Anode effect early prediction method based on adversarial shapelet learning

Abstract

The application provides an anode effect early prediction method based on an adversarial shapelet learning, which is applied to diagnosis and prediction of fault positions and fault signal trends.The steps are as follows: firstly, an adjacency matrix of an anode current signal is calculated according to the position relationship of anodes of an aluminum electrolytic cell; secondly, a shapelet learning classifier with an adaptive spatial structure is acquired according to the adjacency matrix of the anode current signal, and a multi-head adversarial neural network is constructed; finally, the adversarial neural network is optimized by using an adaptive network shapelet learning method, so that an adversarial neural network model is obtained; and a classification result is obtained by inputting an input time sequence into the adversarial neural network model.The self-adjusting adjacency matrix is established as an adaptive network and a diversity regularization constraint for subsequent subsequence feature learning; and the adversarial network is combined with the shapelet learning method, so that the interpretability of the shapelets is strengthened.

Description

Technical Field

[0001] This invention relates to the field of anode effect detection technology, and in particular to an early prediction method for anode effects based on adversarial shapelet learning. Background Technology

[0002] Anodic effects (AEs) are a unique phenomenon in the aluminum electrolysis process. When an AE occurs at one or more anodes, the voltage and current density at these anodes rise sharply in a short period, reducing the current efficiency of the aluminum electrolysis cell and shortening its lifespan. Furthermore, AEs generate two harmful perfluorocarbon gases: CF4 and C2F6, whose global warming potential is 11,100 times and 6,630 times that of CO2, respectively. Therefore, early detection of anodic effects (AEED) is crucial for reducing the frequency of AEs in aluminum electrolysis cells. However, AEs exhibit regional distribution characteristics caused by a decrease in alumina concentration. Moreover, the AEC is a nonlinear, multivariate, coupled, time-varying system. Therefore, constructing an accurate AEED model is a significant challenge.

[0003] In actual aluminum electrolysis, the AE prediction program in the AEC control system is realized through the relationship between alumina concentration and AEC resistance curve. However, AEC resistance is not only affected by alumina concentration but also by other factors, such as fluorides. Therefore, such a curve is difficult to accurately reflect the changes in alumina concentration in AEC. Therefore, other more detailed rate methods need to be studied. In recent years, some data-driven methods for AEEC have been proposed. The literature [Hang R, Liu Q, Hong D, et al. Cascaded recurrent neural networks for hyperspectral image classification[J].IEEE Transactions on Geoscience and Remote Sensing,2019,57(8):5384-5394.] mentions that AE is predicted by the overall signal, including the cumulative slope trend and the cumulative slope of the cell resistance. In addition, the average voltage and other features of the overall signal are also used to predict the anode effect by constructing singular value threshold and extreme value gradient boost models. However, AE is generally generated on one anode first and then propagates to other anodes. Therefore, the characteristics of the anode effect generally appear on other multi-dimensional anode current signals. Therefore, AEED should use distributed signals. Reference [1]-[Wang G, Zhang Z, Bian Z, et al.A short-term voltage stability online prediction method based on graph convolutional networks and long short-term memory networks[J].International Journal of Electrical Power & Energy Systems,2021,127:106647.] analyzed the power spectrum of ACS and used it for the detection of local anode effects. However, Reference [1] treats each anode current signal as a separate signal, which ignores the spatial relationship between different dimensions. Therefore, the key spatiotemporal features hidden in ACS cannot be obtained. This limits the reliability of these methods for early detection of AE. Summary of the Invention

[0004] To address the shortcomings of the aforementioned background technology, this invention proposes an early prediction method for the anode effect based on adversarial shapelet learning.

[0005] The technical solution of this invention is implemented as follows:

[0006] An early prediction method for the anode effect based on adversarial shapelet learning, the steps of which are as follows:

[0007] Step 1: Calculate the adjacency matrix of the anode current signal based on the positional relationship of the anodes in the aluminum electrolysis cell;

[0008] Step 2: Based on the adjacency matrix of the anode current signal from Step 1, obtain the shapelet learning classifier with adaptive spatial structure;

[0009] Step 3: Construct a multi-head adversarial neural network based on a shapelet learning classifier with an adaptive spatial structure;

[0010] Step 4: Optimize the adversarial neural network using the gradient descent optimization algorithm to obtain the final multi-head adversarial neural network model; then input the input time series into the multi-head adversarial neural network model to obtain the classification results.

