A belt roaster working condition recognition method based on a multi-scale time sequence convolutional neural network

Abstract

The application discloses a belt roaster working condition recognition method based on a multi-scale time sequence convolutional neural network. The method first collects offline test data and online detection data, aligns the detection data based on the test data time stamp after cleaning, and generates time sequence data samples in combination with process fluctuation characteristics; then, an unsupervised clustering method is combined with a feature selection method to perform data preprocessing on the samples, and a working condition classification data set is constructed; finally, the data set is used to train and optimize a multi-scale time sequence convolutional neural network model, and a working condition recognition model is obtained, which can output a recognition result of the belt roaster running working condition based on online time sequence data in real time. The method provided by the application has good recognition accuracy and generalization based on the synergistic effect between steps, and is suitable for intelligent recognition of various complex working conditions in the roasting process.

Description

TECHNICAL FIELD

[0001] The application relates to a belt roaster working condition recognition method, in particular to a belt roaster working condition recognition method based on a multi-scale time sequence convolutional neural network, and belongs to the technical field of working condition recognition. BACKGROUND

[0002] The steel industry is an important basic industry of the national economy, and its development level is related to the national economy and people's livelihood. Pellet, as high-quality blast furnace burden, has the advantages of high iron grade, uniform particle size, good reducibility, etc., and is of great importance to the green and intelligent development of modern steel industry. The belt roaster pellet process has become the mainstream technology for pellet production due to its high thermal efficiency, large single machine capacity and strong raw material adaptability. However, in actual production, factors such as equipment performance degradation, green ball feeding fluctuation and external interference will continuously change the system working condition. Even if all process parameters are within the normal range, they often deviate from the initial set optimal working condition. Traditional process monitoring methods are mostly limited to coarse-grained differentiation between normal and abnormal states, and it is difficult to meet the urgent need for fine-grained identification and cognition of the above complex working conditions.

[0003] A Chinese patent (CN117150345A) discloses a rotary kiln abnormal working condition recognition method and system based on multi-modal data fusion. The method includes collecting rotary kiln working condition sample set; a feature extraction module extracts flame video image sequence features; multi-thermal variable data features are extracted; flame video image sequence and multi-thermal variable data fusion features are extracted; a composite loss function is constructed to train each feature extraction module; cascaded flame video dynamic features Ffast, multi-thermal variable dynamic features Fslow and multi-modal fusion features Ffusion are formed to form a unified sintering working condition feature F; a conformal transformation function is constructed to modify the kernel function of the mcODM classifier to obtain the KM-mcODM classifier; the unified sintering working condition feature F and the corresponding class label are input into the KM-mcODM classifier for classification and recognition. It also includes a rotary kiln abnormal working condition recognition system based on multi-modal data fusion. The application realizes multi-modal information complementary fusion, combines long-tail sample classifier model improvement, and can effectively improve the recognition accuracy of tail class sample abnormal working condition. However, this recognition method mainly targets the thermodynamic conditions of the sintering process and recognizes the sintering terminal, which is difficult to realize the working condition recognition in the sintering process, and the recognition result is only for a single sintering feature, which is difficult to meet the fine and accurate working condition recognition of multiple features in the sintering process.

[0004] A sintering process working condition identification method and system considering timing are disclosed in Chinese patent (CN110245850A), which takes time series data of process parameters of the sintering process as input and takes sintering process working condition as output. First, the Spearman rank correlation analysis method and information entropy analysis method are used for parameter selection and combination to obtain combined decision parameters. Then, the fuzzy C-means clustering algorithm based on dynamic time warping distance is used to cluster the time series data to obtain the clustering results of the combined decision parameters. Finally, the working condition is identified by using the Naive Bayes classifier to obtain the identified sintering process working condition. The working condition identification method realizes effective identification of the sintering process working condition and has important economic value and application value. However, the Naive Bayes classifier used in the identification method can improve the classification efficiency, but its accuracy is too low, resulting in a final working condition identification accuracy of only about 75%, which still needs to be further identified and classified by artificial means.

[0005] Therefore, the prior art urgently needs an intelligent method for quickly and accurately identifying the working condition of the sintering process, which can greatly improve the timeliness and accuracy of working condition identification without the need for manual identification. SUMMARY

[0006] In view of the problems existing in the prior art, the present application provides a belt roaster working condition identification method based on a multi-scale time series convolutional neural network. The method first collects offline test data and online detection data, cleans the data, and aligns the process data based on the quality data timestamp, generates time series data samples, then classifies the performance parameters in the samples and assigns working condition class labels using an unsupervised clustering method, and finally realizes automatic identification of the working condition of the roasting process time series data through training and optimization of the multi-scale time series convolutional neural network working condition identification model.

