Deep learning-based rare earth metal electrolysis factory predictive maintenance method
Through deep learning-based autoencoders and gated recurrent unit networks, the problem of rare earth electrolysis plant fault judgment relying on manual observation is solved, accurate fault prediction and early detection are achieved, the level of production automation and product quality consistency are improved, and it is suitable for predictive maintenance in rare earth metal electrolysis plants and other industrial fields.
Patent Information
- Application Number
- CN202510915124.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-03
- Publication Date
- 2025-10-17
AI Technical Summary
In the existing technology, fault diagnosis of rare earth electrolysis plants mainly relies on manual observation, which leads to increased unplanned downtime and affects production efficiency. In addition, existing data analysis methods fail to effectively utilize the large number of process parameters in the complex nonlinear electrolysis process.
A deep learning-based method is adopted, combined with an autoencoder and a gated recurrent unit network. Through data cleaning, feature extraction and time series modeling, intelligent fault diagnosis is achieved. The autoencoder is used for unsupervised feature extraction, the gated recurrent unit is used for fault classification, and a predictive maintenance model is constructed.
It achieves accurate fault prediction and early fault detection, reduces manual intervention, improves the level of production automation, reduces the risk of human operational errors, ensures production stability and product quality consistency, and is suitable for predictive maintenance in rare earth metal electrolysis plants and other industrial fields.
Smart Images
Figure CN120804823A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of rare earth metal electrolysis equipment maintenance, and particularly relates to a rare earth metal electrolysis plant predictive maintenance method based on deep learning. BACKGROUND
[0002] At present, the main method for judging the fault of the electrolysis plant is artificial observation method, which often repairs after the fault occurs, resulting in an increase in unplanned downtime and affecting production efficiency. Therefore, it is necessary to study appropriate methods to discover and locate faults early, predict the working state of the feeding auxiliary machine and diagnose faults according to system operating parameters, so as to discover and locate faults in the early stage of fault occurrence and take corresponding measures.
[0003] The fault prediction method based on data driving uses the methods of degradation process model, machine learning and deep learning to analyze and mine the equipment state and operation data, judge the probability of possible equipment failure, and estimate the time when the equipment system fails or reaches the life threshold. However, a large amount of process parameters are generated in the rare earth electrolysis process. Since the electrolysis process is complex and nonlinear, most of the existing data analysis methods are based on simple statistical models, and a large amount of data generated is not effectively utilized, and lacks deep analysis and mining. SUMMARY
[0004] The purpose of the present application is to provide a rare earth metal electrolysis plant predictive maintenance method based on deep learning, which can obtain more accurate fault prediction results by effectively managing and utilizing the massive data of the factory production line, and give the user a predictive maintenance scheme based on the analysis results.
[0005] To achieve the above purpose, the present application provides a rare earth metal electrolysis plant predictive maintenance method based on deep learning, comprising the following steps:
[0006] Step 1: Collect the historical operation data of the rare earth electrolysis plant, and perform cleaning and preprocessing;
[0007] Step 2: Construct an autoencoder model for feature extraction;
[0008] Step 3: Construct a gated recurrent unit model and perform time series modeling and classification;
[0009] Step 4: Model training;
[0010] Step 5: Model testing, using the trained gated recurrent unit model to classify faults and obtain the prediction result of the fault type.
[0011] Optionally, the historical operation data in step 1 includes normal working condition and known fault working condition data, and the cleaning and preprocessing process includes removing noise and missing values, and finally the data is divided into two types of data set and test set.
[0012] Optionally, the number of nodes in the input layer of the autoencoder model in step 2 is the same as the number of features, the autoencoder model includes an encoding layer and a decoding layer, the encoding layer is used to reduce the dimension of the data, and the encoder hidden layer function is Z = f(W (1) X + b (1) ), wherein X is the multi-source sensor data matrix X = [x1, x2,..., x T ] T ∈R T×n , W (1) ,b (1) are parameters of the autoencoder network.
[0013] The decoding layer attempts to restore the encoded data to the original dimension, and the output layer decoding reconstruction formula is Z = f(W (2) X + b (2) ), wherein X is derived from the previous layer transmission, W (2) ,b (2) are parameters of the autoencoder network, and the parameters are obtained by gradient descent training.
