Condition Diagnosis Method for the Control System of an Electric Fused Magnesia Furnace Based on an Improved Federated Distillation Algorithm

Abstract

The present invention discloses a method for diagnosing the working conditions of an electrofused magnesia furnace control system based on an improved federated distillation algorithm, including: 1. The m-th client constructs an initial electrofused magnesia furnace model F m and downloads the public image set Data0 from the federated central server; 2. The m-th client performs graph preprocessing on the public image set Data m to obtain an ordered residual image sequence V m ; 3. Input V m into the model for training; 4. Input Data0 into the trained local model to obtain a prediction vector #imgabs0# and upload it to the federated central server to complete aggregation, and then the parameters are sent down; 5. The client performs local model training according to the improved federated distillation algorithm; 6. Use the trained improved electrofused magnesia furnace model to judge whether the electrofused magnesia furnace is in an abnormal state through the test image at time t. The present invention can realize the method for diagnosing the working conditions of the electrofused magnesia furnace control system, thereby improving the accuracy of diagnosing the working conditions of the electrofused magnesia furnace control system and enhancing the security of enterprise data privacy.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control of industrial magnesium smelting, and specifically to a method for diagnosing the working conditions of an electrofused magnesia furnace control system based on an improved federated distillation algorithm. Background Art

[0002] Electrofused magnesia has excellent properties such as high melting point, dense structure, strong oxidation resistance, high compressive strength, strong corrosion resistance, and stable chemical properties, and is an indispensable strategic resource in the industrial, military, and even aerospace industries. At present, the technological process of refining electrofused magnesia usually uses a three-phase alternating current electrofused magnesia furnace (referred to as an electrofused magnesia furnace) to heat and melt powdery raw materials mainly composed of magnesite ore through an electric arc. By absorbing the heat generated by the electric arc, the ore powder is heated to nearly 3000 degrees, decomposed into molten magnesium oxide and carbon dioxide gas, and then impurities are removed through a cooling and crystallization process, thereby obtaining high-purity magnesium oxide crystals, that is, electrofused magnesia. The container used for smelting is an iron furnace shell with a limited number of uses, and is cooled through a water circulation system.

[0003] The operation of the electrofused magnesia furnace mainly includes three normal working conditions: heating and melting, feeding, and exhausting. Due to the characteristics of low grade, complex mineral composition, and large composition fluctuations of electrofused magnesia ore in China, the resistance and melting point of the material are unstable during the melting process, and abnormal working conditions such as underburning are likely to occur, which have a great impact on production safety, personnel safety, and product quality. Usually, it is due to the increase in local melting point caused by impurity components in the raw materials, and inappropriate current setting values that make the temperature in the local smelting area lower than the melting temperature of the raw materials, and the molten pool is too viscous, so that carbon dioxide gas cannot be discharged normally, forcing the solution to penetrate the furnace shell protection layer and directly contact the iron furnace shell, resulting in too high a temperature of the furnace shell until it is red-hot and burned through. If not processed in time, it will even lead to phenomena such as burning through and leakage of molten liquid.

[0004] Since the temperature of the molten liquid in the ultra-high temperature electrofused magnesia furnace is not directly measurable, at present, the underburning abnormal working conditions are mainly diagnosed by on-site workers regularly observing the characteristics of the furnace shell. However, the manual inspection work intensity is large, the risk is high, and the accuracy highly depends on the experience and state of the operators. In addition, usually, it only takes 2-3 minutes from the normal working condition to the underburning working condition, and the inspection workers usually need to make round trips to inspect three electrofused magnesia furnaces in each factory area, with a large lag. If not processed in time, it will lead to missed inspections. In view of the above problems, there is an urgent need for an automatic, fast, and effective method for diagnosing the working conditions of an electrofused magnesia furnace. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention aims to solve the problems in the electric fused magnesia furnace control system, such as the high labor intensity, high danger, high dependence on the experience and state of operators for accuracy in manual inspection, and easy leakage of enterprise data. A method for diagnosing the working conditions of an electric fused magnesia furnace control system based on an improved federated distillation algorithm is proposed, with the expectation of optimizing the diagnosis of the working conditions of the electric fused magnesia furnace control system, improving the accuracy of the diagnosis of the working conditions of the electric fused magnesia furnace control system, and enhancing the security of enterprise data privacy.