[0007] In order to achieve the above technical purpose, the present application provides a belt roaster working condition identification method based on a multi-scale time series convolutional neural network, the process of which is as follows:

[0008] Step S1, collecting offline test data and online detection data in the production process of the belt roaster;

[0009] Step S2, data cleaning and data alignment according to the time dimension are performed on the data collected in step S1, and corresponding time series data samples are generated;

[0010] Step S3, feature extraction and clustering analysis of key performance parameters are performed from the time series data samples using an unsupervised learning method, the clustering results are determined, and working condition class labels are assigned to each time series sample;

[0011] Step S4, based on the time sequence samples obtained in step S2, the correlation of each state parameter with the key performance parameter and the operation parameter is calculated, and the redundancy between each state parameter is evaluated; according to the standard of high correlation and low redundancy, the optimal feature subset is screened out, and the corresponding working condition category label determined in step S3 is given to the feature subset, so as to construct the final data set;

[0012] Step S5, a belt roaster working condition recognition model is constructed based on a multi-scale time sequence convolutional neural network, and is trained and optimized through the data set obtained in step S4.

[0013] The recognition method provided by the application is an intelligent modeling method based on process data, aiming to improve the timeliness and accuracy of working condition recognition, and can identify non-optimal working conditions in advance, thereby providing data support for production optimization, and has important application value and popularization prospect.

[0014] As a preferred scheme, the offline test data is the average compressive strength of each batch of pellets, with the dimension of Newton / piece.

[0015] As a preferred scheme, the online detection data includes: machine speed per unit time, with the dimension of m / min, pallet thickness, with the dimension of mm, green ball feeding amount, with the dimension of t / h, temperature of each air box, with the dimension of ℃, pressure of each air box, with the dimension of kPa, temperature of each smoke hood, with the dimension of ℃, pressure of each smoke hood, with the dimension of kPa, temperature of each burner, with the dimension of ℃, gas flow of each burner, with the dimension of Nm³ / h, pellet production, with the dimension of t / h, rotating speed of the main induced draft fan and the regenerative fan, with the dimension of %, and inlet temperature of the main induced draft fan and the regenerative fan, with the dimension of ℃.

[0016] As a preferred scheme, the data cleaning process is as follows:

[0017] Step S2-1, the abnormal interval of the roaster is divided into the following three kinds:

[0018] (1) when the feeding amount is 0 t / h, the machine speed is 0 m / min, and the duration is greater than 30 min, it is determined as shutdown;

[0019] (2) when the feeding amount is 0 t / h, the machine speed is in a low-speed self-circulation state less than or equal to 1.5 m / min, it is determined as material breakage;

[0020] (3) when the feeding amount is 0 t / h, the machine speed is 0 m / min, and the duration is 10-30 min, it is determined as replacing the pallet;

[0021] Step S2-2, analyze the incoming quantity and machine speed in the online detection data, detect the above three types of abnormal intervals, eliminate all data rows in the abnormal intervals, and simultaneously eliminate the offline test data falling into the abnormal intervals and the adjacent invalid window.

[0022] As a preferred scheme, the generation process of the cleaned data corresponding to the time series data sample is:

[0023] Step S2-3-1, according to the pellet sampling time Calculate the time of the pellet reaching the tail of the belt roaster , the calculation process is:

[0024] Formula 1: ;

[0025] Step S2-3-2, from , the machine speed per minute is sequentially accumulated forward , until the accumulated distance is greater than or equal to the total length B of the belt roaster, the length of the sample backtracking window is preliminarily determined :

[0026] Formula 2: ;

[0027] Step S2-3-3, in the interval covered by the window , the minimum value of the machine speed is extracted , and the final backtracking window length is determined according to the minimum machine speed in the backtracking interval , the process is:

[0028] Formula 3: ;

[0029] Formula 4: ;

[0030] Step S2-3-4, take as the time endpoint, backtrace minutes forward, extract the data of the corresponding interval from the cleaned roasting process parameters, and construct the process parameter matrix of the ith time series sample, the process is:

[0031] Formula 5: ;

[0032] In formulas 1-5, is the pellet compressive strength offline sampling time, with the dimension of min; is the transportation delay time of the pellet from the tail of the belt roaster to the sample sampling point, with the dimension of min; is the time of the pellet reaching the tail of the belt roaster, with the dimension of min; V is the machine speed, with the unit of m / min; B is the total length of the belt roaster, with the unit of m; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min;

[0033] As a preferred solution, the key performance parameters include: a yield index, a quality index, an energy consumption index and an operation behavior index;

[0034] The yield index is the output capacity per unit time in the backtracking window period, and the calculation process is as follows:

[0035] Formula 6: ;

[0036] The quality index is the compressive strength, denoted as ;

[0037] The energy consumption index is the total fuel usage in the backtracking window period, and the calculation process is as follows:

[0038] Formula 7: ;

[0039] The operation behavior index includes: the total adjustment amplitude of the fan in the backtracking window period, the main induced draft fan speed statistics and the regenerative fan speed statistics.

