Rotary kiln state recognition method based on Tucker decomposition and related equipment

Abstract

The embodiment of the present application belongs to the field of industrial Internet technology, and relates to a rotary kiln state recognition method based on Tucker decomposition and related equipment, including constructing process combination data corresponding to each time window based on the acquired process parameters, constructing a firing zone splicing image based on the kiln condition video corresponding to each time window, respectively training an initial full convolution model and an initial residual neural model through the process combination data and the firing zone splicing image to obtain a full convolution model and a residual neural model; converting the process combination data into process feature data based on the full convolution model, and converting the firing zone splicing image into image feature data based on the residual neural model; obtaining an initial feature fusion model and an initial long short-term memory model based on Tucker decomposition, training the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data, and obtaining a feature fusion model and a long short-term memory model. The present application improves the accuracy of rotary kiln state recognition.

Description

Technical Field

[0001] The present application relates to the field of industrial Internet technology, and in particular to a rotary kiln state identification method based on Tucker decomposition and related equipment. Background Art

[0002] The control and operation of the rotary kiln of the new dry process cement production line is the most important link to finally realize the design intention, tap the equipment capacity and improve the enterprise efficiency. The cement rotary kiln is a type of kiln that is widely used in my country's cement industry. Its production process is a typical thermal process. The cement burning process has the characteristics of nonlinearity, multivariable, strong coupling, large time lag and time variation. This makes the kiln condition of the rotary kiln system very complicated. At present, most of the domestic cement rotary kilns are still under manual or semi-automatic control, that is, they rely on manual observation of the flame, combined with relevant measuring instruments, comprehensive judgment of the operating status of the rotary kiln, and manual adjustment of various control quantities. Obviously, this control method is greatly affected by human factors and the control is not timely. Even experienced operators find it difficult to keep the rotary kiln in the best working condition for a long time.

[0003] Neural network control simulates the structure and function of the human brain from a microscopic perspective, that is, a control system designed from the perspective of studying and simulating the structure, function, and mechanism of transmitting, processing, and controlling information of human neural networks. Neural networks are not good at displaying and expressing knowledge, but they have a strong ability to approximate nonlinear functions. At present, many studies have used the nonlinear mapping, feature extraction, and fault tolerance of neural networks to solve the optimization control modeling problem of cement kilns, such as BP and RBF. In actual industrial production applications, the working conditions are more complex and extensive, and the expressiveness of these neural networks will be greatly reduced. There is also a way to use the temperature field information and clinker sintering information contained in the video image of the burning belt to judge the working conditions of the clinker sintering of the rotary kiln. However, due to the rotation of the kiln body and the interference of dust and smoke in the kiln, the image quality of the burning belt is not high, the boundaries are unclear, and the coupling of significant areas is strong. The existing static image-based methods are difficult to extract features with distinguishing and robustness, resulting in low precision and accuracy in the recognition of the state of the rotary kiln. Summary of the invention

[0004] The purpose of the embodiments of the present application is to propose a rotary kiln state recognition method, device, computer equipment and storage medium based on Tucker decomposition to improve the accuracy of rotary kiln state recognition.

[0005] In order to solve the above technical problems, the embodiment of the present application provides a rotary kiln state recognition method based on Tucker decomposition, which adopts the following technical solution:

[0006] A rotary kiln state recognition method based on Tucker decomposition comprises the following steps:

[0007] Obtain corresponding process parameters and kiln condition videos according to preset time windows, construct process combination data corresponding to each time window based on the process parameters, construct corresponding firing zone splicing images based on the kiln condition videos corresponding to each time window, and respectively train a preset initial full convolution model and an initial residual neural model through all the process combination data and all the firing zone splicing images to obtain a full convolution model and a residual neural model;

[0008] Performing a feature extraction operation on the process combination data based on the full convolution model to obtain process feature data, and performing a feature extraction operation on the fired belt splicing image based on the residual neural model to obtain image feature data;

[0009] Obtaining a preset initial feature fusion model based on Tucker decomposition and a preset initial long short-term memory model, training the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data, and obtaining a feature fusion model and a long short-term memory model;

[0010] The target process parameters and the target firing zone image are acquired in real time, and feature extraction operations are performed on the target process parameters and the target firing zone image according to the full convolution model and the residual neural model respectively to obtain target process feature data and target image feature data respectively, and the target process feature data and the target image feature data are input into the feature fusion model and the long short-term memory model in sequence to obtain the target kiln condition category.

[0011] Furthermore, the step of constructing process combination data corresponding to each time window based on the process parameters includes:

[0012] Convert the process parameters corresponding to each of the time windows into an initial process matrix respectively;

[0013] splicing the initial process matrix based on the time sequence to obtain a process splicing matrix, and performing a matrix transposition operation on the process splicing matrix to obtain a process transposed matrix;

[0014] The process transposed matrix is ​​segmented according to the dimension of the process parameters to obtain the process combination data.

[0015] Furthermore, the step of converting the process parameters corresponding to each of the time windows into an initial process matrix comprises:

[0016] Check whether the process parameters have null values ​​or abnormal values;

[0017] When the null value or the abnormal value exists, the value of the process parameter at the previous moment is retrieved to fill the null value or replace the abnormal value, so as to obtain the first process parameter;

[0018] The first process parameters are normalized to obtain second process parameters, and the second process parameters are converted into the initial process matrix.

[0019] Furthermore, the step of constructing a corresponding spliced ​​image of the fired zone based on the kiln condition video corresponding to each time window includes:

[0020] According to a preset frame extraction interval, the kiln condition video is subjected to a frame extraction operation to obtain firing band frame data;

[0021] The fired belt frame data are spliced ​​according to a time sequence to generate the fired belt spliced ​​image.