[0014] Optionally, in step 3, the encoding of the autoencoder model is used as input to build a gated recurrent unit model, the number of nodes in the input layer of the gated recurrent unit is the same as the output dimension of the autoencoder, one or more gated recurrent unit layers are connected behind, which are used to learn the long-term dependence relationship in the data, and an output layer is connected at the end, the number of nodes of which is the same as the number of fault types.
[0015] Optionally, in the process of step 4, the autoencoder model is first trained using normal condition data, and then the gated recurrent unit model is trained using the features extracted by the autoencoder model and the fault labels; during the training process, the mean square error is used as the loss function of the autoencoder model, the classification cross-entropy is used as the loss function of the gated recurrent unit model, and the adam optimizer is used for parameter update.
[0016] Optionally, in step 5, the test set is input into the trained gated recurrent unit model for fault classification to obtain the prediction result of the fault type.
[0017] The application provides a rare earth metal electrolysis plant predictive maintenance method based on deep learning, which realizes intelligent fault diagnosis by combining an automatic encoder (AE) and a gated recurrent unit (GRU) network. First, original time series data are collected from various sensors of the electrolysis plant, and data cleaning, normalization and other preprocessing operations are performed to improve data quality. Then, the automatic encoder is used to perform unsupervised feature extraction on the pretreated data, the encoder is reduced in dimension and learns key feature representation, forming a training data set that can be used for supervised learning, and a part of the data is taken as a test set. In the model training stage, the GRU network is used to model the labeled time series feature data, the loss function is calculated and the network parameters are adjusted through back propagation, so that the model has accurate fault classification capability. The trained GRU diagnosis network is used to analyze the real-time monitoring data, and the specific fault type prediction result is output. BRIEF DESCRIPTION OF DRAWINGS
[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, below the drawings needed to be used in the embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.
[0019] Figure 1 is a step flow diagram of a rare earth metal electrolysis plant predictive maintenance method based on deep learning of the present application.
[0020] Figure 2 is a structure diagram of the autoencoder AE.
[0021] Figure 3 is a structure diagram of the GRU. DETAILED DESCRIPTION
[0022] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.
[0023] The present application provides a rare earth metal electrolysis plant predictive maintenance method based on deep learning, comprising the following steps:
[0024] Step 1: Collecting historical operation data of the rare earth electrolysis plant, cleaning and pretreating;
[0025] Step 2: Building an autoencoder model for feature extraction;
[0026] Step 3: Building the gated recurrent unit model and timing modeling and classification;
[0027] Step 4: Model training;
[0028] Step 5: Model testing, using the trained gated recurrent unit model for fault classification, obtaining the prediction results of fault type.
[0029] The specific process is shown in Figure 1 The rare earth metal electrolysis plant predictive maintenance method based on deep learning mainly realizes intelligent fault diagnosis by combining automatic encoder (AE) and gated recurrent unit (GRU) network.
[0030] The following will be further described in combination with specific execution steps:
[0031] Step 1 is the data preprocessing stage;
[0032] Collect the historical operation data of rare earth electrolysis plant, including normal working condition and known fault working condition data. These data usually contain temperature, pressure, flow, current, voltage and other sensor signals. Let the multi-source sensor data matrix generated by the electrolysis plant at T time be: X = [x1, x2,..., xn] ∈ R T ] T ∈ R T×n , where n is the sensor dimension (including temperature, pressure, flow, current, voltage, etc.).
[0033] After collecting the data, we need to clean and preprocess the data to remove noise and missing values to ensure the quality and reliability of the data. Then divide it into two categories: data set and test set. Next, feature extraction is performed on the data, and key features related to faults are selected, [0, 1,..., K] (0 represents normal, 1 ~ K represents fault type) interval, such as the following table. Then normalize the data to the [0, 1] interval for model training.
[0034] KEY Fault type Performance characteristics 0 No fault Data normal 1 Cathode breakage fault Voltage drop increase, local hot spot 2 Anode effect fault Voltage surge, current fluctuation 3 Feeding system fault (jamming, etc.) Electrolyte concentration, yield ... ... ...