[0040] The calculation process of the total adjustment amplitude of the fan is as follows:

[0041] Formula 8: ;

[0042] The main induced draft fan speed statistics include the mean value, the maximum value and the minimum value, denoted as , and ;

[0043] The regenerative fan speed statistics include the mean value, the maximum value and the minimum value, denoted as , and ;

[0044] The process of feature extraction of the key performance parameters from the time series data samples is as follows: according to the key performance parameters of the backtracking time series samples, the key performance parameter vector of the samples is formed , and the key performance parameter matrix is constructed, wherein the performance parameter vector The calculation process of Q is:

[0045] Formula 9: ;

[0046] In formulas 6-9, Q is the average value of instantaneous belt output, and the dimension is ton; Q t represents the output of the t th minute, and the dimension is ton / min; σ is the compressive strength, and the dimension is Newton; Q is the total gas flow, and the dimension is m³; Q t represents the gas flow of the t th minute, and the dimension is m 3 / min; ΔN is the change amount of the main induced draft fan speed, and the dimension is %; ΔN is the change amount of the regenerative fan speed, and the dimension is %; ΔN is the total adjustment range of the fan, and the dimension is %; Z is the key performance parameter matrix, and the dimension is dimensionless; N is the number of time sequence samples, and the dimension is number.

[0047] As a preferred scheme, the process of clustering analysis of the key performance parameters from the time sequence data samples is: performing principal component analysis on the key performance parameter matrix, setting the cumulative variance proportion threshold to , obtaining the reduced key performance parameter matrix as the input of unsupervised clustering; determining the optimal number of clusters by the elbow method, and clustering the feature matrix by using the K-Means algorithm to obtain the working condition category label corresponding to each sample.

[0048] As a preferred scheme, the process of assigning the working condition category label to each time sequence sample is: determining the specific working condition category corresponding to each cluster by manual determination, and corresponding to .

[0049] As a preferred scheme, the process of obtaining the identification method data set is:

[0050] Step S4-1, removing the features in the time sequence samples that have been used for working condition label construction, averaging each state parameter of the time sequence samples in the time dimension, and mapping the time sequence process features into a one-dimensional mean vector with the same dimension as the working condition vector;

[0051] Step S4-2, calculating the mutual information (MI) value of each state parameter and , selecting the state parameters with the top k mutual information, and obtaining the candidate features after deduplication;

[0052] Step S4-3, calculate the Pearson correlation coefficient between the candidate features, set a redundancy threshold, remove redundant features, and obtain a non-redundant state parameter set, denoted as ;

[0053] Step S4-4, for each time series sample , the features with column names belonging to are retained to obtain a time series matrix with unified feature dimension , and the classification label corresponding to each sample is obtained to obtain a recognition method data set:

[0054] Formula 10: ;

[0055] In formula 10, is a variable-length time series sample, is the minimum engine speed in the backtracking interval to determine the final backtracking window length, is the working condition classification label.

[0056] As an preferred scheme, the belt roaster working condition recognition model comprises the following modules: a multi-scale time series convolution module, a time attention module, a channel attention module, a gate fusion module and a classification module.

[0057] As an preferred scheme, the training and optimization process of the belt roaster working condition recognition model is as follows:

[0058] Step S5-1, extract time series data according to the time series length of each sample, and divide the time series sample into a training set and a test set according to 7~8:2~3, and perform standardization processing on the training set and the test set, wherein the training set is used for model parameter fitting, and the test set is used for evaluating the working condition classification performance of the model on unknown samples;

[0059] Step S5-2, define a data set class, encapsulate the time series data with the corresponding length and label, support loading variable-length samples according to the index, and automatically read the time series matrix, length and label of the sample;

[0060] Step S5-3, after setting the model hyperparameters, use the training set data to train the model, adjust the hyperparameters according to the model index, initialize the model with multiple random seeds, train multiple models respectively, and after all the models are trained, select the model that performs best on the training set to save it for the subsequent test stage;

[0061] Step S5-4, test the model that performs best on the training set using the test set data, return the working condition recognition result of the time series data sample, and output the classification report and the confusion matrix diagram.

[0062] Note that, when creating the data loader, the samples are sorted and padded to ensure consistent length of the time series data within a batch.

[0063] Compared with the prior art, the technical scheme provided by the application has the beneficial effects that:

[0064] (1) The belt roaster working condition recognition method provided by the application first collects offline test data and online detection data, aligns the process data based on the quality data timestamp after cleaning the data, generates time series data samples, then classifies the performance parameters in the samples and labels the working condition categories by using an unsupervised clustering method, and finally realizes automatic recognition of the working condition of the roasting process time series data by training and optimizing a multi-scale time series convolutional neural network working condition recognition model.

[0065] (2) The technical scheme provided by the application fully considers the correlation and redundancy between the state parameters and the key performance parameters in the data selection, cleaning and classification process, so as to screen out state parameters with high correlation and low redundancy, and further improve the classification accuracy in the subsequent multi-scale time series model training by optimizing the hyperparameters, thereby giving the method good recognition accuracy and generalization, and realizing the technical effect of intelligent and rapid recognition of various complex working conditions in the roasting process without manual intervention. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 The belt roaster working condition recognition method flowchart provided by the embodiment 1 of the application is provided.

