved, and the fault monitoring effect is greatly improved.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of the TE process of the present invention;
FIG. 3 is a comparison of the failure detection rates in feature space according to the present invention;
FIG. 4 is a comparison diagram of the fault detection rate in the residual error space according to the present invention.
Detailed Description
In order to make the technical problems, technical solutions and advantages of the present invention more apparent, the following detailed description is given with reference to the accompanying drawings and specific embodiments.
The invention provides a process industry nonlinear process monitoring method based on effective feature selection, aiming at the problems that the existing self-encoder model only focuses on feature extraction for reconstructing original input as much as possible and ignores feature information for increasing the difference between fault data and normal data extracted from the fault detection angle.
As shown in fig. 1 to 4, an embodiment of the present invention provides a process industry nonlinear process monitoring method based on active feature selection, including: step 1, acquiring process variables during normal operation of a process industry, constructing a normal data set according to the acquired process variables, and performing data preprocessing on the normal data set; step 2, inputting the preprocessed normal data sample into a preset stacking self-encoder model, extracting hidden layer characteristics of the preprocessed normal data sample, pre-training a plurality of self-encoders, sequentially connecting and stacking input layers and hidden layers of the pre-trained self-encoders, constructing the trained stacking self-encoder model, and finely adjusting parameters of the trained stacking self-encoder model to obtain hyper-parameters of the trained stacking self-encoder model; step 3, selecting effective characteristics in the hidden layer characteristics of the preprocessed normal data sample through an effective characteristic selection strategy, calculating the statistic of the normal data concentrated sample according to the effective characteristics of the hidden layer characteristics of the selected normal data sample, and calculating the control limit of the statistic through kernel density estimation; step 4, acquiring real-time process variables in the process industrial operation process, constructing an online detection data set according to the acquired real-time process variables, and performing data preprocessing on the online detection data set; step 5, inputting the preprocessed online detection data sample into the trained stacked self-encoder model, extracting hidden layer characteristics of the preprocessed online detection data sample, selecting effective characteristics in the hidden layer characteristics through an effective characteristic selection strategy, and calculating statistics of online detection data concentrated samples according to the selected effective characteristics of the hidden layer characteristics; step 6, comparing the statistic of each sample in the online detection data set with the control limit of the statistic, judging whether the statistic of each sample in the online detection data set exceeds the control limit of the statistic, when the statistic of a certain online detection data sample in the online detection data set exceeds the control limit of the statistic, alarming and prompting are carried out, and the current online detection data sample is judged as the online detection data sample of the fault, the number of the online detection data samples judged as the fault is recorded, when the statistic of one online detection data sample in the online detection data set does not exceed the control limit of the statistic, the current online detection data sample is judged as a normal online detection data sample, and calculating the fault detection rate according to the number of the fault data samples in the online detection data set and the number of the online detection data samples which are judged to be faults in the online detection data set.
Wherein, the step 1 specifically comprises: collecting process variables according to a certain time sequence based on normal process industry, and constructing a normal data set according to the collected process variables, wherein the normal data set is represented as X1＝[x1,x2,...,xn]T∈Rn×mWherein n represents the number of normal data sets, and m represents the number of process variables;
for a normal data set X1＝[x1,x2,...,xn]T∈Rn×mData preprocessing was performed as follows:
where i denotes a normal data sample, i 1,2., n,
the mth process variable representing the ith normal data sample,
represents the maximum of the mth process variable in the normal data sample,
representing the minimum of the mth process variable in the normal data sample.