[0022] Furthermore, the step of splicing the burned belt frame data according to the time sequence to generate the burned belt spliced ​​image includes:

[0023] Seamlessly splice the fired strip frame data according to the time sequence to obtain an initial fired strip spliced ​​image of a specified size;

[0024] It is determined whether there is a missing part in the initial fired zone stitching image. If so, a filling operation is performed using a preset color to obtain the fired zone stitching image.

[0025] Furthermore, the step of performing a feature extraction operation on the process combination data based on the full convolution model to obtain process feature data, and performing a feature extraction operation on the fired belt splicing image based on the residual neural model to obtain image feature data includes:

[0026] Removing the fully connected layer of the fully convolutional model to obtain a parameter feature extraction model, and inputting the process combination data into the parameter feature extraction model to obtain the process feature data;

[0027] The fully connected layer of the residual neural model is removed to obtain an image feature extraction model, and the sintered belt spliced ​​image is input into the image feature extraction model to obtain the image feature data.

[0028] Furthermore, the initial full convolution model includes a plurality of full convolution layers, the full convolution layer includes a convolution layer and a nonlinear activation layer, and the steps of respectively training the preset initial full convolution model and the initial residual neural model through all the process combination data and all the fired belt spliced ​​images to obtain the full convolution model and the residual neural model include:

[0029] Passing the process combination data through each of the full convolution layers, performing a downsampling operation on the process combination data through the convolution layer, and performing an activation operation on the process combination data through the Leaky_relu function of the nonlinear activation layer to obtain first classification data;

[0030] Iteratively training the initial full convolution model based on the first classification data to obtain the full convolution model;

[0031] The fired belt spliced ​​image is input into the initial residual neural model to obtain second classification data, and the initial residual neural model is iteratively trained based on the second classification data to obtain the residual neural model.

[0032] Furthermore, the step of training the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data to obtain the feature fusion model and the long short-term memory model includes:

[0033] Inputting the process feature data and the image feature data into the initial feature fusion model, wherein the initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data to obtain fusion features;

[0034] Inputting the fused features into the initial long short-term memory model to obtain an estimated kiln condition category;

[0035] The initial feature fusion model and the initial long short-term memory model are iteratively trained based on the estimated kiln condition category to obtain the feature fusion model and the long short-term memory model.

[0036] Furthermore, the initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data, and the step of obtaining fusion features includes:

[0037] The initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data through the following formula to obtain fusion features:

[0038] y=(T c ×q T W q )×(v T W v ))×W o , where y represents the fusion feature, q T Represents the process characteristic data, v T represents the image feature data, Wq, Wv represent the preset feature mapping matrix, Tc represents the preset fusion matrix, and Wo represents the preset result mapping matrix.

[0039] In order to solve the above technical problems, the embodiment of the present application also provides a rotary kiln state identification device based on Tucker decomposition, which adopts the following technical solution:

[0040] A rotary kiln state recognition device based on Tucker decomposition, comprising:

[0041] A training module, used for acquiring corresponding process parameters and kiln condition videos according to preset time windows, constructing process combination data corresponding to each time window based on the process parameters, constructing corresponding firing zone splicing images based on the kiln condition videos corresponding to each time window, and respectively training a preset initial full convolution model and an initial residual neural model through all the process combination data and all the firing zone splicing images to obtain a full convolution model and a residual neural model;

[0042] A conversion module, used for performing a feature extraction operation on the process combination data based on the full convolution model to obtain process feature data, and performing a feature extraction operation on the fired belt splicing image based on the residual neural model to obtain image feature data;

[0043] A fusion module, used to obtain a preset initial feature fusion model based on Tucker decomposition and a preset initial long short-term memory model, train the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data, and obtain a feature fusion model and a long short-term memory model;

[0044] An input module is used to acquire target process parameters and target firing zone images in real time, perform feature extraction operations on the target process parameters and the target firing zone images according to the full convolution model and the residual neural model, respectively, obtain target process feature data and target image feature data, and input the target process feature data and the target image feature data into the feature fusion model and the long short-term memory model in sequence to obtain the target kiln condition category.

[0045] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0046] A computer device comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the above-mentioned rotary kiln state identification method based on Tucker decomposition when executing the computer-readable instructions.

[0047] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0048] A computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the rotary kiln state identification method based on Tucker decomposition are implemented.

[0049] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0050] This application obtains the corresponding process parameters and kiln condition videos through a preset time window, thereby constructing process combination data and fired zone splicing images corresponding to different time windows, training the corresponding initial full convolution model and initial residual neural model, and obtaining a full convolution model and a residual neural model. Based on the full convolution model and the residual neural model, feature extraction is performed on the process combination data and the fired zone splicing image to obtain process feature data and image feature data, which can extract effective features of process parameters and kiln condition videos. The extracted process feature data and image feature data are modally fused based on Tucker decomposition, which can effectively establish the complex relationship between process parameters and kiln condition videos. Finally, by training the initial long short-term memory model, a long short-term memory model is obtained, and the long short-term memory model is used to extract the time series features of the kiln condition state, obtain the target kiln condition category, and improve the judgment accuracy of the kiln condition state category. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the scheme in the present application, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0052] Figure 1 is an exemplary system architecture diagram to which the present application may be applied;

[0053] Figure 2 is a flow chart of an embodiment of a rotary kiln state identification method based on Tucker decomposition according to the present application;

[0054] Figure 3 is a structural schematic diagram of an embodiment of a rotary kiln state identification device based on Tucker decomposition according to the present application;

[0055] Figure 4 It is a structural diagram of an embodiment of a computer device according to the present application.

