A superheat degree recognition method, system, device and storage medium

Abstract

The present invention discloses a superheat degree recognition method, system, device and storage medium. The method includes obtaining a Huoyan video training sample set; constructing a superheat degree recognition model based on the I3D network layer and the Vision Transformer layer; performing weight initialization on the parameters in the shallow convolutional neural network in the superheat degree recognition I3D network layer based on parameter-sharing transfer learning to obtain initialized weight parameters; training all the parameters in the superheat degree recognition model by using the superheat degree recognition initialized weight parameters and the superheat degree recognition Huoyan video training sample set to obtain a trained superheat degree recognition model; and performing superheat degree recognition on a sample to be predicted by using the trained superheat degree recognition model to obtain a superheat degree recognition result. The present invention can improve the accuracy of superheat degree recognition and the generalization performance of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial control, and in particular, to a superheat degree recognition method, system, device, and storage medium. Background Art

[0002] The superheat degree of the electrolyte in an aluminum electrolysis cell refers to the difference between the electrolyte temperature and the primary crystallization temperature. The superheat degree is one of the key indicators reflecting the current production efficiency and quality of the aluminum electrolysis cell. Maintaining an appropriate superheat degree can improve the current efficiency, stabilize the production process, and reduce production energy consumption. Existing superheat degree measurement methods need to measure the electrolyte temperature and the primary crystallization temperature separately. The electrolyte temperature is generally measured online using a thermocouple or an infrared thermometer, while the measurement of the primary crystallization temperature needs to be carried out offline by sampling and testing. The measurement process is relatively complex, the measurement cost is high, and real-time measurement cannot be carried out. In the actual production process of current aluminum electrolysis plants, the superheat degree is mainly judged by process personnel observing the fire eye video with the naked eye. However, due to the high temperature, large magnetic field, and strong corrosiveness of the production environment of the aluminum electrolysis cell, it will cause great interference to manual judgment. At the same time, the process of manually judging the superheat degree seriously depends on a small number of experienced process personnel, and there are subjective factors in the judgment process. These two factors lead to unstable electrolysis states, making it difficult to adjust to the optimal state, resulting in low electrolysis efficiency and resource waste.

[0003] In recent years, with the development and wide application of machine vision and deep learning, many scholars have proposed methods based on deep learning and fire eye video data to identify the superheat degree state. The basis is to use relevant algorithms to extract the features of the fire eye video for the fire eye video taken during the aluminum electrolysis production process, so as to imitate the decision-making judgment of production personnel during the actual production process to identify the superheat degree state of the current fire eye. The fire eye video is three-dimensional video data, which has an additional time dimension compared with two-dimensional image data. In addition to containing a large amount of position information in the two-dimensional spatial dimension, it also contains rich motion information in the time dimension. However, in the current superheat degree recognition methods based on deep learning, only the static features regarding the position information of the flame video are better considered, but the dynamic features in the time dimension of the fire eye video are not considered perfectly, so it will affect the superheat degree recognition result. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art. For this purpose, the present invention provides a superheat degree recognition method, system, device, and storage medium, which can improve the accuracy of superheat degree recognition and the generalization performance of the model.

[0005] In a first aspect, an embodiment of the present invention provides a superheat degree recognition method, and the method includes:

[0006] Obtain a fire eye video training sample set;

[0007] Construct an overheat recognition model based on the I3D network layer and the Vision Transformer layer;

[0008] Based on transfer learning with parameter sharing, initialize the weights of the parameters in the shallow convolutional neural network in the I3D network layer to obtain initialized weight parameters;

[0009] According to the initialized weight parameters, train all the parameters in the overheat recognition model through the Huoyan video training sample set to obtain a trained overheat recognition model;

[0010] Perform overheat recognition on the sample to be predicted through the trained overheat recognition model to obtain an overheat recognition result.

[0011] Compared with the prior art, the first aspect of the present invention has the following beneficial effects:

[0012] In this method, an overheat recognition model is constructed based on the I3D network layer and the Vision Transformer layer, which not only considers the position information of the Huoyan video data in the two-dimensional spatial dimension, but also considers the rich motion information in the time dimension, and comprehensively considers the spatio-temporal correlation of the depth features of the Huoyan video to model the overheat, improves the dynamic features of the Huoyan video, and improves the accuracy of overheat recognition; based on transfer learning with parameter sharing, initialize the weights of the parameters in the shallow convolutional neural network in the I3D network layer to obtain initialized weight parameters. By introducing transfer learning to initialize the weights of some parameters of the overheat recognition model, the accuracy of overheat recognition can be improved, and the generalization performance of the model can be improved.