Wherein, the step 2 specifically comprises: inputting the preprocessed normal data set into a first self-encoder of a preset stack self-encoder model for model training, wherein the structures of the self-encoder and a decoder are as follows:
hi＝f(xi；θ1)＝f(W1xi+b1),θ1＝{W1,b1} (2)
ri＝g(hi；θ2)＝g(W2hi+b2),θ2＝{W2,b2} (3)
wherein, h ═ f (x) denotes an activation function of the hidden layer, and r ═ g (h) denotes an activation function of the output layer; theta1Set of parameters, θ, representing the hidden layer2A parameter set representing the output layer, k representing the number of nodes of the hidden layer, xi∈RmRepresents the ith normal data sample, hi∈RkRepresenting the hidden layer features extracted by the self-encoder, g (-) represents the decoding function, ri∈RmRepresenting the output vector, W1Representing the weight of the encoder, W1∈Rk×m，W2Representing the weight of the decoder, W2∈Rm×k，b1Representing the bias parameters of the encoder, b1∈Rk，b2Representing the offset parameters of the decoder, b2∈Rm。
Wherein, the step 2 further comprises: inputting the normal data set into a first self-encoder of a preset stacked self-encoder model, and extracting hidden layer characteristics through the first self-encoder; inputting the extracted hidden layer characteristics into a subsequent self-encoder, and extracting the hidden layer characteristics through the subsequent self-encoder; repeatedly inputting the hidden layer characteristic data extracted by the previous self-encoder into the next self-encoder, extracting the hidden layer characteristic through the next self-encoder until no next self-encoder stops, and obtaining the weight of each self-encoder, the bias parameter of each self-encoder and the hidden layer characteristic of each self-encoder; sequentially connecting the self-coders after pre-training according to an input layer and a hidden layer to construct a trained stacked self-coder model; and carrying out global fine adjustment on the parameters of the trained stacked self-encoder model through a back propagation algorithm.
In the method for monitoring a nonlinear process of process industry based on effective feature selection according to the embodiment of the invention, the principle of the self-encoder is to reconstruct the input at the output part through the learning of the hidden layer of the network, i.e. let r beiApproximation xiThen, through a back propagation algorithm, an error back propagation training between the two is utilized to obtain a pre-trained self-encoder, and the training of the l-1 self-encoder is completedAnd (3) taking the (l-1) th hidden layer characteristic as the input of the (l) th self-encoder, obtaining the weight and the bias parameter of the (l) th self-encoder, and extracting the (l) th hidden layer characteristic, wherein the activation function can be a sigmoid, a tanh hyperbolic tangent function or a ReLu linear rectification function.
Wherein, the step 3 specifically comprises: performing effective feature extraction on the features of the hidden layer finally extracted by the stacked self-encoder through an effective feature selection strategy, and constructing a statistic T in a feature space2As follows:
defining a calculation formula of the contribution degree of each feature of the normal data set hidden layer according to the formula (4) and calculating the contribution degree of each feature of the normal data set hidden layer as follows:
wherein, cjRepresenting the contribution of the jth feature of the hidden layer of the normal data set, f representing the self-encoder activation function, X1A normal data set is represented that is,
self-encoder weights representing the jth feature of the hidden layer of the normal data set,
a bias parameter representing the jth characteristic of the hidden layer of the normal data set,
means, σ, representing the jth feature in the hidden layer of normal data set extraction1j 2Representing the variance of the jth feature in the hidden layer extracted from the normal data set.
Wherein, the step 3 further comprises: and performing descending order on the calculated contribution degrees of the characteristics of the hidden layer, selecting d characteristics with the largest contribution degree, wherein the hyper-parameter d of the trained stacked self-encoder model meets the following conditions as follows:
C(d)≥C* (6)
wherein, C*Indicates a predetermined threshold value, C*Set between 80% and 90%;
assuming that the hidden layer of the I-th self-encoder has y features, the contribution degrees of the y features are arranged in a descending order as c1≥c2≥...≥cyThe minimum trained hyper-parameter d of the stacked self-coder model satisfying the condition of equation (6) is derived as follows:
wherein d represents a hyper-parameter of the trained stacked self-encoder model, d is more than or equal to 1 and less than or equal to y, y represents the number of features, k represents the kth hidden layer feature, ckRepresenting the contribution degree of the characteristic of the kth hidden layer;
statistics of samples in the normal dataset were calculated as follows:
wherein, T1 2Statistics representing samples in the normal data set, H1jRepresenting the value of the jth feature in the hidden layer extracted from the normal data set.
