Roller kiln atmosphere prediction method and device, electronic equipment and storage medium

Abstract

The invention provides a roller kiln atmosphere prediction method and device, electronic equipment and a storage medium, and the method comprises the steps: carrying out the feature selection of an atmosphere initial data set based on a preset nonlinear dimensionality reduction algorithm, obtaining a dimensionality-reduced atmosphere initial data set, carrying out the coding of the dimensionality-reduced atmosphere initial data set, and obtaining a dimensionality-reduced atmosphere initial data set; the method comprises the following steps: coding an atmosphere initial data set to obtain a coded atmosphere initial data set, constructing a Transform prediction model based on the coded atmosphere initial data set, training the Transform prediction model based on a training data set to obtain a trained Transform prediction model, and optimizing the Transform prediction model based on a preset Bayesian optimization algorithm to obtain an optimized Transform prediction model. And predicting the atmosphere of the roller kiln based on the optimized Transform prediction model to obtain a prediction result of the atmosphere of the roller kiln.

Description

Technical Field

[0001] The present application relates to the field of smelting technology, and in particular to a roller kiln atmosphere prediction method, device, electronic equipment and storage medium. Background Art

[0002] The atmosphere within a roller hearth kiln is a key control parameter in the cathode material firing process. Its state is determined by the coupling of multiple parameters, including fan frequency, gate valve opening, and inlet air flow rate. Accurately predicting atmosphere changes is crucial for improving product quality, reducing energy consumption, and optimizing production processes. In the field of new energy materials, a quantitative dynamic model of this parameter-atmosphere relationship is urgently needed.

[0003] At present, the prediction of the kiln atmosphere, which is the working condition of releasing reaction heat and carbon dioxide during the firing process of positive electrode materials in the kiln, mainly relies on traditional physical field numerical simulation methods (such as computational fluid dynamics) to analyze the local airflow distribution, or uses a single algorithm (such as classic PID control) to adjust basic parameters.

[0004] However, the existing methods have low computational efficiency and weak generalization capabilities, making it difficult to meet the needs of high-precision prediction. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a roller kiln atmosphere prediction method, device, electronic device and storage medium. By reducing the dimension and encoding the roller kiln atmosphere historical data set, a Transformer prediction model is constructed according to the atmosphere historical data set, and the kiln atmosphere is predicted after the Transformer prediction model is trained and optimized, which improves the computational efficiency, and has stronger adaptability to various scenarios and improved generalization ability. At the same time, the prediction results of the atmosphere based on the constructed Transformer prediction model are more accurate and can meet the high-precision prediction requirements.

[0006] In a first aspect, an embodiment of the present application provides a method for predicting the atmosphere of a roller kiln, the method comprising:

[0007] Acquiring a historical atmosphere data set of the roller kiln atmosphere during the kiln firing process within a preset historical period, and preprocessing the historical atmosphere data set to obtain a preprocessed initial atmosphere data set;

[0008] Performing feature selection on the initial atmosphere data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced-dimensional initial atmosphere data set, and encoding the reduced-dimensional initial atmosphere data set to obtain an encoded initial atmosphere data set; wherein the encoded initial atmosphere data set is divided into a training data set and a test data set;

[0009] constructing a Transformer prediction model based on the encoded initial atmosphere data set, and training the Transformer prediction model based on the training data set to obtain a trained Transformer prediction model;

[0010] The Transformer prediction model is optimized based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model, and the roller kiln atmosphere is predicted based on the optimized Transformer prediction model to obtain a prediction result of the roller kiln atmosphere.

[0011] In a possible implementation, the performing feature selection on the initial atmosphere dataset based on a preset nonlinear dimensionality reduction algorithm includes:

[0012] Determine any first original sample point and second original sample point in the initial atmosphere data set, and obtain a first probability distribution of point-to-point adjacency in a high-dimensional space based on the first original sample point and the second original sample point;

[0013] Determine any third original sample point and fourth original sample point in the initial atmosphere data set, and obtain a second probability distribution of point-to-point proximity in a low-dimensional space based on the third original sample point and the fourth original sample point;

[0014] A cost function is constructed, and the first probability distribution and the second probability distribution are optimized based on the cost function so that the first probability distribution and the second probability distribution are consistent.

[0015] In a possible implementation, encoding the dimensionally reduced initial atmosphere dataset to obtain the encoded initial atmosphere dataset includes:

[0016] Obtaining a preset target encoder; wherein the target encoder includes a preliminary information extractor and a one-dimensional position information embedder, the preliminary information extractor is a feedforward, non-recurrent single-layer perceptron, and the one-dimensional position information embedder is a standard, learnable one-dimensional position information embedder;

[0017] The dimension-reduced initial atmosphere dataset is encoded based on the target encoder to obtain an encoded initial atmosphere dataset.

[0018] In a possible implementation, the target encoder is composed of an encoder and a decoder, and encoding the dimensionally reduced initial atmosphere dataset includes:

[0019] Determining an encoder transform defined by the encoder and a decoder transform defined by the decoder, and obtaining the target encoder based on the encoder transform and the decoder transform; wherein the encoder transform represents a transformation from a first parameter to a second parameter, and the decoder transform represents a transformation from the second parameter to the first parameter;

[0020] The encoder receives a first parameter input of the encoder transformation and maps an output result of the encoder to a second parameter of the encoder transformation, thereby obtaining a mapping result of the encoder; wherein the first parameter and the second parameter are determined by the reduced-dimensional atmosphere initial data set;

[0021] The first parameter is reconstructed to obtain a first reconstruction parameter, and the mapping result of the encoder is mapped to the first reconstruction parameter based on the decoder; wherein the shape of the first reconstruction parameter is the same as that of the first parameter.

[0022] In one possible implementation, the Transformer prediction model is determined by a multi-layer perceptron, a normalization layer, a normalization layer, and a multi-head attention mechanism, wherein the multi-head attention mechanism includes multiple independent attention models; and constructing the Transformer prediction model based on the encoded initial atmosphere dataset includes:

[0023] Determining target input data based on the encoded initial atmosphere data set, calculating a weighted sum of all values ​​in the sequence of the input data and an attention weight based on the attention model, and calculating an attention output of the attention model based on the weighted sum and the attention weight; wherein the target input data is determined by the initial atmosphere data set, the number of sections of the roller kiln during the kiln firing process, and the data dimension of each section;

[0024] The attention outputs of all attention models are concatenated to obtain the final output of the multi-head attention mechanism.

