Coke oven health condition analysis method based on concept drift

Abstract

The invention relates to a coke oven health condition analysis method based on concept drift, which comprises drift type identification, drift point positioning and stage conversion analysis, and the drift type identification comprises the following steps: acquiring an error rate flow, constructing a classifier to classify the concept drift of a data flow, and positioning a prototype vector of each category in a semantic space; the drift point positioning is that a new drift feature vector is obtained by combining a prototype vector and a coding vector of a data stream, the new drift feature vector is input into a multi-layer perceptron regression model, and output indicating a drift occurrence position in the data stream, namely a stage conversion stream, is generated; the stage conversion flow analysis is to predict the stage conversion flow by using a recurrent neural network, obtain a predicted stage conversion flow in a future reversing period, calculate an Euclidean distance between a real conversion flow and the predicted conversion flow, and determine the health condition of the coke oven; the health condition of the coke oven can be reliably and automatically evaluated, and potential fault hidden dangers can be found in time.

Description

Technical Field

[0001] The invention relates to the technical field of industrial equipment fault early warning, and in particular to a coke oven health status analysis method based on concept drift. Background Art

[0002] In the industrial production process, the status of coke ovens is generally evaluated based on workers' experience. It is difficult to objectively judge the health of coke ovens based on coke oven operating data, which may cause production accidents and threaten the interests of manufacturers and the personal safety of workers.

[0003] In the field of machine learning, concept drift refers to a phenomenon in which the statistical properties of the prediction target change in unpredictable ways over time. In order to overcome the concept drift phenomenon, more and more mechanisms for detecting concept drift have emerged. In general, these methods follow a framework that includes 4 stages: Stage 1, data retrieval, aims to retrieve data blocks from the data stream; Stage 2, data modeling, aims to extract the retrieved data and extract key features; the focus of this stage is to reduce the sample size or sample dimension to meet the needs of storage and online learning; Stage 3, test statistic calculation, aims to measure the degree of difference between new and old data. This type of calculation forms a test statistic for hypothesis testing and quantifies the degree of concept drift; Stage 4, hypothesis testing, aims to use hypothesis testing to evaluate the statistical significance of the change between new and old data. Without stage 4, the test statistics obtained in stage 3 are meaningless for drift detection because they cannot determine the drift confidence interval, that is, the change may not be caused by concept drift but by noise or random sample selection bias.

[0004] However, since the accuracy of drift detection is determined by hypothesis testing, this framework faces two disadvantages: first, for hypothesis testing methods, the correct hypothesis testing method should be selected according to the data type and characteristics; however, since the data stream does not provide all data at once, it is difficult to select the most appropriate hypothesis, which affects the accuracy of detecting concept drift; second, although concept drift can be divided into four types, existing methods can only identify whether concept drift has occurred, but cannot identify which type of concept drift has occurred; the reason is that although the change of data distribution can be represented by test statistics such as average error rate, test statistics cannot represent the change pattern of data distribution, that is, it cannot capture the relationship between concept drifts on adjacent timestamps. Therefore, a coke oven health status analysis method based on concept drift is designed for coke oven production process stream data. Summary of the invention

[0005] The present invention provides a coke oven health status analysis method based on concept drift, which can reliably and automatically evaluate the health status of the coke oven, make up for the shortcomings of traditional characteristic parameters that are difficult to adapt to the changes in the stages of the coke oven production process and the changes in the production environment, can monitor the coke oven status in real time, discover health problems in time and deal with them, achieve data dimensionality reduction, improve the recognition rate to a certain extent, and reduce the system modeling time; combined with the coke oven operation status recognition method, it can achieve accurate switching of the model concept drift stage, and greatly improve the accuracy of the coke oven health status analysis.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A coke oven health status analysis method based on concept drift includes drift type identification, drift point positioning and phase transition analysis. The drift type identification is to obtain an error rate flow, build a classifier to classify the concept drift of a data flow, and locate the prototype vector of each category in a semantic space; the drift point positioning is to combine the prototype vector and the encoding vector of the data flow to obtain a new drift feature vector, input the new feature vector into a multi-layer perceptron regression model, and generate an output indicating the location where the drift occurs in the data flow, namely, a phase transition flow; the phase transition flow analysis is to use a recurrent neural network to predict the phase transition flow, obtain the predicted phase transition flow within a future switching cycle, calculate the Euclidean distance between the real transition flow and the predicted transition flow, and determine the health status of the coke oven.