[0056] Attached figures: 200, computer equipment; 201, memory; 202, processor; 203, network interface; 300, rotary kiln state recognition device based on Tucker decomposition; 301, training module; 302, conversion module; 303, fusion module; 304, input module. DETAILED DESCRIPTION

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0058] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0059] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0060] like Figure 1 As shown, the system architecture 100 may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0061] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0062] Terminal devices 101, 102, 103 can be various electronic devices with display screens and supporting web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 (Moving Picture Experts Group Audio Layer IV) players, laptop computers, desktop computers, etc.

[0063] The server 105 may be a server that provides various services, such as a background server that provides support for web pages displayed on the terminal devices 101 , 102 , and 103 .

[0064] It should be noted that the rotary kiln state identification method based on Tucker decomposition provided in the embodiment of the present application is generally executed by a server / terminal device, and accordingly, the rotary kiln state identification device based on Tucker decomposition is generally arranged in a server / terminal device.

[0065] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0066] Continue to refer Figure 2 , shows a flow chart of an embodiment of a rotary kiln state identification method based on Tucker decomposition according to the present application. The rotary kiln state identification method based on Tucker decomposition comprises the following steps:

[0067] S1: Obtain corresponding process parameters and kiln condition videos according to preset time windows, construct process combination data corresponding to each time window based on the process parameters, construct corresponding firing zone splicing images based on the kiln condition videos corresponding to each time window, and respectively train the preset initial full convolution model and initial residual neural model through all the process combination data and all the firing zone splicing images to obtain a full convolution model and a residual neural model.

[0068] In this embodiment, the process parameters and kiln condition videos are collected, stored and converted by the deployed sensors. The computer obtains the process parameters and kiln condition videos in the sensor. The kiln condition parameters include process parameters and kiln condition videos. The process parameters have N dimensions, that is, corresponding to N related process parameters, such as secondary air temperature, kiln current, KH value, etc. This application sets the total duration of data collection, the time window T minutes and the frame extraction interval k seconds. Each time window corresponds to a kiln condition status category, such as low, normal, high, etc. These kiln condition status categories are used as labels for the corresponding process combination data and the spliced ​​images of the fired belt, and are used to train the initial full convolution model and the initial residual neural model. The kiln condition status categories corresponding to the process parameters and kiln condition videos corresponding to the same time window are the same. A process parameter data set is constructed based on all the process combination data, and a kiln fired belt image set is constructed based on all the spliced ​​images of the fired belt. The initial full convolution model and the initial residual neural model are trained respectively by the process parameter data set and the kiln fired belt image set.

[0069] It should be noted that: after training the initial full convolution model and the initial residual neural model respectively through the process parameter data set (i.e., all process combination data structures) and the kiln firing zone image set (i.e., all firing zone splicing images), 80% of the samples in the process parameter data set are randomly selected as process training samples, and the remaining 20% ​​of the samples are used as process test samples. The initial convolution model is trained with process training samples, and the initial full convolution model during the training process is tested with process test samples. Among them, the specific process is to pass the process training samples through each full convolution layer in the initial convolution model in turn, and the full convolution layer includes a convolution layer and a nonlinear activation layer. The convolution layer performs a downsampling operation on the process training samples, and the process training samples are activated by the leaky_relu function of the nonlinear activation layer. And input the data output by the last nonlinear activation layer into the fully connected layer of the model to obtain the output first classification data, until all process training samples are trained to obtain the fully convolutional model to be verified, and test the fully convolutional model to be verified through process test samples. If the accuracy exceeds the accuracy threshold, it is determined that the verification is passed, and the fully convolutional model to be verified is used as the trained full convolutional model. If the verification fails, the initial full convolutional model is re-iteratively trained until the verification passes and the full convolutional model is obtained. 80% of the samples are randomly selected from the kiln firing belt image set as image training samples, and the remaining 20% ​​of the samples are used as image test samples. The initial residual neural model is trained with image training samples, and the initial residual neural model during the training process is tested with image test samples.

[0070] In this embodiment, the rotary kiln state recognition method based on Tucker decomposition is run on an electronic device (e.g. Figure 1 The server / terminal device shown in the figure can obtain the corresponding process parameters and kiln condition video through a wired connection or a wireless connection. It should be noted that the above wireless connection method may include but is not limited to 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or to be developed in the future.

[0071] Specifically, the step of constructing process combination data corresponding to each time window based on the process parameters includes:

[0072] Convert the process parameters corresponding to each of the time windows into an initial process matrix respectively;

[0073] splicing the initial process matrix based on the time sequence to obtain a process splicing matrix, and performing a matrix transposition operation on the process splicing matrix to obtain a process transposed matrix;

[0074] The process transposed matrix is ​​segmented according to the dimension of the process parameters to obtain the process combination data.

[0075] In this embodiment, the dimension of the process parameters is N, that is, it includes N related process parameters, such as secondary air temperature, kiln current, KH value, etc. The process parameters in the time window of T minutes are converted into an initial process matrix, and the size of the initial process matrix is ​​N×T, for example:

[0076]

[0077] , then the N×T initial process matrix is ​​expressed as: Among them, N related process parameters are N rows, time window T is T columns, the value of each minute is a column, and the N related process parameters are arranged in time series. Then, based on the time series splicing of the initial process matrix, a process splicing matrix is ​​obtained, that is, all the initial process matrices are spliced ​​by column to form a process splicing matrix of (S÷T×N)×T, and the process splicing matrix is ​​transposed to obtain a process transposed matrix. The process transposed matrix is ​​segmented according to the dimension of the process parameters to obtain the process combination data. That is, the process transposed matrix is ​​segmented into groups of N rows, representing the data of N related process parameters in the time window, wherein each row is the minute-level data of a certain related process parameter in the time window. Finally, process combination data of size N×T is obtained, and each process combination data corresponds to a time window and a kiln condition category.