[0025] In a possible implementation, optimizing the Transformer prediction model includes:

[0026] Determine a corresponding objective function based on the hyperparameters of the Transformer prediction model, model the objective function using a Gaussian process, and obtain a corresponding posterior distribution function;

[0027] The expected improvement in Bayesian optimization is used as an acquisition function, a plurality of observation points obtained by sampling the initial atmosphere data set by the acquisition function is determined, and the hyperparameters are optimized based on the posterior distribution function and the observation points through the acquisition function.

[0028] In one possible implementation, the method further includes:

[0029] Obtaining a prediction result output by the Transformer prediction model for the test data set, determining an actual result of the test data set, and obtaining a multi-dimensional evaluation index calculation result based on a preset multi-dimensional evaluation index and an error between the prediction result and the actual result;

[0030] The prediction accuracy of the Transformer prediction model is evaluated based on the calculation result of the evaluation index. In response to the prediction accuracy of the Transformer prediction model meeting the preset prediction accuracy requirement, the Transformer prediction model is determined as the target Transformer prediction model, so as to predict the roller kiln atmosphere through the target Transformer prediction model.

[0031] In a second aspect, an embodiment of the present application further provides a roller kiln atmosphere prediction device, the device comprising:

[0032] an acquisition module, configured to acquire a historical atmosphere data set of the roller hearth kiln atmosphere during the kiln firing process within a preset historical period, and preprocess the historical atmosphere data set to obtain a preprocessed initial atmosphere data set;

[0033] a processing module, configured to perform feature selection on the initial atmosphere data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced-dimensional initial atmosphere data set, and encode the reduced-dimensional initial atmosphere data set to obtain an encoded initial atmosphere data set; wherein the encoded initial atmosphere data set is divided into a training data set and a test data set;

[0034] A training module, configured to construct a Transformer prediction model based on the encoded initial atmosphere data set, and train the Transformer prediction model based on the training data set to obtain a trained Transformer prediction model;

[0035] A prediction module is used to optimize the Transformer prediction model based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model, and to predict the roller kiln atmosphere based on the optimized Transformer prediction model to obtain a prediction result of the roller kiln atmosphere.

[0036] In a possible implementation, the processing module is specifically configured to:

[0037] Determine any first original sample point and second original sample point in the initial atmosphere data set, and obtain a first probability distribution of point-to-point adjacency in a high-dimensional space based on the first original sample point and the second original sample point;

[0038] Determine any third original sample point and fourth original sample point in the initial atmosphere data set, and obtain a second probability distribution of point-to-point proximity in a low-dimensional space based on the third original sample point and the fourth original sample point;

[0039] A cost function is constructed, and the first probability distribution and the second probability distribution are optimized based on the cost function so that the first probability distribution and the second probability distribution are consistent.

[0040] In a possible implementation, the processing module is specifically configured to:

[0041] Obtaining a preset target encoder; wherein the target encoder includes a preliminary information extractor and a one-dimensional position information embedder, the preliminary information extractor is a feedforward, non-recurrent single-layer perceptron, and the one-dimensional position information embedder is a standard, learnable one-dimensional position information embedder;

[0042] The dimension-reduced initial atmosphere dataset is encoded based on the target encoder to obtain an encoded initial atmosphere dataset.

[0043] In a possible implementation, the target encoder is composed of an encoder and a decoder, and the processing module is specifically configured to:

[0044] Determining an encoder transform defined by the encoder and a decoder transform defined by the decoder, and obtaining the target encoder based on the encoder transform and the decoder transform; wherein the encoder transform represents a transformation from a first parameter to a second parameter, and the decoder transform represents a transformation from the second parameter to the first parameter;

[0045] The encoder receives a first parameter input of the encoder transformation and maps an output result of the encoder to a second parameter of the encoder transformation, thereby obtaining a mapping result of the encoder; wherein the first parameter and the second parameter are determined by the reduced-dimensional atmosphere initial data set;

[0046] The first parameter is reconstructed to obtain a first reconstruction parameter, and the mapping result of the encoder is mapped to the first reconstruction parameter based on the decoder; wherein the shape of the first reconstruction parameter is the same as that of the first parameter.

[0047] In one possible implementation, the Transformer prediction model is determined by a multi-layer perceptron, a normalization layer, a normalization layer, and a multi-head attention mechanism, wherein the multi-head attention mechanism includes multiple independent attention models; the training module is specifically used to:

[0048] Determining target input data based on the encoded initial atmosphere data set, calculating a weighted sum of all values ​​in the sequence of the input data and an attention weight based on the attention model, and calculating an attention output of the attention model based on the weighted sum and the attention weight; wherein the target input data is determined by the initial atmosphere data set, the number of sections of the roller kiln during the kiln firing process, and the data dimension of each section;

[0049] The attention outputs of all attention models are concatenated to obtain the final output of the multi-head attention mechanism.

[0050] In a possible implementation, the prediction module is specifically configured to:

[0051] Determine a corresponding objective function based on the hyperparameters of the Transformer prediction model, model the objective function using a Gaussian process, and obtain a corresponding posterior distribution function;

[0052] The expected improvement in Bayesian optimization is used as an acquisition function, a plurality of observation points obtained by sampling the initial atmosphere data set by the acquisition function is determined, and the hyperparameters are optimized based on the posterior distribution function and the observation points through the acquisition function.

[0053] In one possible implementation, the roller kiln atmosphere prediction device of the present application further includes:

[0054] a calculation module, configured to obtain a prediction result output by the Transformer prediction model for the test data set, determine an actual result of the test data set, and obtain a calculation result of a multi-dimensional evaluation index based on a preset multi-dimensional evaluation index and an error between the prediction result and the actual result;

[0055] An evaluation module is used to evaluate the prediction accuracy of the Transformer prediction model based on the calculation results of the evaluation index, and in response to the prediction accuracy of the Transformer prediction model meeting the preset prediction accuracy requirements, the Transformer prediction model is determined as the target Transformer prediction model, so as to predict the roller kiln atmosphere through the target Transformer prediction model.

[0056] In a third aspect, an embodiment of the present application provides an electronic device comprising: a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the roller kiln atmosphere prediction method as described in any one of the first aspects.

[0057] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the roller kiln atmosphere prediction method described in any one of the first aspects are executed.

