Production control system and method for polycrystalline mullite fiber
Through deep learning technology, fiber status and disc speed characteristics are extracted from colloidal fiber-forming monitoring videos, and real-time disc speed adjustment is carried out, which solves the problem of disc speed control in the production process of polycrystalline mullite fibers, and improves fiber quality and production efficiency.
Patent Information
- Application Number
- CN202311163680.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-11
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2043-09-11
AI Technical Summary
In the production process of polycrystalline mullite fibers, the rotation speed of the disc is difficult to accurately control, resulting in fiber breakage or the size of the fiber is not up to standard.
Deep learning technology is used to extract fiber state change feature information and disk speed timing dynamic features from colloidal fiber formation monitoring videos, and real-time disk speed adjustment is performed through correlation feature representation to reduce the occurrence of fiber fracture and size failure.
The fiber quality and production efficiency are improved. By adjusting the disc speed in real time, the incidence of fiber breakage and defects in size are reduced.
Smart Images

Figure CN117270467B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of intelligent control technology, and more specifically, to a production control system and method for polycrystalline mullite fiber. Background Art
[0002] Polycrystalline mullite fiber is the latest high-temperature resistant insulation material in the world. It combines the characteristics of fiber materials and crystal materials. It has the characteristics of high temperature resistance, creep resistance, thermochemical stability, low thermal expansion coefficient, low thermal conductivity and high modulus. It is often used in the preparation of high-temperature insulation and refractory materials. Polycrystalline mullite fiber is usually made by chemical colloid method. It is made of high-purity alumina, silica and gelling additives in a certain proportion to prepare a colloid. Then, under the action of high-speed rotating centrifugal force, the colloidal solution is pulled out into fibers, and then the crystal transformation is completed after low-temperature drying and high-temperature heat treatment.
[0003] The fiber-forming process is the key to the production of polycrystalline mullite fiber, and it is also a link with greater technical and operational difficulty. First, the prepared colloid is placed in a pressure vessel and pneumatically transported to a high-speed rotating disc. The colloid is centrifugally ejected and then blown into fibers by hot air. The principle is that centrifugal force stretches the fiber and causes the fiber to form a finer structure. However, too high a disc speed may cause fiber breakage or uneven fiber formation, while too low a disc speed will make the fiber size not meet the requirements.
[0004] Therefore, a production control system and method for polycrystalline mullite fibers are desired. Summary of the invention
[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a production control system and method for polycrystalline mullite fiber, which uses deep learning technology to extract the state change characteristic information of the fiber from the colloid fiberization monitoring video, and extract the time series dynamic characteristics of the disk speed from the disk speed at multiple time points, and use the correlation characteristics of the two to perform real-time control of the disk speed at the current time point. In this way, the disk speed can be adjusted in real time and accurately based on the fiber state to reduce the occurrence of defects such as fiber breakage and substandard size, thereby improving the quality and production efficiency of the fiber.
[0006] Accordingly, according to one aspect of the present application, a production control system for polycrystalline mullite fiber is provided, comprising:
[0007] A data acquisition module, used to obtain the disk rotation speed values at multiple predetermined time points in a predetermined time period during the fiberization process of the colloid and the colloid fiberization monitoring video of the predetermined time period acquired by the camera;
[0008] A key frame extraction module, used to extract multiple fiber state key frames from the colloid fiberization monitoring video;
[0009] A fiber state feature extraction module, configured to pass the plurality of fiber state key frames through a first convolutional neural network model including a salient object detector to obtain a plurality of fiber state feature vectors;
[0010] A transfer calculation module, used for calculating the transfer matrix between the fiber state feature vectors of each two adjacent time points in the plurality of fiber state feature vectors to obtain a plurality of transfer matrices;
[0011] A fiber state temporal change feature extraction module, configured to obtain a fiber state temporal change feature vector by arranging the multiple transfer matrices into a three-dimensional input tensor through a second convolutional neural network model using a channel attention mechanism;
[0012] A disk speed time series encoding module, used for arranging the disk speed values at the plurality of predetermined time points into a speed input vector according to the time dimension and then passing the vector through a time series encoder including a fully connected layer and a one-dimensional convolutional layer to obtain a speed time series feature vector;
[0013] An association coding module, used for performing association coding on the fiber state time series change feature vector and the rotation speed time series feature vector to obtain a classification feature matrix;
[0014] The control result generation module is used to pass the classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the rotation speed of the disk at the current time point should be increased or decreased.
[0015] In the above-mentioned polycrystalline mullite fiber production control system, the key frame extraction module is used to extract a plurality of fiber state key frames from the colloidal fiberization monitoring video at a predetermined sampling frequency.
[0016] In the above-mentioned polycrystalline mullite fiber production control system, the fiber state feature extraction module is used to: use each layer of the first convolutional neural network model containing the salient target detector to respectively perform the following on the input data in the forward pass of the layer: use the first convolution kernel to convolve the input data to obtain a first convolutional feature map; use the second convolution kernel to convolve the first convolutional feature map to obtain a second convolutional feature map, wherein the size of the first convolution kernel is larger than the size of the second convolution kernel; perform global pooling processing on each feature matrix based on the channel dimension on the second convolutional feature map to obtain a pooled feature map; perform activation processing on the pooled feature map to obtain an activated feature map; wherein the output of the last layer of the first convolutional neural network model is the fiber state feature vector, and the input of the first layer of the first convolutional neural network model is the fiber state key frame.
[0017] In the above-mentioned polycrystalline mullite fiber production control system, the transfer calculation module is used to: calculate the transfer matrix between the fiber state feature vectors at every two adjacent time points in the multiple fiber state feature vectors using the following transfer formula to obtain the multiple transfer matrices;
[0018] Wherein, the transfer formula is:
[0019]
[0020] Wherein V1 and V2 represent the fiber state feature vectors at two adjacent time points among the plurality of fiber state feature vectors, represents matrix multiplication, and M represents the transfer matrix.
[0021] In the above-mentioned polycrystalline mullite fiber production control system, the fiber state time series change feature extraction module is used to: use each layer of the second convolutional neural network model to perform on the input data in the forward transfer of the layer: perform convolution processing based on a three-dimensional convolution kernel on the input data to obtain a convolution feature map; calculate the global mean of each feature matrix along the channel dimension of the convolution feature map to obtain a channel feature vector; calculate the ratio of the eigenvalue of each position in the channel feature vector to the weighted sum of the eigenvalues of all positions of the channel feature vector to obtain a channel weighted feature vector; use the eigenvalue of each position of the channel weighted feature vector as a weight to perform point multiplication on the feature matrix along the channel dimension of the convolution feature map to obtain a channel attention map; perform global mean pooling processing on the channel attention map based on each feature matrix of the channel dimension to obtain a pooled feature map; perform activation processing on the pooled feature map to obtain an activated feature map; wherein the output of the last layer of the second convolutional neural network model is the fiber state time series change feature vector, the input from the second layer to the last layer of the second convolutional neural network is the output of the previous layer, and the input of the second convolutional neural network is the three-dimensional input tensor.
