A method and system for identifying the state of a gas-solid fluidized bed

Abstract

The present invention belongs to the fields of chemical process control and mining engineering, and discloses a method and system for identifying the state of a gas-solid fluidized bed, including: obtaining a pressure fluctuation signal during the operation of the fluidized bed, decomposing the pressure fluctuation signal to obtain multi-scale decomposition coefficients, denoising the multi-scale decomposition coefficients to form a denoised pressure fluctuation signal; extracting time-domain characteristic parameters and frequency-domain characteristic parameters from the denoised pressure fluctuation signal to obtain a multi-dimensional feature vector; performing dimensionality reduction processing on the multi-dimensional feature vector to obtain a dimensionality-reduced feature matrix; constructing a bed state identification model according to the dimensionality-reduced feature matrix; and dynamically identifying the bed state according to the bed state identification model. The present invention realizes real-time monitoring and feedback control of the state of the fluidized bed, thereby improving the stability and efficiency of the operation of the gas-solid fluidized bed.

Description

Technical Field

[0001] The present invention relates to the fields of chemical process control and mining engineering, and particularly relates to a method and system for identifying the state of a gas-solid fluidized bed. Background Art

[0002] Gas-solid fluidized beds are widely used in industrial fields such as petrochemical engineering, mineral separation, and coal separation. Their operating states have an important impact on reaction efficiency, material transportation, and product quality. During the operation of a fluidized bed, different bed states (fixed bed, critical fluidization state, bubbling bed) directly affect its performance and stability.

[0003] Existing fluidized bed bed state identification technologies usually rely on operating experience, visual observation, or analysis methods based on single signal features. These methods have the following defects:

[0004] 1. Experience-dependent and highly subjective: Manual judgment cannot be quantified and automated, and the identification accuracy is relatively low.

[0005] 2. Insufficient response to complex states: Single signal features are difficult to comprehensively describe complex fluidization behaviors.

[0006] 3. Lack of real-time performance and adaptability: Existing technologies are difficult to meet the requirements of real-time on-line monitoring in industrial sites.

[0007] In view of the above problems, the present invention proposes a method for identifying the state of a gas-solid fluidized bed by combining multi-source signal processing and pattern recognition technologies, which can efficiently and accurately identify the state of the fluidized bed bed, providing support for process optimization and safe operation. Summary of the Invention

[0008] In order to solve the problems of insufficient accuracy, poor applicability, and weak real-time performance existing in the prior art, the present invention provides a method and system for identifying the state of a gas-solid fluidized bed. Through collecting pressure fluctuation signals, feature extraction, and pattern recognition, real-time monitoring and feedback control of the fluidized bed bed state are realized, thereby improving the stability and efficiency of the operation of the gas-solid fluidized bed.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A method for identifying the state of a gas-solid fluidized bed, the method comprising:

[0011] Obtaining a pressure fluctuation signal during the operation of the fluidized bed, decomposing the pressure fluctuation signal to obtain multi-scale decomposition coefficients, and denoising the multi-scale decomposition coefficients to form a denoised pressure fluctuation signal;

[0012] Extract time-domain feature parameters and frequency-domain feature parameters from the denoised pressure fluctuation signal to obtain a multi-dimensional feature vector, where the time-domain feature parameters include: the mean, variance, skewness, and kurtosis of the pressure fluctuation signal, and the frequency-domain feature parameters include: the power spectral density and the main frequency component of the pressure fluctuation signal;

[0013] Perform dimensionality reduction on the multi-dimensional feature vector to obtain a dimensionality-reduced feature matrix;

[0014] Construct a bed state recognition model based on the dimensionality-reduced feature matrix;

[0015] Dynamically identify the bed state according to the bed state recognition model.

[0016] Preferably, obtain the pressure fluctuation signal during the operation of the fluidized bed, decompose the pressure fluctuation signal to obtain multi-scale decomposition coefficients, and denoise the multi-scale decomposition coefficients to form the denoised pressure fluctuation signal, including:

[0017] Obtain the pressure fluctuation signal during the operation of the fluidized bed, convert the pressure fluctuation signal into a digital signal, and perform preprocessing to obtain the preprocessed pressure fluctuation signal;

[0018] Perform wavelet decomposition on the preprocessed pressure fluctuation signal according to the preset wavelet basis function and decomposition level to obtain wavelet coefficients of each layer;

[0019] Calculate the thresholds of the wavelet coefficients of each layer according to the statistical characteristics of the wavelet coefficients of each layer, and perform threshold processing on the wavelet coefficients of each layer using the soft threshold method to obtain the denoised wavelet coefficients;

[0020] Reconstruct the wavelet coefficients of each layer using the wavelet reconstruction method according to the denoised wavelet coefficients to obtain the denoised pressure fluctuation signal.

[0021] Preferably, extract time-domain feature parameters and frequency-domain feature parameters from the denoised pressure fluctuation signal to obtain a multi-dimensional feature vector, including:

[0022] According to the denoised pressure fluctuation signal, use the time-domain analysis method to calculate the mean, variance, skewness, and kurtosis of the signal to obtain the time-domain feature parameters;

[0023] According to the denoised pressure fluctuation signal, use the frequency-domain analysis method to calculate the power spectral density and the main frequency component of the signal to obtain the frequency-domain feature parameters;

[0024] Combine the time-domain feature parameters and the frequency-domain feature parameters to construct a multi-dimensional feature vector for characterizing the characteristics of the pressure fluctuation signal.