[0058] The embodiments of the present application provide a roller kiln atmosphere prediction method, device, electronic device and storage medium, which obtain an atmosphere historical data set of the roller kiln atmosphere within a preset historical period during the kiln firing process, preprocess the atmosphere historical data set to obtain a preprocessed atmosphere initial data set, perform feature selection on the atmosphere initial data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced atmosphere initial data set, encode the reduced atmosphere initial data set to obtain an encoded atmosphere initial data set, construct a Transformer prediction model based on the encoded atmosphere initial data set, train the Transformer prediction model based on a training data set to obtain a trained Transformer prediction model, optimize the Transformer prediction model based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model, predict the roller kiln atmosphere based on the optimized Transformer prediction model to obtain a prediction result of the roller kiln atmosphere. In this application, by reducing the dimension and encoding the historical data set of roller kiln atmosphere, a Transformer prediction model is constructed based on the historical data set of atmosphere, and the kiln atmosphere is predicted after the Transformer prediction model is trained and optimized, thereby improving the computational efficiency, and being more adaptable to various scenarios and improving the generalization ability. At the same time, the prediction results of the atmosphere based on the constructed Transformer prediction model are more accurate and can meet the high-precision prediction requirements.

[0059] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0061] Figure 1 This is a flow chart of a roller kiln atmosphere prediction method provided in accordance with an embodiment of the present application;

[0062] Figure 2 1 is a schematic structural diagram of a roller kiln atmosphere prediction device provided according to an embodiment of the present application;

[0063] Figure 3 This is a schematic structural diagram of an electronic device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of illustration and description and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowcharts can be implemented out of sequence, and steps without logical context can be reversed or implemented simultaneously. In addition, those skilled in the art, under the guidance of the contents of this application, can add one or more other operations to the flowchart, or remove one or more operations from the flowchart.

[0065] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0066] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the existence of the features declared thereafter, but does not exclude the addition of other features.

[0067] Considering that the atmosphere within a roller hearth kiln is a key control parameter in the cathode material firing process, its state is determined by the coupling of multiple parameters such as fan frequency, gate valve opening, and intake air flow. Accurately predicting atmosphere changes is crucial for improving product quality, reducing energy consumption, and optimizing production processes. In particular, in the field of new energy materials, a quantitative dynamic model of this parameter-atmosphere relationship is urgently needed.

[0068] Currently, the prediction of kiln atmosphere conditions—the release of reaction heat and carbon dioxide during the firing of cathode materials—relies primarily on traditional physical field numerical simulation methods (such as computational fluid dynamics) to analyze local airflow distribution, or on the use of single algorithms (such as classic PID control) to adjust basic parameters. However, existing methods suffer from low computational efficiency and weak generalization capabilities, making them incapable of meeting the requirements for high-precision predictions.

[0069] To address this problem, the present application provides a roller kiln atmosphere prediction method, device, electronic device and storage medium. By reducing the dimension and encoding the roller kiln atmosphere historical data set, a Transformer prediction model is constructed based on the atmosphere historical data set. After training and optimizing the Transformer prediction model, the kiln atmosphere is predicted, which improves the computational efficiency, and has stronger adaptability to various scenarios and improved generalization ability. At the same time, the atmosphere prediction results based on the constructed Transformer prediction model are more accurate and can meet the high-precision prediction requirements.

[0070] Figure 1 It is a flow chart of a roller kiln atmosphere prediction method provided according to an embodiment of the present application.

[0071] like Figure 1 As shown, the roller kiln atmosphere prediction method of the embodiment of the present application may specifically include:

[0072] S101 , obtaining a historical atmosphere dataset of a roller hearth kiln atmosphere during a kiln firing process within a preset historical period, and preprocessing the historical atmosphere dataset to obtain a preprocessed initial atmosphere dataset.

[0073] S102 : performing feature selection on the initial atmosphere data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced-dimensional initial atmosphere data set, and encoding the reduced-dimensional initial atmosphere data set to obtain an encoded initial atmosphere data set.

[0074] S103: constructing a Transformer prediction model based on the encoded initial atmosphere data set, and training the Transformer prediction model based on the training data set to obtain a trained Transformer prediction model.

[0075] S104. Optimizing the Transformer prediction model based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model; predicting the roller hearth kiln atmosphere based on the optimized Transformer prediction model to obtain a prediction result of the roller hearth kiln atmosphere.

[0076] In the above-mentioned roller kiln atmosphere prediction method, the dimension of the roller kiln atmosphere historical data set is reduced and encoded, and a Transformer prediction model is constructed based on the atmosphere historical data set. After the Transformer prediction model is trained and optimized, the kiln atmosphere is predicted, which improves the computational efficiency, makes it more adaptable to various scenarios, and improves the generalization ability. At the same time, the prediction results of the atmosphere based on the constructed Transformer prediction model are more accurate and can meet the high-precision prediction requirements.

[0077] The above exemplary steps of the embodiment of the present application are described below with reference to specific examples:

[0078] S101 , obtaining a historical atmosphere dataset of a roller hearth kiln atmosphere during a kiln firing process within a preset historical period, and preprocessing the historical atmosphere dataset to obtain a preprocessed initial atmosphere dataset.

[0079] It should be noted that this application takes the firing process of positive electrode materials as an example. The sintering process involves many complex factors, including fan frequency (blower, exhaust fan), flow intake in different sections (active, passive, bottom, etc.), gate valve opening, etc. These factors jointly determine the dynamic changes of furnace pressure and oxygen concentration.

[0080] In the embodiment of the present application, the atmosphere historical data set is a historical data set of the atmosphere in each zone of the roller kiln. The atmosphere historical data set includes at least the fan frequency, the air intake flow rate of each section of the kiln and the gate valve opening. Specifically, it is the operating frequency of each fan of the kiln, the air intake flow rate of each temperature zone and the gate valve opening; the atmosphere historical data set of the roller kiln atmosphere during the kiln firing process is obtained, and the atmosphere historical data set is preprocessed to obtain the preprocessed atmosphere initial data set for subsequent processing.

[0081] S102 , performing feature selection on the initial atmosphere data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced-dimensional initial atmosphere data set, and encoding the reduced-dimensional initial atmosphere data set to obtain an encoded initial atmosphere data set.

[0082] In an embodiment of the present application, the encoded initial atmosphere data set is divided into a training data set and a test data set, and a nonlinear dimensionality reduction algorithm, namely the SNE (Distributed Stochastic Neighbor Embedding) algorithm, is used to perform feature selection on the initial atmosphere data set to reduce the dimension of the initial atmosphere data set to obtain the reduced-dimensional initial atmosphere data set, and the reduced-dimensional initial atmosphere data set is encoded to obtain the encoded initial atmosphere data set for subsequent processing.