[0022] In the above-mentioned polycrystalline mullite fiber production control system, the disk speed timing encoding module includes: a normalization mapping unit, which is used to perform normalization processing on the speed input vector based on the maximum value to obtain a normalized speed input vector; a fully connected encoding unit, which is used to use the fully connected layer of the timing encoder to perform fully connected encoding on the normalized speed input vector to extract high-dimensional implicit features of the eigenvalues at each position in the normalized speed input vector; a one-dimensional convolution encoding unit, which is used to use the one-dimensional convolution layer of the timing encoder to perform one-dimensional encoding on the normalized speed input vector to extract high-dimensional implicit correlation features of the correlation between the eigenvalues at each position in the normalized speed input vector.
[0023] In the above-mentioned polycrystalline mullite fiber production control system, the associated coding module includes: a first affine density calculation unit, which is used to calculate the first affine density value of the fiber state time series change feature vector relative to the speed time series feature vector using the following first affine density calculation formula, wherein the first affine density calculation formula is:
[0024]
[0025] Wherein, x1 represents the fiber state time series change characteristic vector, x2 represents the speed time series characteristic vector, σ1 represents the width parameter of the first Gaussian kernel function, ||·|| represents the norm function, exp represents the exponential function with e as the base, and d1 represents the first affine density value; the second affine density calculation unit is used to calculate the second affine density value of the speed time series characteristic vector relative to the fiber state time series change characteristic vector using the following second affine density calculation formula, wherein the second affine density calculation formula is:
[0026]
[0027] Among them, x1 represents the fiber state temporal change characteristic vector, x2 represents the speed temporal characteristic vector, σ2 represents the width parameter of the second Gaussian kernel function, ||·|| represents the norm function, exp represents the exponential function with base e, and d2 represents the second affine density value; a weighting unit is used to weight the fiber state temporal change characteristic vector and the speed temporal characteristic vector respectively with the first affine density value and the second affine density value as weights to obtain the weighted fiber state temporal change characteristic vector and the weighted speed temporal characteristic vector; an association coding unit is used to perform association coding on the weighted fiber state temporal change characteristic vector and the weighted speed temporal characteristic vector to obtain the classification feature matrix.
[0028] According to another aspect of the present application, a production control method for polycrystalline mullite fiber is provided, comprising:
[0029] Obtaining the disk rotation speed values at multiple predetermined time points within a predetermined time period during the fiberization process of the colloid and the colloid fiberization monitoring video of the predetermined time period collected by the camera;
[0030] Extracting a plurality of fiber state key frames from the colloid fiberization monitoring video;
[0031] Passing the plurality of fiber state key frames through a first convolutional neural network model including a salient object detector to obtain a plurality of fiber state feature vectors;
[0032] Calculating a transfer matrix between the fiber state feature vectors at every two adjacent time points in the plurality of fiber state feature vectors to obtain a plurality of transfer matrices;
[0033] Arranging the multiple transfer matrices into a three-dimensional input tensor and then using a second convolutional neural network model with a channel attention mechanism to obtain a fiber state temporal change feature vector;
[0034] Arranging the disk speed values at the plurality of predetermined time points into a speed input vector according to the time dimension and then passing the vector through a temporal encoder including a fully connected layer and a one-dimensional convolutional layer to obtain a speed temporal feature vector;
[0035] Performing association coding on the fiber state time series change feature vector and the rotation speed time series feature vector to obtain a classification feature matrix;
[0036] The classification feature matrix is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the rotation speed of the disk at the current time point should be increased or decreased.
[0037] In the above-mentioned production control method of polycrystalline mullite fiber, extracting a plurality of fiber state key frames from the colloid fiberization monitoring video includes: extracting a plurality of fiber state key frames from the colloid fiberization monitoring video at a predetermined sampling frequency.
[0038] In the above-mentioned production control method of polycrystalline mullite fiber, the multiple fiber state key frames are respectively passed through a first convolutional neural network model including a significant target detector to obtain multiple fiber state feature vectors, including: using each layer of the first convolutional neural network model including the significant target detector to respectively perform the following on the input data in the forward pass of the layer: using a first convolution kernel to perform convolution processing on the input data to obtain a first convolution feature map; using a second convolution kernel to perform convolution processing on the first convolution feature map to obtain a second convolution feature map, wherein the size of the first convolution kernel is larger than the size of the second convolution kernel; performing global pooling processing on each feature matrix based on the channel dimension on the second convolution feature map to obtain a pooled feature map; performing activation processing on the pooled feature map to obtain an activated feature map; wherein the output of the last layer of the first convolutional neural network model is the fiber state feature vector, and the input of the first layer of the first convolutional neural network model is the fiber state key frame.
[0039] Compared with the prior art, the polycrystalline mullite fiber production control system and method provided by the present application adopts deep learning technology to extract the state change characteristic information of the fiber from the colloid fiberization monitoring video, and extract the time series dynamic characteristics of the disk speed from the disk speed at multiple time points, and use the correlation characteristics of the two to perform real-time control of the disk speed at the current time point. In this way, the disk speed can be adjusted in real time and accurately based on the fiber state to reduce the occurrence of defects such as fiber breakage and substandard size, and improve the quality and production efficiency of the fiber. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.
[0041] Figure 1 It is a block diagram of a production control system for polycrystalline mullite fiber according to an embodiment of the present application.
[0042] Figure 2 Schematic diagram of the structure of a production control system for polycrystalline mullite fiber according to an embodiment of the present application.
[0043] Figure 3 It is a block diagram of a disc speed timing encoding module in a production control system of polycrystalline mullite fiber according to an embodiment of the present application.
[0044] Figure 4 The figure is a flow chart of a method for controlling the production of polycrystalline mullite fibers according to an embodiment of the present application. DETAILED DESCRIPTION
[0045] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.
[0046] Application Overview
[0047] As mentioned above, the fiber-forming process is the key to the production of polycrystalline mullite fibers. In this link, the prepared colloid is first placed in a pressure vessel and pneumatically transported to a high-speed rotating disc. The colloid is centrifugally ejected and then blown into fibers by hot air. The principle is that centrifugal force stretches the fiber and causes the fiber to form a finer structure. However, too high a disc speed may cause the fiber to break or form uneven fibers. However, too low a disc speed will make the size of the obtained fiber fail to meet the requirements. Therefore, a production control system and method for polycrystalline mullite fibers are expected to be able to intelligently control the disc speed during the fiber-forming process of polycrystalline mullite fibers.