Wherein, the step 3 further comprises: the control limit of the statistic is calculated by kernel density estimation as follows:
wherein p (X) denotes a normal data set X1Of a probability distribution function of X1iI normal data sample of the normal data set, n normal data sample number, and h normal data sample numberThe bandwidth parameter, K (-) represents the kernel function;
the kernel function K (·) satisfies the following condition:
K(x)≥0 (11)。
wherein, the step 4 specifically comprises: acquiring real-time process variables in the process industry, and constructing an online detection data set according to the acquired real-time process variables, wherein the online detection data set is represented as X2＝[x1,x2,...,xn]T∈Rn×mThe online detection data set comprises normal online detection data samples and fault data samples, and the online detection data set X2＝[x1,x2,...,xn]T∈Rn×mData preprocessing was performed as follows:
where θ represents an online detection data sample, θ ═ 1,2., n,
represents the mth process variable of the theta online test data sample.
The method for monitoring the process industry nonlinear process based on the effective feature selection according to the embodiment of the invention verifies the prediction performance of the trained stacked self-coding model through the online detection data set.
Wherein, the step 5 specifically comprises: inputting the preprocessed online detection data sample into the trained stacked self-encoder model, extracting the characteristics of the hidden layer of the online detection data sample, selecting the effective characteristics of the hidden layer of the online detection data sample through an effective characteristic selection strategy, and calculating the statistic of the online detection data concentrated sample according to the effective characteristics of the selected characteristics of the hidden layer of the online detection data sample, wherein the statistic is as follows:
wherein, T2 2Statistics representing samples in an on-line test dataset, H2jAnd representing the value of the jth feature in the hidden layer extracted by the online detection data set.
Wherein, the step 6 specifically comprises: the failure detection rate was calculated as follows:
wherein, fdr represents the fault detection rate, a represents the number of online detection data samples judged as faults in the online detection data set, and s represents the number of fault data samples in the online detection data set.
According to the method for monitoring the process industry nonlinear process based on the effective feature selection, the training process of the self-encoder model is divided into two steps of pre-training layer by layer and reverse fine tuning, each self-encoder is pre-trained layer by layer, a normal data set is input and passes through a first self-encoder, and first hidden layer feature data are extracted; taking the first hidden layer characteristic data as the input of a second self-encoder, and extracting second hidden layer characteristic data; repeating the steps of taking hidden layer feature data of a previous self-encoder as input of hidden layer feature data of a next self-encoder and extracting hidden layer feature data of the next self-encoder, wherein l represents the position of the self-encoder in the stacked self-encoder model, and l is 1,2, …, n; for the training of a subsequent self-encoder (i.e. when l > -2), taking l-1 hidden layer feature data as the input of the first self-encoder, further obtaining the weight and the bias parameter of the first self-encoder, and extracting the first hidden layer feature data; sequentially connecting the pre-trained self-coders according to the input layer and the hidden layer to form a stacked self-coder model, and finally performing global fine tuning by utilizing back propagation on the basis of the parameters obtained by pre-training to train the stacked self-coder model. In the reverse fine tuning step, parameters of the stacked self-encoder model are fine-tuned by constructing a loss function of the stacked self-encoder model by using a back propagation algorithm until the network converges to a target range.
In this embodiment, the method of monitoring the TE process in the process industry includes the following steps: the control architecture for the TE process (Tennessee Eastman, TE, Tennessee-Eastman) comprises five major operating units, a reactor, a condenser, a separator, a compressor, and a stripper. The method comprises 22 measurement variables, 19 component variables and 12 operation variables, 11 operation variables and 22 continuous variables are selected for process monitoring, 21 fault data sets are used as a test set in the TE process, each data set comprises 960 samples, faults are introduced from the 161 st sample, 500 samples are used as normal data sets, normal data are collected for the TE process, and a stacked self-encoder model is 33->30->27->30->33, i.e. the input dimension is 33, and the number of self-encoders is 2. The parameters from the input layer to the hidden layer of the two coders are w1、w2、b1、b2The hidden layer activation functions of the two autoencoders are g1、g2}。
As shown in table 1, the feature space and residual space fault detection rate of the three models, which are Principal Component Analysis (PCA), Stacked Auto Encoder (SAE), and active feature selection stacked auto encoder (SAE-CS), on the fault in the test set 21 is given.
TABLE 1 Fault detection Rate for three models in TE Process
It can be seen from the table that the fault detection rate of the stacked self-encoder selected by the effective features is very good in 21 types of faults of the TE data set, and the detection rate of the stacked self-encoder selected by the effective features is generally better than that of the original stacked self-encoder in both the feature space and the residual space, and the stacked self-encoder selected by the effective features has higher accuracy.
