Underfiring Condition Identification Method for Electric Fused Magnesia Furnace Based on Active Reinforcement Learning and Multimodality

Abstract

An underburning condition recognition method for an electrofused magnesia furnace based on active reinforcement learning and multi-modalities belongs to the field of intelligent control technology for industrial magnesium smelting, and solves the problem of how to design a fast and accurate discriminant method for the underburning condition of an electrofused magnesia furnace. By collecting historical data of the electrofused magnesia furnace, the most valuable image data samples are selected based on active reinforcement learning for labeling, an image-current sample set is constructed based on time series, a neural network is constructed and initialized based on unlabeled multi-modal data, the multi-modal learning neural network is trained based on labeled multi-modal data, and the underburning condition of the electrofused magnesia furnace is recognized based on the multi-modal neural network. Compared with conventional detection methods, the active reinforcement learning algorithm is introduced. By selecting the most valuable data samples for labeling, the labeled data required for dataset construction is reduced, and the required labor cost is reduced while ensuring the recognition accuracy of the algorithm, thus ensuring safe production.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent control for industrial magnesium smelting, and relates to a method for identifying underburning conditions of an electrofused magnesia furnace based on active reinforcement learning and multi-modalities. Background Art

[0002] Electrofused magnesia has properties such as fire resistance, high temperature resistance, corrosion resistance, and oxidation resistance. It is an important strategic raw material in China and is widely used in many fields such as aviation and military industry. Most of the electrofused magnesia preparation in China uses magnesite ore as raw material, heats it to above 2800°C for smelting using a three-phase alternating current electrofused magnesia furnace, and then cools and crystallizes the obtained magnesium oxide and removes impurities to obtain high-quality electrofused magnesia.

[0003] The smelting process of the electrofused magnesia furnace includes operating conditions such as furnace startup, feeding, normal smelting, and underburning. Among them, the underburning condition is a kind of underburning condition. Its occurrence is usually due to impurities in the raw materials resulting in incomplete melting of the raw materials, and the resulting bubbles cause local overheating in the furnace. If the underburning condition fails to be detected and processed in time, it will not only greatly reduce the product quality, but may also lead to major accidents such as burning through the furnace wall and leakage of molten raw materials, threatening personal safety. Therefore, timely judgment of the underburning condition is very important for the preparation of electrofused magnesia. In actual production, the underburning condition is mainly judged by manual inspection, observing the furnace wall and flame state. However, this method relies on the experience of operators, is prone to misjudgment or missed judgment, and at the same time, the production site environment is harsh, posing a personal safety risk.

[0004] At present, the research on the identification of underfired conditions in electrofused magnesia furnaces can be roughly divided into two categories. One method uses the change patterns of three-phase currents for condition identification. Based on a rule inference algorithm for three-phase current values, by analyzing the historical current statistical characteristics under different conditions, a set of expert rule bases for condition judgment is summarized. Then, the underfired condition is discriminated according to the current values collected in real time on site. However, due to the existence of a large amount of noise in the current values, it is not ideal to judge the condition only based on current characteristics, and this method is only suitable as an auxiliary method. The other method mainly uses the monitoring images of the production of electrofused magnesia furnaces for condition identification. This kind of method uses the intuitive information contained in the furnace wall and furnace mouth flame images to establish a perception model for the condition. However, the thermal imaging device used in this technology has a high cost and it is difficult to achieve large-scale industrial applications. Tests on the above two methods in industrial sites have found that the underfired condition identification technology that solely uses current or images is difficult to achieve satisfactory levels in terms of real-time performance and accuracy. On the one hand, although the change pattern of the current can reflect the production condition to a certain extent, such characteristics are difficult for people to identify. Therefore, the condition marking of current data still needs to be indirectly determined through the monitoring video at the corresponding moment. On the other hand, the cost of manually marking the production images of electrofused magnesia furnaces is relatively high, and it is difficult to accurately mark them in the initial stage or transition state of underfiring. Therefore, how to combine the sensitive and fast characteristics of current characteristics with the accuracy of image characteristics in order to accurately give an early warning when the underfired condition has not been fully formed is an urgent problem to be solved.