[0048] At present, deep learning and neural networks have been widely used in computer vision, natural language processing, speech signal processing and other fields. In addition, deep learning and neural networks have also shown a level close to or even beyond that of humans in image classification, object detection, semantic segmentation, text translation and other fields.
[0049] In recent years, the development of deep learning and neural networks has provided new solutions and plans for the intelligent control of disk speed in the fiber-forming process of polycrystalline mullite fibers.
[0050] Accordingly, when the disk speed is actually controlled intelligently during the fiberization process of polycrystalline mullite fiber, the disk speed should be adaptively adjusted according to the state change of the fiber during the colloid fiberization process. In other words, since different disk speeds lead to different fiber states, it is necessary to dig out the correlation characteristic distribution information between the fiber state change characteristics and the disk speed time series dynamic characteristics, so as to control the disk speed.
[0051] Specifically, in the technical solution of the present application, first, the rotation speed values of the disc at multiple predetermined time points within a predetermined time period in the fiber-forming process of the colloid are obtained by a sensor, and the colloid fiber-forming monitoring video of the predetermined time period is collected by a camera. Next, considering that the state change characteristics of the fiber can be represented by the difference information between adjacent image frames in the colloid fiber-forming monitoring video, but since the adjacent image frames in the colloid fiber-forming monitoring video are almost the same, the difference characteristics are too small. Therefore, if the feature extraction of the colloid fiber-forming monitoring video is directly performed, not only the amount of calculation is too large, but also it will interfere with the extraction of the state change characteristics of the fiber. Based on this, the colloid fiber-forming monitoring video is further sampled and processed at a predetermined sampling frequency to extract multiple fiber state key frames from the colloid fiber-forming monitoring video.
[0052] Then, the plurality of fiber state key frames are subjected to feature mining respectively through a first convolutional neural network model including a significant target detector to obtain a plurality of fiber state feature vectors. It should be understood that a significant target detector is a technology for detecting significant objects or areas in an image or video. In the technical solution of the present application, the detector is used to detect and extract significant fiber structures in fiber state key frames. By inputting the fiber state key frames into a convolutional neural network model including a significant target detector, it is possible to learn to identify and capture key features of the fiber structure, such as fiber refinement, uniformity, and other properties related to the fiber state, providing useful input for subsequent fiber quality monitoring and disc speed control.
[0053] Next, the transfer matrix of each two adjacent time points in the multiple fiber state feature vectors is calculated to reveal the evolution and transfer rules of the fiber state over time. It should be understood that the transfer matrix can describe the probability distribution of the fiber state from one time point to the next time point. By calculating the transfer matrix between the multiple fiber state feature vectors, the correlation degree and transfer probability between different fiber states can be understood, providing more detailed information and basis for fiber quality monitoring and speed control.
[0054] Furthermore, the multiple transfer matrices are arranged as a three-dimensional input tensor, and feature extraction is performed by using a second convolutional neural network model with a channel attention mechanism to obtain a fiber state temporal change feature vector. By arranging the multiple transfer matrices as a three-dimensional input tensor, the time dimension is taken into consideration, so that the model can simultaneously focus on the transfer matrix information at multiple time points. The channel attention mechanism can automatically learn and emphasize the key information in different channels in the input tensor. In other words, by using the second convolutional neural network model with a channel attention mechanism, in the process of extracting the fiber state temporal change features, the channels related to the fiber state change can be automatically selected and emphasized, and the interference of irrelevant information can be filtered out, providing more detailed and fine-grained timing information for fiber quality monitoring and speed control.
[0055] At the same time, the disk speed values at the multiple predetermined time points are arranged as a speed input vector according to the time dimension, and processed by a time series encoder including a fully connected layer and a one-dimensional convolutional layer to capture the changing trend and pattern of the speed over time, thereby obtaining a speed time series feature vector. It should be understood that the time series encoder is a neural network structure for processing time series data, which can extract stable feature representations from time series data. Among them, the fully connected layer can learn the nonlinear relationship between the speed data, and perform dimensional expansion and information fusion on the input to provide a richer and more abstract representation. The one-dimensional convolutional layer can capture the local pattern and time series correlation of the speed by performing a convolution operation on the speed sequence.
[0056] In particular, considering that the fiber state time series change feature vector and the speed time series feature vector may contain a lot of common information. In this case, directly transposing and multiplying them for association coding may lead to redundant feature representation and increase the complexity of the model. At the same time, through association coding, the model needs to process the information of two time series feature vectors at the same time, which will increase the number of parameters and computational complexity of the model. If there is a strong correlation between the two time series feature vectors, directly transposing and multiplying them may not bring obvious benefits, but will increase the computational burden of the model. In order to avoid data redundancy and increased complexity, the fiber state time series change feature vector and the speed time series feature vector are associated based on the affine density domain of the feature space. Here, the affine density domain association of the feature space is a method for dimensionality reduction and feature extraction in machine learning, which is based on mapping high-dimensional data to low-dimensional space while maintaining the local density distribution of the data. The technical effect of this method is that it can effectively process nonlinear and high-dimensional data, improve data visualization and classification performance, and reduce computational complexity and memory consumption.
[0057] Specifically, the fiber state time series change feature vector and the speed time series feature vector are associated and encoded to obtain a classification feature matrix, including: calculating the first affine density value of the fiber state time series change feature vector relative to the speed time series feature vector by the following formula, wherein the formula is
[0058]
[0059] Wherein, x1 and x2 represent the fiber state time series change characteristic vector and the speed time series characteristic vector respectively, and σ1 represents the width parameter of the first Gaussian kernel function; the second affine density value of the speed time series characteristic vector relative to the fiber state time series change characteristic vector is calculated by the following formula, wherein the formula is:
[0060]
[0061] Among them, x1 and x2 represent the fiber state temporal change feature vector and the rotational speed temporal feature vector respectively, and σ2 represents the width parameter of the second Gaussian kernel function; the fiber state temporal change feature vector and the rotational speed temporal feature vector are weighted respectively with the first affine density value and the second affine density value as weights to obtain the weighted fiber state temporal change feature vector and the weighted rotational speed temporal feature vector; the weighted fiber state temporal change feature vector and the weighted rotational speed temporal feature vector are associated encoded to obtain the classification feature matrix.