[0067] Figure 2 The process flowchart of the belt roaster roasting process in the embodiment 1 of the application is provided.

[0068] Figure 3 The K-Means clustering effect point distribution diagram in the embodiment 1 of the application is provided.

[0069] Figure 4 The prediction result confusion matrix result diagram of the belt roaster working condition recognition method provided in the embodiment 1 of the application is provided. DETAILED DESCRIPTION

[0070] The application will be further explained and described in conjunction with specific embodiments, including more detailed embodiments and operational details. The purpose of providing the embodiments is not to limit the scope of the application, but on the contrary, the embodiments are developed according to the technical scheme of the application, and are only illustrative, as a detailed description to better understand the content of the application.

[0071] Those skilled in the art can make various modifications and improvements under the guidance of the disclosure of the present application, and these modifications and improvements shall all belong to the protection of the present application without departing from the general concept and purpose of the present application.

[0072] Embodiment 1

[0073] The present embodiment provides a pellet belt induration machine working condition recognition method based on a multi-scale time sequence convolutional neural network for a pellet belt induration machine, a process flow diagram of a roasting process of the belt induration machine is as shown in Figure 1 The process of the method is as follows:

[0074] Step S1, collecting offline test data and online detection data in the production process of the belt induration machine;

[0075] The offline test data is the average compressive strength of each batch of pellets, with the dimension of Newton / piece;

[0076] The online detection data includes: machine speed per unit time, with the dimension of m / min, pallet thickness, with the dimension of mm, green ball feeding amount, with the dimension of t / h, temperature of each air box, with the dimension of ℃, pressure of each air box, with the dimension of kPa, temperature of each smoke hood, with the dimension of ℃, pressure of each smoke hood, with the dimension of kPa, temperature of each burner, with the dimension of ℃, gas flow of each burner, with the dimension of Nm³ / h, pellet production, with the dimension of t / h, speed of the main induced draft fan and the regenerative fan, with the dimension of %, and inlet temperature of the main induced draft fan and the regenerative fan, with the dimension of ℃;

[0077] Step S2, data cleaning and data alignment are performed on the data collected in step S1 according to the time dimension, and corresponding time sequence data samples are generated;

[0078] The process of data cleaning is as follows:

[0079] Step S2-1, the abnormal interval of the belt induration machine is divided into the following three kinds:

[0080] (1) when the feeding amount is 0 t / h, the machine speed is 0 m / min, and the duration is greater than 30 min, it is determined as stop;

[0081] (2) when the feeding amount is 0 t / h, the machine speed is in a low-speed self-circulation state less than or equal to 1.5 m / min, it is determined as material breakage;

[0082] (3) when the feeding amount is 0 t / h, the machine speed is 0 m / min, and the duration is 10-30 min, it is determined as pallet replacement;

[0083] Step S2-2: Analyze the feed rate and speed in the online roasting process data, detect the above three types of abnormal intervals, remove all data rows in the abnormal intervals, and simultaneously remove offline pellet data whose sampling time falls into the abnormal intervals and adjacent to the invalid window;

[0084] The process of generating the time-series data samples corresponding to the cleaned data is as follows:

[0085] Step S2-3-1: Based on the pellet sampling time Calculate the time it takes for the pellets to reach the tail end of the belt roaster. The calculation process is as follows:

[0086] Formula 1: ;

[0087] Step S2-3-2, from Starting from [the beginning], the machine speed is accumulated sequentially in minutes. The length of the sample backtracking window is initially determined until the cumulative distance is greater than or equal to the total length B of the belt roaster. :

[0088] Formula 2: ;

[0089] Step S2-3-3, in the window Within the covered interval, extract the minimum speed. Then, the final backtracking window length is determined based on the minimum machine speed within the backtracking interval. The process is as follows:

[0090] Formula 3: ;

[0091] Formula 4: ;

[0092] Steps S2-3-4, with As the end of time, look back. Minutes are used to extract data from the calcination process parameters after cleaning, and to construct the process parameter matrix of the i-th time series sample. The process is as follows:

[0093] Formula 5: ;

[0094] In equations 1 to 5, The offline sampling time for the compressive strength of the pellets is expressed in min. The transport delay time of the pellets from the tail of the belt roaster to the sample collection point is expressed in min. The time it takes for the pellets to reach the tail end of the belt roaster, in min; V is the machine speed, with the unit of m / min; B is the total length of the belt-type calciner, with the unit of m; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min; Vmin is the minimum machine speed in the interval covered by the window, with the unit of m / min;

[0095] Step S3, feature extraction and cluster analysis of the key performance parameters from the time series data samples are performed by using an unsupervised learning method, the cluster result is determined, and a working condition category label is assigned to each time series sample;

[0096] The key performance parameters include: yield indicators, quality indicators, energy consumption indicators, and operation behavior indicators;