[0005] Conventional research methods such as supervised learning methods can, based on a large amount of labeled data, combine the image and current characteristics under the same condition to train a model with better performance. However, as a heavy industrial equipment, for the production of the condition dataset of electrofused magnesia furnaces, workers with a certain length of service are required to mark the data. Not only is the production cost high, but most of the marked data may have serious feature coupling problems, which not only do not help the dataset much, but may even cause overfitting of the model and reduce the robustness of the model. Therefore, how to select "valuable" samples from the large amount of data for marking and utilize the remaining unlabeled samples in order to reduce the corresponding labor cost during the model training process, while ensuring the recognition accuracy and performance of the model and achieving cost reduction and efficiency improvement, is an urgent problem to be solved. Summary of the Invention

[0006] The technical problem to be solved by the present invention is how to design a fast and accurate method for discriminating the underfired condition of an electrofused magnesia furnace, liberating workers from dangerous and high-intensity work, and achieving safe production.

[0007] The present invention solves the above technical problems through the following technical solutions:

[0008] An underfired condition identification method for an electrofused magnesia furnace based on active reinforcement learning and multi-modal includes the following steps:

[0009] Step 1, collect historical operating condition data of the electrofused magnesia furnace;

[0010] Step 2, select the most valuable image data samples for labeling based on active reinforcement learning. The specific method is as follows:

[0011] Step 2.1, input unlabeled training data

[0012] Step 2.2, at the beginning of training, randomly select some samples for annotation as the initial annotation training set Then use the annotated sample set to perform classification training to obtain the initial model θ c .

[0013] Step 2.3, calculate the state S according to formula (1). The (i, j) element of S is defined as:

[0014]

[0015] where x u is the unlabeled training sample, is the corresponding unknown label, and σ(·) is the sigmoid function;

[0016] Step 2.4, based on the Actor network, make action = π(S; θ a ), and select K unlabeled training samples;

[0017] Step 2.5, update the labeled training data Based on this, train the classifier

[0018] Step 2.6, based on the updated training data, calculate the state S' and the reward r;

[0019] Step 2.7, repeat steps 2.3, 2.4, 2.5, and 2.6 until the annotation cost reaches the budget;

[0020] Step 3, sample based on the image data set and the three-phase current data set being in the same time series, select the three-phase current data that matches the image from the three-phase current data set D C to construct an image-current data pair data set;

[0021] Step 4, construct and initialize a neural network based on unlabeled multimodal data. The specific method is as follows:

[0022] Step 4.1, use the CNN convolutional layer to extract the features of the unlabeled image data of the electrofused magnesia furnace;

[0023] Step 4.2: Use an autoencoder to extract the three-phase current data features corresponding to the image data, and encode its output dimension to be corresponding to the dimension of the image features extracted by the CNN;

[0024] Step 4.3: Based on the Transformer multi-modal encoder, perform feature fusion on the current features and the image features;

[0025] Step 4.4: Hand over the fused features to a neural network for model pre-training to obtain an initial recognition model for the underburned condition of fused magnesia;

[0026] Step 5: Train a multi-modal learning neural network based on the labeled multi-modal data. The specific method is as follows:

[0027] Step 5.1: Perform feature extraction and feature fusion on the labeled data;

[0028] Step 5.2: Hand over the fused labeled data features to the initial condition recognition model for condition recognition, and calculate its classification result for the condition by the Softmax layer;

[0029] Step 5.3: Introduce the fault label corresponding to the condition, and calculate its cross-entropy loss based on the model classification result and the true label;

[0030] Step 6: Process the video of the operating condition of the fused magnesia furnace to be recognized and the three-phase current at the corresponding moment, and output its condition discrimination result, so as to judge whether the fused magnesia furnace is in an abnormal state at the current moment according to the discrimination result. If it is abnormal, an alarm prompt for abnormal conditions should be given at this time.

[0031] Furthermore, the historical condition data of the fused magnesia furnace collected in Step 1 includes: video data and three-phase current data during the operation of the fused magnesia furnace. Based on this, a data set D=(D I ; D C ) is constructed, where D I =(I1, I2,...I k ,..., I n ), D C =(C1, C2,..., C k ,..., C n ), D I is the image data set extracted based on the video data, D C is the constructed three-phase current data set, n is the total number of moments in the data set, C k is the k-th element in D C , I k is the k-th element in D I , and k = 1, 2, 3...n.