[0062] In the technical solution of the present application, the first affine density value and the second affine density value between the fiber state timing change feature vector and the rotational speed timing feature vector are calculated to represent the relative quantitative characterization comparison value between the feature distribution density domains of the fiber state timing change feature vector and the rotational speed timing feature vector after they are respectively mapped to low-dimensional affine subspaces. In this way, after the fiber state timing change feature vector and the rotational speed timing feature vector are respectively mapped to low-dimensional affine subspaces, the first affine density value and the second affine density value can be used to perform collaborative constraints on the feature distribution of the fiber state timing change feature vector and the rotational speed timing feature vector based on the feature distribution density domain, so that the classification feature matrix can be effectively associated and approximated with the fiber state timing change feature vector and the rotational speed timing feature vector at the density domain level. In this way, the relative distance and density information between the data points of the fiber state timing change feature vector and the rotational speed timing feature vector are retained, that is, the intrinsic structure and distribution of the data are maintained, the main features of the data are effectively extracted, the complexity and redundancy of the data are reduced, and the classification and clustering performance of the data are improved.
[0063] Finally, the classification feature matrix is passed through a classifier to obtain a classification result indicating whether the disc speed should be increased, kept constant, or decreased at the current time point. In this way, a corresponding control strategy can be adopted based on the classification result to adjust and optimize the fiber quality, thereby meeting the process requirements and improving the fiber quality.
[0064] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.
[0065] Exemplary Systems
[0066] Figure 1 FIG. 1 is a block diagram of a production control system for polycrystalline mullite fiber according to an embodiment of the present application. Figure 1As shown, according to the polycrystalline mullite fiber production control system 100 of the embodiment of the present application, it includes: a data acquisition module 110, which is used to obtain the disk rotation speed value of a plurality of predetermined time points within a predetermined time period in the fiber-forming process of the colloid and the colloid fiber-forming monitoring video of the predetermined time period collected by the camera; a key frame extraction module 120, which is used to extract a plurality of fiber state key frames from the colloid fiber-forming monitoring video; a fiber state feature extraction module 130, which is used to respectively pass the plurality of fiber state key frames through a first convolutional neural network model including a significant target detector to obtain a plurality of fiber state feature vectors; a transfer calculation module 140, which is used to calculate the transfer matrix between the fiber state feature vectors of each two adjacent time points in the plurality of fiber state feature vectors to obtain a plurality of transfer matrices; a fiber state time matrix ... A sequential change feature extraction module 150 is used to arrange the multiple transfer matrices into a three-dimensional input tensor and then obtain a fiber state temporal change feature vector by using a second convolutional neural network model with a channel attention mechanism; a disk speed temporal encoding module 160 is used to arrange the disk speed values of the multiple predetermined time points into a speed input vector according to the time dimension and then obtain a speed temporal feature vector by a temporal encoder including a fully connected layer and a one-dimensional convolutional layer; an association encoding module 170 is used to perform association encoding on the fiber state temporal change feature vector and the speed temporal feature vector to obtain a classification feature matrix; a control result generation module 180 is used to pass the classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the disk speed at the current time point should increase or decrease.
[0067] Figure 2 FIG. 1 is a schematic diagram of the structure of a production control system for polycrystalline mullite fiber according to an embodiment of the present application. Figure 2As shown, first, the disk speed values at multiple predetermined time points within a predetermined time period in the fiber-forming process of the colloid and the colloid fiber-forming monitoring video of the predetermined time period collected by the camera are obtained. Then, multiple fiber state key frames are extracted from the colloid fiber-forming monitoring video. Then, the multiple fiber state key frames are respectively passed through a first convolutional neural network model containing a significant target detector to obtain multiple fiber state feature vectors. Secondly, the transfer matrix between the fiber state feature vectors of each two adjacent time points in the multiple fiber state feature vectors is calculated to obtain multiple transfer matrices. Then, the multiple transfer matrices are arranged as a three-dimensional input tensor and then passed through a second convolutional neural network model using a channel attention mechanism to obtain a fiber state time series change feature vector. At the same time, the disk speed values at the multiple predetermined time points are arranged as a speed input vector according to the time dimension and then passed through a time series encoder containing a fully connected layer and a one-dimensional convolutional layer to obtain a speed time series feature vector. Then, the fiber state time series change feature vector and the speed time series feature vector are associated and encoded to obtain a classification feature matrix. Finally, the classification feature matrix is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the rotation speed of the disk at the current time point should be increased or decreased.
[0068] In the production control system 100 of the polycrystalline mullite fiber, the data acquisition module 110 is used to obtain the disc speed value of the colloid at multiple predetermined time points in a predetermined time period during the fiber-forming process and the colloid fiber-forming monitoring video of the predetermined time period collected by the camera. As mentioned in the above background technology, the fiber-forming process is the key to the production of polycrystalline mullite fiber. The principle of the fiber-forming process is to use centrifugal force to stretch the fiber and prompt the fiber to form a thinner structure. However, too high a disc speed may cause the fiber to break or form uneven fibers, but too low a disc speed will make the obtained fiber size fail to meet the requirements. Therefore, it is expected that the disc speed in the fiber-forming process of polycrystalline mullite fiber can be intelligently controlled.
[0069] Accordingly, when the disc speed is actually intelligently controlled during the fiberization process of polycrystalline mullite fibers, the disc speed should be adaptively adjusted according to the change in the state of the fiber during the fiberization process of the colloid. In other words, since different disc speeds lead to different states of the fiber, it is necessary to dig out the characteristic distribution information of the correlation between the fiber state change characteristics and the disc speed timing dynamic characteristics, so as to control the disc speed. Based on this, in the technical solution of the present application, first, the disc speed values at multiple predetermined time points within a predetermined time period in the fiberization process of the colloid are obtained by a sensor, and the colloid fiberization monitoring video of the predetermined time period is collected by a camera.
[0070] In the above-mentioned polycrystalline mullite fiber production control system 100, the key frame extraction module 120 is used to extract multiple fiber state key frames from the colloidal fiberization monitoring video. Considering that the state change characteristics of the fiber can be represented by the difference information between adjacent image frames in the colloidal fiberization monitoring video, but since the adjacent image frames in the colloidal fiberization monitoring video are almost the same, the difference characteristics are too small. Therefore, if the colloidal fiberization monitoring video is directly subjected to feature extraction, not only the amount of calculation is too large, but also it will interfere with the state change feature extraction of the fiber. Based on this, the colloidal fiberization monitoring video is further sampled and processed at a predetermined sampling frequency to extract multiple fiber state key frames from the colloidal fiberization monitoring video.
[0071] In the above-mentioned polycrystalline mullite fiber production control system 100, the fiber state feature extraction module 130 is used to pass the multiple fiber state key frames through the first convolutional neural network model including a significant target detector to obtain multiple fiber state feature vectors. It should be understood that the significant target detector is a technology for detecting significant objects or areas in an image or video. In the technical solution of the present application, the detector is used to detect and extract significant fiber structures in the fiber state key frames. By inputting the fiber state key frames into the convolutional neural network model including the significant target detector, it is possible to learn to identify and capture the key features of the fiber structure, such as the degree of fiber refinement, uniformity and other properties related to the fiber state, and provide useful input for subsequent fiber quality monitoring and disc speed control.