[0025] Preferably, the multi-dimensional feature vector is dimensionally reduced to obtain a dimensionally reduced feature matrix, including:

[0026] According to the multi-dimensional feature vector, dimension information is obtained. If the dimension is greater than a preset threshold, it is determined that there is redundant information and dimensionality reduction processing is required;

[0027] According to the statistical characteristics of the multi-dimensional feature vector, a covariance matrix is constructed, and through eigenvalue decomposition, the eigenvalues and eigenvectors of the covariance matrix are obtained;

[0028] According to the eigenvalue magnitudes, the number of principal components is determined, and the top N eigenvectors with the largest eigenvalues are selected as the principal component vectors;

[0029] The multi-dimensional feature vector is projected onto the principal component vectors to obtain a dimensionally reduced feature vector representation, forming a compressed feature matrix;

[0030] Using the singular value decomposition method, the compressed feature matrix is further dimensionally reduced to remove noise and redundant components and extract feature information;

[0031] Through a combination of principal component analysis and singular value decomposition, the finally dimensionally reduced feature matrix is obtained.

[0032] Preferably, according to the dimensionally reduced feature matrix, a bed state recognition model is constructed, including:

[0033] According to pre-set different bed state labels, a classification model, i.e., a bed state recognition model, is constructed using the support vector machine algorithm, and the dimensionally reduced feature matrix is input into the classification model for training;

[0034] According to the training results of the classification model, the kernel function type and kernel function parameters of the support vector machine algorithm are determined to obtain an optimized classification model;

[0035] Among them, determining the kernel function type and kernel function parameters of the support vector machine algorithm according to the training results of the classification model includes:

[0036] Step l: Using the initialization search center , where is the initial value of m eddy centers, is the initial value of the m-th eddy center, using the initialization standard deviation , the fitness function of the optimal solution is set as , where is the global best solution of all candidate solutions in the current iteration, and the iteration step t = 0;

[0037] Step 2: With a search radius , around the search center , generate candidate solutions that follow a Gaussian distribution , the total number of candidate solutions is 250. If the value of a candidate solution exceeds the boundary range, formula is used, where is the candidate solution that exceeds the range, is a random number that conforms to a uniform distribution, and are both d-dimensional vectors, representing the upper and lower bounds of the d-dimensional search space respectively, and are transformed into the boundary;

[0038] Step 3: Select the best solution from , save it to the matrix , then select the global best solution from and store it in . If is better than the current global optimal solution, then update the optimal solution ;

[0039] Step 4: Take as the m-th new eddy center, and use formula , where is the loss function, is the normal vector of the hyperplane, is the transformation of the original feature, is the penalty factor for misclassified samples, is the slack variable, b is the threshold, update the first m - 1 eddy centers, reduce the radius to get , use the transformed m centers as the eddy centers after reducing the radius, generate candidate solutions that follow a normal distribution, and set the iteration step t = t + 1. s.t. is the standard writing in operations research, meaning "such that", is the transformed feature, and T is the transpose of the normal vector of the hyperplane;

[0040] Step 5: Determine whether the termination condition is met: If the iteration step t ≥ E, where is the maximum number of iterations, used to control the running time of the algorithm and prevent overfitting, then terminate the iteration and output the optimal solution S; if the termination condition is not met, then go back to Step 2.

[0041] Preferably, according to the bed state recognition model, dynamically recognizing the bed state includes:

[0042] Input the feature matrix to be classified into the classification model according to the optimized classification model, and obtain the bed state classification results corresponding to each feature matrix;

[0043] According to the classification results, use the voting method to determine the final state of each bed. If the classification results of multiple feature matrices of a certain bed are the same, then it is judged that the bed is in the state corresponding to this classification result;

[0044] According to the final states of each bed layer, cross-sectional diagrams of the bed layer in different states are drawn to visually display the state distribution of each bed layer in a graphical manner;

[0045] According to the state distribution of each bed layer, an interpolation algorithm is used to calculate the transitional states between each bed layer, obtaining a complete bed layer state distribution diagram to provide data support for subsequent production management.

[0046] The present invention also provides a gas-solid fluidized bed bed layer state recognition system, which is used to implement any of the above methods. The system includes: a decomposition module, an extraction module, a dimensionality reduction module, a construction module, and an identification module;

[0047] The decomposition module is used to obtain the pressure fluctuation signal during the operation of the fluidized bed, decompose the pressure fluctuation signal to obtain multi-scale decomposition coefficients, and denoise the multi-scale decomposition coefficients to form a denoised pressure fluctuation signal;

[0048] The extraction module is used to extract time-domain characteristic parameters and frequency-domain characteristic parameters according to the denoised pressure fluctuation signal to obtain a multi-dimensional feature vector. Among them, the time-domain characteristic parameters include: the mean, variance, skewness, and kurtosis of the pressure fluctuation signal, and the frequency-domain characteristic parameters include: the power spectral density and main frequency component of the pressure fluctuation signal;

[0049] The dimensionality reduction module is used to perform dimensionality reduction processing on the multi-dimensional feature vector to obtain a dimensionality-reduced feature matrix;

[0050] The construction module is used to construct a bed layer state recognition model according to the dimensionality-reduced feature matrix;

[0051] The identification module is used to dynamically identify the bed layer state according to the bed layer state recognition model.