[0011] Preferably, the adjacency matrix of the anode current signal is calculated as follows:

[0012] Define the graphical structure of the anode current signal; and calculate the distance representation matrix M of the anode current signal based on the planar graph of the anode current signal. a :

[0013]

[0014] in, Represents the anode v i and v j The horizontal distance between them, V a ={v1 a v2 a ,…,v 12 a},V b ={v1 b v2 b ,…,v 12 b} is the collection of anodes on the same anode generatrix;

[0015] For M a Filtering yields matrix M r :

[0016]

[0017] For matrix M r After standardization, the adjacency matrix M of the anode current signal is obtained:

[0018]

[0019] in, It is a diagonal matrix. It is a degree matrix;

[0020] If the adjacency matrix M of the anode current signal is defined as a fixed adjacency matrix M, and the adaptive adjacency matrix is ​​defined as a dynamic adjacency matrix P, then the comprehensive adjacency matrix A = (1-τ)M + τP, where, W1 = XM1, W2 = XM2. T is the dimension of the input features, C out M1 and M2 are the dimensions of the output features, and M1 and M2 are the parameters of two mapping spaces. M1 and M2 map the input X to two spaces W1 = XM1 and W2 = XM2, where...

[0021] Preferably, the multi-head adversarial neural network includes a multi-head discriminator and a multi-head generator;

[0022] Generator G k The multi-head generator consists of three deconvolutional layers and batch normalization layers, with each batch normalization layer followed by a ReLU activation function. The loss function is defined as follows:

[0023] LossG k =λ g BCELoss(D(s k ),label real )+LossC;

[0024] Among them, LossG k It is generator G k The loss function, λ g The weights of the discriminator loss are represented by D(s). k ) represents the input subsequence s corresponding to the discriminator D. k Predicted label real Representing the true label, BCELoss is the linear cross-entropy loss function, s k =G k (z) is the k-th shapelet, k = 1, 2, ..., K, where K is the number of shapelets, and LossC is the loss function that minimizes the classifier C;

[0025] The BCELoss expression is:

[0026]

[0027] Here, V is a classifier with X as input and Y as output.

[0028] The discriminator consists of three convolutional layers, and the loss function of the discriminator D is LossD, expressed as:

[0029]

[0030] Where, label fake A label indicating false.

[0031] Preferably, the implementation method of the adaptive spatial structure shapelet learning classifier is as follows:

[0032] Calculate s k Minimum distance from the original sequence X:

[0033]

[0034] in, s k Compared with univariate time series X n The distance, where α represents the adjustment parameter for the soft distance approximation; The expression is:

[0035]

[0036] Where d(s) k ,X j”,n ) represents the subsequence s k With X j”,n The Euclidean distance between them;

[0037] By defining the parameter matrix W k Get the predicted label of X

[0038]

[0039] in, D k s k The minimum distance from the original sequence X;

[0040] We obtain shapelets with higher accuracy by minimizing the loss function LossC.

[0041]

[0042] in, λ w λ represents the parameter for classifier regularization. A Denotes the parameters for regularization of the self-adjusting spatial network, and W represents the parameter matrix of the classifier. This represents the distance value of the u-th variable in the input time series corresponding to the k-th shapelet. Let denot represent the distance value of the v-th variable in the input time series corresponding to the k-th shapelet, A represent the network spatial relationship of the input data X, and δ represent the adjustment. The coefficient, B, represents the inverse network space relationship of the input data X.

[0043] Preferably, the method for inputting the input time series data into the adversarial neural network model to obtain the classification result is as follows:

[0044] Noise z is input into a multi-head generator to generate shapelet-like subsequences. The generated shapelet and the real time series X are then input into a discriminator to determine their similarity. The generated shapelet and X are then input into a classifier to determine if the generator-generated shapelet has the ability to distinguish the input X. Finally, the discriminator, generator, and classifier are iteratively updated to obtain shapelets that are more similar to the subsequences and have stronger discriminative power, as well as a classifier with better classification performance.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] 1) This invention combines the spatial relationships between anode current signals of different dimensions to extract interpretable shapelet features;

[0047] 2) The adjacency matrix defines the overall correlation and dynamic graph structure between anodes, which can be used to obtain dynamic spatial structure changes;

[0048] 3) A self-adjusting adjacency matrix is ​​established as a constraint for the adaptive network and diversity regularization for subsequent subsequence feature learning;

[0049] 4) By combining adversarial networks with shapelet learning methods, the interpretability of shapelets is enhanced. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of the present invention.