[0035] Step 2 is the process of automatic encoder AE feature extraction;
[0036] First, build an AE model with the same number of input layer nodes as the number of features. AE is an unsupervised data dimension compression and data feature expression method. In most cases where automatic encoders are mentioned, compression and decompression functions are achieved through neural networks. Generally, the input data is reconstructed to generate an output with the same dimension as the input data. For example Figure 2As shown, AE structure only has one hidden layer, from input to hidden layer is Encoder, from hidden layer to output is Decoder. The number of hidden layer neurons is much lower than the input layer, then you can use less features (neurons) to represent the input data, so as to reduce the dimensionality.
[0037] AE includes encoding layer and decoding layer, encoding layer is used to reduce the dimensionality of data, capture the main features of data belongs to nonlinear dimensionality reduction, encoder hidden layer function is Z = f(W (1) X + b (1) ), wherein X into electrolytic plant T time generated by multiple source sensor data matrix X = [x1, x2,..., x T ] T ∈R T×n , W (1) ,b (1) is the parameter of autoencoder network, which is obtained by gradient descent training.
[0038] Decoding layer tries to restore the encoded data to the original dimension. Through training AE, it can minimize the reconstruction error, so as to learn the effective representation of normal working condition data. The output layer decoding reconstruction formula is Z = f(W (2) X + b (2) ), wherein X is derived from the last layer transmission, W (2) ,b (2) is the parameter of autoencoder network, which is obtained by gradient descent training.
[0039] Step 3 is c.GRU time series modeling and classification process;
[0040] After getting the encoding representation of AE, it is used as input to build a gated recurrent unit (GRU) model.
[0041] GRU (Gate Recurrent Unit) is a kind of recurrent neural network, which can solve the problems of long-term memory and gradient in back propagation in RNN, and has high computational efficiency. As shown in Figure 3 x t is the input information at the current time;
[0042] h t-1 is the hidden state at the last time, which acts as neural network memory and contains the data information seen by the previous nodes;
[0043] h t is the hidden state passed to the next time, and its update formula is h t = (1-z t )*h t-1 +z t *h t';
[0044] h t is the candidate hidden state, whose update formula is h t ′=tanh(W*[r t *h t-1 ,x t ]);
[0045] r t is the reset gate, which determines to what extent the hidden state of the previous time step is ignored. When the output of the reset gate is close to 0, the network "forgets" the information of the previous time step and only relies on the current input; when the output is close to 1, on the contrary. Its update formula is r t =σ(W r *[h t-1 ,x t ]);
[0046] z t is the update gate, which determines how much information of the previous time step needs to be retained in the hidden state of the current time step. The output value is between 0 and 1, and the greater the value, the more past information is retained, and the smaller the value, the less past information is retained. Its update formula is z t =σ(W z *[h t-1 ,x t ]);
[0047] σ is a sigmoid function that can change data to a value between 0 and 1;
[0048] tanh is a tanh function that can change the value to -1~1.
[0049] The use of GRU in the present application can effectively process time series data and capture the time sequence features therein. The number of input layer nodes of the GRU is the same as the output dimension of the AE, and one or more GRU layers are connected behind, which are used to learn the long-term dependence relationship in the data. An output layer is connected at the end, and the number of nodes is the same as the number of fault types. Its state update formula is h t =(1-z t )*h t-1 +z t *h t ′. The output layer uses Softmax as the activation function to convert the output into a probability distribution.
[0050] Step 4 is the model training process;
[0051] In step 4, the AE is trained using normal operating data to enable it to effectively reconstruct normal data. Then, the GRU model is trained using the features extracted by the AE and the fault labels to enable it to accurately identify fault types. During training, the mean square error (MSE) is used as the loss function for the AE, the classification cross-entropy is used as the loss function for the GRU, and the adam optimizer is used for parameter update.
[0052] In step 5, the test set X' is input into the trained GRU model for fault classification, obtaining the predicted results of fault types, ensuring that the model can effectively distinguish different fault categories. In this way, the running state of the rare earth electrolysis plant can be monitored in real time, and potential faults can be found in time, providing protection for the safe production of the plant.