[0052] Preferably, the decomposition module includes: a conversion unit, a preprocessing unit, a calculation unit, and a reconstruction unit;

[0053] The conversion unit is used to obtain the pressure fluctuation signal during the operation of the fluidized bed, convert the pressure fluctuation signal into a digital signal, and perform preprocessing to obtain a preprocessed pressure fluctuation signal;

[0054] The preprocessing unit is used to perform wavelet decomposition on the preprocessed pressure fluctuation signal according to a preset wavelet basis function and decomposition level to obtain wavelet coefficients of each layer;

[0055] The calculation unit is used to calculate the thresholds of the wavelet coefficients of each layer according to the statistical characteristics of the wavelet coefficients of each layer, and perform threshold processing on the wavelet coefficients of each layer using a soft threshold method to obtain denoised wavelet coefficients;

[0056] The reconstruction unit is configured to reconstruct the wavelet coefficients of each layer according to the denoised wavelet coefficients by using a wavelet reconstruction method to obtain a denoised pressure fluctuation signal.

[0057] Preferably, the extraction module includes: a time-domain feature unit, a frequency-domain feature unit, and a combination unit;

[0058] The time-domain feature unit is configured to calculate the mean, variance, skewness, and kurtosis of the signal by using a time-domain analysis method according to the denoised pressure fluctuation signal to obtain time-domain feature parameters;

[0059] The frequency-domain feature unit is configured to calculate the power spectral density and the main frequency component of the signal by using a frequency-domain analysis method according to the denoised pressure fluctuation signal to obtain frequency-domain feature parameters;

[0060] The combination unit is configured to combine the time-domain feature parameters and the frequency-domain feature parameters to construct a multi-dimensional feature vector for characterizing the features of the pressure fluctuation signal.

[0061] Preferably, the dimensionality reduction module includes: an acquisition unit, an eigenvalue unit, a selection unit, a projection unit, a denoising unit, and a combination unit;

[0062] The acquisition unit is configured to obtain dimension information according to the multi-dimensional feature vector. If the dimension is greater than a preset threshold, it is determined that there is redundant information and dimensionality reduction processing is required;

[0063] The eigenvalue unit is configured to construct a covariance matrix according to the statistical characteristics of the multi-dimensional feature vector, and obtain the eigenvalues and eigenvectors of the covariance matrix through eigenvalue decomposition;

[0064] The selection unit is configured to determine the number of principal components according to the eigenvalue magnitudes, and select the top N eigenvectors with the largest eigenvalues as the principal component vectors;

[0065] The projection unit is configured to project the multi-dimensional feature vector onto the principal component vectors to obtain a dimensionality-reduced feature vector representation, forming a compressed feature matrix;

[0066] The denoising unit is configured to further reduce the dimension of the compressed feature matrix by using a singular value decomposition method, remove noise and redundant components, and extract feature information;

[0067] The combination unit is configured to obtain a finally dimensionally-reduced feature matrix by combining principal component analysis and singular value decomposition.

[0068] Compared with the prior art, the beneficial effects of the present invention are:

[0069] The present invention discloses a method and system for identifying the bed state of a gas-solid fluidized bed. The method obtains the pressure fluctuation signal during the operation of the fluidized bed, uses wavelet transform for denoising and extracts time-domain and frequency-domain characteristic parameters to form a multi-dimensional feature vector. Subsequently, principal component analysis is used for dimensionality reduction to remove redundant information. A classification model is constructed using the support vector machine algorithm, and the feature matrix is trained to obtain a preliminary classification result. For the boundary fuzzy samples with low classification confidence, the K-nearest neighbor algorithm is used for secondary classification to improve the classification accuracy. Finally, a bed state identification report is generated, including the distribution of characteristic parameters and the classification confidence. The present invention can also monitor the pressure fluctuation signal in real time, dynamically identify the bed state, output a real-time state report, and provide a basis for process optimization. Through multiple algorithm optimizations, the method effectively improves the accuracy and real-time performance of the gas-solid fluidized bed bed state identification, and is applicable to gas-solid fluidized beds under different scales and process conditions. Based on the high-efficiency algorithm computing power, the present invention can achieve online real-time monitoring. The present invention supports the adaptive adjustment of the system, improves the operation optimization ability, and provides reliable support for the operation management and process optimization of the fluidized bed. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0071] Figure 1 It is a schematic flowchart of a method for identifying the bed state of a gas-solid fluidized bed according to an embodiment of the present invention;

[0072] Figure 2 It is a schematic flowchart of SVM parameter optimization based on an improved eddy current search algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0073] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0074] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present disclosure should have the ordinary meanings understood by those of ordinary skill in the field to which the present disclosure belongs. The "first", "second" and similar terms used in the embodiments of the present disclosure do not denote any order, quantity or importance, but are only used to distinguish different components. Words such as "including" or "comprising" mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Upper", "lower", "left", "right", etc. are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0075] To make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0076] Embodiment 1

[0077] As Figure 1 shown, this embodiment provides a method for identifying the state of a gas-solid fluidized bed, and the method includes:

[0078] Obtain the pressure fluctuation signal during the operation of the fluidized bed, decompose the pressure fluctuation signal to obtain multi-scale decomposition coefficients, and denoise the multi-scale decomposition coefficients to form a denoised pressure fluctuation signal;

[0079] According to the denoised pressure fluctuation signal, extract time-domain characteristic parameters and frequency-domain characteristic parameters to obtain a multi-dimensional feature vector, where the time-domain characteristic parameters include: the mean, variance, skewness and kurtosis of the pressure fluctuation signal, and the frequency-domain characteristic parameters include: the power spectral density and the main frequency component of the pressure fluctuation signal;

[0080] Perform dimensionality reduction processing on the multi-dimensional feature vector to obtain a dimensionality-reduced feature matrix;

[0081] Construct a bed state recognition model according to the dimensionality-reduced feature matrix;

[0082] Dynamically identify the bed state according to the bed state recognition model.