[0097] The yield indicator is the output capacity per unit time in the backtracking window period, and the calculation process is as follows:

[0098] Formula 6: ;

[0099] The quality indicator is the compressive strength, denoted as ;

[0100] The energy consumption indicator is the total fuel usage in the backtracking window period, and the calculation process is as follows:

[0101] Formula 7: ;

[0102] The operation behavior indicators include: total adjustment amplitude of the fan in the backtracking window period, main induced fan speed statistics, and regenerative fan speed statistics;

[0103] The calculation process of the total adjustment amplitude of the fan is as follows:

[0104] Formula 8: ;

[0105] The main induced fan speed statistics include mean value, maximum value, and minimum value, denoted as , and ;

[0106] The regenerative fan speed statistics include mean value, maximum value, and minimum value, denoted as , and ;

[0107] The process of feature extraction of the key performance parameters from the time series data samples is: forming a key performance parameter vector of the sample according to the key performance parameters of the backtracking time series sample , constructing a key performance parameter matrix , wherein the calculation process of the performance parameter vector is:

[0108] Formula 9: ;

[0109] In formulas 6-9, is the average value of the instantaneous output of the belt, with the dimension of tons; represents the output of the tth minute, with the dimension of tons / minute; is the compressive strength, with the dimension of Newton / individual; is the total gas flow, with the dimension of m³; represents the gas flow of the tth minute of the burner, with the dimension of m 3 / min; is the change amount of the speed of the main induced draft fan, with the dimension of %; is the change amount of the speed of the regenerative air fan, with the dimension of %; is the total adjustment range of the fan, with the dimension of %; is the performance parameter vector, dimensionless; Z is the key performance parameter matrix, dimensionless; N is the number of time series samples, with the dimension of pieces.

[0110] The process of clustering analysis of the key performance parameters from the time series data samples is: performing principal component analysis on the key performance parameter matrix, setting the cumulative variance ratio threshold value as , obtaining the reduced key performance parameter matrix as the input of unsupervised clustering; determining the optimal clustering number by using the elbow method, and clustering the feature matrix by using the K-Means algorithm to obtain the working condition class labels corresponding to each sample The process of assigning working condition class labels to each time series sample is: determining the specific working condition classes corresponding to each cluster by manual determination, and corresponding to respectively;

[0111] In this embodiment, the working condition label semantic definition of the belt-type roaster is shown in Table 1:

[0112] ;

[0113] Step S4, based on the time sequence samples obtained in step S2, the correlation of each state parameter with the key performance parameter and the operation parameter is calculated, and the redundancy between each state parameter is evaluated; according to the criteria of high correlation and low redundancy, the optimal feature subset is screened out, and the corresponding working condition category label determined in step S3 is given to the feature subset, thereby constructing the final data set;

[0114] The acquisition process of the identification method data set is as follows:

[0115] Step S4-1, the features used for working condition label construction in the time sequence samples are removed, the mean value of each state parameter of the time sequence sample in the time dimension is calculated, and the time sequence feature is mapped into a one-dimensional mean vector with the same dimension as the working condition vector;

[0116] Step S4-2, the mutual information (MI) value of each state parameter with is calculated respectively, the state parameters with the top k mutual information are selected, and the candidate features are obtained after deduplication;

[0117] Step S4-3, the Pearson correlation coefficient between the candidate features is calculated, the redundant features are removed by setting a redundancy threshold, and a set of state parameters without redundancy is obtained, denoted as ;

[0118] Among them, {“roasting machine speed”, “green ball into machine quantity”, “average value of bottom material thickness”, “heat recovery fan inlet temperature”, “Ⅱ cold section smoke cover temperature”, “preheating section pressure difference”, “average value of 1-4 burner temperature”, “average value of 5-8 burner temperature”, “average value of 21-24 burner temperature”, “average value of 25-28 burner temperature”, “average value of 29-32 burner temperature”, “4th air box temperature”, “7th air box temperature”, “8th air box temperature”, “10th air box temperature”, “12th air box temperature”, “13th air box temperature”, “14A air box temperature”}, the dimension number of the screened features M=| |=21, is the input for subsequent classification model modeling;

[0119] Step S4-4, the time sequence samples are retained, the features with column names belonging to are retained, the time sequence matrix with uniform feature dimension is obtained, and the identification method data set is obtained in combination with the category labels corresponding to each sample:

[0120] Formula 10: ;

[0121] In formula 10, is a variable-length time sequence sample, To determine the final backtracking window length for the minimum speed in the backtracking interval, To determine the working condition category label;

[0122] Step S5, a belt roaster working condition recognition model is constructed based on a multi-scale time sequence convolutional neural network, and is trained and optimized through the data set obtained in step S4;

[0123] The belt roaster working condition recognition model comprises the following modules: a multi-scale time sequence convolution module, a time attention module, a channel attention module, a gated fusion module, and a classification module.