[0052] Figure 2 This is a flowchart of the adaptive network shapelet learning algorithm of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] like Figure 1 As shown, this embodiment of the invention provides an early prediction method for the anode effect based on adversarial shapelet learning, and the specific steps are as follows:

[0055] Step 1: Calculate the adjacency matrix of the anode current signal based on the positional relationship of the anodes in the aluminum electrolysis cell;

[0056] This invention defines the adjacency matrix M∈R of the anode current signal. N×N The graph structure of the anode current signal is shown. The graph structure illustrates the relationships between anodes, and the adjacency matrix represents these relationships in a matrix format. Each element in the adjacency matrix indicates whether there is an edge connecting the anodes. The distance representation matrix M of the anode current signal is calculated from the planar graph of the anode current signal. a :

[0057]

[0058] in, Represents the anode v i and v j The horizontal distance between them, V a ={v1 a v2 a ,…,v 12 a},V b ={v1 b v2 b ,…,v 12 b A set of anodes is a collection of anodes on the same anode generatrix. Since anodes in the same set are connected in parallel, adjacent anodes have a closer connection. As shown in Formula 1: anodes belonging to the same anode generatrix are... i ,v j The corresponding M a [v i ,v j The value of ] is greater than that of M, which does not belong to the same anode busbar. a [v i ,v j The value of ] should be large. M a [v i ,v jThe relevant distances between the anodes were calculated, illustrating the strength of the connecting edges. All M... a The elements in the array are restricted to the range [0,1].

[0059] For M a Filtering yields matrix M r :

[0060]

[0061] According to formula M r Formula 2 is defined to define the error correction function to correct the previous M. a The values ​​inside are filtered. According to Formula 2, weak connections can be masked, and the most relevant connections can be strengthened. Matrix M r The values ​​in are all 0 or 1.

[0062] Since different nodes have different degrees, standardization is important. For matrix M... r After standardization, the adjacency matrix M of the anode current signal is obtained:

[0063]

[0064] in, It is a diagonal matrix. It is the degree matrix; since the graph is undirected,

[0065] If the adjacency matrix M of the anode current signal is defined as a fixed adjacency matrix M, and the adaptive adjacency matrix is ​​defined as a dynamic adjacency matrix P, then the comprehensive adjacency matrix A = (1-τ)M + τP, where, T is the dimension of the input features, C out M1 and M2 are the dimensions of the output features, and M1 and M2 are the parameters of two mapping spaces. M1 and M2 map the input X to two spaces W1 = XM1 and W2 = XM2, where... All elements in P belong to the range [0,1], representing X. i With X j The level of similarity. The definition of P makes it easy to calculate similarity and gradient.

[0066] Step 2: Based on the adjacency matrix of the anode current signal from Step 1, obtain the shapelet learning classifier with adaptive spatial structure;

[0067] Step 3: Construct a multi-head adversarial neural network based on a shapelet learning classifier with an adaptive spatial structure;

[0068] Multi-head adversarial neural networks (MANs) are used to improve the interpretability of adaptive graph structure shapelet learning. They mainly consist of two parts: 1. a multi-head discriminator and 2. a multi-head generator. The multi-head generator learns to generate various shapelets, while the discriminator adjusts the similarity between the shapelets and the real subsequences. Figure 2 The classifier in the code is used to classify time series with the help of shapelets. In general, the classifier aims to find discriminative subsequences, while the discriminator helps maintain the interpretability of the shapelets. The generator is used to generate shapelets that maintain both discriminativity and interpretability.

[0069] like Figure 2 As shown, a multi-head adversarial neural network mainly consists of three parts: the generator G. k The system consists of a discriminator (D) and a classifier (C). The classifier classifies the original time series. The generator produces discriminative subsequences. The discriminator determines whether shapelets resemble real subsequences. To make the generated shapelets more representative of the real categories, spatial features between dimensions are incorporated. The method mainly comprises three parts: 1. By designing fixed and adaptive adjacency matrices, the spatial relationships and graph structure of the anode current signal are obtained. 2. The overall network framework: the design of a multi-head discriminative generative adversarial network. A multi-adversarial neural network is added to the original shapelet learning network to increase the interpretability of the learned shapelets. 3. A learning method for network shapelets that incorporates an adaptive graph structure. The learning process of the adaptive network shapelet is shown in Table 1.