[0083] In this embodiment, obtaining the pressure fluctuation signal during the operation of the fluidized bed, decomposing the pressure fluctuation signal to obtain multi-scale decomposition coefficients, and denoising the multi-scale decomposition coefficients to form a denoised pressure fluctuation signal includes:

[0084] Obtain the pressure fluctuation signal during the operation of the fluidized bed, convert the pressure fluctuation signal into a digital signal, and perform preprocessing, including removing outliers and linear trends in the signal, to obtain the preprocessed pressure fluctuation signal;

[0085] According to the preset wavelet basis function and decomposition level, perform wavelet decomposition on the preprocessed pressure fluctuation signal to obtain wavelet coefficients of each layer. The wavelet coefficients include low-frequency approximation coefficients and high-frequency detail coefficients;

[0086] According to the statistical characteristics of the wavelet coefficients of each layer, calculate the thresholds of the wavelet coefficients of each layer, and use the soft threshold method to perform threshold processing on the wavelet coefficients of each layer to remove the noise components in the high-frequency detail coefficients and obtain the denoised wavelet coefficients;

[0087] Specifically: In the present invention, it is necessary to perform noise reduction processing on the noise-dominated components. Therefore, the wavelet soft threshold noise reduction method is improved. In the original noise reduction method, the wavelet coefficients of different layers are processed, and each order of IMF is used to replace the wavelet coefficients of different layers, and the following calculation formula is obtained:

[0088]

[0089]

[0090] Among them, and are the IMF after the improved wavelet soft threshold denoising process and the initial IMF respectively. n is the order of the IMF. i is the i-th data point of the IMF. is the wavelet transform coefficient of the original signal. is the threshold of the n-th order data point of the IMF. The calculation formula of

[0091]

[0092] Among them, N is the number of data points.

[0093] According to the denoised wavelet coefficients, use the wavelet reconstruction method to reconstruct the wavelet coefficients of each layer to obtain the denoised pressure fluctuation signal.

[0094] In this embodiment, according to the denoised pressure fluctuation signal, time-domain characteristic parameters and frequency-domain characteristic parameters are extracted, and the multi-dimensional feature vector obtained includes:

[0095] According to the denoised pressure fluctuation signal, use the time-domain analysis method to calculate the mean, variance, skewness and kurtosis of the signal to obtain the time-domain characteristic parameters;

[0096] According to the denoised pressure fluctuation signal, use the frequency-domain analysis method to calculate the power spectral density and main frequency component of the signal to obtain the frequency-domain characteristic parameters;

[0097] Combine the time-domain feature parameters and frequency-domain feature parameters to construct a multi-dimensional feature vector for characterizing the features of the pressure fluctuation signal.

[0098] Specifically, construct an 8-dimensional feature vector, which includes 4 time-domain features and 4 frequency-domain features, reflecting both the statistical properties of the signal and containing frequency information, laying a foundation for subsequent classification.

[0099] In this embodiment, perform dimensionality reduction processing on the multi-dimensional feature vector, and the obtained feature matrix after dimensionality reduction includes:

[0100] For the multi-dimensional feature vector, obtain its dimension information. If the dimension is greater than the preset threshold, it is determined that there is redundant information and dimensionality reduction processing is required;

[0101] According to the statistical characteristics of the multi-dimensional feature vector, construct a covariance matrix, and through eigenvalue decomposition, obtain the eigenvalues and eigenvectors of the covariance matrix;

[0102] According to the eigenvalue magnitudes, determine the number of principal components, and select the top N eigenvectors with the largest eigenvalues as the principal component vectors;

[0103] Project the multi-dimensional feature vector onto the principal component vectors to obtain the dimensionality-reduced feature vector representation, forming a compressed feature matrix;

[0104] Adopt the singular value decomposition method to further reduce the dimensionality of the feature matrix, remove noise and redundant components, and extract the most important feature information;

[0105] By combining principal component analysis and singular value decomposition, the essential features of the multi-dimensional feature vector are retained to the greatest extent, while significantly reducing the feature dimension;

[0106] Use the feature matrix after dimensionality reduction as the input of the machine learning algorithm, which can effectively improve the training efficiency and generalization ability of the algorithm, and achieve efficient and accurate feature learning and pattern recognition.

[0107] In this embodiment, according to the feature matrix after dimensionality reduction, constructing a bed state recognition model includes:

[0108] According to the pre-set different bed state labels, use the support vector machine algorithm to construct a classification model, that is, a bed state recognition model, and input the feature matrix after dimensionality reduction into the classification model for training;

[0109] According to the training results of the classification model, determine the kernel function type and kernel function parameters of the support vector machine algorithm to obtain an optimized classification model;

[0110] Among them, as Figure 2As shown, determining the kernel function type and kernel function parameters of the support vector machine algorithm according to the training results of the classification model includes:

[0111] Step l: Use the initialization search center , where is the initial value of m eddy centers, is the initial value of the m-th eddy center, use the initialization standard deviation , and the fitness function of the optimal solution is set to , where is the global best solution of all candidate solutions in the current iteration, and the iteration step number t = 0;

[0112] Step 2: With the search radius , around the search center , generate candidate solutions that follow a Gaussian distribution. The total number of candidate solutions is 250. If the value of a candidate solution exceeds the boundary range, use the formula , where is the candidate solution that exceeds the range, is a random number that conforms to a uniform distribution, and are both d-dimensional vectors, representing the upper and lower bounds of the d-dimensional search space respectively, and transform them inside the boundary;