[0083] It should be noted that the corresponding partitioning strategy can be determined based on the actual operating condition information of the roller kiln atmosphere, and the initial atmosphere dataset after dimensionality reduction can be divided into a training dataset and a test dataset based on the partitioning strategy. For example, in order to accurately evaluate the generalization performance of the model, the initial atmosphere dataset is divided into a training dataset and a test dataset in a certain ratio (such as 8:2 or 7:3). It can also be divided in chronological order, for example, the first 70% of the initial atmosphere dataset is used for training, and the last 30% of the initial atmosphere dataset is used for testing to prevent the leakage of time series data. In short, the appropriate partitioning strategy is selected based on factors such as the scale and distribution of the actual operating condition data.

[0084] Optionally, performing feature selection on the initial atmosphere dataset based on a preset nonlinear dimensionality reduction algorithm may include the following steps:

[0085] a1. Determine any first original sample point and second original sample point in the initial atmosphere data set, and obtain a first probability distribution of point-to-point proximity in a high-dimensional space based on the first original sample point and the second original sample point.

[0086] For example, determine any two original sample points x in the initial atmosphere data set i and x j , respectively represent the first original sample point and the second original sample point, based on the two original sample points x i and x j Define its first probability distribution p about the adjacency of points in high-dimensional space.

[0087] In addition, in high-dimensional space, Gaussian kernel is generally used to calculate conditional probability. The first probability distribution p is calculated by the following formula:

[0088]

[0089] Among them, p i,j represents the first probability distribution, x i represents the first original sample point (i.e., for high-dimensional space, the i-th sample point in the initial atmosphere data set), x j represents the second original sample point (i.e., the jth sample point in the initial atmosphere data set), di,j Represents x i and x j The degree of dissimilarity in high-dimensional space. Specifically, d i,j Same as d i,k , represents the dissimilarity between the i-th sample point and the j / k-th sample point of the initial atmosphere dataset, and k is used to represent the accumulation.

[0090] It can be added here that d i,j The specific method is to use the scaled squared Euclidean distance. Therefore, x i and x j The degree of dissimilarity d in high-dimensional space i,j Calculated by the following formula:

[0091]

[0092] Among them, σ i It represents the width of the Gaussian kernel centered on sample i. It is an adjustable hyperparameter that can be adjusted individually for each sample point i. It is usually obtained by setting a "perplexity" value and performing a binary search to make the neighborhood size around different points adaptive to balance the preservation of local and global structures.

[0093] a2. Determine any third original sample point and fourth original sample point in the initial atmosphere data set, and obtain a second probability distribution of point-to-point proximity in the low-dimensional space based on the third original sample point and the fourth original sample point.

[0094] At the same time, define any third original sample point y in the initial atmosphere data set i and the fourth original sample point y j , based on point y i and point y j The second probability distribution q of the adjacent points in the low-dimensional space is obtained. In the low-dimensional space, SNE also defines the conditional probability with a Gaussian kernel, but no scaling is performed at this time, that is, the Gaussian kernel width is set to 0. The second probability distribution q is calculated by the following formula:

[0095]

[0096] Among them, y i represents the third original sample point (i.e., the i-th sample point in the initial atmosphere dataset for low-dimensional space), y j represents the fourth original sample point (i.e., the jth sample point in the atmosphere initial data set for low-dimensional space), q i,j represents the second probability distribution.

[0097] a3. Construct a cost function, and optimize the first probability distribution and the second probability distribution based on the cost function to make the first probability distribution and the second probability distribution consistent.

[0098] Specifically, by constructing a cost function so that p i,j The distribution and q i,j As consistent as possible, the kernel function used is KL divergence, also known as relative entropy. The cost function mentioned above is the following formula (where KL represents the kernel function of KL divergence):

[0099]

[0100] Among them, P i and Q i Represents the distribution of sample points.

[0101] It can be added that the above cost function needs to be calculated with respect to y i The first-order partial derivative of is optimized, so its partial derivative is:

[0102]

[0103] Obviously, this cost function is non-convex. Therefore, for large amounts of data, the Gauss-Newton algorithm or the Levenberg-Marquardt algorithm is used, taking into account computational speed requirements. For smaller amounts of data, or when computing resources are sufficient, the simulated annealing algorithm can be used to find the global optimal solution. In short, the above algorithms achieve data dimensionality reduction.

[0104] Specifically, the process of the SNE algorithm is as follows:

[0105] 1. For each point x in high-dimensional space i , pre-set σ i , then traverse all other points (as j), calculate p i,j .

[0106] 2. Randomly initialize all low-dimensional points {y i}.

[0107] 3. Calculate the low-dimensional similarity q i,j .

[0108] 4. Use the above formula to calculate the cost C, as well as the first-order partial derivative, and use the above-mentioned algorithm for all {y i} for optimization.

[0109] 5. Repeat steps 3 and 4 until convergence.

[0110] It should be noted that encoding the dimensionally reduced initial atmosphere dataset to obtain the encoded initial atmosphere dataset includes the following steps:

[0111] b1. Get the preset target encoder.

[0112] Among them, the target encoder (AutoEncoder) includes a preliminary information extractor and a one-dimensional position information embedder. The preliminary information extractor is a feedforward, non-cyclic single-layer perceptron, and the one-dimensional position information embedder is a standard, learnable one-dimensional position information embedder.

[0113] b2. Encode the reduced-dimensional initial atmosphere dataset based on the target encoder to obtain an encoded initial atmosphere dataset.

[0114] Among them, the target encoder consists of an encoder and a decoder.

[0115] Optionally, when encoding the initial atmosphere data set after dimensionality reduction, determine the encoder transform defined by the encoder and the decoder transform defined by the decoder, and obtain the target encoder based on the encoder transform and the decoder transform; based on the encoder, accept the first parameter input of the encoder transform and map the output result of the encoder to the second parameter of the encoder transform to obtain the mapping result of the encoder; reconstruct the first parameter to obtain a first reconstruction parameter, and map the mapping result of the encoder to the first reconstruction parameter based on the decoder.

[0116] The encoder transformation representation is a transformation from a first parameter to a second parameter, and the decoder transformation representation is a transformation from a second parameter to a first parameter; the first parameter and the second parameter are determined by the reduced-dimensional atmosphere initial data set; and the shape of the first reconstruction parameter is the same as the first parameter.