[0078] The step of converting the process parameters corresponding to each of the time windows into an initial process matrix comprises:

[0079] Check whether the process parameters have null values ​​or abnormal values;

[0080] When the null value or the abnormal value exists, the value of the process parameter at the previous moment is retrieved to fill the null value or replace the abnormal value, so as to obtain the first process parameter;

[0081] The first process parameters are normalized to obtain second process parameters, and the second process parameters are converted into the initial process matrix.

[0082] In this embodiment, it is checked whether the process parameters have null values ​​or abnormal values. When the null values ​​or abnormal values ​​exist, the values ​​at the previous moment are used for filling or replacement. For the determination of abnormal values, the variances of all process parameters are calculated, and a preset multiple of the variance, such as 3 times the variance, is used as an abnormal threshold. When the value of the process parameter exceeds the abnormal threshold, the value is determined to be an abnormal value. The dimension of the process parameter is N, and the relevant process parameters of each dimension of the process parameter are normalized to zero mean to eliminate the dimension effect.

[0083] In addition, the step of constructing a corresponding spliced ​​image of the fired zone based on the kiln condition video corresponding to each time window includes:

[0084] According to a preset frame extraction interval, the kiln condition video is subjected to a frame extraction operation to obtain firing band frame data;

[0085] The fired belt frame data are spliced ​​according to a time sequence to generate the fired belt spliced ​​image.

[0086] In this embodiment, the present application sets a frame extraction interval of k seconds, wherein T(min)×60=number of extracted images (pictures)×frame extraction interval k(s), that is, an image is extracted at an interval of k seconds in each time window, wherein 60T>k.

[0087] The step of splicing the burned belt frame data according to the time sequence to generate the burned belt spliced ​​image includes:

[0088] Seamlessly splice the fired strip frame data according to the time sequence to obtain an initial fired strip spliced ​​image of a specified size;

[0089] It is determined whether there is a missing part in the initial fired zone stitching image. If so, a filling operation is performed using a preset color to obtain the fired zone stitching image.

[0090] In this embodiment, the kiln condition video is subjected to frame extraction and cropping operations to obtain the burning belt frame data of m×m size. The burning belt frame data is seamlessly spliced ​​according to the time sequence to obtain the initial burning belt splicing image of the specified size, and it is determined whether there is a missing part in the initial burning belt splicing image. If there is, it is filled with a preset color, such as black or white, to obtain the burning belt splicing image, that is, the image extracted in each time window T constitutes an M×M (specified size) burning belt splicing image. The specific splicing process is: splicing in time sequence, first splicing a row horizontally, splicing to the specified number of the row, and then switching to the next row to continue splicing, where M÷m=integer. For example, M=3, m=1, then 9 1×1 small pictures are required to form a 3×3 large picture. When 7 small pictures are obtained by frame extraction through the time window, the remaining parts are spliced ​​in order and filled with black or white. If there is no missing part, there is no need to perform the filling operation, and the initial burning belt splicing image is directly used as the burning belt splicing image. Among them, a stitched image of the firing zone corresponds to a time window and a kiln condition category.

[0091] In addition, the initial full convolution model includes a plurality of full convolution layers, the full convolution layer includes a convolution layer and a nonlinear activation layer, and the steps of respectively training the preset initial full convolution model and the initial residual neural model through all the process combination data and all the fired belt splicing images to obtain the full convolution model and the residual neural model include:

[0092] Passing the process combination data through each of the full convolution layers, performing a downsampling operation on the process combination data through the convolution layer, and performing an activation operation on the process combination data through the Leaky_relu function of the nonlinear activation layer to obtain first classification data;

[0093] Iteratively training the initial full convolution model based on the first classification data to obtain the full convolution model;

[0094] The fired belt spliced ​​image is input into the initial residual neural model to obtain second classification data, and the initial residual neural model is iteratively trained based on the second classification data to obtain the residual neural model.

[0095] In this embodiment, it can be seen from the above that the process combination data is an N×T matrix. The kiln condition is not only related to the process combination data corresponding to each time window, but also has a certain correlation between the process combination data. Therefore, the initial full convolution model (i.e., one-dimensional convolutional neural network) is used to perform downsampling feature extraction to extract contextual relationship features. The initial full convolution model is a full convolutional network, and the network composition is shown in the following table:

[0096]

[0097]

[0098] , where the Leaky_relu layer is a nonlinear activation layer. The nonlinear activation function (Leaky ReLU) is a variant of the ReLu activation function, and the output of this function has a small slope for negative inputs. Since the derivative is always non-zero, this can reduce the appearance of silent neurons, allowing gradient-based learning (although it will be slow), and solving the problem that neurons do not learn after the Relu function enters the negative interval. The initial full convolution model includes multiple convolutional layers and multiple nonlinear activation layers (Leaky_relu layers) in the above table, as well as a fully connected layer. The process combination data (N×T size) is input into the first convolutional layer, where the convolution kernel is 2 and the number of channels is 8, and the output is (N-1,8). Input (N-1,8) into the first Leaky_relu layer, where the number of channels is 8, and the output is (N-1,8). The output (N-1,8) of the first Leaky_relu layer is input into the second convolution layer. The convolution kernel of the second convolution layer is 2 and the number of channels is 16, and the output is (N-2,16). The output (N-2,16) is input into the second Leaky_relu layer. The number of channels of the second Leaky_relu layer is 16, and the output is (N-2,16). And so on, in which the number of convolution kernels in the convolution layer remains unchanged, and the number of channels doubles successively. The number of channels in the Leaky_relu layer doubles successively. Finally, the output (1,c) of the last Leaky_relu layer is obtained, and the (1,c) is input into the fully connected layer, that is, the first classification data is output.