[0113] Step 3: Select the best solution from , save it to the matrix , then select the global best solution from and store it in . If is better than the current global optimal solution, then update the optimal solution ;

[0114] Step 4: Take as the m-th new eddy center, and use the formula , where is the loss function, is the normal vector of the hyperplane, is the transformation of the original feature, is the penalty factor for misclassified samples, is the slack variable, b is the threshold, update the first m - 1 eddy centers, reduce the radius to get , use the transformed m centers as the eddy centers after reducing the radius, generate candidate solutions that follow a normal distribution, and transform the iteration step number t = t + 1. s.t. is the standard writing in operations research, meaning "such that", is the transformed feature, and T is the transpose of the normal vector of the hyperplane;

[0115] Step 5: Determine whether the termination condition is met: If the number of iteration steps \(t\geq E\), where \(E\) is the maximum number of iterations, which is used to control the running time of the algorithm and prevent overfitting, then terminate the iteration and output the optimal solution \(S\); if the termination condition is not met, go to Step 2.

[0116] In this embodiment, according to the bed state recognition model, dynamically recognizing the bed state includes:

[0117] Input the feature matrix to be classified into the classification model according to the optimized classification model, and obtain the bed state classification results corresponding to each feature matrix;

[0118] According to the classification results, use the voting method to determine the final state of each bed. If the classification results of multiple feature matrices of a certain bed are consistent, it is determined that the bed is in the state corresponding to the classification result;

[0119] According to the final state of each bed, draw the cross-sectional view of the bed in different states, and visually display the state distribution of each bed in a graphical way;

[0120] According to the state distribution of each bed, use the interpolation algorithm to calculate the transition state between each bed, and obtain the complete bed state distribution map, providing data support for subsequent production management.

[0121] Embodiment 2

[0122] The present invention also provides a gas-solid fluidized bed bed state recognition system, which is used to implement any one of the above methods. The system includes: a decomposition module, an extraction module, a dimensionality reduction module, a construction module, and an identification module;

[0123] The decomposition module is used to obtain the pressure fluctuation signal during the operation of the fluidized bed, decompose the pressure fluctuation signal to obtain multi-scale decomposition coefficients, and denoise the multi-scale decomposition coefficients to form a denoised pressure fluctuation signal;

[0124] The extraction module is used to extract time-domain feature parameters and frequency-domain feature parameters according to the denoised pressure fluctuation signal to obtain a multi-dimensional feature vector, where the time-domain feature parameters include: the mean, variance, skewness, and kurtosis of the pressure fluctuation signal, and the frequency-domain feature parameters include: the power spectral density and main frequency component of the pressure fluctuation signal;

[0125] The dimensionality reduction module is used to perform dimensionality reduction processing on the multi-dimensional feature vector to obtain a dimensionality-reduced feature matrix;

[0126] The construction module is used to construct a bed state recognition model according to the dimensionality-reduced feature matrix;

[0127] The recognition module is used to dynamically recognize the bed state according to the bed state recognition model.

[0128] In this embodiment, the decomposition module includes: a conversion unit, a preprocessing unit, a calculation unit, and a reconstruction unit;

[0129] The conversion unit is used to obtain the pressure fluctuation signal during the operation of the fluidized bed, convert the pressure fluctuation signal into a digital signal, and perform preprocessing to obtain the preprocessed pressure fluctuation signal;

[0130] The preprocessing unit is used to perform wavelet decomposition on the preprocessed pressure fluctuation signal according to the preset wavelet basis function and decomposition level to obtain wavelet coefficients of each layer;

[0131] The calculation unit is used to calculate the threshold of the wavelet coefficients of each layer according to the statistical characteristics of the wavelet coefficients of each layer, and perform threshold processing on the wavelet coefficients of each layer by using the soft threshold method to obtain the denoised wavelet coefficients;

[0132] The reconstruction unit is used to reconstruct the wavelet coefficients of each layer by using the wavelet reconstruction method according to the denoised wavelet coefficients to obtain the denoised pressure fluctuation signal.

[0133] In this embodiment, the extraction module includes: a time-domain feature unit, a frequency-domain feature unit, and a combination unit;

[0134] The time-domain feature unit is used to calculate the mean, variance, skewness, and kurtosis of the signal by using the time-domain analysis method according to the denoised pressure fluctuation signal to obtain time-domain feature parameters;

[0135] The frequency-domain feature unit is used to calculate the power spectral density and main frequency component of the signal by using the frequency-domain analysis method according to the denoised pressure fluctuation signal to obtain frequency-domain feature parameters;

[0136] The combination unit is used to combine the time-domain feature parameters and the frequency-domain feature parameters to construct a multi-dimensional feature vector for characterizing the features of the pressure fluctuation signal.

[0137] In this embodiment, the dimensionality reduction module includes: an acquisition unit, an eigenvalue unit, a selection unit, a projection unit, a denoising unit, and a combination unit;

[0138] The acquisition unit is used to obtain dimension information according to the multi-dimensional feature vector. If the dimension is greater than the preset threshold, it is determined that there is redundant information and dimensionality reduction processing is required;

[0139] The eigenvalue unit is used to construct a covariance matrix according to the statistical characteristics of the multi-dimensional feature vector, and obtain the eigenvalues and eigenvectors of the covariance matrix through eigenvalue decomposition;

[0140] The selection unit is used to determine the number of principal components according to the eigenvalue magnitudes, and select the top N eigenvectors with the largest eigenvalues as the principal component vectors;

[0141] The projection unit is used to project the multi-dimensional feature vector onto the principal component vectors to obtain the reduced-dimensional feature vector representation and form a compressed feature matrix;

[0142] The denoising unit is used to further reduce the dimension of the compressed feature matrix by using the singular value decomposition method, remove noise and redundant components, and extract feature information;

[0143] The combination unit is used to obtain the finally reduced-dimensional feature matrix by combining principal component analysis and singular value decomposition.