[0013] According to some embodiments of the present invention, the obtaining of the Huoyan video training sample set includes:

[0014] Collect Huoyan video samples of aluminum electrolytic cells, and obtain the Huoyan centers corresponding to the Huoyan video samples of the aluminum electrolytic cells through the YOLO model;

[0015] Perform sample amplification on the Huoyan video samples of the aluminum electrolytic cells through the Mixup data augmentation method to obtain a Huoyan video training sample set.

[0016] According to some embodiments of the present invention, the overheat recognition method further includes:

[0017] Extract features from the samples in the Huoyan video training sample set through the deep convolutional neural network in the I3D network layer to obtain deep abstract features;

[0018] Adopt three-dimensional adaptive max pooling to reduce the dimension of the deep abstract features to obtain dimension-reduced deep abstract features;

[0019] Merge and transpose the last two dimensions of the dimensionality-reduced depth abstract features to obtain multiple groups of one-dimensional feature vectors;

[0020] Concatenate the CLS Token vector with the multiple groups of one-dimensional feature vectors to obtain a concatenated vector;

[0021] Input the concatenated vector into the Vison Transformer layer to obtain an output vector at the position of the CLS Token vector and output vectors at the positions of multiple groups of one-dimensional feature vectors; wherein, the Vison Transformer layer includes a self-attention mechanism and a feed-forward neural network;

[0022] Input the output vector at the position of the CLS Token vector into the classification module to obtain an overheat degree recognition result, and the classification module includes two linear layers with activation functions and a Softmax classifier.

[0023] According to some embodiments of the present invention, the step of inputting the concatenated vector into the Vison Transformer layer to obtain an output vector at the position of the CLS Token vector and output vectors at the positions of multiple groups of one-dimensional feature vectors includes:

[0024] Input the concatenated vector into the Vison Transformer layer, so that the Vison Transformer layer performs weighted aggregation on the multiple groups of one-dimensional feature vectors through the self-attention mechanism to obtain a globally relevant feature vector between the CLS Token vector and the multiple groups of one-dimensional feature vectors;

[0025] Input the globally relevant feature vector into the feed-forward neural network, so that the feed-forward neural network performs non-linear projection on the CLS Token vector and the multiple groups of one-dimensional feature vectors to obtain an output vector at the position of the CLS Token vector and output vectors at the positions of multiple groups of one-dimensional feature vectors.

[0026] According to some embodiments of the present invention, the step of inputting the output vector at the position of the CLS Token vector into the classification module to obtain an overheat degree recognition result includes:

[0027] Input the output vector at the position of the CLS Token vector into the classification module, so that the classification module performs non-linear mapping through the two linear layers with activation functions to obtain a mapped vector;

[0028] Classify the mapped vector through a Softmax classifier to obtain an overheat degree recognition result.

[0029] According to some embodiments of the present invention, for the transfer learning based on parameter sharing, weight initialization is performed on the parameters of the shallow convolutional neural network in the I3D network layer to obtain initialized weight parameters, including:

[0030] Using the training video classification model of the Kinetics dataset as a pre-trained model;

[0031] Transferring the weight parameters of the pre-trained model to the I3D network layer, and performing weight initialization on the parameters of the shallow convolutional neural network in the I3D network layer to obtain initialized weight parameters.

[0032] According to some embodiments of the present invention, for training all the parameters in the superheat degree recognition model through the Huoyan video training sample set according to the initialized weight parameters to obtain a trained superheat degree recognition model, including:

[0033] Extracting texture features through the shallow convolutional neural network according to the initialized weight parameters to obtain the texture features extracted after the initialization of the weights;

[0034] Training the abstract features extracted by the deep convolutional neural network based on the texture features extracted after the initialization of the weights to obtain the trained abstract features;

[0035] Training all the parameters in the superheat degree recognition model through the Huoyan video training sample set according to the trained abstract features to obtain a trained superheat degree recognition model.