[0006] The present invention adopts the following technical solutions to solve the technical problems:

[0007] A method for diagnosing the working conditions of an electric fused magnesia furnace control system based on an improved federated distillation algorithm of the present invention is characterized in that it is applied to a network environment composed of M clients and a federated central server, and the federated central server stores an unlabeled public image set Data 0 and the M clients store a labeled image set Data m ; The method for diagnosing the working conditions of the electric fused magnesia furnace control system is carried out according to the following steps:

[0008] Step 1: The m-th client constructs an initial electric fused magnesia furnace model F m and downloads the public image set Data 0 from the federated central server;

[0009] Step 2: Define the local training period of the model as E and the number of training rounds as i, and initialize i = 1; Define the parameters of the initial electric fused magnesia furnace model in the (i - 1)-th round of training as and randomly initialize

[0010] Step 3: The m-th client performs image gray-scale consistency transformation processing on the labeled image set Data m and then extracts the temporal residual images to obtain a temporal residual image sequence V m ;

[0011] Step 4: The m-th client inputs the temporal residual image sequence V m into the initial electric fused magnesia furnace model in the (i - 1)-th round of training, and uses the stochastic gradient descent algorithm to update the parameters of the electric fused magnesia furnace model to obtain the parameters of the local electric fused magnesia furnace model

[0012] Step 5: After assigning i + 1 to i, judge whether i > E holds. If it holds, it means that the local training of the model is completed, and the trained local electric fused magnesia furnace model is obtained Otherwise, return to step 4 and execute sequentially;

[0013] Step 6: The m-th client inputs the public image set Data 0 into the trained local electric fused magnesia furnace model and outputs the prediction vectors of all n electric fused magnesia furnace condition sample data in the public image set Data 0 and uploads them to the federated central server; where, x represents the j-th electric fused magnesia furnace condition sample data, j and represents the prediction vector of the m-th client and represents the prediction probability of the j-th electric fused magnesia furnace condition sample data x in the prediction vector of the m-th client j ;

[0014] Step 7: The federated central server uses Equation (1) to aggregate the prediction vectors of the j-th electric fused magnesia furnace condition sample data x 0 in the public image set Data j from the M clients to obtain the prediction probability j of the j-th electric fused magnesia furnace condition sample data x and thus obtains the globally shared prediction vector according to Equation (2)

[0015]

[0016]

[0017] Step 8: The federated central server broadcasts the globally shared prediction vector to all M clients through a multi-broadcast channel

[0018] Step 9: Initialize the training round number i = 1; and randomly initialize

[0019] Step 10: The m-th client inputs the globally shared prediction vector and the public image set Data 0 into the initial electric fused magnesia furnace model trained in the (i - 1)-th round and updates the parameters of the local electric fused magnesia furnace model using Equations (3) - (6), thereby obtaining the parameters of the local electric fused magnesia furnace model trained in the i-th round

[0020]

[0021]

[0022]

[0023]

[0024] In formula (3), S is the softmax function with temperature coefficient T, and D(S(·))||S(·)) is the K-L divergence;

[0025] In formula (4), is the sign function, indicating whether the j-th sample data x of the marked image set Data of the m-th client m in the m,j belongs to class c. If then it means that the j-th sample data x of the electric fused magnesia furnace working condition m,j belongs to class c. If it means that the j-th sample data x of the electric fused magnesia furnace working condition m,j does not belong to class c; C is the number of classes;

[0026] In formula (5), α is the weight of the loss function, and T is the temperature coefficient;

[0027] In formula (6), η KD is the learning rate of knowledge distillation; is the gradient;

[0028] Step 11: After assigning i + 1 to i, determine whether i > T holds. If it holds, it means that the local training based on the improved federated distillation algorithm is completed, and the trained improved model F m ′ of the electric fused magnesia furnace is obtained. Otherwise, return to Step 10 and execute sequentially;

[0029] Step 12: The m-th client collects the original video segment at time t through the camera and obtains the local image set Data′ to be classified of the m-th client m ;

[0030] Step 13: Input the local image set Data′ to be classified m into the trained improved model F′ m of the electric fused magnesia furnace for working condition diagnosis, and output the prediction result. Then, judge whether the electric fused magnesia furnace is in an abnormal state at time t according to the prediction result. If it is abnormal, an alarm prompt is given through the alarm.

[0031] Compared with the prior art, the beneficial effects of the present invention are embodied in:

[0032] 1. The present invention realizes the self-diagnosis of the working conditions of the electric fused magnesia furnace control system through a neural network. Compared with traditional methods, it does not require operators to monitor regularly, greatly improving the safety of the working condition diagnosis of the electric fused magnesia furnace.