[0070] Table 1 Adaptive Network Shapelet Learning

[0071]

[0072]

[0073] Generator G k It consists of three deconvolutional layers and batch normalization layers. Each batch normalization layer is followed by a ReLU activation function. The loss function of the multi-head generator is defined as shown in Equation 4, where if the number of shapelets is K, then K generators G are needed. k The k-th generator is trained using the loss function in Formula 4.

[0074] LossG k =λ g BCELoss(D(s k ),label real )+LossC (4);

[0075] Among them, LossG k It is generator G k The loss function, λ g The weights of the discriminator loss are represented by D(s). k ) represents the input subsequence s corresponding to the discriminator D. k Predicted label real Representing the true label, BCELoss is the linear cross-entropy loss function, used to solve linear problems; s k =G k (z) is the k-th shapelet, k = 1, 2, ..., K, where K is the number of shapelets, and LossC is the loss function that minimizes the classifier C.

[0076] The BCELoss expression is:

[0077]

[0078] Where V is a classifier, with X as input and Y as output. real The actual label is set to 1. fake The label representing false is set to 0.

[0079] The multi-head discriminator D is used to discriminate multiple generators G. k Does the generated subsequence resemble a real one? The discriminator consists of convolutional layers, such as... Figure 2 As shown. The loss function of the discriminator D is LossD. By reducing the loss function, we can find s that are more like the real sequence X. k The expression for LossD is:

[0080]

[0081] Where, label fake A label indicating false.

[0082] The implementation method of the shapelet learning classifier based on adaptive spatial structure is as follows:

[0083] Calculate s k Minimum distance from the original sequence X:

[0084]

[0085] in, s k Compared with univariate time series X n The distance, where α represents the adjustment parameter for the soft distance approximation; The expression is:

[0086]

[0087] Where d(s) k ,X j”,n ) represents the subsequence s k With X j”,n The Euclidean distance between them;

[0088] By defining the parameter matrix W k Get the predicted label of X

[0089]

[0090] in, D k s k The minimum distance from the original sequence X;

[0091] We obtain shapelets with higher accuracy by minimizing the loss function LossC.

[0092]

[0093] in, λ w λ represents the parameter for classifier regularization. A Denotes the parameters for regularization of the self-adjusting spatial network, and W represents the parameter matrix of the classifier. This represents the distance value of the u-th variable in the input time series corresponding to the k-th shapelet. Let denot represent the distance value of the v-th variable in the input time series corresponding to the k-th shapelet, A represent the network spatial relationship of the input data X, and δ represent the adjustment. The coefficient, B, represents the inverse network space relationship of the input data X.

[0094] In the loss function C, spatial constraints are added: an adjacency matrix A and a reversed adjacency matrix B, where each element of the reversed adjacency matrix B is 1 minus the corresponding element in A. The resulting shapelet not only resembles the true subsequence but also exhibits good discriminative power, while also taking into account the spatial constraints of the original time series X.

[0095] Step 4: Optimize the adversarial neural network using adaptive network shapelet learning to obtain the adversarial neural network model; and input the input time series into the adversarial neural network model to obtain the classification result.

[0096] The method for inputting time series data into an adversarial neural network model to obtain classification results is as follows:

[0097] The noise z is input into the multi-head generator to generate a shapelet-like subsequence. The input time series X is input into the shapelet to learn the classifier. The multi-head discriminator is used to judge the similarity between the generated shapelet subsequence and the real time series X to update the parameters of the generator, classifier and discriminator.

[0098] The updated shapelet learning classifier uses the generated shapelet subsequences to classify the input time series X, and outputs the classification results and shapelets with strong discriminative power.

[0099] The anode current signal classification method based on shapelet transformation proposed in this invention provides important support for local cell condition analysis of aluminum electrolysis cells and intelligent production of aluminum electrolysis. This method will be applied to the diagnosis and prediction of fault location and fault signal trends.