[0036] In a second aspect, an embodiment of the present invention further provides a superheat degree recognition system, and the system includes:

[0037] A data acquisition unit, configured to acquire a Huoyan video training sample set;

[0038] A model construction unit, configured to construct a superheat degree recognition model based on the I3D network layer and the Vision Transformer layer;

[0039] A transfer learning unit, configured to perform weight initialization on the parameters of the shallow convolutional neural network in the I3D network layer based on transfer learning with parameter sharing to obtain initialized weight parameters;

[0040] A model training unit, configured to train all the parameters in the superheat degree recognition model through the Huoyan video training sample set according to the initialized weight parameters to obtain a trained superheat degree recognition model;

[0041] A superheat degree recognition unit, configured to perform superheat degree recognition on a sample to be predicted through the trained superheat degree recognition model to obtain a superheat degree recognition result.

[0042] In a third aspect, an embodiment of the present invention further provides a superheat degree identification device, including at least one control processor and a memory communicatively connected to the at least one control processor; the memory stores instructions executable by the at least one control processor, and the instructions are executed by the at least one control processor to enable the at least one control processor to execute a superheat degree identification method as described above.

[0043] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium storing computer-executable instructions for causing a computer to execute a superheat degree identification method as described above.

[0044] It can be understood that the beneficial effects of the above second to fourth aspects compared with the related art are the same as those of the above first aspect compared with the related art. For relevant descriptions, reference can be made to the relevant descriptions in the above first aspect, and details are not described herein again. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the following description of the embodiments in conjunction with the accompanying drawings, where:

[0046] Figure 1 is a flowchart of a superheat degree identification method according to an embodiment of the present invention;

[0047] Figure 2 is a schematic structural diagram of a Vison Transformer layer according to an embodiment of the present invention;

[0048] Figure 3 is a structural diagram of a superheat degree identification system according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

[0050] In the description of the present invention, if the first, second, etc. are described only for the purpose of distinguishing technical features, they should not be construed as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features or implicitly indicating the sequence of the indicated technical features.

[0051] In the description of the present invention, it should be understood that regarding the orientation description, such as the orientation or positional relationship indicated by up, down, etc., is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation to the present invention.

[0052] In the description of the present invention, it should be noted that unless otherwise clearly defined, terms such as "set", "install", "connect", etc. should be understood in a broad sense. Those skilled in the art can reasonably determine the specific meanings of the above terms in the present invention in combination with the specific content of the technical solution.

[0053] First, analyze several terms involved in this application:

[0054] Superheat degree: It refers to the difference between the superheat temperature and the saturation temperature of the refrigerant under the same evaporation pressure in the refrigeration cycle. In terms of the properties of water and water vapor, the superheat degree refers to the superheated steam temperature minus the dry saturation temperature corresponding to the pressure. For water and water vapor, its saturation curve is an ascending curve on the water vapor diagram, that is, as the pressure increases, the saturation temperature of water also increases. Similarly, when water vapor is already in a superheated state, if the pressure is increased, its corresponding saturation temperature also increases, and the degree by which its temperature is higher than the saturation temperature decreases, that is, the superheat degree of the steam decreases.

[0055] Feedforward neural network: Abbreviated as feedforward network, it is a type of artificial neural network. In this kind of neural network, each neuron starts from the input layer, receives the input from the previous level, and inputs it to the next level until the output layer. There is no feedback in the whole network, and it can be represented by a directed acyclic graph. The feedforward neural network is the earliest proposed artificial neural network and also the simplest type of artificial neural network. According to the number of layers of the feedforward neural network, it can be divided into single-layer feedforward neural network and multi-layer feedforward neural network. Common feedforward neural networks include Perceptrons, BP (Back Propagation) network, RBF (Radial Basis Function) network, etc.

[0056] Transfer Learning: It is to use the knowledge learned from one environment to help with the learning task in a new environment.