[0033] 2. The present invention uses an improved federated distillation algorithm to train the electric fused magnesia furnace network, improving the accuracy of the working condition diagnosis of the electric fused magnesia furnace control system and the security of enterprise data privacy. Brief Description of the Drawings

[0034] Figure 1 It is a flowchart of a method for diagnosing the working conditions of an electric fused magnesia furnace control system based on an improved federated distillation algorithm of the present invention. Detailed Embodiments

[0035] In this embodiment, referring to Figure 1 , a method for diagnosing the working conditions of an electric fused magnesia furnace control system based on an improved federated distillation algorithm is to train with the constructed initial electric fused magnesia furnace model to obtain the local model of the electric fused magnesia furnace. The unlabeled public image set downloaded from the federated central server is input into the local model of the electric fused magnesia furnace for prediction to obtain a prediction vector. The federated central server aggregates the prediction vectors of all clients to obtain a globally shared prediction vector and distributes it to the clients. The clients use the globally shared prediction vector and the public image set to train the electric fused magnesia furnace model, and use the trained electric fused magnesia furnace model to diagnose the working conditions of the electric fused magnesia furnace at time t. Specifically, the following steps are carried out:

[0036] Step 1. The m-th client constructs an initial electric fused magnesia furnace model F m , and downloads the public image set Data 0 from the federated central server;

[0037] Step 2. Define the local training period of the model as E and the number of training rounds as i, and initialize i = 1; define the parameters of the initial electric fused magnesia furnace model of the (i - 1)-th round of training as and randomly initialize

[0038] Step 3. The m-th client performs image gray-scale consistency transformation processing on the labeled image set Data m and then extracts the temporal residual images to obtain the temporal residual image sequence V m ;

[0039] Step 4. The m-th client inputs the temporal residual image sequence V m into the initial electric fused magnesia furnace model of the (i - 1)-th round of training, and uses the stochastic gradient descent algorithm to update the parameters of the electric fused magnesia furnace model Update to obtain the initial electric fused magnesia furnace model for the i-th round of training parameters

[0040] Step 5: After assigning i + 1 to i, determine whether i > E holds. If it holds, it means that the local model training is completed, and the trained local electric fused magnesia furnace model is obtained Otherwise, return to Step 4 and execute sequentially;

[0041] Step 6: The m-th client inputs the public image set Data 0 into the trained local electric fused magnesia furnace model and outputs the prediction vectors of the local electric fused magnesia furnace model for all n electric fused magnesia furnace working condition sample data in the public image set Data 0 and uploads them to the federated central server; where x represents the j-th electric fused magnesia furnace working condition sample data, j and represents the prediction probability of the j-th electric fused magnesia furnace working condition sample data in the prediction vector of the m-th client;

[0042] Step 7: The federated central server uses Equation (1) to aggregate the local model prediction vectors of the M clients for the j-th electric fused magnesia furnace working condition sample data x 0 in the public image set Data j to obtain the prediction probability j of the j-th electric fused magnesia furnace working condition sample data x and thus obtains the globally shared prediction vector according to Equation (2)

[0043]

[0044]

[0045] Step 8: The federated central server broadcasts the globally shared prediction vector to all M clients through a multi-broadcast channel

[0046] Step 9: Initialize the training round number i = 1; and randomly initialize

[0047] Step 10: The m-th client inputs the globally shared prediction vector and the public image set Data 0 into the initial electric fused magnesia furnace model for the (i - 1)-th round of training and updates the parameters of the local electric fused magnesia furnace model using Equations (3) - (6), thereby obtaining the local electric fused magnesia furnace model for the i-th round of training parameters

[0048]

[0049]

[0050]

[0051]

[0052] In Equation (3), S is the softmax function with temperature coefficient T, and D(S(·))||S(·)) is the K-L divergence;

[0053] In Equation (4), is the sign function, indicating the labeled image set Data of the m-th client m for the j-th electrofused magnesia furnace operating condition sample data x m,j whether it belongs to class c. If then it indicates that the j-th electrofused magnesia furnace operating condition sample data x m,j belongs to class c. If it indicates that the j-th electrofused magnesia furnace operating condition sample data x m,j does not belong to class c; C is the number of classes;

[0054] In Equation (5), α is the weight of the loss function, and T is the temperature coefficient;

[0055] In Equation (6), η KD is the learning rate of knowledge distillation; is the gradient;

[0056] Step 11: After assigning i + 1 to i, determine whether i > T holds. If it holds, it means that the local training based on the improved federated distillation algorithm is completed, and the trained improved electrofused magnesia furnace model F′ m is obtained. Otherwise, return to Step 10 and execute sequentially;

[0057] Step 12: The m-th client collects the original video segment at time t through the camera and obtains the local image set Data′ to be classified of the m-th client m ;

[0058] Step 13: Input the local image set Data′ to be classified m into the trained improved electrofused magnesia furnace model F′ m for operating condition diagnosis, and output the prediction result. Then, determine whether the electrofused magnesia furnace is in an abnormal state at time t according to the prediction result. If it is abnormal, an alarm prompt is given through the alarm.