[0008] Furthermore, the drift type identification specifically includes the following steps:

[0009] (1) Obtain the error rate e at timestamp t t ; Using a window W of length n i , indicating W i Contains n samples, so W i The average error rate is expressed as:

[0010]

[0011] in, is the average error rate at timestamp t, e t is the error rate, n is the window W i The number of samples in ;

[0012] (2) Calculate the difference in average error rate between the two data streams. The formula for calculating the difference is:

[0013]

[0014] Among them, Gap i is the difference between the average error rate of the coke oven predicted data flow and the actual data flow, is the average error rate at timestamp t+1, is the average error rate at timestamp t;

[0015] (3) Each data stream contains l×n samples, where l is the window count and l-1 is the available gap. i Therefore, in the pre-training stage, a sample is represented as {X i ,y i}∈R l , where X i = {Gap 1 ,...,Gap l-1},y i is a label, and y i ∈{1,...,4}; therefore, the underlying data distribution of a data stream is mapped to {X i ,y i};

[0016] (4) Based on the dataset G: {(X 1 ,y 1 ),...,(X L ,y L ), L>0}, using an embedding neural network t with learning parameters θ θ The error feature X i Mapping to the embedding space t θ (X i );

[0017] (5) Calculate the average value of the support set of each category k in the latent space as a single prototype vector; for each category k, the set of data streams marked as k is represented as S k , calculate S k The hidden representation t corresponding to each sample in θ (X i ), and sum them up and average them to get the average hidden representation of category k, that is, a single prototype vector c k , the formula for calculating the prototype vector is:

[0018]

[0019] where c k is a single prototype vector for category k, S k is a set of data streams labeled k, |S k | is the number of support sets, t θ (X i ) represents the sample (X i ,y i )’s hidden representation.

[0020] Furthermore, the drift point positioning specifically includes the following steps:

[0021] (1) L training data streams D: {(X 1 ,z 1 ),...,(X L ,z L ), L>0}, where X i It is also the label z of the corresponding drift point i The relevant error characteristics, and z i ∈{1,...,l-1};

[0022] (2) Using an embedding neural network with learning parameters θ, the error feature X i Mapped to embedding space; defines an embedding neural network v θ , where θ represents the learnable parameters, the network converts the error feature X i Mapped to the embedding space; the mapping process is completed through a linear transformation and ReLU activation function, and its formula is:

[0023] v θ (X i )=ReLU(W e X i +b e ) (4)

[0024] where v θ (X i ) is the error feature X i The representation in the embedding space, W e is the weight parameter of the embedding network, with a dimension of d×d, where d is the feature dimension and b e is the bias parameter of the embedding network, with a dimension of d×n, where n is the number of training data streams and ReLU is the rectified linear unit activation function;

[0025] (3) The prototype vector c k and the encoding vector v θ After concatenation, a new feature vector is obtained through linear transformation and ReLU activation function. The formula is:

[0026] j θ (X i )=ReLU(W c [c k (X i ),v θ (X i )]+b c ) (5)

[0027] where j θ (X i) is the new feature vector for sample X i , W c is the weight parameter of the linear transformation, with a dimension of d×2d, where d is the feature dimension, and b c is the bias parameter of the linear transformation, with a dimension of d×N, where N is the number of training data streams, [c k (X i ), v θ (X i )] is to concatenate the prototype vector c k (X i ) and the encoding vector, v θ (X i ); ReLU is the rectified linear unit activation function;

[0028] (4) Use a multi-layer perceptron for prediction. Input the embedding space vector j θ (X i ) that combines the prototype vector and the input data into the multi-layer perceptron, and obtain the final prediction result through a linear transformation. The linear transformation formula is:

[0029]

[0030] where is the prediction result for sample X i , j θ (X i ) is the embedding space vector that combines the prototype vector and the input data, W o represents the weight parameter of the multi-layer perceptron, with a dimension of 1×d, and d represents the feature dimension, b o represents the bias parameter of the multi-layer perceptron, with a dimension of 1×N, where N represents the number of training data streams;