[0072] Correspondingly, in a specific example, the fiber state feature extraction module 130 is used to: use each layer of the first convolutional neural network model containing the salient target detector to respectively perform the following on the input data in the forward pass of the layer: use a first convolution kernel to convolve the input data to obtain a first convolutional feature map; use a second convolution kernel to convolve the first convolutional feature map to obtain a second convolutional feature map, wherein the size of the first convolutional kernel is larger than the size of the second convolutional kernel; perform global pooling processing on each feature matrix based on the channel dimension on the second convolutional feature map to obtain a pooled feature map; perform activation processing on the pooled feature map to obtain an activated feature map; wherein the output of the last layer of the first convolutional neural network model is the fiber state feature vector, and the input of the first layer of the first convolutional neural network model is the fiber state key frame.
[0073] In the above-mentioned polycrystalline mullite fiber production control system 100, the transfer calculation module 140 is used to calculate the transfer matrix between the fiber state feature vectors of each two adjacent time points in the multiple fiber state feature vectors to obtain multiple transfer matrices. It should be understood that the transfer matrix can describe the probability distribution of the fiber state from one time point to the next time point. By calculating the transfer matrix between the multiple fiber state feature vectors, the evolution and transfer law of the fiber state in time can be revealed, the correlation degree and transfer probability between different fiber states can be understood, and more detailed information and basis can be provided for fiber quality monitoring and speed control.
[0074] Accordingly, in a specific example, the transfer calculation module 140 is used to: calculate the transfer matrix between the fiber state feature vectors of each two adjacent time points in the multiple fiber state feature vectors using the following transfer formula to obtain the multiple transfer matrices;
[0075] Wherein, the transfer formula is:
[0076]
[0077] Wherein V1 and V2 represent the fiber state feature vectors at two adjacent time points among the plurality of fiber state feature vectors, represents matrix multiplication, and M represents the transfer matrix.
[0078] In the above-mentioned polycrystalline mullite fiber production control system 100, the fiber state time series change feature extraction module 150 is used to arrange the multiple transfer matrices into a three-dimensional input tensor and then obtain the fiber state time series change feature vector by using the second convolutional neural network model of the channel attention mechanism. It should be understood that the multiple transfer matrices are arranged as a three-dimensional input tensor to take the time dimension into consideration, so that the model can simultaneously focus on the transfer matrix information at multiple time points. In addition, the channel attention mechanism can automatically learn and emphasize the key information in different channels in the input tensor. In other words, by using the second convolutional neural network model of the channel attention mechanism, in the process of extracting the fiber state time series change features, the channels related to the fiber state change can be automatically selected and emphasized, and the interference of irrelevant information can be filtered out, so as to provide more detailed and fine-grained timing information for fiber quality monitoring and speed control.
[0079] Correspondingly, in a specific example, the fiber state temporal change feature extraction module 150 is used to: use each layer of the second convolutional neural network model to perform the following on the input data in the forward transfer of the layer: perform convolution processing based on a three-dimensional convolution kernel on the input data to obtain a convolution feature map; calculate the global mean of each feature matrix along the channel dimension of the convolution feature map to obtain a channel feature vector; calculate the ratio of the eigenvalue of each position in the channel feature vector to the weighted sum of the eigenvalues of all positions of the channel feature vector to obtain a channel weighted feature vector; perform point multiplication on the feature matrix along the channel dimension of the convolution feature map using the eigenvalue of each position of the channel weighted feature vector as a weight to obtain a channel attention map; perform global mean pooling processing based on each feature matrix in the channel dimension on the channel attention map to obtain a pooled feature map; perform activation processing on the pooled feature map to obtain an activated feature map; wherein the output of the last layer of the second convolutional neural network model is the fiber state temporal change feature vector, the input from the second layer to the last layer of the second convolutional neural network is the output of the previous layer, and the input of the second convolutional neural network is the three-dimensional input tensor.
[0080] In the above-mentioned polycrystalline mullite fiber production control system 100, the disk speed timing encoding module 160 is used to arrange the disk speed values at the multiple predetermined time points into a speed input vector according to the time dimension and then pass it through a timing encoder including a fully connected layer and a one-dimensional convolution layer to obtain a speed time series feature vector. It should be understood that the timing encoder is a neural network structure for processing time series data, which can extract stable feature representations from time series data to capture the trend and pattern of speed changes over time. Among them, the fully connected layer can learn the nonlinear relationship between speed data, and perform dimensional expansion and information fusion on the input to provide a richer and more abstract representation. The one-dimensional convolution layer can capture the local pattern and time series correlation of the speed by performing a convolution operation on the speed sequence.
[0081] Figure 3 FIG. 1 is a block diagram of a disk speed timing encoding module in a production control system for polycrystalline mullite fibers according to an embodiment of the present application. Figure 3As shown, the disk speed time series encoding module 160 includes: a normalization mapping unit 161, which is used to perform normalization processing on the speed input vector based on the maximum value to obtain a normalized speed input vector; a fully connected encoding unit 162, which is used to use the fully connected layer of the time series encoder to perform fully connected encoding on the normalized speed input vector to extract high-dimensional implicit features of the eigenvalues at each position in the normalized speed input vector; a one-dimensional convolution encoding unit 163, which is used to use the one-dimensional convolution layer of the time series encoder to perform one-dimensional encoding on the normalized speed input vector to extract high-dimensional implicit correlation features of the correlation between the eigenvalues at each position in the normalized speed input vector.
[0082] In the above-mentioned polycrystalline mullite fiber production control system 100, the association coding module 170 is used to perform association coding on the fiber state time series change feature vector and the speed time series feature vector to obtain a classification feature matrix. In particular, considering that the fiber state time series change feature vector and the speed time series feature vector may contain a large amount of common information. In this case, directly transposing and multiplying them for association coding may lead to redundant feature representation and increase the complexity of the model. At the same time, through association coding, the model needs to process the information of two time series feature vectors at the same time, which will increase the number of parameters and computational complexity of the model. If there is a strong correlation between the two time series feature vectors, directly transposing and multiplying them may not bring obvious benefits, but will increase the computational burden of the model. In order to avoid data redundancy and increased complexity, the fiber state time series change feature vector and the speed time series feature vector are associated based on the affine density domain of the feature space. Here, the affine density domain association of the feature space is a method for dimensionality reduction and feature extraction in machine learning, which is based on mapping high-dimensional data to low-dimensional space while maintaining the local density distribution of the data. The technical effect of this method is that it can effectively process nonlinear and high-dimensional data, improve data visualization and classification performance, and reduce computational complexity and memory consumption.