[0057] The superheat degree of the electrolyte in an aluminum electrolysis cell refers to the difference between the electrolyte temperature and the primary crystallization temperature. The superheat degree is one of the key indicators reflecting the current production efficiency and quality of the aluminum electrolysis cell. Maintaining an appropriate superheat degree can improve the current efficiency, stabilize the production process, and reduce production energy consumption. Existing superheat measurement methods require separate measurements of the electrolyte temperature and the primary crystallization temperature. The electrolyte temperature is generally measured online using a thermocouple or an infrared thermometer, while the measurement of the primary crystallization temperature needs to be carried out offline by sampling and testing. The measurement process is relatively complex, the measurement cost is high, and real-time measurement cannot be carried out. In the actual production process of current aluminum electrolysis plants, the superheat degree is mainly judged by technicians observing the video of the flame eye with the naked eye. However, due to the high temperature, strong magnetic field, and strong corrosiveness in the production environment of the aluminum electrolysis cell, it will cause great interference to manual judgment. At the same time, the process of manually judging the superheat degree highly depends on a small number of experienced technicians, and there are subjective factors in the judgment process. These two factors lead to unstable electrolysis states, making it difficult to adjust to the optimal state, resulting in low electrolysis efficiency and resource waste.

[0058] In recent years, with the development and wide application of machine vision and deep learning, many scholars have proposed methods based on deep learning and fire eye video data to identify the superheat state. The basis is to use relevant algorithms to extract the features of the fire eye video from the fire eye video taken during the aluminum electrolysis production process, and to imitate the decision-making judgment of production personnel during the actual production process to identify the superheat state of the current fire eye. The fire eye video is three-dimensional video data, which has an additional time dimension compared to two-dimensional image data. In addition to containing a large amount of position information in the two-dimensional spatial dimension, it also contains rich motion information in the time dimension. However, in the current superheat identification methods based on deep learning, only the static features regarding the position information of the flame video are better considered, but the dynamic features in the time dimension of the fire eye video are not considered perfectly, thus affecting the superheat identification result.

[0059] To solve the above problems, this application constructs a superheat identification model based on the I3D network layer and the Vision Transformer layer. It not only considers the position information of the fire eye video data in the two-dimensional spatial dimension, but also considers the rich motion information in the time dimension, and comprehensively considers the spatio-temporal correlation of the deep features of the fire eye video to model the superheat degree, improving the dynamic features of the fire eye video and the accuracy of superheat identification; based on transfer learning with parameter sharing, the parameters in the shallow convolutional neural network in the I3D network layer are initialized with weights to obtain the initialized weight parameters. By introducing transfer learning to initialize the weights of some parameters of the superheat identification model, the accuracy of superheat identification can be improved and the generalization performance of the model can be enhanced.

[0060] Refer to Figure 1, an embodiment of the present invention provides a superheat degree recognition method, and the method includes:

[0061] Step S100, obtain a training sample set of the fire eye video.

[0062] Specifically, collect the fire eye video samples of the aluminum electrolysis cell, and obtain the fire eye center corresponding to the fire eye video samples of the aluminum electrolysis cell through the YOLO model;

[0063] Since the number of fire eye video samples of aluminum electrolysis is relatively small, in order to increase the diversity of training data to avoid the overfitting phenomenon of the superheat degree recognition model, during the training process of this embodiment, in addition to using the methods of random cropping and random flipping to expand the fire eye video samples of the aluminum electrolysis cell, the Mixup data augmentation method is also used to amplify the samples of the fire eye video samples of the aluminum electrolysis cell, so as to obtain a training sample set of the fire eye video; among them, the Mixup data augmentation method is to randomly select and mix two samples and their corresponding labels to generate new random samples, and the Mixup data augmentation method is:

[0064]

[0065] where x mixup , y mixup respectively represent the new sample and the label corresponding to the new sample after mixing by the Mixup data augmentation method, x1, x2 represent two randomly selected samples, y1, y2 represent the corresponding one-hot labels, and λ represents the sampling value of the beta distribution with parameters (α, β).

[0066] In this embodiment, considering that the number of fire eye video samples of aluminum electrolysis is relatively small, in order to increase the diversity of training data to avoid the overfitting phenomenon of the superheat degree recognition model, in addition to using the methods of random cropping and random flipping to expand the fire eye video samples of the aluminum electrolysis cell, the Mixup data augmentation method is also used to enrich the training process data. By randomly selecting and mixing two samples and the corresponding labels of the two samples through the Mixup data augmentation method to generate new random samples, the diversity of training data is increased. Therefore, this embodiment can better train the superheat degree recognition model and prevent the superheat degree recognition model from overfitting.

[0067] Step S200, construct a superheat degree recognition model based on the I3D network layer and the Vision Transformer layer.