[0099] The size of the spliced ​​image of the fired belt is M×M, and the spliced ​​image of the fired belt includes a series of changes in the fired belt in the kiln within the corresponding time window. The initial residual neural model is used to extract image features, and a series of images in the time window are used for extraction and kiln condition judgment, which reduces the interference of the rotation of the kiln body and the dust and smoke in the kiln on classification in a single image. The initial residual neural model adopts the method of transfer learning, removes the original fully connected layer of the current ResNet (residual network), and adds a self-designed fully connected layer that meets the needs of the scene. Among them, the self-involved fully connected layer is set to be finally input into it, and the data is normalized. Its different values ​​correspond to the classification category of this scene, such as 0.5 corresponds to a low kiln condition, 0.7 corresponds to a high kiln condition, etc. After the residual neural model is built, the labels of the spliced ​​image of the fired belt and the corresponding kiln condition category are put into the initial residual neural model for training, and the trained model is saved.

[0100] S2: performing feature extraction operation on the process combination data based on the full convolution model to obtain process feature data, and performing feature extraction operation on the fired belt splicing image based on the residual neural model to obtain image feature data.

[0101] In this embodiment, the two modal data, process combination data and fired belt splicing image, are converted into high-dimensional feature expressions, namely process feature data and image feature data, through a full convolution model and a residual neural model, respectively, for subsequent training of the initial feature fusion model and the preset initial long short-term memory model.

[0102] The steps of performing feature extraction on the process combination data based on the full convolution model to obtain process feature data, and performing feature extraction on the fired belt splicing image based on the residual neural model to obtain image feature data include:

[0103] Removing the fully connected layer of the fully convolutional model to obtain a parameter feature extraction model, and inputting the process combination data into the parameter feature extraction model to obtain the process feature data;

[0104] The fully connected layer of the residual neural model is removed to obtain an image feature extraction model, and the sintered belt spliced ​​image is input into the image feature extraction model to obtain the image feature data.

[0105] In this embodiment, the process feature data and image feature data are subsequently used as training samples for the initial feature fusion model and the initial long short-term memory model. Its label is still the kiln condition category of the corresponding time window. Feature extraction of process parameters: remove the fully connected layer of the full convolution model, that is, the classification layer, to obtain a parameter feature extraction model. The process combination data is put into the parameter feature extraction model for calculation to obtain the corresponding 1×C process feature data, that is, the 1×C feature vector. It represents the characteristics of N related process parameters in the corresponding time window. Image feature extraction of the fired belt in the kiln: remove the fully connected layer of the residual neural model to obtain an image feature extraction model. Put the spliced ​​image of the fired belt into the image feature extraction model for calculation to obtain the corresponding 7×7 size image feature data, that is, a 7×7 feature map. Among them, the size of the image feature data is obtained based on the setting of the initial residual neural model. When the initial residual model changes, the size of the image feature data may also change accordingly.

[0106] S3: Obtain a preset initial feature fusion model based on Tucker decomposition and a preset initial long short-term memory model, train the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data, and obtain a feature fusion model and a long short-term memory model.

[0107] In this embodiment, since the initial feature fusion model based on Tucker decomposition is adopted, and the time domain connection between training samples in the time window is taken into consideration, a long-term time series prediction method, namely a long short-term memory model (LSTM), is used to extract time domain features. A fully connected layer based on a softmax activation function is used for final classification and recognition. The initial feature fusion model and the initial long short-term memory model are trained based on the process feature data and the image feature data, and the trained models are saved.

[0108] The step of training the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data to obtain the feature fusion model and the long short-term memory model comprises:

[0109] Inputting the process feature data and the image feature data into the initial feature fusion model, wherein the initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data to obtain fusion features;

[0110] Inputting the fused features into the initial long short-term memory model to obtain an estimated kiln condition category;

[0111] The initial feature fusion model and the initial long short-term memory model are iteratively trained based on the estimated kiln condition category to obtain the feature fusion model and the long short-term memory model.

[0112] In this embodiment, the process feature data and the image feature data are subjected to Tucker decomposition and feature fusion through the initial feature fusion model to obtain fusion features, thereby realizing feature fusion based on Tucker decomposition of two modal features in the same time window. The features are then input into the classification model, i.e., the initial long short-term memory model, to obtain the estimated kiln condition category.

[0113] Furthermore, the initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data, and the step of obtaining fusion features includes:

[0114] The initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data through the following formula to obtain fusion features:

[0115] y=(T c ×q T W q )×(v T W v ))×W o , where y represents the fusion feature, q T Represents the process characteristic data, v Trepresents the image feature data, Wq, Wv represent the preset feature mapping matrix, Tc represents the preset fusion matrix, and Wo represents the preset result mapping matrix.

[0116] In this embodiment, feature fusion based on Tucker decomposition: the extracted process feature data and image feature data are subjected to Tucker decomposition and feature fusion. After decomposition, each matrix is ​​interpretable, Wq, Wv represent the preset feature mapping matrix, Tc is the fusion matrix of the image feature data, after calculating Tc, it contains the relationship information between any image feature data and process feature data, and finally Wo maps these relationships to the fusion feature y. Each time window sample has its corresponding fusion feature. The initial feature fusion model and the initial long short-term memory model are trained together, and the trained Wq, Wv, Tc, Wo are obtained through optimization iteration, and the initial values ​​are random or customized.

[0117] S4: Acquire the target process parameters and the target firing zone image in real time, perform feature extraction operations on the target process parameters and the target firing zone image according to the full convolution model and the residual neural model, respectively, obtain target process feature data and target image feature data, respectively, and input the target process feature data and the target image feature data into the feature fusion model and the long short-term memory model in sequence to obtain the target kiln condition category.

[0118] In this embodiment, in the application process of the present application, the target process parameters and the target fired zone image are subjected to feature extraction operations respectively through the full convolution model and the residual neural model, and target process feature data and target image feature data are obtained respectively. In fact, the target process parameters and the target fired zone image are subjected to feature extraction operations respectively through the parameter feature extraction model and the image feature extraction model, and target process feature data and target image feature data are obtained respectively.