[0117] Specifically, the target encoder AutoEncoder consists of an encoder and a decoder, which are defined as transformations φ and ψ, respectively, such that:

[0118]

[0119] The encoder takes an input x∈χ of the first parameter of the transformation φ and maps it to the second parameter of the transformation ψ get:

[0120] h=σ(Wx+b)

[0121] It should be noted that W here is the model parameter, b is the additional learnable bias, and σ is the activation function. h is the output of the encoder and serves as the input of the decoder during training.

[0122] The decoder maps the encoder's mapping result h to the first reconstruction parameter, i.e., reconstructing x' (which has the same shape as x), and obtains:

[0123] x′=σ′(W′h+b′)

[0124] The decoder's σ′, W′, b′ is independent of the encoder's σ, W, b. Here, σ′, W′, b′ are the same as the encoder's, but the internal parameters are independent of the encoder.

[0125] It can be added here that a one-dimensional position information embedder is used to indicate to which segment the current input data belongs. A standard information embedder is usually an additional learnable bias.

[0126] S103, constructing a Transformer prediction model based on the encoded initial atmosphere data set, and training the Transformer prediction model based on the training data set to obtain a trained Transformer prediction model.

[0127] It should be noted that factors such as the air flow rate and gate valve opening in different sections during sintering will jointly influence the final prediction results. Conventional neural networks usually consider each input separately. Therefore, a network model that can consider the interactions between different sections will have better performance.

[0128] In an embodiment of the present application, the required Transformer prediction model is constructed based on the encoded atmosphere initial data set, and the Transformer prediction model is trained based on the training data set divided by the atmosphere initial data set to obtain a trained Transformer prediction model for subsequent processing.

[0129] Among them, the Transformer prediction model is determined by a multi-layer perceptron, a standardization layer, a normalization layer and a multi-head attention mechanism, which includes multiple independent attention models.

[0130] Optionally, the target input data is determined based on the encoded initial atmosphere dataset, the weighted sum of all values ​​in the input data sequence and the attention weights are calculated based on the attention model, and the attention output of the attention model is calculated based on the weighted sum and the attention weights. The attention outputs of all attention models are concatenated to obtain the final output of the multi-head attention mechanism. The target input data is determined by the initial atmosphere dataset, the number of roller kiln sections during the kiln firing process, and the data dimension of each section.

[0131] Specifically, the Transformer network's attention mechanism effectively meets these requirements. The attention mechanism comprehensively evaluates input data from different segments, assesses their mutual influence, and ultimately presents a comprehensive assessment. The following definitions are given for the Transformer formula:

[0132] Multilayer Perceptron MLP(z)=σ(Wz+b)

[0133] Normalization layer

[0134] Normalization layer

[0135] Multi-head attention mechanism MSA(z)

[0136] Among them, z represents the input of this layer, and the multi-head attention mechanism (MSA) is based on the qkv self-attention model (SA). For an input data z∈R D×D , N is the number of segments (different temperature segments), D is the data dimension of each segment, R represents the real number space, that is, the dimension of the input z, and the attention model (SA) calculates the weighted sum of all values ​​(value) v in the sequence of input data. Attention weight A ij Two elements of a sequence based on the input data and their respective queries q i and keyword (key) k j Indicates that D h is the dimension of the attention model, thus the attention weight A ij The weighted sum v is the output of the attention model (SA). In general, it is as follows:

[0137]

[0138] SA(z)=Av

[0139] Among them, U represents the learnable parameter, and its dimension has been explained here, which is D*3D h In other words, here we multiply the input z (z is N*D) and U by matrix to get D*3D h , and then split into three groups q, k, v, each group is N*D h , where N is the number of segments mentioned above, so each group can be regarded as N segments of length D h This vector is called a token.

[0140] It can be understood that the multi-head attention mechanism is an extension of the attention mechanism. The multi-head attention mechanism contains k independent attention modules (or attention models SA), called "heads", and their outputs are spliced ​​together as the final output of the multi-head attention mechanism. Therefore, the formula of the multi-head attention mechanism is:

[0141]

[0142] It can be added here that, assuming the input data is [x1; x2; ...; x N ], then the first layer after the target encoder AutoEncoder is:

[0143] z0=[E(x1);E(x2);…;E(x N )]+E pos

[0144] Among them, E(x) represents the encoder and decoder in AutoEncoder, E pos Represents the position information embedder. In the remaining layers of the Transformer model, the output of each layer can be generally expressed as:

[0145] z′ l =MSA(LN(z l-1 ))+z l-1

[0146] z l =MLPLLN(z′ l ))+z′ l

[0147] The output of the last layer of the Transformer model is

[0148]

[0149] It should be noted that the Transformer structure is built by many layers of the same "block". Specifically, x is the input, then x->autoeEncoder->z_0->block1->z_1->block2->...->output, and the block structure of each layer is the same. Here, the block structure of each layer is the above-mentioned z' and z. Given the output z_l-1 of the previous layer, first calculate z_l', and then calculate z_l as the output. Among them, the last layer of the Transformer model, that is, the output of the last layer of the last block, only takes the first token. Simply put, the output dimension is N*D h , here we take the first D h .

[0150] After the Transformer model outputs the result y, an additional multi-layer perceptron is used to produce the final result, namely:

[0151] prediction = MLP(y)

[0152] In short, this is an overall description of the Transformer model's processing flow for input data.

[0153] S104, optimizing the Transformer prediction model based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model, and predicting the roller hearth kiln atmosphere based on the optimized Transformer prediction model to obtain a prediction result of the roller hearth kiln atmosphere.

[0154] In an embodiment of the present application, the prediction result includes at least the predicted furnace pressure and the predicted oxygen concentration. The hyperparameter optimization of the Transformer prediction model is performed based on the Bayesian Optimization algorithm to obtain an optimized Transformer prediction model. The roller kiln atmosphere is predicted based on the optimized Transformer prediction model to obtain a prediction result of the roller kiln atmosphere, thereby realizing the prediction of the roller kiln atmosphere.

[0155] It should be noted that the Bayesian optimization algorithm is used to tune the hyperparameters of the model. Compared with the traditional GridSearch or genetic algorithm, Bayesian optimization is an approximation idea, which can play a good role when the calculation is very complex and the number of iterations is high.