[0068] Specifically, the samples in the Huoyan video training sample set are subjected to feature extraction through the deep convolutional neural network in the I3D network layer to obtain deep abstract features; three-dimensional adaptive max pooling is used to reduce the dimensions of the deep abstract features to obtain dimension-reduced deep abstract features; the last two dimensions of the dimension-reduced deep abstract features are merged and transposed to obtain multiple groups of one-dimensional feature vectors; the CLS Token vector is concatenated with the multiple groups of one-dimensional feature vectors to obtain a concatenated vector; the concatenated vector is input into the Vison Transformer layer, so that the Vison Transformer layer performs weighted aggregation on the multiple groups of one-dimensional feature vectors through the self-attention mechanism to obtain a feature vector globally related between the CLS Token vector and the multiple groups of one-dimensional feature vectors; the globally related feature vector is input into the feed-forward neural network, so that the feed-forward neural network performs non-linear projection on the CLS Token vector and the multiple groups of one-dimensional feature vectors to obtain an output vector at the position of the CLS Token vector and output vectors at the positions of the multiple groups of one-dimensional feature vectors; the output vector at the position of the CLS Token vector is input into the classification module, so that the classification module performs non-linear mapping through two linear layers with activation functions to obtain a mapped vector. The classification module includes two linear layers with activation functions and a Softmax classifier; the mapped vector is classified through the Softmax classifier to obtain the superheat degree recognition result. For the convenience of those skilled in the art to understand, a set of optimal embodiments is provided below:

[0069] Construct a superheat degree recognition model based on the I3D network layer and the Vison Transformer layer, and recognize the superheat degree of the Huoyan in the aluminum electrolysis cell Huoyan video through the superheat degree recognition model. The superheat degree recognition model includes: a feature extraction part, a VT encoding part, and a classification module.

[0070] The superheat degree recognition model uses a deep convolutional neural network in the I3D network layer to extract features from the samples in the input Huoyan video training sample set, obtaining deep abstract features with fewer dimensions. The dimension of the deep abstract features is 2×1024×7×7. To reduce the dimension of the Huoyan deep features and remove some features with relatively low activation levels, in this embodiment, the deep abstract features extracted by the I3D network layer are input into a three-dimensional adaptive max pooling (Adaptive MaxPool) to reduce the dimension of the deep abstract features to 1024×4×4, obtaining dimension-reduced deep abstract features; then, the last two dimensions of the dimension-reduced deep abstract features are merged and transposed to obtain 16 groups of one-dimensional feature vectors with 1024 dimensions. To obtain the relationship between this group of one-dimensional vector sequences and the superheat degree state and obtain the current superheat degree state, this embodiment introduces a CLS Token vector that is independent of the feature vectors. The CLS Token vector is concatenated with the vector sequence to obtain a concatenated vector; after the concatenated vector is position-encoded, it is input into the encoding part and classification part of the subsequent VT layer (i.e., the Vision Transformer layer) to obtain the superheat degree result.

[0071] The VT encoding part in the superheat degree recognition model consists of N stacked VT layers. The structural schematic diagram of each VT layer is referred to Figure 2 Figure. The VT layer includes a self-attention mechanism and an FFN (i.e., a feed-forward neural network) layer. The self-attention mechanism is used to obtain the global correlation between the CLS Token vector and multiple groups of one-dimensional feature vectors, that is, the multiple groups of one-dimensional feature vectors are weighted and aggregated through the self-attention mechanism to obtain a feature vector that is globally correlated between the CLS Token vector and multiple groups of one-dimensional feature vectors; the globally correlated feature vector is input into the FFN layer to enable the feed-forward neural network to perform non-linear projection on the CLS Token vector and multiple groups of one-dimensional feature vectors to obtain an output vector at the position of the CLS Token vector and output vectors at the positions of multiple groups of one-dimensional feature vectors. The FFN layer includes a layer normalization layer with a residual connection and two linear layers, and a ReLU activation function is added between the two linear layers.

[0072] In this embodiment, the residual connection reduces the loss in the feature transmission process and avoids the problem of gradient disappearance; compared with Transformer that places the layer normalization layer after the residual output, this embodiment places the layer normalization layer before the input of the linear layer to avoid the drastic increase of the gradient norm during the training process and improve the stability during the model training process.