[0144] In this embodiment, constructing a bed state recognition model according to the reduced-dimensional feature matrix includes:

[0145] According to the preset different bed state labels, a classification model, i.e., the bed state recognition model, is constructed using the support vector machine algorithm, and the reduced-dimensional feature matrix is input into the classification model for training;

[0146] According to the training result of the classification model, determine the kernel function type and kernel function parameters of the support vector machine algorithm to obtain an optimized classification model;

[0147] Among them, determining the kernel function type and kernel function parameters of the support vector machine algorithm according to the training result of the classification model includes:

[0148] Step 1: Use the initial search center , where is the initial value of the m eddy centers, is the initial value of the m-th eddy center, use the initial standard deviation , and set the fitness function of the optimal solution as , where is the global best solution of all candidate solutions in the current iteration, and the iteration step t = 0;

[0149] Step 2: With a search radius , around the search center , generate candidate solutions that follow a Gaussian distribution. The total number of candidate solutions is 250. If the value of a candidate solution exceeds the boundary range, then use Equation , where is the candidate solution that exceeds the range, is a random number that conforms to a uniform distribution, and are both d-dimensional vectors, representing the upper and lower bounds of the d-dimensional search space respectively, and transform it inside the boundary;

[0150] Step 3: Select the best solution from and save it to matrix , then select the global best solution from and store it in . If is better than the current global optimal solution, then update the optimal solution ;

[0151] Step 4: Take as the m-th new eddy center, and use Equation , where is the loss function, is the normal vector of the hyperplane, is the transformation of the original feature, is the penalty factor for misclassified samples, is the slack variable, b is the threshold, update the first m - 1 eddy centers, reduce the radius to get , use the transformed m centers as the eddy centers after reducing the radius, generate candidate solutions that follow a normal distribution, and set the transformation iteration step t = t + 1. S.t. is the standard notation in operations research, meaning "such that", is the transformed feature, and T is the transpose of the normal vector of the hyperplane;

[0152] Step 5: Determine whether the termination condition is met: If the iteration step t ≥ E, where is the maximum number of iterations, which is used to control the running time of the algorithm and prevent overfitting, then terminate the iteration and output the optimal solution S; if the termination condition is not met, then go back to Step 2.

[0153] In this embodiment, according to the bed layer state recognition model, dynamically recognizing the bed layer state includes:

[0154] Input the feature matrix to be classified into the classification model according to the optimized classification model, and obtain the bed layer state classification results corresponding to each feature matrix;

[0155] According to the classification results, use the voting method to determine the final state of each bed layer. If the classification results of multiple feature matrices of a certain bed layer are the same, then it is determined that the bed layer is in the state corresponding to this classification result;

[0156] According to the final state of each bed layer, draw the cross-sectional view of the bed layer in different states, and visually display the state distribution of each bed layer in a graphical way;

[0157] According to the state distribution of each bed layer, use the interpolation algorithm to calculate the transition state between each bed layer, and obtain the complete bed layer state distribution map, providing data support for subsequent production management.

[0158] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the spirit of the present invention's design, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for identifying the state of a gas-solid fluidized bed, characterized in that, The method includes: Obtain the pressure fluctuation signal during the operation of the fluidized bed, decompose the pressure fluctuation signal to obtain multi-scale decomposition coefficients, and denoise the multi-scale decomposition coefficients to form a denoised pressure fluctuation signal; According to the denoised pressure fluctuation signal, extract time-domain characteristic parameters and frequency-domain characteristic parameters to obtain a multi-dimensional feature vector, where the time-domain characteristic parameters include: the mean, variance, skewness, and kurtosis of the pressure fluctuation signal, and the frequency-domain characteristic parameters include: the power spectral density and the main frequency component of the pressure fluctuation signal; Perform dimensionality reduction processing on the multi-dimensional feature vector to obtain a dimensionality-reduced feature matrix; Construct a bed state recognition model according to the dimensionality-reduced feature matrix; Dynamically identify the bed state according to the bed state recognition model; Constructing a bed state recognition model according to the dimensionality-reduced feature matrix includes: According to preset different bed state labels, use the support vector machine algorithm to construct a classification model, that is, the bed state recognition model, and input the dimensionality-reduced feature matrix into the classification model for training; According to the training results of the classification model, determine the kernel function type and kernel function parameters of the support vector machine algorithm to obtain an optimized classification model; Among them, determining the kernel function type and kernel function parameters of the support vector machine algorithm according to the training results of the classification model includes: Step 1: Utilize the initialized search center , where is the initial value of m eddy centers, is the initial value of the m-th eddy center, and utilize the initialized standard deviation . The fitness function of the optimal solution is set as , where is the global best solution of all candidate solutions in the current iteration, and the iteration step number t = 0; Step 2: With a search radius , around the search center , generate candidate solutions that follow a Gaussian distribution . The total number of candidate solutions is 250. If the value of a candidate solution exceeds the boundary range, use Equation where is the candidate solution that exceeds the range, is a random number that conforms to a uniform distribution, and are both d-dimensional vectors, representing the upper and lower bounds of the d-dimensional search space respectively, and transform them inside the boundary; Step 3: Select the best solution from and save it to matrix , then select the global best solution from and store it in . If is better than the current global optimal solution, update the optimal solution ; Step 4: Take as the m-th new eddy current center, and use the formula , where is the loss function, is the normal vector of the hyperplane, is the transformation of the original feature, is the penalty factor for misclassified samples, is the slack variable, b is the threshold. Update the first m - 1 eddy current centers, reduce the radius to get . Use the transformed m centers as the eddy current centers after reducing the radius, generate candidate solutions that follow a normal distribution, and update the iteration step t = t + 1. s.t. is the standard notation in operations research, meaning "such that". is the transformed feature, and T is the transpose of the normal vector of the hyperplane; Step 5: Determine whether the termination condition is satisfied: If the number of iteration steps \(t\geq E\), where is the maximum number of iterations, which is used to control the running time of the algorithm and prevent overfitting, then terminate the iteration and output the optimal solution \(S\); if the termination condition is not satisfied, go to Step 2; Dynamically identifying the bed state according to the bed state recognition model includes: According to the optimized classification model, input the feature matrix to be classified into the classification model to obtain the bed state classification results corresponding to each feature matrix; According to the classification results, use the voting method to determine the final state of each bed. If the classification results of multiple feature matrices of a certain bed are the same, it is determined that the bed is in the state corresponding to the classification result; According to the final state of each bed, draw the cross-sectional view of the bed in different states, and visually display the state distribution of each bed in a graphical way; According to the state distribution of each bed, use the interpolation algorithm to calculate the transition state between each bed to obtain a complete bed state distribution map, providing data support for subsequent production management.