[0083] Accordingly, in a specific example, the association coding module 170 includes: a first affine density calculation unit, used to calculate the first affine density value of the fiber state time series change feature vector relative to the speed time series feature vector using the following first affine density calculation formula, wherein the first affine density calculation formula is:
[0084]
[0085] Wherein, x1 represents the fiber state time series change characteristic vector, x2 represents the speed time series characteristic vector, σ1 represents the width parameter of the first Gaussian kernel function, ||·|| represents the norm function, exp represents the exponential function with e as the base, and d1 represents the first affine density value; the second affine density calculation unit is used to calculate the second affine density value of the speed time series characteristic vector relative to the fiber state time series change characteristic vector using the following second affine density calculation formula, wherein the second affine density calculation formula is:
[0086]
[0087] Among them, x1 represents the fiber state temporal change characteristic vector, x2 represents the speed temporal characteristic vector, σ2 represents the width parameter of the second Gaussian kernel function, ||·|| represents the norm function, exp represents the exponential function with base e, and d2 represents the second affine density value; a weighting unit is used to weight the fiber state temporal change characteristic vector and the speed temporal characteristic vector respectively with the first affine density value and the second affine density value as weights to obtain the weighted fiber state temporal change characteristic vector and the weighted speed temporal characteristic vector; an association coding unit is used to perform association coding on the weighted fiber state temporal change characteristic vector and the weighted speed temporal characteristic vector to obtain the classification feature matrix.
[0088] In the technical solution of the present application, the first affine density value and the second affine density value between the fiber state timing change feature vector and the rotational speed timing feature vector are calculated to represent the relative quantitative characterization comparison value between the feature distribution density domains of the fiber state timing change feature vector and the rotational speed timing feature vector after they are respectively mapped to low-dimensional affine subspaces. In this way, after the fiber state timing change feature vector and the rotational speed timing feature vector are respectively mapped to low-dimensional affine subspaces, the first affine density value and the second affine density value can be used to perform collaborative constraints on the feature distribution of the fiber state timing change feature vector and the rotational speed timing feature vector based on the feature distribution density domain, so that the classification feature matrix can be effectively associated and approximated with the fiber state timing change feature vector and the rotational speed timing feature vector at the density domain level. In this way, the relative distance and density information between the data points of the fiber state timing change feature vector and the rotational speed timing feature vector are retained, that is, the intrinsic structure and distribution of the data are maintained, the main features of the data are effectively extracted, the complexity and redundancy of the data are reduced, and the classification and clustering performance of the data are improved.
[0089] In the above-mentioned polycrystalline mullite fiber production control system 100, the control result generation module 180 is used to pass the classification feature matrix through the classifier to obtain a classification result, and the classification result is used to indicate whether the rotation speed of the disk at the current time point should be increased or decreased. In this way, a corresponding control strategy can be adopted based on the classification result to adjust and optimize the fiber quality, thereby meeting the process requirements and improving the fiber quality.
[0090] In summary, the production control system of polycrystalline mullite fiber according to the embodiment of the present application is explained, which uses deep learning technology to extract the state change characteristic information of the fiber from the colloid fiberization monitoring video, and extracts the time series dynamic characteristics of the disk speed from the disk speed at multiple time points, and uses the correlation characteristics of the two to perform real-time control of the disk speed at the current time point. In this way, the disk speed can be adjusted in real time and accurately based on the fiber state to reduce the occurrence of defects such as fiber breakage and substandard size, and improve the quality and production efficiency of the fiber.
[0091] Exemplary Methods
[0092] Figure 4 FIG. 1 is a flow chart of a method for controlling the production of polycrystalline mullite fibers according to an embodiment of the present application. Figure 4 As shown, according to the production control method of polycrystalline mullite fiber of the embodiment of the present application, the method comprises the following steps: S110, obtaining the disk rotation speed values at multiple predetermined time points within a predetermined time period in the fiber-forming process of the colloid and the colloid fiber-forming monitoring video of the predetermined time period collected by the camera; S120, extracting multiple fiber state key frames from the colloid fiber-forming monitoring video; S130, respectively passing the multiple fiber state key frames through a first convolutional neural network model including a significant target detector to obtain multiple fiber state feature vectors; S140, calculating the transfer matrix between the fiber state feature vectors of each two adjacent time points in the multiple fiber state feature vectors to obtain multiple transfer matrices; S141, obtaining multiple fiber state feature vectors; S142, obtaining multiple fiber state feature vectors; S143, obtaining multiple fiber state feature vectors; S144, obtaining multiple fiber state feature vectors; S145, obtaining multiple fiber state feature vectors; S146, obtaining multiple fiber state feature vectors; S147, obtaining multiple fiber state feature vectors; S148, obtaining multiple fiber state feature vectors; S149, obtaining multiple fiber state feature vectors; S150, obtaining multiple fiber state feature vectors; S151, obtaining multiple fiber state feature vectors; S152, obtaining multiple fiber state feature vectors; S153, obtaining multiple fiber state feature vectors; S154, obtaining multiple fiber state feature vectors; S155, obtaining multiple fiber state feature vectors; S156, obtaining multiple fiber state feature vectors; S157, obtaining multiple fiber state feature vectors; S158, obtaining multiple fiber state feature vectors; S159, obtaining multiple fiber state feature vectors; S160, obtaining multiple fiber state feature vectors; S161, obtaining multiple fiber state feature vectors; S162, obtaining multiple fiber state feature vectors; 150, after arranging the multiple transfer matrices into a three-dimensional input tensor, a second convolutional neural network model using a channel attention mechanism is used to obtain a fiber state temporal change feature vector; S160, after arranging the disk speed values at the multiple predetermined time points into a speed input vector according to the time dimension, a temporal encoder including a fully connected layer and a one-dimensional convolutional layer is used to obtain a speed temporal feature vector; S170, associatively encoding the fiber state temporal change feature vector and the speed temporal feature vector to obtain a classification feature matrix; S180, passing the classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the disk speed at the current time point should increase or decrease.
[0093] In a specific example, in the above-mentioned polycrystalline mullite fiber production control method, the step S120 of extracting a plurality of fiber state key frames from the colloid fiberization monitoring video includes: extracting a plurality of fiber state key frames from the colloid fiberization monitoring video at a predetermined sampling frequency.