[0156] Optionally, when optimizing the Transformer prediction model, the corresponding objective function is determined based on the hyperparameters of the Transformer prediction model, and the objective function is modeled using a Gaussian process to obtain the corresponding posterior distribution function; the expected improvement in Bayesian optimization is used as the acquisition function, and multiple observation points obtained by sampling the initial atmosphere data set by the acquisition function are determined, and the hyperparameters are optimized based on the posterior distribution function and the observation points through the acquisition function.

[0157] Specifically, Bayesian optimization uses a Gaussian process to model the objective function of the hyperparameters, thereby obtaining their posterior distribution. Bayesian optimization requires the use of an acquisition function to find the function for the next sampling point. Assuming that the posterior distribution, i.e., the distribution function f(x), has been obtained from the Gaussian process, the expected improvement (EI) is used as the acquisition function. Assuming that the hyperparameter to be optimized is θ, the function can be written as:

[0158]

[0159] Among them, {x n ,y n} is the first N observation points that have been sampled based on the initial atmosphere data set, σ 2 (x; {x n ,y n},θ) is the prediction variance function, μ(x; {x n ,y n},θ) is the predicted mean function, and Φ(·) is the cumulative distribution function of the distribution. refers to the Gaussian distribution. In addition, γ(x) is defined as:

[0160]

[0161] The roller kiln atmosphere prediction method provided in the embodiment of the present application obtains an atmosphere historical data set of the roller kiln atmosphere within a preset historical period during the kiln firing process, and preprocesses the atmosphere historical data set to obtain a preprocessed atmosphere initial data set, performs feature selection on the atmosphere initial data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced dimensionality atmosphere initial data set, encodes the reduced dimensionality atmosphere initial data set to obtain an encoded atmosphere initial data set, constructs a Transformer prediction model based on the encoded atmosphere initial data set, trains the Transformer prediction model based on the training data set to obtain a trained Transformer prediction model, optimizes the Transformer prediction model based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model, predicts the roller kiln atmosphere based on the optimized Transformer prediction model to obtain a prediction result of the roller kiln atmosphere. The roller kiln atmosphere prediction method of the present application reduces the dimension and encodes the roller kiln atmosphere historical data set, constructs a Transformer prediction model based on the atmosphere historical data set, and predicts the kiln atmosphere after training and optimizing the Transformer prediction model, thereby improving computational efficiency, being more adaptable to various scenarios, and improving generalization ability. At the same time, the prediction results of the atmosphere based on the constructed Transformer prediction model are more accurate and can meet the high-precision prediction requirements.

[0162] It should be noted that the prediction accuracy of the Transformer prediction model can be evaluated.

[0163] Furthermore, the prediction results output by the Transformer prediction model for the test data set are obtained, the actual results of the test data set are determined, and the multi-dimensional evaluation index calculation results are obtained based on the preset multi-dimensional evaluation indicators and the error between the prediction results and the actual results; the prediction accuracy of the Transformer prediction model is evaluated based on the evaluation index calculation results, and in response to the prediction accuracy of the Transformer prediction model meeting the preset prediction accuracy requirements, the Transformer prediction model is determined as the target Transformer prediction model, so as to predict the roller kiln furnace atmosphere through the target Transformer prediction model.

[0164] Optionally, in response to the prediction accuracy of the Transformer prediction model meeting a preset prediction accuracy requirement, the training of the Transformer prediction model continues.

[0165] Among them, the multi-dimensional evaluation indicators include at least:

[0166] MAE, MAE is defined as the average of the absolute values ​​of the errors:

[0167]

[0168] MSE, MSE is defined as the average of the squares of the errors:

[0169]

[0170] RMSE, RMSE is the square root of MSE:

[0171]

[0172] R2 is used to measure how well the model fits the data. It is defined as 1 minus the ratio of the sum of squared residuals (SSR) to the sum of squares (SST):

[0173]

[0174] It should be noted that this application builds a prediction model based on a deep neural network (mainly composed of CNN Embedding, AutoEncoder, Transformer and subsequent MLP predictor) for the firing process of positive electrode materials to predict the final furnace pressure and oxygen concentration. In order to better capture these complex relationships, this application uses a combination of the following modules:

[0175] A) CNN Embedding Layer: solves the strong correlation problem between "active-passive-bottom";

[0176] B) Auto Encoding Layer: Further compresses and abstracts features through nonlinear dimensionality reduction methods (AutoEncoder);

[0177] C) Transformer: Uses a multi-head self-attention mechanism to capture the interaction between different segments;

[0178] D) MLP predictor: maps the output of the Transformer to furnace pressure and oxygen concentration.

[0179] After determining the CNN Embedding, AutoEncoder, Transformer, and MLP predictor, the performance or prediction accuracy of the entire Transformer prediction model was evaluated. The above multi-dimensional evaluation indicators were calculated based on the errors between the predicted and actual furnace pressure and oxygen concentration.

[0180] To verify the contribution of each module to prediction performance, this application designed the following ablation experiment:

[0181] 1. Remove CNN Embedding: Directly flatten the "active-passive-bottom" parameters and input them into the subsequent network, replacing convolutional feature extraction with simple vector inputs. Compare the MSE / MAE differences between this model and the full model under the same hyperparameter settings.

[0182] 2. Remove the AutoEncoder: Instead of using the AutoEncoder for nonlinear dimensionality reduction, the original data is directly fed into the Transformer. Compare the prediction performance of this model with that of the full model at the same training time.

[0183] 3. Using only a single attention head (Single-Head Attention): Compare the results of the multi-head attention mechanism (Multi-Head Attention) with single-head attention to examine the impact of multi-head attention on capturing segment interactions.

[0184] As a result, we obtained the experimental results shown in Tables 1 and 2 below, which prove the impact of each module on model accuracy under various evaluation indicators and declare the importance of these modules in improving model prediction accuracy:

[0185] Table 1

[0186]

[0187] Table 2

[0188]

[0189] Figure 2 Schematic diagram of the structure of the roller kiln atmosphere prediction device provided in accordance with the embodiment of the present application; Figure 2 As shown, the roller kiln atmosphere prediction device 200 of the embodiment of the present application may specifically include:

[0190] The acquisition module 201 is used to acquire a historical atmosphere dataset of the roller hearth kiln atmosphere during the kiln firing process within a preset historical period, and preprocess the historical atmosphere dataset to obtain a preprocessed initial atmosphere dataset.