[0124] Further, the functions and limitations of each module in the belt roaster working condition recognition model constructed by the multi-scale time sequence convolutional neural network in the embodiment are as follows:

[0125] (1) Multi-scale time sequence convolution module: multiple time sequence convolution branches are used to extract multi-scale features in time sequence data, and different expansion factors are used in each convolution branch to capture information at different time scales; each convolution branch comprises a one-dimensional convolution layer, an activation function, a normalization layer, and a residual connection, the size of the convolution kernel is controlled by the hyperparameter kernel_size, and the expansion factor is realized by adjusting the convolution stride and zero padding;

[0126] (2) Time attention module: this module generates a time attention weight by weighting the features at each time step, and then calculates the time attention weight by a Softmax activation function, which is applied to each time sequence output, thereby weighting the importance of different time steps and improving the attention ability of the network to time sequence features;

[0127] (3) Channel attention module: the channel attention module uses two fully connected layers, which first compress the input channel features to a lower dimension, and then recover to the original channel number through a Sigmoid activation function, thereby generating a weight for each channel;

[0128] (4) Gated fusion module: by calculating the gating weight of each scale feature, the network can automatically select and weight fuse features of different scales, and the weighted fusion result of all features is normalized to obtain the final feature representation, ensuring effective integration of information of different scales;

[0129] (5) Classification module: the fused features are input into a fully connected layer, and a series of linear transformations and activation functions are used to output the final classification result; this module comprises two fully connected layers, and the output class number is 4; the classification module enhances the nonlinear representation through the activation function, and finally outputs the probability of each class through the Softmax activation function;

[0130] The training and optimization process of the belt roaster working condition recognition model is:

[0131] Step S5-1, according to the time sequence length of each sample, extract the time sequence data, and divide the time sequence samples into training set and test set according to 8:2, and carry out standardization processing on the training set and test set, wherein the training set is used for model parameter fitting, and the test set is used for evaluating the working condition classification performance of the model on unknown samples;

[0132] Wherein, the evaluation index of the model includes accuracy, confusion matrix and F1 value.

[0133] Step S5-2, define the data set class, encapsulate the time sequence data with the corresponding length and label, support loading variable length samples according to the index, and automatically read the time sequence matrix, length and label of the sample;

[0134] Step S5-3, after setting the model hyperparameters, use the training set data to train the model, adjust the hyperparameters according to the model indicators, initialize the model with multiple random seeds, train multiple models respectively, and after all the models are trained, select the model with the best performance on the training set for saving, which is used for the subsequent test stage;

[0135] In the training process, set the model training rounds, batch size, initial learning rate, optimizer and loss function; in each training period, the model performs forward propagation, loss calculation, back propagation and parameter update; by observing the loss, accuracy and F1 value and other indicators on the training set, adjust the hyperparameters at the end of each period to optimize the model performance; in addition, initialize the model with multiple random seeds, train multiple models respectively, and after all the models are trained, select the model with the best performance on the training set for saving, which is used for the subsequent test stage;

[0136] Step S5-4, test the model with the best performance on the training set using the test set data, return the data label, and output the classification report and confusion matrix diagram, wherein the test set model working condition result is shown in Table 2, the model evaluation result is shown in Table 3, and the prediction result confusion matrix result is shown in Figure 4 .

[0137] ;

[0138] ;

[0139] To evaluate the generalization performance of the model on unknown data, an independent test set is used to evaluate the model with the best performance in the training stage. The classification report (Table 2, Table 3) and confusion matrix (Table 4) are combined to evaluate the model performance on the test set. Figure 4The analysis results show that the model achieved good performance on the test set, with an overall accuracy of 91%, and both the macro average and weighted average F1 scores were not lower than 0.90, which confirms that the model has excellent generalization ability and class balance performance.

[0140] Confusion matrix ( Figure 4 The results clearly show that the predictions are highly concentrated along the main diagonal, indicating that the model's predictions are highly consistent with the true labels. This intuitive impression is corroborated by the 91% accuracy and the balanced macro-average F1 score (0.90), confirming the overall reliability and stability of the model.

[0141] The "Poor" (0) category: High recall ensures anomaly detection. This category boasts a recall of 0.92 (meaning 92% of genuine "Poor" samples were correctly identified), and the confusion matrix shows that 36 out of 39 genuine samples were correctly classified. Although its precision (0.78) is relatively low, indicating a small number of other category samples (mainly "Excellent" and "Good") were misclassified as "Poor," primarily due to the small differences in quantity among the three categories, this performance characteristic, ensuring a high detection rate, is extremely valuable in industrial applications. It minimizes the risk of false negatives and provides a reliable guarantee for timely intervention.

[0142] "Medium" (2) category: robust recognition ability. This category achieved the highest recall rate (0.93), with 38 out of 41 real samples in the confusion matrix correctly classified, while maintaining a high precision rate of 0.90. The main misclassification direction was "Good" (1), with a total of 3 cases, which reflects the proximity of adjacent quality levels in the feature space, consistent with actual cognition.