[0031] (5) Minimize the mean squared error MSE loss at all drift points of all error features X i existing during training; Given a set of error features in the data stream X i , each drift point of each data stream corresponds to an object value. Therefore, the learning objective is to minimize the MSE loss at all drift points of all error features X i existing during training. The calculation formula of the MSE loss is:

[0032]

[0033] where L(θ) is the loss function, that is, the mean squared error MSE, N is the number of training data streams, z θ (X i ) is the true value of the drift point in the data stream X i , is the regression model for the data stream Xi The predicted value of the drift point in ;

[0034] (6) The drift point locator converts the data stream X i As input, it produces an output indicating where the drift occurs in the data stream, namely the stage transition stream S: {(k 1 ,t 1 ),...,(k M ,t M ), M≥0}.

[0035] Furthermore, the stage conversion flow analysis specifically includes the following steps:

[0036] (1) Use a recurrent neural network to receive the output, i.e., the phase transition flow S, and generate a predicted phase transition flow S in the next switching cycle p : Where S p is the predicted stage transition flow, (k i ,t i ) is an element in the stage transition stream, k i is the stage identifier, t i is the time point, M p The number of elements in the stream transformed for the prediction phase;

[0037] (2) In the next switching cycle, the real phase conversion flow S is generated according to the real data flow r : Where S r is the real stage transition flow, (k i ,t i ) is an element in the stage transition stream, k i is the stage identifier, t i is the time point, M r The number of elements in the stream transformed for the real stage;

[0038] (3) Calculate S p With S r The mean square error between them is:

[0039]

[0040] Where MSE is the mean square error, M p is the number of elements in the stage transition stream, is the phase identifier of the i-th element in the prediction phase transition stream, is the phase identifier of the i-th element in the real phase transition stream;

[0041] (4) Take the maximum mean square error (MSE) of multiple commutation cycles of normal data flow max, if MSE>1.2×MSE in the coke oven production process max , it means that the health of the coke oven is poor.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention can reliably and automatically evaluate the health status of the coke oven by performing concept drift-based analysis on the data stream of the coke oven production process, can monitor the coke oven status in real time, promptly discover health problems and handle them, thereby reducing downtime in production and improving production efficiency. A reliable evaluation system can accurately judge the health status of the coke oven, promptly discover potential faults or safety hazards, help prevent the occurrence of production accidents, and ensure the personal safety of workers. It can help manufacturers to perform maintenance and care in a timely manner, reduce equipment damage caused by failure to discover problems in a timely manner, and thus reduce maintenance costs. By accurately evaluating the health status of the coke oven and promptly adjusting production parameters and operations, the stability and consistency of product quality can be ensured, and the production quality can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the method of the present invention.

[0045] Figure 2 It is a partial schematic diagram of the neural network mapping described in the present invention. DETAILED DESCRIPTION

[0046] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0047] See Figure 1-2 The present invention discloses a method for analyzing the health status of a coke oven based on concept drift, including drift type identification, drift point positioning and phase transition analysis. The drift type identification is to obtain an error rate flow, build a classifier to classify the concept drift of a data flow, and locate the prototype vector of each category in a semantic space; the drift point positioning is to combine the prototype vector and the encoding vector of the data flow to obtain a new drift feature vector, input the vector into a multi-layer perceptron regression model, and generate an output indicating the location where the drift occurs in the data flow, namely, a phase transition flow; the phase transition flow analysis is to use a recurrent neural network to predict the phase transition flow, obtain the predicted phase transition flow within a future switching cycle, calculate the Euclidean distance between the real transition flow and the predicted transition flow, and determine the health status of the coke oven.

[0048] The specific steps include:

[0049] 1) Drift type identification:

[0050] Step A: First, the prediction model is applied directly to the data stream to obtain the error rate e at timestamp t t, then, using a window W of length n i , indicating W i Contains n samples, so W i The average error rate can be expressed as:

[0051]

[0052] in, represents the average error rate at timestamp t, e t represents the error rate, n represents the window W i The number of samples in .