[0191] The processing module 202 is used to perform feature selection on the initial atmosphere data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced-dimensional initial atmosphere data set, and encode the reduced-dimensional initial atmosphere data set to obtain an encoded initial atmosphere data set; wherein the encoded initial atmosphere data set is divided into a training data set and a test data set.

[0192] The training module 203 is used to build a Transformer prediction model based on the encoded initial atmosphere data set, and train the Transformer prediction model based on the training data set to obtain a trained Transformer prediction model.

[0193] The prediction module 204 is used to optimize the Transformer prediction model based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model, and predict the roller kiln atmosphere based on the optimized Transformer prediction model to obtain a prediction result of the roller kiln atmosphere.

[0194] In a possible implementation, the processing module is specifically configured to:

[0195] Determine any first original sample point and second original sample point in the initial atmosphere data set, and obtain a first probability distribution of point-to-point proximity in a high-dimensional space based on the first original sample point and the second original sample point;

[0196] Determine any third original sample point and fourth original sample point in the initial atmosphere data set, and obtain a second probability distribution of point-to-point proximity in the low-dimensional space based on the third original sample point and the fourth original sample point;

[0197] A cost function is constructed, and the first probability distribution and the second probability distribution are optimized based on the cost function so that the first probability distribution and the second probability distribution are consistent.

[0198] In a possible implementation, the processing module is specifically configured to:

[0199] Obtain a preset target encoder; wherein the target encoder includes a preliminary information extractor and a one-dimensional position information embedder, the preliminary information extractor is a feedforward, non-recurrent single-layer perceptron, and the one-dimensional position information embedder is a standard, learnable one-dimensional position information embedder;

[0200] The dimensionally reduced initial atmosphere dataset is encoded based on the target encoder to obtain an encoded initial atmosphere dataset.

[0201] In one possible implementation, the target encoder is composed of an encoder and a decoder, and the processing module is specifically configured to:

[0202] Determining an encoder transform defined by an encoder and a decoder transform defined by a decoder, and obtaining a target encoder based on the encoder transform and the decoder transform; wherein the encoder transform represents a transformation from a first parameter to a second parameter, and the decoder transform represents a transformation from the second parameter to the first parameter;

[0203] Based on the encoder accepting a first parameter input of the encoder transformation and mapping the output result of the encoder to a second parameter of the encoder transformation, a mapping result of the encoder is obtained; wherein the first parameter and the second parameter are determined by the initial atmosphere data set after dimensionality reduction;

[0204] The first parameter is reconstructed to obtain a first reconstruction parameter, and a mapping result of the encoder is mapped to the first reconstruction parameter based on the decoder; wherein the shape of the first reconstruction parameter is the same as that of the first parameter.

[0205] In one possible implementation, the Transformer prediction model is determined by a multi-layer perceptron, a normalization layer, a normalization layer, and a multi-head attention mechanism, wherein the multi-head attention mechanism includes multiple independent attention models; and a training module is specifically configured to:

[0206] Determine target input data based on the encoded initial atmosphere data set, calculate a weighted sum of all values ​​in the sequence of the input data and an attention weight based on the attention model, and calculate an attention output of the attention model based on the weighted sum and the attention weight; wherein the target input data is determined by the initial atmosphere data set, the number of sections of the roller kiln during the kiln firing process, and the data dimension of each section;

[0207] The attention outputs of all attention models are concatenated to obtain the final output of the multi-head attention mechanism.

[0208] In a possible implementation, the prediction module is specifically configured to:

[0209] The corresponding objective function is determined based on the hyperparameters of the Transformer prediction model, and the Gaussian process is used to model the objective function to obtain the corresponding posterior distribution function;

[0210] The expected improvement in Bayesian optimization is used as the acquisition function, and multiple observation points obtained by sampling the initial atmosphere data set are determined. The hyperparameters are optimized based on the posterior distribution function and the observation points through the acquisition function.

[0211] In one possible implementation, the roller kiln atmosphere prediction device of the present application further includes:

[0212] The calculation module is used to obtain the prediction results output by the Transformer prediction model for the test data set, determine the actual results of the test data set, and obtain the multi-dimensional evaluation index calculation results based on the preset multi-dimensional evaluation indicators and the error between the prediction results and the actual results;

[0213] An evaluation module is used to evaluate the prediction accuracy of the Transformer prediction model based on the calculation results of the evaluation index. In response to the prediction accuracy of the Transformer prediction model meeting the preset prediction accuracy requirements, the Transformer prediction model is determined as the target Transformer prediction model to predict the roller kiln atmosphere through the target Transformer prediction model.

[0214] The roller kiln atmosphere prediction device provided in the embodiment of the present application obtains an atmosphere historical data set of the roller kiln atmosphere within a preset historical period during the kiln firing process, and preprocesses the atmosphere historical data set to obtain a preprocessed atmosphere initial data set, performs feature selection on the atmosphere initial data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced dimensionality atmosphere initial data set, encodes the reduced dimensionality atmosphere initial data set to obtain an encoded atmosphere initial data set, constructs a Transformer prediction model based on the encoded atmosphere initial data set, trains the Transformer prediction model based on the training data set to obtain a trained Transformer prediction model, optimizes the Transformer prediction model based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model, predicts the roller kiln atmosphere based on the optimized Transformer prediction model to obtain a prediction result of the roller kiln atmosphere. The roller kiln atmosphere prediction device of the present application reduces the dimension and encodes the roller kiln atmosphere historical data set, constructs a Transformer prediction model based on the atmosphere historical data set, and predicts the kiln atmosphere after training and optimizing the Transformer prediction model, thereby improving computing efficiency, being more adaptable to various scenarios, and improving generalization ability. At the same time, the prediction results of the atmosphere based on the constructed Transformer prediction model are more accurate and can meet the high-precision prediction needs.

[0215] like Figure 3 As shown, an electronic device 300 provided in an embodiment of the present application includes: a processor 301, a memory 302 and a bus, wherein the memory 302 stores machine-readable instructions executable by the processor 301. When the electronic device is running, the processor 301 communicates with the memory 302 through the bus, and the processor 301 executes the machine-readable instructions to perform the steps of the above-mentioned roller kiln atmosphere prediction method.

[0216] Specifically, the memory 302 and processor 301 can be general-purpose memories and processors, which are not specifically limited here. When the processor 301 runs the computer program stored in the memory 302, the roller kiln atmosphere prediction method can be executed.

[0217] Corresponding to the above-mentioned roller kiln atmosphere prediction method, an embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned roller kiln atmosphere prediction method are executed.