[0032] Further, the calculation formula of the Softmax layer described in step 5.2 is as follows:

[0033]

[0034] where f i is the corresponding feature vector input to the Softmax layer.

[0035] Further, the calculation formula of the cross-entropy loss described in step 5.3 is:

[0036]

[0037] where N is the total number of samples, y represents the true label of the sample, and p represents the probability that the network predicts that the sample belongs to this category, that is, the output of each sample through the Softmax layer.

[0038] The advantages of the present invention are as follows:

[0039] By collecting the historical data of the electric fused magnesia furnace, the present invention selects the most valuable image data samples for labeling based on active reinforcement learning, constructs an image-current sample set based on time series, constructs and initializes a neural network based on unlabeled multimodal data, trains a multimodal learning neural network based on labeled multimodal data, and identifies the underfiring condition of the electric fused magnesia furnace based on the multimodal neural network. Compared with conventional detection methods, the active reinforcement learning algorithm is introduced, and by selecting the most valuable data samples for labeling, the labeled data required for dataset construction is reduced, the labor cost required is reduced while ensuring the recognition accuracy of the algorithm, and safe production is ensured. Description of the Drawings

[0040] Figure 1 is a flowchart of the method for identifying the underfiring condition of the electric fused magnesia furnace based on active reinforcement learning and multimodality of the present invention. Detailed Embodiments

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0042] The technical solutions of the present invention will be further described below in conjunction with the drawings of the specification and specific embodiments:

[0043] Embodiment 1

[0044] As Figure 1As shown in the figure, the method for identifying underfiring conditions of an electrofused magnesia furnace based on active reinforcement learning and multi-modal includes the following steps:

[0045] Step 1: Collect historical condition data of the electrofused magnesia furnace, including video data and three-phase current data during the operation of the electrofused magnesia furnace, and construct a data set D = (D I , D C ), where D I = (I1, I2,... I k ,..., I n ), D C = (C1, C2,..., C k ,..., C n ), D I is the image data set extracted based on the video data, D C is the constructed three-phase current data set, n is the total number of moments of the data set, C k is the k-th element in D C , I k is the k-th element in D I , k = 1, 2, 3... n; since the image data set and the three-phase current data set are sampled at the same time series, I k and C k are the image-current sample pairs at the same moment.

[0046] Step 2: Select the most valuable image data samples based on Active Reinforcement Learning (ARL) and submit them to human experts for labeling. The basic concepts of active reinforcement learning are explained as follows:

[0047] State: To select samples that are beneficial to improving the classification performance, it is necessary to consider the prediction results of the current classifier for the samples to be selected. Based on this, the state is designed as a matrix, including all the predicted values of the unlabeled training samples X U , where is the number of unlabeled training samples, and C is the number of categories. Mathematically, the (i, j) element of S is defined as:

[0048]

[0049] where x u is the unlabeled training sample, is the corresponding unknown label, and σ(·) is the sigmoid function.

[0050] Action: Let θ aDefined as the parameters of the Actor network. Since the goal of the Actor network is to select samples from the unlabeled training dataset for annotation, we define the action as a vector where each element corresponds to an unlabeled training sample. The sigmoid function is used as the activation function for each element to obtain a value between 0 and 1. The learning policy π(S; θ a ) generates an action a based on the state S. After obtaining the action vector, all candidate samples except the already selected samples are sorted in descending order, and the top K samples with the highest values are selected for annotation. The selected samples and the labels provided by the annotator are denoted as The augmented annotated training data is denoted as where i represents the current training round.

[0051] State Transition: After the selected samples are annotated and added to the labeled training data, the classifier can be updated with the augmented training dataset (X L , Y L ) by minimizing the loss in the equation. After that, a new state matrix S' can be obtained using the new classifier according to formula (1). The minimization equation is:

[0052]

[0053] where, y c is the true label, is the predicted label.