2. The method according to claim 1, characterized in that, Obtaining the pressure fluctuation signal during the operation of the fluidized bed, decomposing the pressure fluctuation signal to obtain multi-scale decomposition coefficients, and denoising the multi-scale decomposition coefficients to form a denoised pressure fluctuation signal includes: Obtain the pressure fluctuation signal during the operation of the fluidized bed, convert the pressure fluctuation signal into a digital signal, and perform preprocessing to obtain a preprocessed pressure fluctuation signal; According to the preset wavelet basis function and decomposition level, perform wavelet decomposition on the preprocessed pressure fluctuation signal to obtain wavelet coefficients of each layer; According to the statistical characteristics of the wavelet coefficients of each layer, calculate the thresholds of the wavelet coefficients of each layer, and use the soft threshold method to perform threshold processing on the wavelet coefficients of each layer to obtain denoised wavelet coefficients; According to the denoised wavelet coefficients, use the wavelet reconstruction method to reconstruct the wavelet coefficients of each layer to obtain a denoised pressure fluctuation signal.

3. The method according to claim 1, characterized in that, Extracting time-domain characteristic parameters and frequency-domain characteristic parameters according to the denoised pressure fluctuation signal to obtain a multi-dimensional feature vector includes: Based on the denoised pressure fluctuation signal, using the time-domain analysis method, calculate the mean, variance, skewness, and kurtosis of the signal to obtain the time-domain characteristic parameters; Based on the denoised pressure fluctuation signal, using the frequency-domain analysis method, calculate the power spectral density and main frequency component of the signal to obtain the frequency-domain characteristic parameters; Combine the time-domain characteristic parameters and the frequency-domain characteristic parameters to construct a multi-dimensional feature vector for characterizing the characteristics of the pressure fluctuation signal.

4. The method according to claim 1, wherein Perform dimensionality reduction processing on the multi-dimensional feature vector to obtain the dimensionally reduced feature matrix, including: According to the multi-dimensional feature vector, obtain the dimension information. If the dimension is greater than the preset threshold, it is determined that there is redundant information and dimensionality reduction processing is required; According to the statistical characteristics of the multi-dimensional feature vector, construct a covariance matrix, and through eigenvalue decomposition, obtain the eigenvalues and eigenvectors of the covariance matrix; According to the magnitude of the eigenvalues, determine the number of principal components, and select the top N eigenvectors with the largest eigenvalues as the principal component vectors; Project the multi-dimensional feature vector onto the principal component vectors to obtain the dimensionally reduced feature vector representation, forming a compressed feature matrix; Use the singular value decomposition method to further reduce the dimension of the compressed feature matrix, remove noise and redundant components, and extract feature information; Through the combination of principal component analysis and singular value decomposition, obtain the finally dimensionally reduced feature matrix.