[0143] The model performed particularly well on the "Good" (1) and "Excellent" (3) categories, which had the largest sample sizes. The corresponding (1,1) and (3,3) grids in the confusion matrix had the largest values ​​(109 and 94, respectively) and the darkest colors. The misclassifications between the two categories were very limited (5 cases of "Good" → "Excellent" and 3 cases of "Excellent" → "Good"). This limited confusion precisely confirms that the model can effectively learn the subtle differences between adjacent but different quality levels, rather than producing random errors.

[0144] Comprehensive evaluation on the test set shows that the model not only has high overall accuracy, but more importantly, it maintains extremely high sensitivity to key abnormal operating conditions ("poor") while accurately distinguishing other categories. The misjudgment patterns revealed by the confusion matrix are logical, further corroborating the rationality of the model's decisions. These characteristics collectively demonstrate that the model possesses the reliability, practicality, and significant value required for industrial application.

Claims

1. A belt roaster working condition recognition method based on a multi-scale time series convolutional neural network, characterized by: Step S1, collecting offline test data and online detection data in the production process of the belt roaster; Step S2, performing data cleaning and data alignment on the data collected in step S1 according to the time dimension, and generating corresponding time series data samples; Step S3, using an unsupervised learning method to extract features and perform cluster analysis on the key performance parameters from the time series data samples, determine the clustering results, and assign working condition category labels to each time series sample; Step S4, based on the time series samples obtained in step S2, calculate the correlation of each state parameter with the key performance parameters and operating parameters, and evaluate the redundancy between each state parameter; According to the criteria of high correlation and low redundancy, the optimal feature subset is selected, and the corresponding working condition category label determined in step S3 is assigned to the feature subset, thereby constructing the final data set; Step S5, based on the multi-scale time series convolutional neural network, a belt roaster working condition recognition model is constructed, and the data set obtained in step S4 is used for training and optimization; The process of generating corresponding time series data samples is: Step S2-3-1, according to the pellet sampling time Calculating the time of the pellet reaching the tail of the belt induration machine The calculation process is as follows: Formula 1: ; Step S2-3-2, from the speed of the machine per minute is added sequentially forward until the accumulated distance is greater than or equal to the total length B of the belt furnace, the length of the sample backtracking window is preliminarily determined : Formula 2: ; Step S2-3-3, in the window The minimum value of the machine speed in the covered interval is extracted The final backtracking window length is determined according to the minimum machine speed in the backtracking interval The process is as follows: Formula 3: ; Formula 4: ; Step S2-3-4, taking as the time endpoint, backtracking minutes, extracting the data of the corresponding interval from the roasting process parameters after cleaning, constructing the process parameter matrix of the first i time sequence sample, the process is as follows: Formula 5: ; In formulas 1-5, is the off-line sampling time for the pellet compressive strength, dimension min; is the transport delay time for the pellet from the end of the belt induration machine to the sampling point, dimension min; is the time for the pellet to reach the end of the belt induration machine, dimension min; is the machine speed, dimension m / min; B is the total length of the belt induration machine, dimension m; is the length of the preliminary determined sample backtracking window, dimension min; is the minimum machine speed in the interval covered by the window , dimension m / min; is the final backtracking window length determined by the minimum machine speed in the backtracking interval, dimension min; is the process parameter matrix, dimensionless.

2. The belt roaster working condition recognition method based on a multi-scale time sequence convolutional neural network according to claim 1, characterized in that: The offline test data is the average compressive strength of each batch of pellets, with a dimension of Newton / piece; The online detection data includes: machine speed per unit time, dimension m / min, pallet thickness, dimension mm, green ball feeding amount, dimension t / h, temperature of each air box, dimension ℃, pressure of each air box, dimension kPa, temperature of each hood, dimension ℃, pressure of each hood, dimension kPa, temperature of each burner, dimension ℃, gas flow of each burner, dimension Nm³ / h, pellet production, dimension t / h, speed of the main induced draft fan and the regenerative fan, dimension %, and inlet temperature of the main induced draft fan and the regenerative fan, dimension ℃.

3. The belt roaster working condition recognition method based on a multi-scale time sequence convolution neural network according to claim 1, characterized in that: The process of data cleaning is: Step S2-1, the abnormal interval of the roaster is divided into the following three types: 1) When the feeding amount is 0 t / h, the machine speed is 0 m / min, and the duration is greater than 30 min, it is determined as stop; (2) When the feeding amount is 0 t / h, the machine speed is less than or equal to 1.5 m / min in the low-speed self-circulation state, it is determined as material breakage; (3) When the feeding amount is 0 t / h, the machine speed is 0 m / min, and the duration is 10-30 min, it is determined as replacing the pallet; Step S2-2, analyze the feeding amount and machine speed in the online detection data, detect the abnormal interval in step S2-1, remove all data rows in the abnormal interval, and simultaneously remove the offline test data whose sampling time falls within the abnormal interval and the adjacent invalid window.