[0073] The classification model of the superheat degree recognition model includes two linear layers with activation functions and a Softmax classifier. The input of the classification model is the output vector at the position of the CLS Token vector in the VT encoding part. The CLS Token vector position is a vector independent of the one-dimensional feature vector. In the VT encoding part, the CLS Token vector considers the differences caused by different superheat degree states among the one-dimensional feature vectors and aggregates the one-dimensional feature vectors by weighting. Therefore, the CLS Token vector position contains feature information related to the superheat degree state. The output vector at the CLS Token vector position is input into the classification model of the superheat degree recognition model, so that the classification model performs a non-linear mapping on the output vector at the CLS Token vector position through two linear layers with activation functions to obtain a mapped vector; the mapped vector is classified through the Softmax classifier to obtain the superheat degree situation reflected by the current Huoyan video, that is, the superheat degree recognition result is obtained.

[0074] In this embodiment, in the current superheat degree recognition method based on deep learning, only the static features of the aluminum electrolysis Huoyan video are considered well, and the dynamic features of the aluminum electrolysis Huoyan video are not considered perfectly. In addition to considering the position information of the aluminum electrolysis Huoyan video data in the two-dimensional space dimension, this embodiment also considers the rich motion information in the time dimension and comprehensively considers the spatio-temporal correlation of the deep features of the aluminum electrolysis Huoyan video. Therefore, this embodiment can better train the superheat degree recognition model, prevent the overfitting of the superheat degree recognition model, and improve the generalization ability of the superheat degree recognition model.

[0075] Step S300: Based on transfer learning with parameter sharing, initialize the weights of the parameters in the shallow convolutional neural network in the I3D network layer to obtain the initialized weight parameters.

[0076] Specifically, use the video classification model trained on the Kinetics dataset as the pre-trained model; transfer the weight parameters of the pre-trained model to the I3D network layer to initialize the weights of the parameters in the shallow convolutional neural network in the I3D network layer to obtain the initialized weight parameters.

[0077] The Kinetics dataset is a large video recognition dataset, which includes about 240,000 training videos in 400 categories. Generally, it is considered that the shallow convolutional neural network extracts the texture features of the input, and the deep convolutional neural network extracts the abstract features of the input. The texture feature extraction part has a certain generality in different tasks.

[0078] In this embodiment, considering that the number of parameters of the I3D network is too large, training the I3D network starting from random weight initialization requires a large amount of training data. Otherwise, it will cause model overfitting and affect the generalization performance of the model. Therefore, this embodiment introduces transfer learning with parameter sharing in the shallow convolutional neural network, and uses the weight parameters of the pre-trained model as the parameters of the shallow convolutional neural network in the overheat recognition model for weight initialization, thereby preventing the overheat recognition model from overfitting and improving the generalization ability of the overheat recognition model.

[0079] Step S400: According to the initialized weight parameters, train all the parameters in the overheat recognition model through the Huoyan video training sample set to obtain a trained overheat recognition model.

[0080] Specifically, according to the initialized weight parameters, extract texture features through the shallow convolutional neural network to obtain the texture features extracted after the initialization of the weights; based on the texture features extracted after the initialization of the weights, train the abstract features extracted by the deep convolutional neural network to obtain the trained abstract features; according to the trained abstract features, train all the parameters in the overheat recognition model through the Huoyan video training sample set to obtain a trained overheat recognition model.

[0081] It should be noted that in this embodiment, the deep convolutional neural network is used as a feature extractor to extract features from the samples in the Huoyan video training sample set.

[0082] In this embodiment, the parameters of the deep convolutional neural network are trained through the initialized weight parameters of the shallow convolutional neural network to obtain the trained abstract features. Therefore, this embodiment can prevent the overheat recognition model from overfitting and improve the generalization ability of the overheat recognition model.

[0083] Step S500: Use the trained overheat recognition model to perform overheat recognition on the sample to be predicted to obtain an overheat recognition result.

[0084] Specifically, for the newly obtained Huoyan video data in the industry, use it as the sample to be predicted, and use the trained overheat recognition model to perform overheat recognition on the sample to be predicted to obtain an overheat recognition result. This embodiment can improve the accuracy of overheat recognition and the generalization performance of the model.