[0094] In a specific example, in the above-mentioned production control method of polycrystalline mullite fiber, the step S130, respectively passes the multiple fiber state key frames through a first convolutional neural network model including a significant target detector to obtain multiple fiber state feature vectors, including: using each layer of the first convolutional neural network model including the significant target detector to respectively perform the following on the input data in the forward pass of the layer: using a first convolution kernel to perform convolution processing on the input data to obtain a first convolution feature map; using a second convolution kernel to perform convolution processing on the first convolution feature map to obtain a second convolution feature map, wherein the size of the first convolution kernel is larger than the size of the second convolution kernel; performing global pooling processing on each feature matrix based on the channel dimension on the second convolution feature map to obtain a pooled feature map; performing activation processing on the pooled feature map to obtain an activated feature map; wherein the output of the last layer of the first convolutional neural network model is the fiber state feature vector, and the input of the first layer of the first convolutional neural network model is the fiber state key frame.
[0095] In a specific example, in the above-mentioned production control method of polycrystalline mullite fiber, the step S140, calculating the transfer matrix between the fiber state feature vectors at each two adjacent time points in the multiple fiber state feature vectors to obtain multiple transfer matrices, includes: calculating the transfer matrix between the fiber state feature vectors at each two adjacent time points in the multiple fiber state feature vectors using the following transfer formula to obtain the multiple transfer matrices;
[0096] Wherein, the transfer formula is:
[0097]
[0098] Wherein V1 and V2 represent the fiber state feature vectors at two adjacent time points among the plurality of fiber state feature vectors, represents matrix multiplication, and M represents the transfer matrix.
[0099] In a specific example, in the above-mentioned production control method of polycrystalline mullite fiber, the step S150, after arranging the multiple transfer matrices into a three-dimensional input tensor, obtains the fiber state time series change feature vector by using the second convolutional neural network model of the channel attention mechanism, including: using each layer of the second convolutional neural network model to perform on the input data in the forward pass of the layer: performing convolution processing on the input data based on a three-dimensional convolution kernel to obtain a convolution feature map; calculating the global mean of each feature matrix along the channel dimension of the convolution feature map to obtain a channel feature vector; calculating the feature value of each position in the channel feature vector relative to all positions of the channel feature vector. The method comprises the following steps: calculating the ratio of the weighted sum of the eigenvalues to obtain a channel weighted eigenvector; taking the eigenvalues of each position of the channel weighted eigenvector as weights to perform dot multiplication on the feature matrix along the channel dimension of the convolution feature map to obtain a channel attention map; performing global mean pooling processing on the channel attention map based on each feature matrix of the channel dimension to obtain a pooled feature map; performing activation processing on the pooled feature map to obtain an activated feature map; wherein the output of the last layer of the second convolution neural network model is the fiber state temporal change eigenvector, the input from the second layer to the last layer of the second convolution neural network is the output of the previous layer, and the input of the second convolution neural network is the three-dimensional input tensor.
[0100] In a specific example, in the above-mentioned production control method of polycrystalline mullite fiber, the step S160, after arranging the disk speed values at the multiple predetermined time points into a speed input vector according to the time dimension, passes through a timing encoder including a fully connected layer and a one-dimensional convolutional layer to obtain a speed time series feature vector, including: normalizing the speed input vector based on the maximum value to obtain a normalized speed input vector; using the fully connected layer of the timing encoder to fully connect encode the normalized speed input vector to extract high-dimensional implicit features of the feature values at each position in the normalized speed input vector; using the one-dimensional convolutional layer of the timing encoder to one-dimensionally encode the normalized speed input vector to extract high-dimensional implicit correlation features of the correlation between the feature values at each position in the normalized speed input vector.
[0101] In a specific example, in the above-mentioned polycrystalline mullite fiber production control method, the step S170, the fiber state time series change feature vector and the speed time series feature vector are associated and encoded to obtain a classification feature matrix, including: calculating the first affine density value of the fiber state time series change feature vector relative to the speed time series feature vector using the following first affine density calculation formula, wherein the first affine density calculation formula is;
[0102]
[0103] Wherein, x1 represents the fiber state time series change characteristic vector, x2 represents the speed time series characteristic vector, σ1 represents the width parameter of the first Gaussian kernel function, ||·|| represents the norm function, exp represents the exponential function with e as the base, and d1 represents the first affine density value; the second affine density value of the speed time series characteristic vector relative to the fiber state time series change characteristic vector is calculated by the following second affine density calculation formula, wherein the second affine density calculation formula is:
[0104]
[0105] Among them, x1 represents the fiber state temporal change characteristic vector, x2 represents the speed temporal characteristic vector, σ2 represents the width parameter of the second Gaussian kernel function, ||·|| represents the norm function, exp represents the exponential function with base e, and d2 represents the second affine density value; the fiber state temporal change characteristic vector and the speed temporal characteristic vector are weighted respectively with the first affine density value and the second affine density value as weights to obtain the weighted fiber state temporal change characteristic vector and the weighted speed temporal characteristic vector; the weighted fiber state temporal change characteristic vector and the weighted speed temporal characteristic vector are associated encoded to obtain the classification feature matrix.
[0106] Here, those skilled in the art will appreciate that the specific operations of each step in the above-mentioned polycrystalline mullite fiber production control method have been described in detail above. Figures 1 to 3 The description of the production control system of the polycrystalline mullite fiber has been described in detail, and therefore, its repeated description will be omitted.
Claims
1. A production control system for polycrystalline mullite fiber, characterized in that: include: A data acquisition module, used to obtain the disk rotation speed values at multiple predetermined time points in a predetermined time period during the fiberization process of the colloid and the colloid fiberization monitoring video of the predetermined time period acquired by the camera; A key frame extraction module, used to extract multiple fiber state key frames from the colloid fiberization monitoring video; A fiber state feature extraction module, configured to pass the plurality of fiber state key frames through a first convolutional neural network model including a salient object detector to obtain a plurality of fiber state feature vectors; A transfer calculation module, used for calculating the transfer matrix between the fiber state feature vectors of each two adjacent time points in the multiple fiber state feature vectors to obtain multiple transfer matrices; A fiber state temporal change feature extraction module, configured to obtain a fiber state temporal change feature vector by arranging the multiple transfer matrices into a three-dimensional input tensor through a second convolutional neural network model using a channel attention mechanism; A disk speed time series encoding module, used for arranging the disk speed values at the plurality of predetermined time points into a speed input vector according to the time dimension and then passing the vector through a time series encoder including a fully connected layer and a one-dimensional convolutional layer to obtain a speed time series feature vector; An association coding module, used for performing association coding on the fiber state time series change feature vector and the rotation speed time series feature vector to obtain a classification feature matrix; A control result generation module, used for passing the classification feature matrix through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the rotation speed of the disk at the current time point should be increased or decreased; Wherein, the key frame extraction module is used to: extract a plurality of fiber state key frames from the colloid fiberization monitoring video at a predetermined sampling frequency; The fiber state feature extraction module is used to: use each layer of the first convolutional neural network model including the salient target detector to perform the following operations on the input data in the forward pass of the layer: Using a first convolution kernel to perform convolution processing on the input data to obtain a first convolution feature map; Using a second convolution kernel to perform convolution processing on the first convolution feature map to obtain a second convolution feature map, wherein the size of the first convolution kernel is larger than the size of the second convolution kernel; Performing global pooling processing on each feature matrix based on the channel dimension on the second convolutional feature map to obtain a pooled feature map; Performing activation processing on the pooled feature map to obtain an activated feature map; Among them, the output of the last layer of the first convolutional neural network model is the fiber state feature vector, and the input of the first layer of the first convolutional neural network model is the fiber state key frame.