[0119] The present application obtains the corresponding process parameters and kiln condition video through a preset time window, thereby constructing the process combination data and the spliced ​​image of the burning belt corresponding to different time windows, training the corresponding initial full convolution model and the initial residual neural model, and obtaining the full convolution model and the residual neural model. And based on the full convolution model and the residual neural model, the process combination data and the spliced ​​image of the burning belt are feature extracted to obtain the process feature data and the image feature data, and the effective features of the process parameters and the kiln condition video can be extracted. The extracted process feature data and the image feature data are modally fused based on Tucker decomposition, which can effectively establish the complex relationship between the process parameters and the kiln condition video, and finally, by training the initial long short-term memory model, the long short-term memory model is obtained, and the kiln condition state time series features are extracted through the long short-term memory model to obtain the target kiln condition category, thereby improving the judgment accuracy of the kiln condition state category. The present application can effectively identify the state category of the rotary kiln, which lays the foundation for further realizing the optimization control of the rotary kiln, and can improve the automation level of cement production, reduce energy consumption and save energy.

[0120] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the computer-readable instructions are executed, they can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0121] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0122] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of a rotary kiln state recognition device based on Tucker decomposition, and the device embodiment is similar to Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0123] like Figure 3 As shown, the rotary kiln state recognition device 300 based on Tucker decomposition described in this embodiment includes: a training module 301, a conversion module 302, a fusion module 303 and an input module 304. Among them: the training module 301 is used to obtain the corresponding process parameters and kiln condition videos according to a preset time window, construct the process combination data corresponding to each time window based on the process parameters, construct the corresponding fired zone spliced ​​image based on the kiln condition video corresponding to each time window, and train the preset initial full convolution model and the initial residual neural model respectively through all the process combination data and all the fired zone spliced ​​images to obtain a full convolution model and a residual neural model; the conversion module 302 is used to perform feature extraction operations on the process combination data based on the full convolution model to obtain process feature data, and perform feature extraction operations on the fired zone spliced ​​image based on the residual neural model to obtain image feature data; the fusion module 303 is used to extract the process combination data based on the full convolution model to obtain process feature data, and to extract the image feature data based on the residual neural model ... To obtain a preset initial feature fusion model based on Tucker decomposition and a preset initial long short-term memory model, train the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data, and obtain a feature fusion model and a long short-term memory model; an input module 304 is used to obtain target process parameters and target firing zone images in real time, perform feature extraction operations on the target process parameters and the target firing zone image according to the full convolution model and the residual neural model, respectively, obtain target process feature data and target image feature data, and input the target process feature data and the target image feature data into the feature fusion model and the long short-term memory model in sequence to obtain a target kiln condition category.

[0124] In this embodiment, the present application obtains the corresponding process parameters and kiln condition videos through a preset time window, thereby constructing process combination data and fired zone splicing images corresponding to different time windows, training the corresponding initial full convolution model and initial residual neural model, and obtaining a full convolution model and a residual neural model. Based on the full convolution model and the residual neural model, feature extraction is performed on the process combination data and the fired zone splicing image to obtain process feature data and image feature data, and effective features of process parameters and kiln condition videos can be extracted. The extracted process feature data and image feature data are modally fused based on Tucker decomposition, which can effectively establish the complex relationship between process parameters and kiln condition videos. Finally, by training the initial long short-term memory model, a long short-term memory model is obtained, and the kiln condition state time series features are extracted through the long short-term memory model to obtain the target kiln condition category, thereby improving the judgment accuracy of the kiln condition state category.

[0125] The training module 301 includes a conversion submodule, a splicing submodule and a slicing submodule, wherein the conversion submodule is used to convert the process parameters corresponding to each time window into an initial process matrix respectively; the splicing submodule is used to splice the initial process matrix based on time sequence to obtain a process splicing matrix, and perform a matrix transposition operation on the process splicing matrix to obtain a process transposed matrix; the slicing submodule is used to perform a slicing operation on the process transposed matrix according to the dimension of the process parameters to obtain the process combination data.

[0126] The conversion submodule includes a checking unit, a calling unit and a normalization unit. The checking unit is used to check whether there are null values ​​or abnormal values ​​in the process parameters; the calling unit is used to call the value of the process parameter at the previous moment to fill the null value or replace the abnormal value when the null value or the abnormal value exists, so as to obtain the first process parameter; the normalization unit is used to normalize the first process parameter, obtain the second process parameter, and convert the second process parameter into the initial process matrix.

[0127] The training module 301 also includes a frame extraction submodule and an image stitching submodule. The frame extraction submodule is used to perform a frame extraction operation on the kiln condition video according to a preset frame extraction interval to obtain the firing belt frame data; the image stitching submodule is used to stitch the firing belt frame data according to a time sequence to generate the firing belt stitching image.

[0128] The image stitching submodule includes a stitching unit and a filling unit, wherein the stitching unit is used to seamlessly stitch the fired belt frame data according to the time sequence to obtain an initial fired belt stitching image of a specified size. The filling unit is used to determine whether there is a missing part in the initial fired belt stitching image, and if so, fill it with a preset color to obtain the fired belt stitching image.

[0129] The initial full convolution model includes multiple full convolution layers, and the full convolution layer includes a convolution layer and a non-linear activation layer. The training module 301 also includes an activation submodule, an iteration submodule and a training submodule. The activation submodule is used to pass the process combination data through each of the full convolution layers, downsample the process combination data through the convolution layer, and activate the process combination data through the Leaky_relu function of the non-linear activation layer to obtain first classification data; the iteration submodule is used to iteratively train the initial full convolution model based on the first classification data to obtain the full convolution model; the training submodule is used to input the fired belt spliced ​​image into the initial residual neural model to obtain second classification data, and iteratively train the initial residual neural model based on the second classification data to obtain the residual neural model.