[0053] In summary, the present application has the following beneficial effects:
[0054] 1. Improved automation level: Through intelligent data analysis and fault prediction, the dependence on manual observation and intervention is reduced, the automation level of the production process is improved, and the risk of human operation errors is reduced.
[0055] 2. Accurate fault diagnosis: AE can automatically extract key features from complex time series data, combined with the time series modeling capability of GRU network, to achieve accurate classification of fault types (such as electrode wear, current anomaly, etc.). This helps to take targeted measures at the early stage of failure to avoid product quality fluctuations caused by equipment abnormalities.
[0056] 3. Early detection of abnormalities: The system can capture the performance degradation trend of the equipment before the fault occurs, and timely issue a warning to ensure the stability of the production process, thereby improving the consistency of product quality and reducing the high cost of emergency maintenance.
[0057] 4. Unsupervised learning adaptability: In actual production, fault samples are rare, and the unsupervised learning feature of AE enables it to be effectively trained even in the case of insufficient labeled data, making it suitable for industrial scenarios.
[0058] 5. Scalability: This method is not only suitable for rare earth metal electrolysis plants, but also can be extended to predictive maintenance scenarios in other industrial fields (such as chemical industry, power industry, etc.).
[0059] The above disclosure is only one or more preferred embodiments of the present application, and of course cannot limit the scope of the rights of the present application, and those of ordinary skill in the art can understand that all or part of the above-mentioned embodiments can be implemented, and equivalent changes made in accordance with the claims of the present application still fall within the scope of the present application.
Claims
1. A predictive maintenance method for rare earth metal electrolysis plants based on deep learning, characterized in that: The following steps are involved: Step 1: Collect historical operating data of the rare earth electrolysis plant and perform cleaning and pretreatment; Step 2: Build an autoencoder model for feature extraction; Step 3: Build a gated recurrent unit model and perform time series modeling and classification; Step 4: Model training; Step 5: Model testing: Use the trained gated recurrent unit model to perform fault classification and obtain the prediction results of the fault type.
2. The method for predictive maintenance of a rare earth metal electrolysis plant based on deep learning according to claim 1, wherein: The historical operating data in step 1 includes data of normal operating conditions and known fault conditions. The cleaning and preprocessing process includes removing noise and missing values. Finally, the data is divided into two categories: data set and test set.
3. The predictive maintenance method for a rare earth metal electrolysis plant based on deep learning according to claim 1, characterized in that: The number of nodes in the input layer of the autoencoder model in step 2 is the same as the number of features. The autoencoder model includes an encoding layer and a decoding layer. The encoding layer is used to reduce the dimension of the data. The encoder hidden layer function is Z = f(W (1) X+b (1) ), where X is the multi-source sensor data matrix X = [x1, x2, ..., x T ] T ∈R T×n , W (1) ,b (1) are the parameters of the autoencoder network; The decoding layer attempts to restore the encoded data to its original dimension. The output layer decoding reconstruction formula is Z = f(W (2) X+b (2) ), where X comes from the previous layer, W (2) ,b (2) are the parameters of the autoencoder network, which are obtained by gradient descent training.
4. The method for predictive maintenance of a rare earth metal electrolysis plant based on deep learning according to claim 1, wherein: In step 3, the encoding of the autoencoder model is used as input to construct a gated recurrent unit model. The number of nodes in the input layer of the gated recurrent unit is the same as the output dimension of the autoencoder. One or more gated recurrent unit layers are connected to it to learn long-term dependencies in the data. At the end, an output layer is connected, and the number of nodes is the same as the number of fault types.
5. The predictive maintenance method for a rare earth metal electrolysis plant based on deep learning according to claim 1, characterized in that: In step 4, the autoencoder model is first trained using normal operating data, and then the gated recurrent unit model is trained using the features and fault labels extracted by the autoencoder model. During the training process, the mean square error is used as the loss function of the autoencoder model, the categorical cross entropy is used as the loss function of the gated recurrent unit model, and the Adam optimizer is used for parameter update.
6. The method for predictive maintenance of a rare earth metal electrolysis plant based on deep learning according to claim 2, wherein: In step 5, the test set is input into the trained gated recurrent unit model for fault classification to obtain the prediction result of the fault type.