2. The production control system of polycrystalline mullite fiber according to claim 1, characterized in that: The transfer calculation module is used to calculate the transfer matrix between the fiber state feature vectors of each two adjacent time points in the multiple fiber state feature vectors using the following transfer formula to obtain the multiple transfer matrices; Wherein, the transfer formula is: Wherein V1 and V2 represent the fiber state feature vectors at two adjacent time points among the plurality of fiber state feature vectors, represents matrix multiplication, and M represents the transfer matrix.
3. The production control system of polycrystalline mullite fiber according to claim 2, characterized in that: The fiber state temporal change feature extraction module is used to: use each layer of the second convolutional neural network model to perform the following on the input data in the forward transfer of the layer: Performing convolution processing based on a three-dimensional convolution kernel on the input data to obtain a convolution feature map; Calculate the global mean of each feature matrix along the channel dimension of the convolution feature map to obtain a channel feature vector; Calculating the ratio of the eigenvalue at each position in the channel eigenvector to the weighted sum of the eigenvalues at all positions of the channel eigenvector to obtain a channel weighted eigenvector; Performing point multiplication on the feature matrix along the channel dimension of the convolution feature map using the eigenvalues of each position of the channel weighted feature vector as weights to obtain a channel attention map; Performing global mean pooling processing on each feature matrix based on the channel dimension on the channel attention map to obtain a pooled feature map; Performing activation processing on the pooled feature map to obtain an activated feature map; Among them, the last layer output of the second convolutional neural network model is the fiber state temporal change feature vector, the input from the second layer to the last layer of the second convolutional neural network is the output of the previous layer, and the input of the second convolutional neural network is the three-dimensional input tensor.
4. The production control system of polycrystalline mullite fiber according to claim 3, characterized in that: The disk speed timing encoding module comprises: a normalization mapping unit, configured to perform normalization processing on the speed input vector based on a maximum value to obtain a normalized speed input vector; A fully connected encoding unit, configured to perform fully connected encoding on the normalized speed input vector using the fully connected layer of the temporal encoder to extract high-dimensional implicit features of feature values at each position in the normalized speed input vector; A one-dimensional convolutional encoding unit is used to use the one-dimensional convolutional layer of the temporal encoder to perform one-dimensional encoding on the normalized speed input vector to extract high-dimensional implicit correlation features of the correlation between the feature values at each position in the normalized speed input vector.
5. The production control system of polycrystalline mullite fiber according to claim 4, characterized in that: The associated coding module comprises: A first affine density calculation unit is used to calculate a first affine density value of the fiber state time series change characteristic vector relative to the speed time series characteristic vector using the following first affine density calculation formula, wherein the first affine density calculation formula is: Wherein, x1 represents the fiber state time series change characteristic vector, x2 represents the speed time series characteristic vector, σ1 represents the width parameter of the first Gaussian kernel function, ‖·‖ represents the norm function, exp represents the exponential function with e as the base, and d1 represents the first affine density value; The second affine density calculation unit is used to calculate the second affine density value of the speed time series characteristic vector relative to the fiber state time series change characteristic vector using the following second affine density calculation formula, wherein the second affine density calculation formula is: Wherein, x1 represents the fiber state time series change characteristic vector, x2 represents the speed time series characteristic vector, σ2 represents the width parameter of the second Gaussian kernel function, ‖·‖ represents the norm function, exp represents the exponential function with e as the base, and d2 represents the second affine density value; A weighting unit, used for weighting the fiber state time series change feature vector and the rotation speed time series feature vector respectively by using the first affine density value and the second affine density value as weights to obtain a weighted fiber state time series change feature vector and a weighted rotation speed time series feature vector; The associative coding unit is used to perform associative coding on the weighted fiber state time series change feature vector and the weighted rotation speed time series feature vector to obtain the classification feature matrix.
6. A production control method for polycrystalline mullite fiber, characterized in that: include: Obtaining the disk rotation speed values at multiple predetermined time points within a predetermined time period during the fiberization process of the colloid and the colloid fiberization monitoring video of the predetermined time period collected by the camera; Extracting a plurality of fiber state key frames from the colloid fiberization monitoring video; Passing the plurality of fiber state key frames through a first convolutional neural network model including a salient object detector to obtain a plurality of fiber state feature vectors; Calculating a transfer matrix between the fiber state feature vectors at every two adjacent time points in the plurality of fiber state feature vectors to obtain a plurality of transfer matrices; Arranging the multiple transfer matrices into a three-dimensional input tensor and then using a second convolutional neural network model with a channel attention mechanism to obtain a fiber state temporal change feature vector; Arranging the disk speed values at the plurality of predetermined time points into a speed input vector according to the time dimension and then passing the vector through a temporal encoder including a fully connected layer and a one-dimensional convolutional layer to obtain a speed temporal feature vector; Performing association coding on the fiber state time series change feature vector and the rotation speed time series feature vector to obtain a classification feature matrix; Passing the classification feature matrix through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the rotation speed of the disk at the current time point should be increased or decreased; Wherein, extracting a plurality of fiber state key frames from the colloid fiberization monitoring video comprises: extracting a plurality of fiber state key frames from the colloid fiberization monitoring video at a predetermined sampling frequency; The method of passing the plurality of fiber state key frames through a first convolutional neural network model including a salient target detector to obtain a plurality of fiber state feature vectors comprises: using each layer of the first convolutional neural network model including a salient target detector to perform the following operations on the input data in the forward pass of the layer: Using a first convolution kernel to perform convolution processing on the input data to obtain a first convolution feature map; Using a second convolution kernel to perform convolution processing on the first convolution feature map to obtain a second convolution feature map, wherein the size of the first convolution kernel is larger than the size of the second convolution kernel; Performing global pooling processing on each feature matrix based on the channel dimension on the second convolutional feature map to obtain a pooled feature map; Performing activation processing on the pooled feature map to obtain an activated feature map; Among them, the output of the last layer of the first convolutional neural network model is the fiber state feature vector, and the input of the first layer of the first convolutional neural network model is the fiber state key frame.