[0053] Step B: Calculate the difference in the average error rate between the two data streams. First, calculate the average error rate at timestamp t and timestamp t+1, and then find the difference between them. The formula for calculating the difference is:

[0054]

[0055] Among them, Gap i is the difference between the average error rates of the two data streams, represents the average error rate at timestamp t+1, represents the average error rate at timestamp t.

[0056] Step C: Assume that each data stream contains l×n samples, then l is the window count and l-1 is the available Gap i Therefore, in the pre-training stage, a sample can be represented as {X i ,y i}∈R l , where X i = {Gap 1 ,...,Gap l-1},y i is a label, and y i ∈{1,...,4}; therefore, the underlying data distribution of a data stream is mapped to {X i ,y i}.

[0057] Step D: Based on the dataset G: {(X 1 ,y 1 ),...,(X L ,y L ), L>0}, using an embedding neural network t with learnable parameters θ θ The error feature X i Mapping to the embedding space t θ (X i ).

[0058] Step E: Calculate the average value of the support set of each category k in the latent space as a single prototype vector; First, for each category k, the set of data streams labeled k is represented as S k ; Then, calculate S k The hidden representation corresponding to each sample in is denoted as t θ (X i ), and sum them up and average them to get the average hidden representation of category k, that is, a single prototype vector c k ; The formula for calculating the prototype vector is:

[0059]

[0060] where c k Represents a single prototype vector for category k, S k represents the set of data streams marked as k, |S k | represents the number of support sets, t θ (X i ) represents the sample (X i ,y i ) is a hidden representation of .

[0061] 2) Drift point positioning:

[0062] Step F: Assume L training data streams D: {(X 1 ,z 1 ),...,(X L ,z L ), L>0}, where X i It is also the label z of the corresponding drift point i The relevant error characteristics, and z i ∈{1,...,l-1}.

[0063] Step G: Use an embedding neural network with learnable parameters θ to transform the error features X i Mapped to the embedding space; first, an embedding neural network v is defined θ , where θ represents the learnable parameters, the network converts the error feature X i Mapped to the embedding space; the mapping process is completed through a linear transformation and ReLU activation function, and its formula is:

[0064] v θ (X i )=ReLU(W e X i +b e ) (12)

[0065] where v θ (X i ) indicates the error feature Xi The representation in the embedding space, W e represents the weight parameter of the embedding network, with a dimension of d×d, where d represents the feature dimension and b e Represents the bias parameter of the embedding network, with dimension d×n, where n represents the number of training data streams and ReLU represents the rectified linear unit activation function.

[0066] Step H: Combine prototype vector c k and the encoding vector v θ , get a new feature vector; the prototype vector c k and the encoding vector v θ After concatenation, a new feature vector is obtained through a linear transformation and ReLU activation function. The formula is:

[0067] j θ (X i )=ReLU(W c [c k (X i ),v θ (X i )]+b c ) (13)

[0068] where j θ (X i ) represents sample X i The new eigenvector of c Represents the weight parameter of the linear transformation, with a dimension of d×2d, where d represents the feature dimension and b c represents the bias parameter of the linear transformation, with a dimension of d×N, where N represents the number of training data streams, [c k (X i ),v θ (X i )] represents the prototype vector c k (X i ) and the encoding vector, v θ (X i ) are concatenated, and ReLU represents the rectified linear unit activation function.

[0069] Step I: Use a multilayer perceptron to make predictions, combining the prototype vector and the embedding space vector j of the input data θ (X i ) is input into the multilayer perceptron, and the final prediction result is obtained through a linear transformation. The linear transformation formula is:

[0070]

[0071] in Represents sample X iThe prediction results of θ (X i ) represents the embedding space vector combining the prototype vector and the input data, W o represents the weight parameter of the multilayer perceptron, with a dimension of 1×d, where d represents the feature dimension, and b o Represents the bias parameters of the multilayer perceptron, with a dimension of 1×N, where N represents the number of training data streams.

[0072] Step J: All erroneous features X present during training i Minimize the mean square error MSE loss at all drift points; given data stream X i A set of error features in X, with each drift point in each data stream corresponding to an object value; therefore, the learning target is all the error features X present during training i Minimize the MSE loss at all drift points. The calculation formula of MSE loss is:

[0073]

[0074] Where L(θ) represents the loss function, namely the mean square error MSE, N represents the number of training data streams, and z θ (X i ) represents data stream X i The true value of the drift point in Represents the regression model for the data stream X i The predicted value of the drift point in .