[0100] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early prediction of anode effects based on adversarial shapelet learning, characterized in that, The steps are as follows: Step 1: Calculate the adjacency matrix of the anode current signal based on the positional relationship of the anodes in the aluminum electrolysis cell; Step 2: Based on the adjacency matrix of the anode current signal from Step 1, obtain the shapelet learning classifier with adaptive spatial structure; Step 3: Construct a multi-head adversarial neural network based on a shapelet learning classifier with an adaptive spatial structure; Step 4: Optimize the adversarial neural network using the gradient descent optimization algorithm to obtain the final multi-head adversarial neural network model; then input the input time series data into the multi-head adversarial neural network model to obtain the classification results; The adjacency matrix of the anode current signal is calculated as follows: Define the graphical structure of the anode current signal, and calculate the distance representation matrix M of the anode current signal based on the planar graph of the anode current signal. a : in, Represents the anode v i and v j The horizontal distance between them, V a ={v1 a v2 a ,…,v 12 a },V b ={v1 b v2 b ,…,v 12 b } is the collection of anodes on the same anode generatrix; For the distance representation matrix M a Filtering yields matrix M r : For matrix M r After standardization, the adjacency matrix M of the anode current signal is obtained: in, It is a diagonal matrix. It is a degree matrix; The implementation method of the adaptive spatial structure shapelet learning classifier is as follows: Calculate the k-th shapelet s k Minimum distance from the original sequence X: in, s k Compared with univariate time series X n The distance, where α represents the adjustment parameter for the soft distance approximation; distance The expression is: Wherein d(s) k ,X j”,n ) represents the subsequence s k With X j”,n The Euclidean distance between them; By defining the parameter matrix W k Obtain the predicted label of sequence X in, D k s k The minimum distance from the original sequence X; Obtain shapelets with high accuracy by minimizing the loss function LossC: in, λ w λ represents the parameter for classifier regularization. A Denotes the parameters for regularization of the self-adjusting spatial network, and W represents the parameter matrix of the classifier. This represents the distance value between the k-th shapelet and the u-th variable in the input time series. The k-th shapelet represents the distance value of the v-th variable in the input time series. The combined adjacency matrix A represents the network spatial relationships of the input X, and δ represents the adjustment... The coefficient of , B represents the inverse network space connection of input X.

2. The method for early prediction of anode effects based on adversarial shapelet learning according to claim 1, characterized in that, If the adjacency matrix M of the anode current signal is defined as a fixed adjacency matrix M, and the adaptive adjacency matrix is ​​defined as a dynamic adjacency matrix P, then the comprehensive adjacency matrix A = (1-τ)M + τP, where, W1 = XM1, W2 = XM2; T is the dimension of the input features, C out M1 and M2 are the dimensions of the output features, and M1 and M2 are the parameters of two mapping spaces. The input X is mapped to two spaces W1 = XM1 and M2 through M1 and M2. In W2 = XM2, where 3. The method for early prediction of anode effects based on adversarial shapelet learning according to claim 1, characterized in that, The multi-head adversarial neural network includes a multi-head discriminator and a multi-head generator; Generator G k The multi-head generator consists of three deconvolutional layers and batch normalization layers, with each batch normalization layer followed by a ReLU activation function. The loss function is defined as follows: LossG k [μλ] g BCELoss(D(s) k ),label real )+LossC; Among them, LossG k It is generator G k The loss function, λ g The weights of the discriminator loss are represented by D(s). k ) represents the input subsequence s corresponding to the discriminator D. k Predicted label real Representing the true label, BCELoss is the linear cross-entropy loss function, s k =G k (z) is the k-th shapelet, k = 1, 2, ..., K, where K is the number of shapelets, and LossC is the loss function that minimizes the classifier C; The BCELoss expression is: Where V is a classifier, with X as input and Y as output; The discriminator consists of three convolutional layers, and the loss function of the discriminator D is LossD, expressed as: Where, label fake A label indicating false.

4. The method for early prediction of anode effects based on adversarial shapelet learning according to claim 1, characterized in that, The method for inputting the input time series into the adversarial neural network model to obtain the classification result is as follows: The noise z is input into the multi-head generator to generate a subsequence similar to a shapelet. The generated shapelet and the real time series X are input into the discriminator to determine the similarity between the shapelet and the real subsequence. The generated shapelet and X are input into the classifier to determine whether the shapelet generated by the generator has the ability to distinguish the input X. Finally, the discriminator, generator and classifier are iteratively updated to obtain a shapelet that is more like a subsequence and has stronger discriminative power, and a classifier with better classification performance.

Images