5. A gas-solid fluidized bed bed state recognition system, which is used to implement the method described in any one of claims 1-4, and is characterized in that, The system includes: a decomposition module, an extraction module, a dimensionality reduction module, a construction module, and an identification module; The decomposition module is used to obtain the pressure fluctuation signal during the operation of the fluidized bed, decompose the pressure fluctuation signal to obtain multi-scale decomposition coefficients, and denoise the multi-scale decomposition coefficients to form a denoised pressure fluctuation signal; The extraction module is used to extract time-domain characteristic parameters and frequency-domain characteristic parameters according to the denoised pressure fluctuation signal to obtain a multi-dimensional feature vector, where the time-domain characteristic parameters include: the mean, variance, skewness, and kurtosis of the pressure fluctuation signal, and the frequency-domain characteristic parameters include: the power spectral density and main frequency component of the pressure fluctuation signal; The dimensionality reduction module is used to perform dimensionality reduction processing on the multi-dimensional feature vector to obtain a dimensionally reduced feature matrix; The construction module is used to construct a bed state identification model according to the dimensionally reduced feature matrix; The identification module is used to dynamically identify the bed state according to the bed state identification model; Constructing a bed state identification model according to the dimensionally reduced feature matrix includes: According to the preset different bed state labels, use the support vector machine algorithm to construct a classification model, that is, a bed state identification model, and input the dimensionally reduced feature matrix into the classification model for training; According to the training results of the classification model, determine the kernel function type and kernel function parameters of the support vector machine algorithm to obtain an optimized classification model; Among them, determining the kernel function type and kernel function parameters of the support vector machine algorithm according to the training results of the classification model includes: Step 1: Use the initialization search center , where is the initial value of the m eddy current centers, is the initial value of the m-th eddy current center. Use the initialization standard deviation . The fitness function of the optimal solution is set to , where is the global best solution of all candidate solutions in the current iteration, and the iteration step number t = 0; Step 2: With a search radius , around the search center , generate candidate solutions that follow a Gaussian distribution . The total number of candidate solutions is 250. If the value of a candidate solution exceeds the boundary range, use Equation where is the candidate solution that exceeds the range, is a random number that conforms to a uniform distribution, and are both d-dimensional vectors, representing the upper and lower bounds of the d-dimensional search space respectively, and transform them within the boundaries; Step 3: Select the best solution from and save it to matrix . Then select the global best solution from and store it in . If is better than the current global optimal solution, then update the optimal solution . Step 4: Take as the m-th new eddy current center, and use Equation , where is the loss function, is the normal vector of the hyperplane, is the transformation of the original feature, is the penalty factor for misclassified samples, is the slack variable, b is the threshold, update the first m - 1 eddy current centers, reduce the radius to get , use the transformed m centers as the eddy current centers after reducing the radius, generate candidate solutions that follow a normal distribution, update the iteration step t = t + 1, s.t. is the standard notation in operations research, meaning such that, is the transformed feature, T is the transpose of the normal vector of the hyperplane; Step 5: Determine whether the termination condition is satisfied: If the number of iteration steps \(t\geq E\), where, is the maximum number of iterations, which is used to control the running time of the algorithm and prevent overfitting, then terminate the iteration and output the optimal solution \(S\); if the termination condition is not satisfied, go to Step 2; Dynamically identifying the bed state according to the bed state identification model includes: According to the optimized classification model, input the feature matrix to be classified into the classification model to obtain the bed state classification results corresponding to each feature matrix; According to the classification results, the voting method is used to determine the final state of each bed layer. If the classification results of multiple feature matrices of a certain bed layer are consistent, it is determined that the bed layer is in the state corresponding to the classification result. According to the final state of each bed layer, cross-sectional diagrams of the bed layer in different states are drawn to visually display the state distribution of each bed layer in a graphical way. According to the state distribution of each bed layer, an interpolation algorithm is used to calculate the transition state between each bed layer, and a complete bed layer state distribution diagram is obtained to provide data support for subsequent production management.

6. The system according to claim 5, wherein The decomposition module includes: a conversion unit, a preprocessing unit, a calculation unit, and a reconstruction unit. The conversion unit is used to obtain the pressure fluctuation signal during the operation of the fluidized bed, convert the pressure fluctuation signal into a digital signal, and perform preprocessing to obtain the preprocessed pressure fluctuation signal. The preprocessing unit is used to perform wavelet decomposition on the preprocessed pressure fluctuation signal according to the preset wavelet basis function and decomposition level to obtain wavelet coefficients of each layer. The calculation unit is used to calculate the threshold of the wavelet coefficients of each layer according to the statistical characteristics of the wavelet coefficients of each layer, and perform threshold processing on the wavelet coefficients of each layer by using the soft threshold method to obtain the denoised wavelet coefficients. The reconstruction unit is used to reconstruct the wavelet coefficients of each layer by using the wavelet reconstruction method according to the denoised wavelet coefficients to obtain the denoised pressure fluctuation signal.

7. The system according to claim 5, characterized in that, The extraction module includes: a time-domain feature unit, a frequency-domain feature unit, and a combination unit. The time-domain feature unit is used to calculate the mean, variance, skewness, and kurtosis of the signal by using the time-domain analysis method according to the denoised pressure fluctuation signal to obtain time-domain feature parameters. The frequency-domain feature unit is used to calculate the power spectral density and main frequency component of the signal by using the frequency-domain analysis method according to the denoised pressure fluctuation signal to obtain frequency-domain feature parameters. The combination unit is used to combine the time-domain feature parameters and the frequency-domain feature parameters to construct a multi-dimensional feature vector for characterizing the features of the pressure fluctuation signal.

8. The system according to claim 5, wherein, The dimensionality reduction module includes: an acquisition unit, an eigenvalue unit, a selection unit, a projection unit, a denoising unit, and a combination unit. The acquisition unit is used to obtain the dimension information according to the multi-dimensional feature vector. If the dimension is greater than the preset threshold, it is determined that there is redundant information and dimensionality reduction processing is required. The eigenvalue unit is used to construct a covariance matrix according to the statistical characteristics of the multi-dimensional feature vector, and obtain the eigenvalues and eigenvectors of the covariance matrix through eigenvalue decomposition. The selection unit is used to determine the number of principal components according to the eigenvalue size, and select the first N eigenvectors with the largest eigenvalues as the principal component vectors. The projection unit is used to project the multi-dimensional feature vector onto the principal component vectors to obtain a reduced-dimensional feature vector representation and form a compressed feature matrix. The denoising unit is used to further reduce the dimension of the compressed feature matrix by using the singular value decomposition method to remove noise and redundant components and extract feature information. The combination unit is used to obtain the finally reduced-dimensional feature matrix by combining the principal component analysis and the singular value decomposition.

Images