TABLE 2TE Process failure categories
The method for monitoring the nonlinear process of the process industry based on the effective feature selection according to the embodiment of the invention includes acquiring process variables when the process industry normally operates to form a normal data set, inputting the normal data set into a stack self-encoder model which is originally set and formed by sequentially stacking l self-encoders, wherein l is 1,2, n to obtain a plurality of hidden layer features, training the stack self-encoder model to obtain the trained stack self-encoder model, selecting effective features which can distinguish a fault sample from a normal sample from the plurality of hidden layer features from the fault detection angle through an effective feature selection strategy in the stack self-encoder model, calculating statistics of the samples in the normal data set according to the selected effective features, and calculating a control limit of sample control quantity in the normal data set through kernel density estimation; acquiring real-time process variables in the process industry, constructing an online detection data set according to the real-time process variables, preprocessing the online detection data set, inputting the preprocessed online detection data set into a trained stacked self-encoder to extract hidden layer characteristics, selecting effective characteristics in the hidden layer characteristics through an effective characteristic selection strategy, calculating statistics of samples in the online detection data set according to the effective characteristics of the selected hidden layer characteristics, comparing the statistics of each sample in the online detection data set with control limits of a normal data set, judging whether the current online detection data sample is a failed online detection data sample, recording the number of the online detection data samples judged to be failed, calculating a failure detection rate according to the number of the failed data samples in the online detection data set and the number of the online detection data samples judged to be failed, and judging the prediction performance of the trained stacked self-encoder model according to the fault detection rate.
According to the method for monitoring the process industry nonlinear process based on the effective feature selection, the extracted hidden layer features are selected through an effective feature selection strategy, the features more effective for fault monitoring are selected, the effective features for improving the fault detection rate can be selected in a self-adaptive mode, the monitoring effect of the stacked self-encoder model is greatly improved, and the fault monitoring accuracy is greatly improved.
While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined in the appended claims.
Claims (10)
1. A process industry nonlinear process monitoring method based on effective feature selection is characterized by comprising the following steps:
step 1, acquiring process variables during normal operation of a process industry, constructing a normal data set according to the acquired process variables, and performing data preprocessing on the normal data set;
step 2, inputting the preprocessed normal data sample into a preset stacking self-encoder model, extracting hidden layer characteristics of the preprocessed normal data sample, pre-training a plurality of self-encoders, sequentially connecting and stacking input layers and hidden layers of the pre-trained self-encoders, constructing the trained stacking self-encoder model, and finely adjusting parameters of the trained stacking self-encoder model to obtain hyper-parameters of the trained stacking self-encoder model;
step 3, selecting effective characteristics in the hidden layer characteristics of the preprocessed normal data sample through an effective characteristic selection strategy, calculating the statistic of the normal data concentrated sample according to the effective characteristics of the hidden layer characteristics of the selected normal data sample, and calculating the control limit of the statistic through kernel density estimation;
step 4, acquiring real-time process variables in the process industrial operation process, constructing an online detection data set according to the acquired real-time process variables, and performing data preprocessing on the online detection data set;
step 5, inputting the preprocessed online detection data sample into the trained stacked self-encoder model, extracting hidden layer characteristics of the preprocessed online detection data sample, selecting effective characteristics in the hidden layer characteristics through an effective characteristic selection strategy, and calculating statistics of online detection data concentrated samples according to the selected effective characteristics of the hidden layer characteristics;
step 6, comparing the statistic of each sample in the online detection data set with the control limit of the statistic, judging whether the statistic of each sample in the online detection data set exceeds the control limit of the statistic, when the statistic of a certain online detection data sample in the online detection data set exceeds the control limit of the statistic, alarming and prompting are carried out, and the current online detection data sample is judged as the online detection data sample of the fault, the number of the online detection data samples judged as the fault is recorded, when the statistic of one online detection data sample in the online detection data set does not exceed the control limit of the statistic, the current online detection data sample is judged as a normal online detection data sample, and calculating the fault detection rate according to the number of the fault data samples in the online detection data set and the number of the online detection data samples which are judged to be faults in the online detection data set.