[0054] Reward: To improve the performance of the classifier, ARL expects the Actor network to pay more attention to those samples that are very likely to be misclassified by the classifier. To achieve this, ARL designs a novel reward function by considering the predicted values obtained from the classifier and the true labels given by the experts. Specifically, for the selected sample Define k ij as the true label for class j given by the annotation expert, as the predicted label for this sample for class j obtained from the classifier. For the classification network θ c , the reward is defined as:

[0055]

[0056] If the sample Among them, for each category, the difference between the true label and the predicted probability value is calculated in sequence, and the differences of all categories are summed. If the final summed difference is smaller, the predicted probability value is closer to the true label. Therefore, in the representative formula (3), the higher the reward r, it indicates that the gap between the predicted result of the selected sample and the true label is larger, which means that these samples with incorrect predictions should receive more attention from the classifier. Therefore, during the learning process, the Actor network will be encouraged to select these samples with relatively poor classification by the current model.

[0057] Based on the above definitions, the training process is as follows:

[0058] Step 2.1: Input unlabeled training data

[0059] Step 2.2: At the beginning of training, randomly select some samples for annotation as the initial annotated training set Then use the annotated sample set to perform classification training to obtain the initial model θ c .

[0060] Step 2.3: Calculate the state S according to formula (1);

[0061] Step 2.4: Based on the Actor network, make action = π(S; θ a ), and select K unlabeled training samples;

[0062] Step 2.5: Update the labeled training data Based on this, train the classifier

[0063] Step 2.6: Based on the updated training data, calculate the state S' and the reward r;

[0064] Step 2.7: Repeat steps 2.3, 2.4, 2.5, and 2.6 until the annotation cost reaches the budget.

[0065] Step 3: Based on the time series alignment method, select the three-phase current data that matches the image from the three-phase current dataset D C to construct an image-current data pair dataset.

[0066] Step 4: Construct and initialize a neural network based on unlabeled multimodal data. For the multimodal data of the fused magnesia furnace, the multimodal training method designed by this method includes the following modules:

[0067] CNN convolutional layer, AE autoencoder layer, Transformer multimodal encoder, and Softmax classification layer:

[0068] The CNN convolutional layer therein includes A convolutional blocks, denoted as Conv1K, Conv a K, Conv A K, where Conv A K represents the A-th convolutional block;

[0069] The AE encoding layer described above contains B autoencoders, denoted as AutoEncoder1K, AutoEncoder b K, AutoEncoder B K, where AutoEncoder B K represents the B-th autoencoder;

[0070] The Transformer multi-modal encoder contains C encoders, denoted as Encoder1K, Encoder c K, Encoder C K, where Encoder C K represents the C-th encoder. Each encoder contains a Self Attention layer, a Cross Attention layer and a forward propagation layer.

[0071] Step 4.1: Use the CNN convolutional layer to extract the features of the unlabeled image data of the electrofused magnesia furnace;

[0072] Step 4.2: Use the autoencoder to extract the features of the three-phase current data corresponding to the image data, and encode its output dimension to the dimension corresponding to the image features extracted by the CNN;

[0073] Step 4.3: Based on the Transformer multi-modal encoder, perform feature fusion on the current features and the image features;

[0074] Step 4.4: Hand over the fused features to the neural network for model pre-training to obtain an initial recognition model for the underfired condition of electrofused magnesia.

[0075] Step 5: Train the multi-modal learning neural network based on the labeled multi-modal data. The specific process is as follows:

[0076] Step 5.1: Based on Step 4.1, Step 4.2, and Step 4.3, perform feature extraction and feature fusion on the labeled data;

[0077] Step 5.2: Hand over the fused features of the labeled data to the initial condition recognition model for condition recognition, and calculate its classification result for the condition by the Softmax layer. The calculation formula of Softmax is as follows:

[0078]

[0079] wherein f i is the corresponding feature vector input to the Softmax layer.

[0080] Step 5.3: Introduce the fault labels corresponding to the corresponding working conditions, and calculate the cross-entropy loss based on the model classification results and the true labels. The calculation formula of the cross-entropy loss is:

[0081]

[0082] where N is the total number of samples, y represents the true label of the sample, and p represents the probability that the network predicts the sample belongs to this category, that is, the output of each sample through the Softmax layer. Through continuous iterative training in Step 5, a neural network model that can accurately identify the underfiring working conditions of the electrofused magnesia furnace can be finally obtained.