[0218] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and device described above can refer to the corresponding process in the method embodiment, and will not be repeated in this application. In the several embodiments provided in this application, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0219] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network elements. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0220] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0221] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the deployment method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a ROM, a RAM, a magnetic disk, or an optical disk.

[0222] The above are only specific embodiments of the present application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A roller kiln atmosphere prediction method, characterized in that: The method comprises: Acquiring a historical atmosphere data set of the roller kiln atmosphere during the kiln firing process within a preset historical period, and preprocessing the historical atmosphere data set to obtain a preprocessed initial atmosphere data set; Performing feature selection on the initial atmosphere data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced-dimensional initial atmosphere data set, and encoding the reduced-dimensional initial atmosphere data set to obtain an encoded initial atmosphere data set; wherein the encoded initial atmosphere data set is divided into a training data set and a test data set; constructing a Transformer prediction model based on the encoded initial atmosphere data set, and training the Transformer prediction model based on the training data set to obtain a trained Transformer prediction model; The Transformer prediction model is optimized based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model, and the roller kiln atmosphere is predicted based on the optimized Transformer prediction model to obtain a prediction result of the roller kiln atmosphere.

2. The method according to claim 1, characterized in that The feature selection of the initial atmosphere data set based on a preset nonlinear dimensionality reduction algorithm includes: Determine any first original sample point and second original sample point in the initial atmosphere data set, and obtain a first probability distribution of point-to-point adjacency in a high-dimensional space based on the first original sample point and the second original sample point; Determine any third original sample point and fourth original sample point in the initial atmosphere data set, and obtain a second probability distribution of point-to-point proximity in a low-dimensional space based on the third original sample point and the fourth original sample point; A cost function is constructed, and the first probability distribution and the second probability distribution are optimized based on the cost function so that the first probability distribution and the second probability distribution are consistent.

3. The method according to claim 2, characterized in that The step of encoding the dimensionally reduced initial atmosphere dataset to obtain the encoded initial atmosphere dataset comprises: Obtaining a preset target encoder; wherein the target encoder includes a preliminary information extractor and a one-dimensional position information embedder, the preliminary information extractor is a feedforward, non-recurrent single-layer perceptron, and the one-dimensional position information embedder is a standard, learnable one-dimensional position information embedder; The dimension-reduced initial atmosphere dataset is encoded based on the target encoder to obtain an encoded initial atmosphere dataset.

4. The method according to claim 3, characterized in that The target encoder is composed of an encoder and a decoder, and encoding the initial atmosphere dataset after dimension reduction includes: Determining an encoder transform defined by the encoder and a decoder transform defined by the decoder, and obtaining the target encoder based on the encoder transform and the decoder transform; wherein the encoder transform represents a transformation from a first parameter to a second parameter, and the decoder transform represents a transformation from the second parameter to the first parameter; The encoder receives a first parameter input of the encoder transformation and maps an output result of the encoder to a second parameter of the encoder transformation, thereby obtaining a mapping result of the encoder; wherein the first parameter and the second parameter are determined by the reduced-dimensional atmosphere initial data set; The first parameter is reconstructed to obtain a first reconstruction parameter, and the mapping result of the encoder is mapped to the first reconstruction parameter based on the decoder; wherein the shape of the first reconstruction parameter is the same as that of the first parameter.

5. The method according to claim 4, characterized in that The Transformer prediction model is determined by a multi-layer perceptron, a standardization layer, a normalization layer, and a multi-head attention mechanism, wherein the multi-head attention mechanism includes multiple independent attention models; the Transformer prediction model is constructed based on the encoded initial atmosphere dataset, including: Determining target input data based on the encoded initial atmosphere data set, calculating a weighted sum of all values ​​in the sequence of the input data and an attention weight based on the attention model, and calculating an attention output of the attention model based on the weighted sum and the attention weight; wherein the target input data is determined by the initial atmosphere data set, the number of sections of the roller kiln during the kiln firing process, and the data dimension of each section; The attention outputs of all attention models are concatenated to obtain the final output of the multi-head attention mechanism.

6. The method according to claim 5, characterized in that The optimizing the Transformer prediction model includes: Determine a corresponding objective function based on the hyperparameters of the Transformer prediction model, model the objective function using a Gaussian process, and obtain a corresponding posterior distribution function; The expected improvement in Bayesian optimization is used as an acquisition function, a plurality of observation points obtained by sampling the initial atmosphere data set by the acquisition function is determined, and the hyperparameters are optimized based on the posterior distribution function and the observation points through the acquisition function.

7. The method according to claim 6, characterized in that The method further comprises: Obtaining a prediction result output by the Transformer prediction model for the test data set, determining an actual result of the test data set, and obtaining a multi-dimensional evaluation index calculation result based on a preset multi-dimensional evaluation index and an error between the prediction result and the actual result; The prediction accuracy of the Transformer prediction model is evaluated based on the calculation result of the evaluation index. In response to the prediction accuracy of the Transformer prediction model meeting the preset prediction accuracy requirement, the Transformer prediction model is determined as the target Transformer prediction model, so as to predict the roller kiln atmosphere through the target Transformer prediction model.

8. A roller kiln atmosphere prediction device, characterized in that: The device comprises: an acquisition module, configured to acquire a historical atmosphere data set of the roller hearth kiln atmosphere during the kiln firing process within a preset historical period, and preprocess the historical atmosphere data set to obtain a preprocessed initial atmosphere data set; a processing module, configured to perform feature selection on the initial atmosphere data set based on a preset nonlinear dimensionality reduction algorithm to obtain a reduced-dimensional initial atmosphere data set, and encode the reduced-dimensional initial atmosphere data set to obtain an encoded initial atmosphere data set; wherein the encoded initial atmosphere data set is divided into a training data set and a test data set; A training module, configured to construct a Transformer prediction model based on the encoded initial atmosphere data set, and train the Transformer prediction model based on the training data set to obtain a trained Transformer prediction model; A prediction module is used to optimize the Transformer prediction model based on a preset Bayesian optimization algorithm to obtain an optimized Transformer prediction model, and to predict the roller kiln atmosphere based on the optimized Transformer prediction model to obtain a prediction result of the roller kiln atmosphere.

9. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the steps of the roller kiln atmosphere prediction method according to any one of claims 1 to 7 are performed.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the steps of the roller kiln atmosphere prediction method according to any one of claims 1 to 7.

Images