Claims

1. A method for diagnosing the working conditions of an electrofused magnesia furnace control system based on an improved federated distillation algorithm, characterized in that Applied to a network environment composed of M clients and a federated central server, the federated central server stores an unlabeled public image set and the M clients store labeled image sets ; The working condition diagnosis method of the electrofused magnesia furnace control system is carried out according to the following steps: Step 1. The m-th client constructs an initial fused magnesia furnace model and downloads a public image set from the federal central server ; Step 2: Define the local training period of the model as E and the number of training epochs as i, and initialize i = 1; define the initial model of the electrofused magnesia furnace for the (i - 1)-th round of training The parameters of are ; Step 3. The m-th client performs gray-scale consistency transformation processing on the marked image set and then extracts the temporal residual images to obtain a temporal residual image sequence ; Step 4. The m-th client inputs the temporal residual image sequence into the initial model of the fused magnesia furnace in the (i - 1)-th round of training and uses the stochastic gradient descent algorithm to update the parameters of the fused magnesia furnace model to obtain the parameters of the local model of the fused magnesia furnace in the i-th round of training ; ​ Step 5: After assigning i + 1 to i, determine whether i > E holds. If it holds, it means that the local training of the model is completed, and the trained local model of the electrofused magnesia furnace is obtained. Otherwise, return to Step 4 and execute sequentially. Step 6. The m-th client inputs the public image set into the trained local electric fused magnesia furnace model and outputs the prediction vectors of all n electric fused magnesia furnace condition sample data in the public image set , and uploads them to the federal central server; where represents the j-th electric fused magnesia furnace condition sample data, represents the prediction vector of the m-th client and is the prediction probability of the j-th electric fused magnesia furnace condition sample data in the prediction vector ; Step 7: The federal central server uses Equation (1) to aggregate the prediction vectors of the M client devices for the j-th fused magnesia furnace condition sample data in the public image set to obtain the prediction probability of the j-th fused magnesia furnace condition sample data , and then uses Equation (2) to obtain the globally shared prediction vector ; ​ (1) (2) Step 8: The federal central server broadcasts the globally shared prediction vector to all M clients through multiple broadcast channels ; Step 9: Initialize the number of training rounds \(i = 1\); and randomly initialize ; Step 10, the m-th client inputs the globally shared prediction vector and the public image set into the initial model of the fused magnesia furnace for the (i - 1)-th round of training and updates the parameters of the local model of the fused magnesia furnace by using Equations (3) - (6), so as to obtain the parameters of the local model of the fused magnesia furnace for the i-th round of training ; ​​ (3) (4) (5) (6) In Equation (3), S is the softmax function with temperature coefficient T, is the K-L divergence; In formula (4), is a symbolic function, representing the set of labeled images of the mth client Sample data of the jth fused magnesium furnace operating condition Whether it belongs to category c, if , then it represents the jth fused magnesium furnace working condition sample data belongs to category c, if , represents the jth fused magnesium furnace operating condition sample data The category does not belong to category c; C is the number of categories; In Equation (5), is the weight of the loss function, and T is the temperature coefficient; In formula (6), is the learning rate of knowledge distillation; is the gradient; Step 11: After assigning i + 1 to i, determine whether i > E holds. If it holds, it means that the local training based on the improved federated distillation algorithm is completed, and the trained improved model of the electrofused magnesia furnace is obtained. Otherwise, return to Step 10 and execute sequentially. Step 12: The m-th client collects the original video clip at time t through a camera and obtains the local image set to be classified of the m-th client ; Step 13: Input the local image set to be classified into the trained improved electric fused magnesia furnace model for working condition diagnosis and output the prediction result, so as to judge whether the electric fused magnesia furnace is in an abnormal state at time t according to the prediction result. If it is abnormal, an alarm prompt will be given through the alarm

Images