[0130] The conversion module 302 includes a first extraction submodule and a second extraction submodule, wherein the first extraction submodule is used to remove the fully connected layer of the full convolution model to obtain a parameter feature extraction model, and input the process combination data into the parameter feature extraction model to obtain the process feature data; the second extraction submodule is used to remove the fully connected layer of the residual neural model to obtain an image feature extraction model, and input the fired belt spliced ​​image into the image feature extraction model to obtain the image feature data.

[0131] The fusion module 303 includes a fusion submodule, an estimation submodule and an acquisition submodule, wherein the fusion submodule is used to input the process feature data and the image feature data into the initial feature fusion model, and the initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data to obtain fusion features; the estimation submodule is used to input the fusion features into the initial long-short-term memory model to obtain an estimated kiln condition category; the acquisition submodule is used to iteratively train the initial feature fusion model and the initial long-short-term memory model based on the estimated kiln condition category to obtain the feature fusion model and the long-short-term memory model.

[0132] This application obtains the corresponding process parameters and kiln condition videos through a preset time window, thereby constructing process combination data and fired zone splicing images corresponding to different time windows, training the corresponding initial full convolution model and initial residual neural model, and obtaining a full convolution model and a residual neural model. Based on the full convolution model and the residual neural model, feature extraction is performed on the process combination data and the fired zone splicing image to obtain process feature data and image feature data, which can extract effective features of process parameters and kiln condition videos. The extracted process feature data and image feature data are modally fused based on Tucker decomposition, which can effectively establish the complex relationship between process parameters and kiln condition videos. Finally, by training the initial long short-term memory model, a long short-term memory model is obtained, and the long short-term memory model is used to extract the time series features of the kiln condition state, obtain the target kiln condition category, and improve the judgment accuracy of the kiln condition state category.

[0133] To solve the above technical problems, the present application also provides a computer device. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0134] The computer device 200 includes a memory 201, a processor 202, and a network interface 203 that are interconnected and communicated through a system bus. It should be noted that the figure only shows a computer device 200 having components 201-203, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field-programmable gate arrays (Field-Programmable Gate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.

[0135] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device.

[0136] The memory 201 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 201 can be an internal storage unit of the computer device 200, such as a hard disk or memory of the computer device 200. In other embodiments, the memory 201 can also be an external storage device of the computer device 200, such as a plug-in hard disk equipped on the computer device 200, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. Of course, the memory 201 can also include both the internal storage unit of the computer device 200 and its external storage device. In this embodiment, the memory 201 is generally used to store the operating system and various application software installed in the computer device 200, such as computer-readable instructions of the rotary kiln state recognition method based on Tucker decomposition, etc. In addition, the memory 201 can also be used to temporarily store various data that have been output or will be output.

[0137] The processor 202 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor 202 is generally used to control the overall operation of the computer device 200. In this embodiment, the processor 202 is used to run the computer-readable instructions or process data stored in the memory 201, such as running the computer-readable instructions of the rotary kiln state identification method based on Tucker decomposition.

[0138] The network interface 203 may include a wireless network interface or a wired network interface. The network interface 203 is generally used to establish a communication connection between the computer device 200 and other electronic devices.

[0139] In this embodiment, the extracted process feature data and image feature data are modally fused based on Tucker decomposition, which can effectively establish the complex relationship between the process parameters and the kiln condition video. Finally, the initial long short-term memory model is trained to obtain the long short-term memory model. The long short-term memory model is used to extract the time series characteristics of the kiln condition state, and the target kiln condition category is obtained, thereby improving the accuracy of the judgment of the kiln condition state category.

[0140] The present application also provides another embodiment, namely, providing a computer-readable storage medium, which stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to enable the at least one processor to perform the steps of the rotary kiln state identification method based on Tucker decomposition as described above.

[0141] In this embodiment, the extracted process feature data and image feature data are modally fused based on Tucker decomposition, which can effectively establish the complex relationship between the process parameters and the kiln condition video. Finally, the initial long short-term memory model is trained to obtain the long short-term memory model. The long short-term memory model is used to extract the time series characteristics of the kiln condition state, and the target kiln condition category is obtained, thereby improving the accuracy of the judgment of the kiln condition state category.

[0142] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0143] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A rotary kiln state recognition method based on Tucker decomposition, characterized in that: The steps include: Obtain corresponding process parameters and kiln condition videos according to preset time windows, construct process combination data corresponding to each time window based on the process parameters, construct corresponding firing zone splicing images based on the kiln condition videos corresponding to each time window, and respectively train a preset initial full convolution model and an initial residual neural model through all the process combination data and all the firing zone splicing images to obtain a full convolution model and a residual neural model; Performing a feature extraction operation on the process combination data based on the full convolution model to obtain process feature data, and performing a feature extraction operation on the fired belt splicing image based on the residual neural model to obtain image feature data; Obtaining a preset initial feature fusion model based on Tucker decomposition and a preset initial long short-term memory model, training the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data, and obtaining a feature fusion model and a long short-term memory model; Acquire target process parameters and target firing zone images in real time, perform feature extraction operations on the target process parameters and the target firing zone images according to the full convolution model and the residual neural model, respectively, obtain target process feature data and target image feature data, respectively, and input the target process feature data and the target image feature data into the feature fusion model and the long short-term memory model in sequence to obtain the target kiln condition category; The step of performing a feature extraction operation on the process combination data based on the full convolution model to obtain process feature data, and performing a feature extraction operation on the fired belt splicing image based on the residual neural model to obtain image feature data comprises: Removing the fully connected layer of the fully convolutional model to obtain a parameter feature extraction model, and inputting the process combination data into the parameter feature extraction model to obtain the process feature data; The fully connected layer of the residual neural model is removed to obtain an image feature extraction model, and the sintered belt spliced ​​image is input into the image feature extraction model to obtain the image feature data.