[0083] Step 6: Identify the underfiring working conditions of the electrofused magnesia furnace based on the multi-modal neural network: Process the video of the operating conditions of the electrofused magnesia furnace to be identified and the three-phase current at the corresponding moment through Steps 4.1, 4.2, and 4.3, and output the working condition discrimination result through the model obtained in Step 5, so as to judge whether the electrofused magnesia furnace is in an abnormal state at the current moment according to the discrimination result. If it is abnormal, an alarm prompt for abnormal working conditions should be given at this time.

[0084] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An underfiring condition recognition method for an electrofused magnesia furnace based on active reinforcement learning and multi-modal, characterized in that, It includes the following steps: Step 1: Collect historical operating condition data of the electrofused magnesia furnace; Step 2: Select the most valuable image data samples for labeling based on active reinforcement learning. The specific method is as follows: Step 2.1: Input unlabeled training data Step 2.

2. At the beginning of training, randomly select some samples for annotation as the initial annotated training set Then use the annotated sample set to perform classification training to obtain the initial model θ c ; Step 2.3: Calculate the state S according to Equation (1). The (i, j) element of S is defined as: Among them, x u is an unlabeled training sample, is the corresponding unknown label, and σ(·) is the sigmoid function; Step 2.4, based on the Actor network, make an action = π(S; θ a ), and select K unlabeled training samples; Step 2.5, Update the labeled training data Based on this, train a classifier Step 2.6: Calculate the state S' and the reward r based on the updated training data; Step 2.7: Repeat Steps 2.3, 2.4, 2.5, and 2.6 until the labeling cost reaches the budget; Step 3: Based on the fact that the image dataset and the three-phase current dataset are sampled at the same time series, select the three-phase current data that matches the image from the three-phase current dataset D C to construct an image-current data pair dataset; Step 4: Construct and initialize a neural network based on the unlabeled multimodal data. The specific method is as follows: Step 4.1: Use the CNN convolutional layer to extract the features of the unlabeled image data of the electrofused magnesia furnace; Step 4.2: Use the autoencoder to extract the features of the three-phase current data corresponding to the image data and encode its output dimension to the dimension corresponding to the image features extracted by the CNN; Step 4.3: Perform feature fusion on the current features and the image features based on the Transformer multimodal encoder; Step 4.4: Hand over the fused features to the neural network for model pre-training to obtain an initial recognition model for the underburned condition of electrofused magnesia; Step 5: Train the multimodal learning neural network based on the labeled multimodal data. The specific method is as follows: Step 5.1: Perform feature extraction and feature fusion on the labeled data; Step 5.2: Hand over the fused features of the labeled data to the initial condition recognition model for condition recognition, and calculate its classification result for the condition by the Softmax layer; Step 5.3: Introduce the fault label of the corresponding condition, and calculate its cross-entropy loss based on the model classification result and the true label; Step 6: Process the operation condition video of the electrofused magnesia furnace to be recognized and the three-phase current at the corresponding moment, and output its condition discrimination result, so as to judge whether the electrofused magnesia furnace is in an abnormal state at the current moment according to the discrimination result. If it is abnormal, an alarm prompt for the abnormal condition should be given at this time.

2. The method for identifying the underfiring condition of an electrofused magnesia furnace based on active reinforcement learning and multi-modal according to claim 1, wherein The historical operating condition data of the electric fused magnesia furnace described in Step 1 includes: video data and three-phase current data during the operation of the electric fused magnesia furnace. Based on this, a data set D = (D I , D C ) is constructed, where D I = (I1, I2,..., I k ,..., I n ), D C = (C1, C2,..., C k ,..., C n ), D I is an image data set extracted based on the video data, D C is a constructed three-phase current data set, n is the total number of moments in the data set, C k is the k-th element in D C , I k is the k-th element in D I , and k = 1, 2, 3... n.

3. The method for identifying the underfiring condition of an electrofused magnesia furnace based on active reinforcement learning and multi-modal according to claim 2, characterized in that The calculation formula of the Softmax layer described in Step 5.2 is as follows: where f i is the corresponding feature vector input to the Softmax layer.

4. The method for identifying the underfiring condition of an electrofused magnesia furnace based on active reinforcement learning and multi-modal according to claim 3, wherein The calculation formula of the cross-entropy loss described in Step 5.3 is: Among them, N is the total number of samples, y represents the true label of the sample, and p represents the probability that the network predicts the sample belongs to this category, that is, the output of each sample through the Softmax layer.

Images