[0085] Refer to Figure 3 , this embodiment of the present invention also provides an overheat recognition system, which includes:

[0086] A data acquisition unit 100, configured to acquire a Huoyan video training sample set;

[0087] A model construction unit 200, configured to construct a superheat degree recognition model based on an I3D network layer and a Vision Transformer layer;

[0088] A transfer learning unit 300, configured to perform weight initialization on the parameters in the shallow convolutional neural network in the I3D network layer based on transfer learning with parameter sharing, to obtain initialized weight parameters;

[0089] A model training unit 400, configured to train all the parameters in the superheat degree recognition model according to the initialized weight parameters through a Huoyan video training sample set, to obtain a trained superheat degree recognition model;

[0090] A superheat degree recognition unit 500, configured to perform superheat degree recognition on a sample to be predicted through the trained superheat degree recognition model, to obtain a superheat degree recognition result.

[0091] It should be noted that since a superheat degree recognition system in this embodiment and the above-mentioned superheat degree recognition method are based on the same inventive concept, accordingly, the corresponding content in the method embodiment is equally applicable to this system embodiment, and will not be elaborated herein.

[0092] An embodiment of the present invention further provides a superheat degree recognition device, including: at least one control processor and a memory communicatively connected to the at least one control processor.

[0093] The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory may include a high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor through a network. Examples of the above network include but are not limited to the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0094] The non-transitory software programs and instructions required to implement the superheat degree recognition method in the above embodiment are stored in the memory, and when executed by the processor, execute the superheat degree recognition method in the above embodiment. For example, execute the method steps S100 to step S500 described above. Figure 1 in.

[0095] The system embodiments described above are merely illustrative, where the units described as separate components may or may not be physically separated, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.

[0096] An embodiment of the present invention further provides a computer-readable storage medium storing computer-executable instructions, which are executed by one or more control processors, enabling the one or more control processors to execute a superheat degree recognition method in the above method embodiment. For example, execute the functions of method steps S100 to S500 described above. Figure 1

[0097] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform. Those skilled in the art can understand that all or part of the processes of implementing the above embodiment methods can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above method embodiments. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0098] The above has described the embodiments of the present invention in detail with reference to the drawings, but the present invention is not limited to the above embodiments. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those of ordinary skill in the art.​

Claims

1. A superheat degree recognition method, characterized in that The method includes: Obtain a Huoyan video training sample set; Based on the I3D network layer and the Vision Transformer layer, construct an overheat degree recognition model. Specifically, Extract features from the samples in the Huoyan video training sample set through the deep convolutional neural network in the I3D network layer to obtain deep abstract features; Adopt three-dimensional adaptive max pooling to reduce the dimension of the deep abstract features to obtain dimension-reduced deep abstract features; Merge and transpose the last two dimensions of the dimension-reduced deep abstract features to obtain multiple groups of one-dimensional feature vectors; Concatenate the CLS Token vector with the multiple groups of one-dimensional feature vectors to obtain a concatenated vector; Input the concatenated vector into the Vision Transformer layer to obtain an output vector at the position of the CLS Token vector and output vectors at the positions of multiple groups of one-dimensional feature vectors; wherein, the Vision Transformer layer includes a self-attention mechanism and a feed-forward neural network; Input the output vector at the position of the CLS Token vector into the classification module to obtain an overheat degree recognition result, and the classification module includes two linear layers with activation functions and a Softmax classifier; Based on transfer learning with parameter sharing, initialize the weights of the parameters in the shallow convolutional neural network in the I3D network layer to obtain initialized weight parameters; According to the initialized weight parameters, train all the parameters in the overheat degree recognition model through the Huoyan video training sample set to obtain a trained overheat degree recognition model; Perform overheat degree recognition on the sample to be predicted through the trained overheat degree recognition model to obtain an overheat degree recognition result.

2. The superheat degree identification method according to claim 1, characterized in that The obtaining of the Huoyan video training sample set includes: Collect Huoyan video samples of an aluminum electrolysis cell, and obtain the Huoyan center corresponding to the Huoyan video samples of the aluminum electrolysis cell through the YOLO model; Perform sample amplification on the Huoyan video samples of the aluminum electrolysis cell through the Mixup data augmentation method to obtain a Huoyan video training sample set.