2. The rotary kiln state identification method based on Tucker decomposition according to claim 1 is characterized in that: The step of constructing process combination data corresponding to each time window based on the process parameters includes: Convert the process parameters corresponding to each of the time windows into an initial process matrix respectively; splicing the initial process matrix based on the time sequence to obtain a process splicing matrix, and performing a matrix transposition operation on the process splicing matrix to obtain a process transposed matrix; The process transposed matrix is ​​segmented according to the dimension of the process parameters to obtain the process combination data.

3. The rotary kiln state identification method based on Tucker decomposition according to claim 2 is characterized in that: The step of converting the process parameters corresponding to each of the time windows into an initial process matrix comprises: Check whether the process parameters have null values ​​or abnormal values; When the null value or the abnormal value exists, the value of the process parameter at the previous moment is retrieved to fill the null value or replace the abnormal value, so as to obtain the first process parameter; The first process parameters are normalized to obtain second process parameters, and the second process parameters are converted into the initial process matrix.

4. The rotary kiln state identification method based on Tucker decomposition according to claim 1 is characterized in that: The step of constructing a corresponding spliced ​​image of the firing zone based on the kiln condition video corresponding to each time window comprises: According to a preset frame extraction interval, the kiln condition video is subjected to a frame extraction operation to obtain firing band frame data; The fired belt frame data are spliced ​​according to a time sequence to generate the fired belt spliced ​​image.

5. The rotary kiln state identification method based on Tucker decomposition according to claim 4 is characterized in that: The step of splicing the burned belt frame data according to the time sequence to generate the burned belt spliced ​​image includes: Seamlessly splice the fired strip frame data according to the time sequence to obtain an initial fired strip spliced ​​image of a specified size; It is determined whether there is a missing part in the initial fired zone stitching image. If so, a filling operation is performed using a preset color to obtain the fired zone stitching image.

6. The rotary kiln state identification method based on Tucker decomposition according to claim 1 is characterized in that: The initial full convolution model includes a plurality of full convolution layers, and the full convolution layer includes a convolution layer and a nonlinear activation layer. The steps of respectively training the preset initial full convolution model and the initial residual neural model through all the process combination data and all the fired belt splicing images to obtain the full convolution model and the residual neural model include: Passing the process combination data through each of the full convolution layers, performing a downsampling operation on the process combination data through the convolution layer, and performing an activation operation on the process combination data through the Leaky_relu function of the nonlinear activation layer to obtain first classification data; Iteratively training the initial full convolution model based on the first classification data to obtain the full convolution model; The fired belt spliced ​​image is input into the initial residual neural model to obtain second classification data, and the initial residual neural model is iteratively trained based on the second classification data to obtain the residual neural model.

7. The rotary kiln state identification method based on Tucker decomposition according to claim 1 is characterized in that: The step of training the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data to obtain the feature fusion model and the long short-term memory model comprises: Inputting the process feature data and the image feature data into the initial feature fusion model, wherein the initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data to obtain fusion features; Inputting the fused features into the initial long short-term memory model to obtain an estimated kiln condition category; The initial feature fusion model and the initial long short-term memory model are iteratively trained based on the estimated kiln condition category to obtain the feature fusion model and the long short-term memory model.

8. The rotary kiln state identification method based on Tucker decomposition according to claim 7 is characterized in that: The initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data, and the step of obtaining fusion features includes: The initial feature fusion model performs Tucker decomposition and feature fusion on the process feature data and the image feature data through the following formula to obtain fusion features: y=((T c ×q T W q )×(v T W v ))×W o , where y represents the fusion feature, q T Represents the process characteristic data, v T represents the image feature data, Wq, Wv represent the preset feature mapping matrix, Tc represents the preset fusion matrix, and Wo represents the preset result mapping matrix.

9. A rotary kiln state recognition device based on Tucker decomposition, characterized in that: include: A training module, used for acquiring corresponding process parameters and kiln condition videos according to preset time windows, constructing process combination data corresponding to each time window based on the process parameters, constructing corresponding firing zone splicing images based on the kiln condition videos corresponding to each time window, and respectively training a preset initial full convolution model and an initial residual neural model through all the process combination data and all the firing zone splicing images to obtain a full convolution model and a residual neural model; A conversion module, used for performing a feature extraction operation on the process combination data based on the full convolution model to obtain process feature data, and performing a feature extraction operation on the fired belt splicing image based on the residual neural model to obtain image feature data; A fusion module, used to obtain a preset initial feature fusion model based on Tucker decomposition and a preset initial long short-term memory model, train the initial feature fusion model and the initial long short-term memory model based on the process feature data and the image feature data, and obtain a feature fusion model and a long short-term memory model; An input module, used for acquiring target process parameters and target firing zone images in real time, performing feature extraction operations on the target process parameters and the target firing zone images according to the full convolution model and the residual neural model, respectively, obtaining target process feature data and target image feature data, respectively, and inputting the target process feature data and the target image feature data into the feature fusion model and the long short-term memory model in sequence to obtain a target kiln condition category; The conversion module includes: a first extraction submodule and a second extraction submodule, wherein: The first extraction submodule is used to remove the fully connected layer of the fully convolutional model to obtain a parameter feature extraction model, and input the process combination data into the parameter feature extraction model to obtain the process feature data; The second extraction submodule is used to remove the fully connected layer of the residual neural model to obtain an image feature extraction model, and input the fired belt spliced ​​image into the image feature extraction model to obtain the image feature data.

10. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the rotary kiln state identification method based on Tucker decomposition as described in any one of claims 1 to 8 are implemented.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the rotary kiln state identification method based on Tucker decomposition as described in any one of claims 1 to 8.

Images