2. The method for monitoring the non-linear process of the process industry based on the effective characteristic selection according to claim 1, wherein the step 1 specifically comprises:
collecting process variables according to a certain time sequence based on normal process industry, and constructing a normal data set according to the collected process variables, wherein the normal data set is represented as X1＝[x1，x2，...，xn]T∈Rn×mWherein n represents the number of normal data sets, and m represents the number of process variables;
for a normal data set X1＝[x1，x2，...，xn]T∈Rn×mData preprocessing was performed as follows:
where i denotes a normal data sample, i 1,2., n,
the mth process variable representing the ith normal data sample,
represents the maximum value of the mth process variable in the normal data sample,
representing the minimum of the mth process variable in the normal data sample.
3. The method for process industry nonlinear process monitoring based on active feature selection according to claim 2, wherein the step 2 specifically comprises:
inputting the preprocessed normal data set into a first self-encoder of a preset stacked self-encoder model for model training, wherein the structures of the self-encoder and a decoder are as follows:
hi＝f(xi；θ1)＝f(W1xi+b1)，θ1＝{W1，b1} (2)
ri＝g(hi；θ2)＝g(W2hi+b2)，θ2＝{W2，b2} (3)
wherein, h ═ f (x) denotes an activation function of the hidden layer, and r ═ g (h) denotes an activation function of the output layer; theta1Set of parameters, θ, representing the hidden layer2A parameter set representing the output layer, k representing the number of nodes of the hidden layer, xi∈RmRepresents the ith normal data sample, hi∈RkRepresenting the hidden layer features extracted by the self-encoder, g (-) represents the decoding function, ri∈RmRepresenting the output vector, W1Representing the weight of the encoder, W1∈Rk×m，W2Representing the weight of the decoder, W2∈Rm×k，b1Representing the bias parameters of the encoder, b1∈Rk，b2Representing the offset parameters of the decoder, b2∈Rm。
4. The method of claim 3, wherein the step 2 further comprises:
inputting the normal data set into a first self-encoder of a preset stacked self-encoder model, and extracting hidden layer characteristics through the first self-encoder;
inputting the extracted hidden layer characteristics into a subsequent self-encoder, and extracting the hidden layer characteristics through the subsequent self-encoder;
repeatedly inputting the hidden layer characteristic data extracted by the previous self-encoder into the next self-encoder, extracting the hidden layer characteristic through the next self-encoder until no next self-encoder stops, and obtaining the weight of each self-encoder, the bias parameter of each self-encoder and the hidden layer characteristic of each self-encoder;
sequentially connecting the self-coders after pre-training according to an input layer and a hidden layer to construct a trained stacked self-coder model;
and carrying out global fine adjustment on the parameters of the trained stacked self-encoder model through a back propagation algorithm.
5. The method for flow industrial nonlinear process monitoring based on active signature selection as recited in claim 4 wherein said step 3 specifically comprises:
carrying out effective feature extraction on the features of the hidden layer finally extracted by the stacked self-encoder through an effective feature selection strategy, and constructing a statistic T in a feature space2As follows:
defining a calculation formula of the contribution degree of each feature of the normal data set hidden layer according to the formula (4) and calculating the contribution degree of each feature of the normal data set hidden layer as follows:
wherein, cjRepresenting the contribution of the jth feature of the hidden layer of the normal data set, f representing the self-encoder activation function, X1A normal data set is represented that is,
self-encoder weights representing the jth feature of the hidden layer of the normal data set,
a bias parameter representing the jth characteristic of the hidden layer of the normal data set,
represents the mean of the jth feature in the hidden layer extracted from the normal data set,
representing the variance of the jth feature in the hidden layer extracted from the normal data set.
6. The method of claim 5, wherein the step 3 further comprises:
and performing descending order arrangement on the calculated contribution degrees of the characteristics of the hidden layer, selecting d characteristics with the maximum contribution degree, and enabling the hyper-parameter d of the trained stacked self-encoder model to meet the following conditions as follows:
C(d)≥C* (6)
wherein, C*Indicates a predetermined threshold value, C*Set between 80% and 90%;
assuming that the hidden layer of the I-th self-encoder has y features, the contribution degrees of the y features are arranged in a descending order as c1≥c2≥...≥cyThe minimum trained hyper-parameter d of the stacked self-coder model satisfying the condition of equation (6) is derived as follows:
wherein d represents a hyper-parameter of the trained stacked self-encoder model, d is more than or equal to 1 and less than or equal to y, y represents the number of features, k represents the kth hidden layer feature, ckRepresenting the contribution degree of the characteristic of the kth hidden layer;
statistics of samples in the normal dataset were calculated as follows:
wherein, T1 2Represents the statistics of the samples in the normal data set,
representing the value of the jth feature in the hidden layer extracted from the normal data set.