[0075] Step K: The drift point locator converts the data stream X i As input, it produces an output indicating where the drift occurs in the data stream, namely the stage transition stream S: {(k 1 ,t 1 ),...,(k M ,t M ), M≥0}.

[0076] 3) Phase transition flow analysis:

[0077] Step L: Use a recurrent neural network to receive the output, i.e., the phase transition flow S, and generate a predicted phase transition flow S within a future switching cycle p : Where S p represents the predicted stage transition flow, (k i ,t i ) represents an element in the phase transition stream, k i Indicates the stage identifier, t i Indicates the time point, M p Indicates the number of elements in the prediction phase transformation stream.

[0078] Step M: Generate the real phase conversion flow S according to the real data flow in the next commutation cycle r : Where S r represents the actual stage transition flow, (k i ,t i ) represents an element in the phase transition stream, k i Indicates the stage identifier, t i Indicates the time point, M r Indicates the number of elements in the real stage transition stream.

[0079] Step N: Calculate S p With S r The mean square error between them is:

[0080]

[0081] Where MSE stands for mean square error, M p Indicates the number of elements in the stage transition stream, Represents the phase identifier of the i-th element in the prediction phase transformation stream, Represents the phase identifier of the i-th element in the real phase transition stream.

[0082] Step O: Take the maximum mean square error (MSE) of multiple commutation cycles of normal data flow max , if MSE>1.2×MSE in the coke oven production process max , it means that the health of the coke oven is poor.

[0083] The above embodiments are implemented based on the technical solution of the present invention, and detailed implementation methods and specific operation processes are given, but the protection scope of the present invention is not limited to the above embodiments. The methods used in the above embodiments are conventional methods unless otherwise specified.

Claims

1. A coke oven health status analysis method based on concept drift, characterized in that: The method includes drift type identification, drift point location and phase transition analysis. The drift type identification is to obtain the error rate flow, build a classifier to classify the concept drift of the data flow, and locate the prototype vector of each category in the semantic space; the drift point location is to combine the prototype vector and the encoding vector of the data flow to obtain a new drift feature vector, input the vector into a multi-layer perceptron regression model, and generate an output indicating the location where the drift occurs in the data flow, namely, the phase transition flow; the phase transition flow analysis is to use a recurrent neural network to predict the phase transition flow, obtain the predicted phase transition flow within a future switching cycle, calculate the Euclidean distance between the real transition flow and the predicted transition flow, and determine the health status of the coke oven.

2. The coke oven health status analysis method based on concept drift according to claim 1 is characterized in that: The drift type identification specifically includes the following steps: (1) Obtain the error rate e at timestamp t t ; Using a window W of length n i , indicating W i Contains n samples, so W i The average error rate is expressed as: in, is the average error rate at timestamp t, e t is the error rate, n is the window W i The number of samples in ; (2) Calculate the difference in average error rate between the two data streams. The formula for calculating the difference is: Among them, Gapi is the difference between the average error rate of the coke oven prediction data flow and the actual data flow, is the average error rate at timestamp t+1, is the average error rate at timestamp t; (3) Each data stream contains l×n samples, where l is the window count and l-1 is the available gap. i Therefore, in the pre-training stage, a sample is represented as {X i ,y i }∈R l , where X i = {Gap1, ..., Gap l-1 },y i is a label, and y i ∈{1,...,4}; therefore, the underlying data distribution of a data stream is mapped to {X i ,y i }; (4) Based on the data set G: {(X1, y1), ..., (X L ,y L ), L>0}, using an embedding neural network t with learning parameters θ θ The error feature X i Mapping to the embedding space t θ (X i ); (5) Calculate the average value of the support set of each category k in the latent space as a single prototype vector; for each category k, the set of data streams marked as k is represented as S k , calculate S k The hidden representation t corresponding to each sample in θ (X i ), and sum them up and average them to get the average hidden representation of category k, that is, a single prototype vector c k , the formula for calculating the prototype vector is: where c k is a single prototype vector for category k, S k is a set of data streams labeled k, |S k | is the number of support sets, t θ (X i ) represents the sample (X i ,y i )’s hidden representation.