3. The superheat degree identification method according to claim 1, characterized in that The inputting of the concatenated vector into the Vision Transformer layer to obtain an output vector at the position of the CLS Token vector and output vectors at the positions of multiple groups of one-dimensional feature vectors includes: Input the concatenated vector into the Vision Transformer layer, so that the Vision Transformer layer performs weighted aggregation on the multiple groups of one-dimensional feature vectors through the self-attention mechanism to obtain a feature vector globally related between the CLS Token vector and the multiple groups of one-dimensional feature vectors; Input the globally related feature vector into the feed-forward neural network, so that the feed-forward neural network performs non-linear projection on the CLS Token vector and the multiple groups of one-dimensional feature vectors to obtain an output vector at the position of the CLS Token vector and output vectors at the positions of multiple groups of one-dimensional feature vectors.

4. The superheat degree recognition method according to claim 3, wherein The inputting of the output vector at the position of the CLS Token vector into the classification module to obtain an overheat degree recognition result includes: Input the output vector at the position of the CLS Token vector into the classification module, so that the classification module performs non-linear mapping through the two layers of linear layers with activation functions to obtain a mapped vector; Classify the mapped vector through a Softmax classifier to obtain an overheat degree recognition result.

5. The superheat degree identification method according to claim 4, wherein For the transfer learning based on parameter sharing, perform weight initialization on the parameters in the shallow convolutional neural network in the I3D network layer to obtain initialized weight parameters, including: Use the training video classification model of the Kinetics dataset as a pre-trained model; Transfer the weight parameters of the pre-trained model to the I3D network layer, and perform weight initialization on the parameters of the shallow convolutional neural network in the I3D network layer to obtain initialized weight parameters.

6. The superheat degree identification method according to claim 5, characterized in that, For training all the parameters in the overheat degree recognition model according to the initialized weight parameters through the Huoyan video training sample set to obtain a trained overheat degree recognition model, including: Extract texture features through the shallow convolutional neural network according to the initialized weight parameters to obtain the texture features extracted after the initialized weights; Based on the texture features extracted after the initialized weights, train the abstract features extracted by the deep convolutional neural network to obtain the trained abstract features; According to the trained abstract features, train all the parameters in the overheat degree recognition model through the Huoyan video training sample set to obtain a trained overheat degree recognition model.

7. An overheat degree recognition system, characterized in that The system includes: A data acquisition unit for acquiring a Huoyan video training sample set; A model construction unit for constructing an overheat degree recognition model based on the I3D network layer and the Vison Transformer layer. Specifically, Extract features from the samples in the Huoyan video training sample set through the deep convolutional neural network in the I3D network layer to obtain deep abstract features; Use three-dimensional adaptive max pooling to reduce the dimension of the deep abstract features to obtain dimension-reduced deep abstract features; Merge and transpose the last two dimensions of the dimension-reduced deep abstract features to obtain multiple groups of one-dimensional feature vectors; Concatenate the CLS Token vector with the multiple groups of one-dimensional feature vectors to obtain a concatenated vector; Input the concatenated vector into the Vison Transformer layer to obtain an output vector at the position of the CLS Token vector and output vectors at the positions of multiple groups of one-dimensional feature vectors; where the Vison Transformer layer includes a self-attention mechanism and a feed-forward neural network; Input the output vector at the position of the CLS Token vector into the classification module to obtain an overheat degree recognition result, and the classification module includes two layers of linear layers with activation functions and a Softmax classifier; A transfer learning unit for performing weight initialization on the parameters in the shallow convolutional neural network in the I3D network layer based on transfer learning with parameter sharing to obtain initialized weight parameters; A model training unit, configured to train all parameters in the superheat degree recognition model according to the initialized weight parameters through the Huoyan video training sample set, and obtain a trained superheat degree recognition model; A superheat degree recognition unit, configured to perform superheat degree recognition on a sample to be predicted through the trained superheat degree recognition model, and obtain a superheat degree recognition result.

8. An overheat degree recognition device, characterized in that, It includes at least one control processor and a memory for communicatively connecting with the at least one control processor; the memory stores instructions executable by the at least one control processor, and the instructions are executed by the at least one control processor so that the at least one control processor can execute the superheat degree recognition method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to cause a computer to execute the superheat degree recognition method according to any one of claims 1 to 6.

Images