7. The method of claim 6, wherein the step 3 further comprises:
the control limit of the statistic is calculated by kernel density estimation as follows:
wherein p (X) denotes a normal data set X1Of a probability distribution function of X1iThe ith normal data sample of the normal data set is represented, n represents the number of the normal data samples, h represents a bandwidth parameter, and K (·) represents a kernel function;
the kernel function K (·) satisfies the following condition:
K(x)≥0 (11)。
8. the method for process industry nonlinear process monitoring based on active feature selection according to claim 7, wherein the step 4 specifically comprises:
acquiring real-time process variables in the process industry, and constructing an online detection data set according to the acquired real-time process variables, wherein the online detection data set is represented as X2＝[x1,x2,...,xn]T∈Rn×mThe online detection data set comprises normal online detection data samples and fault data samples, and the online detection data set X2＝[x1,x2,...,xn]T∈Rn×mData preprocessing was performed as follows:
where θ represents an online detection data sample, θ ═ 1,2., n,
represents the mth process variable of the theta online test data sample.
9. The method for process industry nonlinear process monitoring based on active feature selection as recited in claim 8 wherein said step 5 specifically comprises:
inputting the preprocessed online detection data sample into the trained stacked self-encoder model, extracting the characteristics of the hidden layer of the online detection data sample, selecting the effective characteristics of the hidden layer of the online detection data sample through an effective characteristic selection strategy, and calculating the statistic of the online detection data concentrated sample according to the effective characteristics of the selected characteristics of the hidden layer of the online detection data sample, wherein the statistic is as follows:
wherein, T2 2A statistic representing a sample in the online test data set,
and representing the value of the jth feature in the hidden layer extracted by the online detection data set.
10. The method for process industry nonlinear process monitoring based on active signature selection as recited in claim 9 wherein said step 6 specifically comprises:
the failure detection rate was calculated as follows:
wherein, fdr represents the fault detection rate, a represents the number of online detection data samples judged to be fault in the online detection data set, and s represents the number of fault data samples in the online detection data set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110044130.5A CN112904810B (en) | 2021-01-13 | 2021-01-13 | Process industry nonlinear process monitoring method based on effective feature selection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110044130.5A CN112904810B (en) | 2021-01-13 | 2021-01-13 | Process industry nonlinear process monitoring method based on effective feature selection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112904810A CN112904810A (en) | 2021-06-04 |
| CN112904810B true CN112904810B (en) | 2022-07-01 |
Family
ID=76112879
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110044130.5A Active CN112904810B (en) | 2021-01-13 | 2021-01-13 | Process industry nonlinear process monitoring method based on effective feature selection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112904810B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113468537B (en) * | 2021-06-15 | 2024-04-09 | 江苏大学 | Feature extraction and exploit attack detection method based on improved self-encoder |
| CN115877223B (en) * | 2023-03-03 | 2023-07-04 | 浙江天能新材料有限公司 | Lithium battery detection method, device, equipment and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109978079A (en) * | 2019-04-10 | 2019-07-05 | 东北电力大学 | A kind of data cleaning method of improved storehouse noise reduction self-encoding encoder |
| CN110096810A (en) * | 2019-05-05 | 2019-08-06 | 中南大学 | A kind of industrial process flexible measurement method based on layer-by-layer data extending deep learning |
| CN111324110A (en) * | 2020-03-20 | 2020-06-23 | 北京工业大学 | Fermentation process fault monitoring method based on multiple shrinkage automatic encoders |
| CN112034793A (en) * | 2020-08-10 | 2020-12-04 | 华东理工大学 | Method for monitoring production running state space of oxidation working section in PTA process flow |