3. The coke oven health status analysis method based on concept drift according to claim 1 is characterized in that: The drift point positioning specifically includes the following steps: (1) L training data streams D: {(X1, z1), ..., (X L , z L ), L>0}, where X i It is also the label z of the corresponding drift point i The relevant error characteristics, and z i ∈{1,…,l-1}; (2) Using an embedding neural network with learning parameters θ, the error feature X i Mapped to embedding space; defines an embedding neural network v θ , where θ represents the learnable parameters, the network converts the error feature X i Mapped to the embedding space; the mapping process is completed through a linear transformation and ReLU activation function, and its formula is: v θ (X i )=ReLU(W e X i +b e ) (4) where v θ (X i ) is the error feature X i The representation in the embedding space, W e is the weight parameter of the embedding network, with a dimension of d×d, where d is the feature dimension and b e is the bias parameter of the embedding network, with a dimension of d×n, where n is the number of training data streams and ReLU is the rectified linear unit activation function; (3) The prototype vector c k and the encoding vector v θ After concatenation, a new feature vector is obtained through linear transformation and ReLU activation function. The formula is: j θ (X i )=ReLU(W c [c k (X i ),v θ (X i )]+b c ) (5) where j θ (X i ) is the sample X i The new eigenvector of c is the weight parameter of the linear transformation, with a dimension of d×2d, where d is the feature dimension and b c is the bias parameter of the linear transformation, with a dimension of d×N, where N is the number of training data streams, [c k (X i ), v θ (X i )] is to convert the prototype vector c k (X i ) and the encoding vector, v θ (X i ) for splicing, ReLU is the rectified linear unit activation function; (4) Use a multi-layer perceptron to make predictions, combining the prototype vector and the embedding space vector j of the input data θ (X i ) is input into the multilayer perceptron, and the final prediction result is obtained through linear transformation. The linear transformation formula is: in For sample X i The prediction results of θ (X i ) is the embedding space vector combining the prototype vector and the input data, W o represents the weight parameter of the multilayer perceptron, with a dimension of 1×d, where d represents the feature dimension, and b o Represents the bias parameters of the multilayer perceptron, with a dimension of 1×N, where N represents the number of training data streams; (5) All erroneous features X present during training i Minimize the mean square error MSE loss at all drift points; given data stream X i Each drift point in each data stream corresponds to an object value, so the learning target is to find all the error features X present during training. i Minimize the MSE loss at all drift points. The calculation formula of MSE loss is: Where L(θ) is the loss function, i.e., mean square error MSE, N is the number of training data streams, z θ (X i ) is the data stream X i The true value of the drift point in For the regression model on the data stream X i The predicted value of the drift point in ; (6) The drift point locator converts the data stream X i As input, it generates an output indicating where the drift occurs in the data stream, namely the phase transition stream S: {(k1, t1), ..., (k M , t M ), M≥0}.

4. The method for analyzing the health status of a coke oven based on concept drift according to claim 1, characterized in that: The stage conversion flow analysis specifically includes the following steps: (1) Use a recurrent neural network to receive the output, i.e., the phase transition flow S, and generate a predicted phase transition flow in the next commutation cycle Where S p is the predicted stage transition flow, (k i , t i ) is an element in the stage transition stream, k i is the stage identifier, t i is the time point, M p The number of elements in the stream transformed for the prediction phase; (2) Generate the real phase conversion flow according to the real data flow in the next switching cycle Where S r is the real stage transition flow, (k i , t i ) is an element in the stage transition stream, k i is the stage identifier, t i is the time point, M r The number of elements in the stream transformed for the real stage; (3) Calculate S p With S r The mean square error between them is: Where MSE is the mean square error, M p is the number of elements in the stage transition stream, is the phase identifier of the i-th element in the prediction phase transition stream, is the phase identifier of the i-th element in the real phase transition stream; (4) Take the maximum mean square error (MSE) of multiple commutation cycles of normal data flow max , if MSE>1.2×MSE in the coke oven production process max , it means that the health of the coke oven is poor.

Images