4. The belt roaster working condition recognition method based on a multi-scale time sequence convolution neural network according to claim 1, characterized in that: The key performance parameters include: yield indicators, quality indicators, energy consumption indicators, and operation behavior indicators; The yield indicator is the output capacity per unit time in the backtracking window period, and its calculation process is: Formula 6: ; The quality indicator is the compressive strength, denoted by ; The energy consumption indicator is the total amount of fuel used in the backtracking window period, and its calculation process is: Formula 7: ; The operation behavior indicator includes: total adjustment amplitude of the fan in the backtracking window period, main induced draft fan speed statistics, and regenerative fan speed statistics; The calculation process of the total adjustment range of the fan is as follows: Formula 8: ; The main induced draft fan rotating speed statistics include mean value, maximum value and minimum value, respectively denoted as , and ; The statistics of the rotation speed of the regenerative air blower include mean value, maximum value and minimum value, respectively denoted as , and ; The process of feature extraction of the key performance parameter from the time series data sample is: forming a key performance parameter vector of the sample according to the key performance parameter of the backtracking time series sample , constructing a key performance parameter matrix , wherein the calculation process of the performance parameter vector is: Formula 9: ; In formula 6~9, is the average value of instantaneous production of the belt, with the dimension of ton; represents the production of the tth minute, with the dimension of ton / min; is the compressive strength, with the dimension of Newton; is the total gas flow, with the dimension of m³; represents the gas flow of the tth minute of the burner, with the dimension of m 3 / min; is the change amount of the main induced draft fan speed, with the dimension of %; is the change amount of the regenerative fan speed, with the dimension of %; is the total adjustment range of the fan, with the dimension of %; is the performance parameter vector, dimensionless; Z is the key performance parameter matrix, dimensionless; N is the number of time sequence samples, with the dimension of pieces.

5. The belt roaster working condition recognition method based on a multi-scale time sequence convolution neural network according to claim 4, characterized in that: The process of the clustering analysis on the key performance parameters from the time sequence data samples is: performing principal component analysis on the key performance parameter matrix, setting the cumulative explained variance ratio threshold value as , obtaining the reduced key performance parameter matrix as the input of the unsupervised clustering; determining the optimal clustering number by using the elbow method, and clustering the feature matrix by using the K-Means algorithm to obtain the working condition category labels corresponding to each sample . The process of assigning the working condition category label to each time sequence sample is as follows: the specific working condition category corresponding to each cluster is determined by manual determination, and the working condition category label is assigned to each time sequence sample in the cluster respectively .

6. The belt roaster working condition recognition method based on a multi-scale time sequence convolution neural network according to claim 4, characterized in that: The acquisition process of the identification method data set is as follows: Step S4-1, the features used for the construction of the working condition label are removed from the time sequence sample, the mean value of each state parameter of the time sequence sample in the time dimension is calculated, and the time sequence process feature is mapped into a one-dimensional mean vector with the same dimension as the working condition vector; Step S4-2, respectively calculate the mutual information (MI) value of each state parameter and , select the top k state parameters in mutual information, and obtain the candidate features after deduplication; Step S4-3, calculate the Pearson correlation coefficient between the candidate features, set a redundancy threshold, eliminate redundant features, and obtain a set of state parameters without redundancy, denoted as ; Step S4-4, for each time sequence sample , the column name belongs to the characteristics of the unified feature dimension time sequence matrix , combined with the corresponding category label of each sample, the recognition method data set is obtained: Formula 10: ; In formula 10, is a variable length time series sample, determines the final backtracking window length as the minimum Mach number in the backtracking interval, is a working condition category label.

7. The belt roaster working condition recognition method based on a multi-scale time sequence convolution neural network according to claim 1, characterized in that: The belt roaster working condition identification model comprises the following modules: Multi-scale time sequence convolution module, time attention module, channel attention module, gate fusion module and classification module.

8. The belt roaster working condition recognition method based on a multi-scale time sequence convolution neural network according to claim 7, characterized in that: The training and optimization process of the belt roaster working condition identification model is as follows: Step S5-1, according to the time sequence length of each sample, the time sequence data is extracted, and the time sequence sample is divided into a training set and a test set according to 7-8:2-3, and the training set and the test set are standardized, wherein the training set is used for model parameter fitting, and the test set is used for evaluating the working condition classification performance of the model on unknown samples; Step S5-2, define the data set class, encapsulate the time sequence data with the corresponding length and label, support to load variable-length samples according to the index, and automatically read the time sequence matrix, length and label of the sample; Step S5-3, after setting the model hyperparameters, the model is trained using the training set data, the hyperparameters are adjusted according to the model index, a plurality of models are trained respectively by using a plurality of random seeds, and after the training of all the models is completed, the model with the best performance on the training set is selected and saved for the subsequent test stage; Step S5-4, the model with the best performance on the training set is tested using the test set data, the working condition identification result of the time sequence data sample is returned, and a classification report and a confusion matrix diagram are outputted.

Images