| CN112101445A (en) * | 2020-09-09 | 2020-12-18 | 浙江大学 | Method for forecasting continuous casting billet subsurface slag inclusion defect in real time based on supervised neural network |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7844558B2 (en) * | 2006-10-05 | 2010-11-30 | Siemens Corporation | Incremental learning of nonlinear regression networks for machine condition monitoring |
| CN108227653B (en) * | 2017-12-28 | 2021-02-05 | 湖州师范学院 | Large-scale nonlinear process monitoring method based on randomized kernel principal component analysis |
| US11568198B2 (en) * | 2018-09-12 | 2023-01-31 | Applied Materials, Inc. | Deep auto-encoder for equipment health monitoring and fault detection in semiconductor and display process equipment tools |
-
2021
- 2021-01-13 CN CN202110044130.5A patent/CN112904810B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109978079A (en) * | 2019-04-10 | 2019-07-05 | 东北电力大学 | A kind of data cleaning method of improved storehouse noise reduction self-encoding encoder |
| CN110096810A (en) * | 2019-05-05 | 2019-08-06 | 中南大学 | A kind of industrial process flexible measurement method based on layer-by-layer data extending deep learning |
| CN111324110A (en) * | 2020-03-20 | 2020-06-23 | 北京工业大学 | Fermentation process fault monitoring method based on multiple shrinkage automatic encoders |
| CN112034793A (en) * | 2020-08-10 | 2020-12-04 | 华东理工大学 | Method for monitoring production running state space of oxidation working section in PTA process flow |
| CN112101445A (en) * | 2020-09-09 | 2020-12-18 | 浙江大学 | Method for forecasting continuous casting billet subsurface slag inclusion defect in real time based on supervised neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112904810A (en) | 2021-06-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109146246B (en) | Fault detection method based on automatic encoder and Bayesian network | |
| Jang et al. | Adversarial autoencoder based feature learning for fault detection in industrial processes | |
| CN108762228B (en) | Distributed PCA-based multi-working-condition fault monitoring method | |
| CN113642754B (en) | Complex industrial process fault prediction method based on RF noise reduction self-coding information reconstruction and time convolution network | |
| CN108875772B (en) | Fault classification model and method based on stacked sparse Gaussian Bernoulli limited Boltzmann machine and reinforcement learning | |
| Tian et al. | Identification of abnormal conditions in high-dimensional chemical process based on feature selection and deep learning | |
| CN112904810B (en) | Process industry nonlinear process monitoring method based on effective feature selection | |
| CN111340110B (en) | Fault early warning method based on industrial process running state trend analysis | |
| CN105607631B (en) | The weak fault model control limit method for building up of batch process and weak fault monitoring method | |
| CN109675935A (en) | A kind of IPCA operation of rolling on-line fault diagnosis method becoming control limit | |
| CN112000081B (en) | Fault monitoring method and system based on multi-block information extraction and Mahalanobis distance | |
| CN111914897A (en) | Fault diagnosis method based on twin long-short time memory network | |
| CN108830006B (en) | Linear-nonlinear industrial process fault detection method based on linear evaluation factor | |
| CN111414943B (en) | Anomaly detection method based on mixed hidden naive Bayes model | |
| CN114611067A (en) | Chemical process slow-change fault detection method based on typical variable dissimilarity analysis | |
| CN113703422B (en) | Gas turbine pneumatic actuator fault diagnosis method based on feature analysis processing | |
| CN112947649A (en) | Multivariate process monitoring method based on mutual information matrix projection | |
| CN113159088A (en) | Fault monitoring and diagnosis method based on multi-feature fusion and width learning | |
| CN113253682B (en) | Nonlinear chemical process fault detection method | |
| CN115017978A (en) | Fault classification method based on weighted probability neural network | |
| CN110674461B (en) | Chemical production process monitoring method based on multi-block projection non-negative matrix decomposition | |
| CN109947082B (en) | Process monitoring method based on integrated core independent component analysis model | |
| CN109523195B (en) | Sample multi-modal classification-based fault detection method for hierarchical non-Gaussian algorithm | |
| Huang et al. | A fault detection method based on stacking the SAE-SRBM for nonstationary and stationary hybrid processes | |
| CN113485269B (en) | Industrial process monitoring method based on hidden variable model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |