A multi-modal fault monitoring method and system for tobacco flavoring process

Abstract

This invention provides a multimodal fault monitoring method and system for the tobacco flavoring process, belonging to the field of fault diagnosis technology. The multimodal fault monitoring method includes: collecting multidimensional sensor signal data during normal operation of multiple modes in the tobacco flavoring process and constructing a training sample dataset; constructing an adaptive discriminant principal component analysis model based on the training sample dataset and solving for the projection matrix and projection coefficients using an offline iterative algorithm; inputting the projection matrix and projection coefficients into a PLC; collecting online monitored multidimensional sensor signal data and constructing a test dataset; calculating the low-dimensional projection of the test dataset and then performing modal identification; and calculating sample anomalies in the test dataset based on the modal identification results. This multimodal fault monitoring method can solve the above problems and achieve real-time intelligent monitoring of multimodal data in the flavoring process on a PLC with highly limited computing resources.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and specifically to a multimodal fault monitoring method and system for the tobacco flavoring process. Background Technology

[0002] In recent years, with the continuous improvement of sensor technology and information processing capabilities, multimodal fault monitoring technology has made significant progress. By fusing heterogeneous data from different sensors, the system's operating status can be more comprehensively characterized, improving the accuracy and robustness of fault detection. In the tobacco flavoring process, the equipment operating environment is complex, involving multiple sensing modes such as flavoring spray flow rate, nozzle pressure, feed liquid temperature, tobacco moisture content, drum speed, motor current, and conveyor belt load. Moreover, the sampling characteristics and signal distribution of different sensors vary significantly, making multimodal fusion monitoring particularly important in industrial settings.

[0003] Currently, industrial control systems are primarily PLCs. While these systems offer high reliability, their limited computing resources make it difficult to execute complex matrix operations, deep learning models, or large-scale iterative calculations. The multimodal data from the incense repair process is characterized by high dimensionality, strong correlation, and inconsistent modal distributions, rendering most multimodal algorithms proposed in academia unsuitable for direct real-time execution on PLCs.

[0004] Against this backdrop, Principal Component Analysis (PCA) has become one of the few feature extraction and fault diagnosis techniques that can stably operate on PLCs due to its linearity, algorithmic simplicity, and extremely low computational overhead. PCA can project high-dimensional multimodal measurements such as temperature, humidity, and electrical signals into a low-dimensional principal component space, effectively achieving noise reduction, decorrelation, and data compression, meeting the requirements of real-time monitoring and lightweight computation in the fragrance application process. However, PCA's global dimensionality reduction approach cannot distinguish between different modes (e.g., fragrance flow mode versus pressure mode, motor mode, and moisture content mode), easily leading to mode aliasing problems. For example, small fluctuations in nozzle pressure may mask fragrance flow distortion, easily resulting in misjudgment or missed judgment during multi-condition switching. Furthermore, traditional PCA does not consider intra-class compactness and inter-class separability, making stable monitoring under complex disturbances in the fragrance application process difficult to achieve. Therefore, an adaptive discriminative principal component analysis (AD-PCA) method is needed to achieve real-time intelligent monitoring of multimodal data in the fragrance application process on PLCs with highly limited computational resources. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal fault monitoring method and system for the tobacco flavoring process. This multimodal fault monitoring method can solve the above-mentioned problems and realize real-time intelligent monitoring of multimodal data of the flavoring process on a PLC with highly limited computing resources.

[0006] To achieve the above objectives, embodiments of the present invention provide a multimodal fault monitoring method for the tobacco flavoring process, the multimodal fault monitoring method comprising: Collect multidimensional sensor signal data during normal operation of multiple modes in the tobacco flavoring process, and construct a training sample dataset; An adaptive discriminant principal component analysis model is constructed based on the training sample dataset, and the projection matrix and projection coefficients are solved by an offline iterative algorithm. The projection matrix and projection coefficients are then input into the PLC; Collect multi-dimensional sensor signal data for online monitoring and construct a test dataset; Based on the principal component analysis model constructed from the projection matrix and projection coefficients in the PLC, the low-dimensional projection of the test dataset is calculated, and then modal recognition is performed. Based on the results of the modality recognition, sample anomaly calculation is performed on the test dataset, and the fault detection rate and false alarm rate are calculated based on the sample anomaly calculation results to evaluate the monitoring performance.

[0007] Optionally, an adaptive discriminative principal component analysis model is constructed based on the training sample dataset, and the projection matrix and projection coefficients are solved using an offline iterative algorithm, including: Construct an adaptive discriminant principal component analysis model based on formula (1): Formula (1) in, This indicates that the training sample dataset belongs to a certain modality. A subset of samples, each subset containing Sample , It is modal The adaptive transformation matrix, Indices representing other modalities , for modality The corresponding projection matrix, and It is derived from the empirical covariance matrix of all training data through singular value decomposition, and then the top contributors are selected based on the cumulative variance contribution rate. A shared fundamental projection matrix consisting of principal component vectors, where , Indicates that for the first For a specific mode, the number of principal components required to ensure that the main information of that mode is not lost is estimated separately using the cumulative variance contribution rate method. It is modal Projection coefficients in the principal component space. and They are modal projection coefficient The mean of and the global mean of all modes, and For model parameters, It is the identity matrix. Denotes the square of the Frobenius norm. Representing modes The transpose of the adaptive transformation matrix, the same below. Representing modes The adaptive transformation matrix, The projection coefficient matrix of all modes in the principal component space is also known as the global projection coefficient matrix. This represents the Frobenius norm.

[0008] Optionally, an adaptive discriminative principal component analysis model is constructed based on the training sample dataset, and the projection matrix and projection coefficients are solved using an offline iterative algorithm, including: Set learning rate , This represents the number of iterations. Calculate the information matrix: ,in For the objective function with respect to gradient: ,in Indicates the t-th The adaptive transformation matrix in the first iteration. This indicates the state before mode i is updated, i.e., the t-th time. The adaptive transformation matrix in the first iteration. express transpose, express transpose, express transpose, express transpose, express transpose; The adaptive transformation matrix is ​​updated according to the information matrix using formula (2): Formula (2) in, This represents the updated adaptive transformation matrix. The adaptive transformation matrix before the update Indicates the learning rate; If the updated adaptive transformation matrix satisfies formula (3), stop iterating on the adaptive transformation matrix: Formula (3) in, Representing scientific notation; After the iteration is complete, based on the iterated adaptive transformation matrix, the formula is used... Calculate the corresponding projection matrix.

[0009] Optionally, an adaptive discriminative principal component analysis model is constructed based on the training sample dataset, and the projection matrix and projection coefficients are solved using an offline iterative algorithm, including: Set learning rate , For the number of iterations, As auxiliary parameters, This indicates the initialization of the auxiliary matrix; The objective function is calculated using formula (4) with respect to... gradient: Formula (4) in, This represents the global mean matrix after expansion and concatenation. Representing the A column vector of the mean values ​​of each modality. Indicates mode 1 to... The transpose of the adaptive transformation matrix. Indicates the corresponding mode arrive a subset of samples It is a row vector consisting entirely of 1s. Indicates the first The global projection coefficient matrix at the next iteration Representing modes The transpose of the corresponding projection matrix 1 represents a column vector of all 1s with appropriate dimensions; Initial update : ,in, This represents the temporary update matrix generated by the regular gradient descent step; Calculate parameters and update again in, This represents the acceleration parameters calculated in the current iteration step. This indicates the acceleration parameters of the previous iteration step; Update the remaining parameters: ; Based on updated parameters and ,judge Convergence: If convergence reaches a set threshold, iteration stops, and the corresponding value is obtained. ,according to It can obtain the corresponding projection coefficients.

[0010] Optionally, based on the principal component analysis model constructed from the projection matrix and projection coefficients in the PLC, a low-dimensional projection of the test dataset is calculated, and then modal recognition is performed, including: The test dataset is obtained and fed into the principal component analysis model constructed based on the projection matrix and projection coefficients in the PLC to obtain the corresponding parameters; Collect test samples and corresponding parameters from the test dataset, and obtain the corresponding low-dimensional representation based on formula (5): Formula (5) in, For the first The mean of the data type. Representing modes The transpose of the adaptive transformation matrix. Indicates the first One test sample.

[0011] Optionally, based on the principal component analysis model constructed from the projection matrix and projection coefficients in the PLC, the low-dimensional projection of the test dataset is calculated, and then modal identification is performed, including: obtaining the modality corresponding to the test sample through formula (6): Formula (6) in, Preset weighting parameters are used to balance the contributions of the two error terms. This represents the total number of modes.

[0012] Optionally, based on the results of the modality recognition, sample anomaly calculation is performed on the test dataset, and the fault detection rate and false alarm rate are calculated based on the sample anomaly calculation results to evaluate the monitoring performance, including: obtaining the Hotelling statistic and squared prediction error of the test samples through formula (7): Formula (7) in, This represents the Hotelling statistic. Indicates the squared prediction error. Indicates test sample When identified as a modality Low-dimensional projection of time.

[0013] Optionally, based on the modality recognition results, sample anomaly calculation is performed on the test dataset, and the fault detection rate and false alarm rate are calculated based on the sample anomaly calculation results to evaluate the monitoring performance, including: Obtain the Hotelling statistic and squared prediction error for the test sample; If either the Hotling statistic or the squared prediction error of the test sample exceeds the corresponding preset threshold, it is determined that the corresponding test sample may be faulty. The number of faults in the test samples and the total number of samples were statistically analyzed, and the corresponding fault detection rate and false alarm rate were obtained according to formula (8): Formula (8) The performance of the monitoring is evaluated based on the calculated fault detection rate and false alarm rate.

[0014] On the other hand, the present invention can also provide a multimodal fault monitoring system for the tobacco flavoring process, the multimodal fault monitoring system comprising: Multidimensional sensors are used to collect multidimensional sensor signal data during the normal operation of multiple modes in the tobacco flavoring process; The PLC module is connected to the multidimensional sensor and is used to execute a multimodal fault monitoring method for the tobacco flavoring process as described above.

[0015] Through the above technical solution, this invention provides a multimodal fault monitoring method and system for the tobacco flavoring process. It collects multidimensional sensor signal data during the normal operation of multiple modes in the tobacco flavoring process, and then constructs a training sample dataset based on this data. After obtaining the training sample dataset, an adaptive discriminant principal component analysis (PCA) model can be constructed, and the projection matrix and projection coefficients can be solved using an offline iterative algorithm. After obtaining the projection matrix and projection coefficients, they can be input into a PLC. Online monitoring of multidimensional sensor signal data can be collected, and a test dataset can be constructed. Based on the PCA model constructed using the projection matrix and projection coefficients in the PLC, the test dataset is input into the PCA model to calculate the low-dimensional projection of the test dataset, and then modal identification can be performed. Based on the results of the modal identification, sample anomaly analysis can be performed on the test dataset, and the fault detection rate and false alarm rate can be calculated based on the anomaly calculation results, thereby evaluating the detection performance. This multimodal fault monitoring method enables real-time intelligent monitoring of multimodal data in the flavoring process on a PLC with highly limited computing resources.

[0016] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a multimodal fault monitoring method for tobacco flavoring process according to an embodiment of the present invention; Figure 2 This is a comparison diagram of modal identification results and mixed principal component analysis method for a multimodal fault monitoring method for tobacco flavoring process according to an embodiment of the present invention; Figure 3 This is the result of fault monitoring in a multimodal fault monitoring method for tobacco flavoring process according to an embodiment of the present invention. Figure 4 This is the result of fault monitoring using a hybrid principal component analysis method according to an embodiment of the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0019] In the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used the solution.

[0020] Figure 1 This is a flowchart of a multimodal fault monitoring method for a tobacco flavoring process according to an embodiment of the present invention. In this invention, the flowchart of the multimodal fault monitoring method may include: In step S1, multidimensional sensor signal data of multiple modes are collected during normal operation of the tobacco flavoring process, and a training sample dataset is constructed.

[0021] In step S2, an adaptive discriminant principal component analysis model is constructed based on the training sample dataset, and the projection matrix and projection coefficients are solved by an offline iterative algorithm.

[0022] In step S3, the projection matrix and projection coefficients are input into the PLC.

[0023] In step S4, multi-dimensional sensor signal data from online monitoring are collected, and a test dataset is constructed.

[0024] In step S5, based on the principal component analysis model constructed from the projection matrix and projection coefficients in the PLC, the low-dimensional projection of the test dataset is calculated, and then modal recognition is performed.

[0025] In step S6, sample anomaly calculation is performed on the test dataset based on the modality recognition results, and the fault detection rate and false alarm rate are calculated based on the sample anomaly calculation results to evaluate the monitoring performance.

[0026] In this invention, during fault monitoring, multi-dimensional sensor signal data from multiple modes of normal operation during the tobacco flavoring process can be collected. A training sample dataset can then be constructed based on this multi-dimensional sensor signal data. After obtaining the training sample dataset, an adaptive discriminative principal component analysis (PCA) model can be constructed based on it, and the projection matrix and projection coefficients can be solved using an offline iterative algorithm. After obtaining the projection matrix and projection coefficients, they can be input into a PLC. Online monitoring multi-dimensional sensor signal data can be collected, and a test dataset can be constructed. Based on the PCA model constructed using the projection matrix and projection coefficients in the PLC, the test dataset can be input into the PCA model to calculate the low-dimensional projection of the test dataset, and then modal recognition can be performed. Based on the modal recognition results, sample anomaly analysis can be performed on the test dataset, and the fault detection rate and false alarm rate can be calculated based on the anomaly calculation results, thereby evaluating the detection performance. This multi-modal fault monitoring method enables real-time intelligent monitoring of multi-modal data during the flavoring process on a PLC with highly limited computing resources.

[0027] In this invention, an adaptive discriminant principal component analysis model can be constructed using formula (1): Formula (1) in, This indicates that the training sample dataset belongs to a certain modality. A subset of samples, each subset containing Sample , It is modal The adaptive transformation matrix, Indices representing other modalities , for modality The corresponding projection matrix, and It is derived from the empirical covariance matrix of all training data through singular value decomposition, and then the top contributors are selected based on the cumulative variance contribution rate. A shared fundamental projection matrix consisting of principal component vectors, where , Indicates that for the first For a specific mode, the number of principal components required to ensure that the main information of that mode is not lost is estimated separately using the cumulative variance contribution rate method. It is modal Projection coefficients in the principal component space. and They are modal projection coefficient The mean of and the global mean of all modes, and For model parameters, It is the identity matrix. Denotes the square of the Frobenius norm. Representing modes The transpose of the adaptive transformation matrix, the same below. Representing modes The adaptive transformation matrix, The projection coefficient matrix of all modes in the principal component space is also known as the global projection coefficient matrix. This represents the Frobenius norm. In mathematical expression... Subscript This indicates that the Frobenius norm is being used.

[0028] The adaptive discriminant principal component analysis model can be composed of three parts, among which... It is the data reconstruction error, that is, the error in the training sample dataset that belongs to the modality sample subset Through projection matrix Reconstruct the data after performing a low-dimensional representation. By penalizing the similarity between the projection matrices of different modes, the coupling between modes is reduced, and the coupling is enhanced. The ability to discriminate, at the same time They all contain factors This is compatible with sharing the same industrial configuration between different modes. Fisher's discrimination criterion improves discrimination ability by reducing intra-modal dispersion and increasing inter-modal separation, while additional... This helps enhance the convexity of the model. Constraints make sure It is an orthogonal matrix, which ensures that the projection variance is interpreted without repetition, thus further improving the stability and interpretability of the model.

[0029] In one embodiment of the present invention, in the model The offline training and iterative update are based on the Stiefel manifold gradient descent algorithm, which then yields the projection matrix, specifically including: Set learning rate , This represents the number of iterations.

[0030] Calculate the information matrix: ,in For the objective function with respect to gradient: ,in Indicates the t-th The adaptive transformation matrix in the first iteration, in the gradient descent algorithm based on the Stiefel manifold, is often constructed in the standard form of an information matrix (skew-symmetric matrix) as follows: (in For the current variable, (For the gradient). In the formula here, In fact, it refers to the current optimization variable, which is equivalent to... . This indicates the state before mode i is updated, i.e., the t-th time. The adaptive transformation matrix in the first iteration. express transpose, express transpose, express transpose, express transpose, express The transpose of .

[0031] The adaptive transformation matrix is ​​updated according to the information matrix using formula (2): Formula (2) in, This represents the updated adaptive transformation matrix. The adaptive transformation matrix before the update This represents the learning rate.

[0032] If the updated adaptive transformation matrix satisfies formula (3), stop iterating on the adaptive transformation matrix: Formula (3) in, To express scientific notation.

[0033] After the iteration is complete, based on the iterated adaptive transformation matrix, the formula is used... Calculate the corresponding projection matrix.

[0034] In one embodiment of the present invention, regarding the model The iterative update is based on the first-order accelerated gradient descent algorithm, specifically including: Set learning rate , For the number of iterations, As auxiliary parameters, This indicates the initialization of the auxiliary matrix. In accelerated gradient descent algorithms, to achieve momentum acceleration, it is usually necessary to introduce and maintain an additional auxiliary state matrix. Here, it means that the auxiliary matrix is ​​initially assigned a matrix of all zeros. The objective function is calculated using formula (4) with respect to... gradient: Formula (4) in, This represents the global mean matrix after expansion and concatenation. Representing the A column vector of the mean values ​​of each modality. Indicates mode 1 to... The transpose of the adaptive transformation matrix. Indicates the corresponding mode arrive a subset of samples These are row vectors all equal to 1. The physical meaning of this formula is to horizontally copy and expand the mean vector of each modality according to the number of samples in that modality, and finally concatenate them into a large matrix with the same dimension as the global data, so as to facilitate the calculation of intra-class or inter-class distances in matrix operations. Indicates the first The global projection coefficient matrix at the next iteration is the starting state before the current iteration update. Representing modes The transpose of the corresponding projection matrix 1 represents a column vector of all 1s with appropriate dimensions.

[0035] Initial update : ,in, This represents the temporary update matrix generated by the regular gradient descent step. It is the result of updating only one step along the current gradient direction, without adding the acceleration (momentum) terms from the historical state.

[0036] Calculate parameters and update again ,in, This represents the acceleration parameter calculated in the current iteration step. It is a dynamically updated scalar used to control the proportion of influence of historical information (momentum) on the current matrix update. (This represents the acceleration parameter from the previous iteration, i.e., the old value before the update), used to substitute into the formula to calculate the current value. .

[0037] Update the remaining parameters: ; Based on updated parameters and ,judge Convergence: If convergence reaches a set threshold, iteration stops, and the corresponding value is obtained. ,according to It can obtain the corresponding projection coefficients.

[0038] In this invention, the process for performing modal recognition may include: In step S7, the test dataset is obtained and fed into the principal component analysis model constructed based on the projection matrix and projection coefficients in the PLC to obtain the corresponding parameters.

[0039] In step S8, test samples and corresponding parameters are collected from the test dataset, and their corresponding low-dimensional representations are obtained based on formula (5): Formula (5) in, For the first The mean of the data type. Representing modes The transpose of the adaptive transformation matrix. Indicates the first One test sample.

[0040] When performing modal recognition, a test dataset can be obtained and fed into a principal component analysis model constructed based on the projection matrix and projection coefficients in the PLC, and then the corresponding parameters can be obtained. The corresponding parameters can be data such as the mean in formula (5). Based on the test sample, the corresponding low-dimensional representation can be obtained through formula (5).

[0041] After obtaining the corresponding low-dimensional representation, the modality corresponding to the test sample can be obtained through formula (6): Formula (6) in, Preset weighting parameters are used to balance the contributions of the two error terms. This represents the total number of modes. After calculating the discrimination formula, as shown... Figure 2 As shown, the model achieves a recognition accuracy of 98.26%, while the traditional hybrid principal component analysis method only achieves 53.59%, fully demonstrating the superior performance of this invention in multimodal monitoring. This significant improvement is attributed to the innovations of this invention in low-dimensional data representation, adaptive modal transformation, and discrimination criterion design. By combining principal component analysis with an adaptive discrimination strategy, not only are the main features of each modality's data effectively extracted, but the distinguishing ability between different modes is also enhanced, and coupling interference is reduced, thereby significantly improving the accuracy and robustness of fault detection.

[0042] In one embodiment of the present invention, when obtaining the fault detection rate and false alarm rate, the Hotelling statistic and squared prediction error of the test sample can be obtained first by formula (7): Formula (7) in, This represents the Hotelling statistic. Indicates the squared prediction error. Indicates test sample When identified as a modality The low-dimensional projection of time. It represents substituting the current test sample into the identified optimal matching mode. The low-dimensional feature representation obtained after the projection model is calculated will be used to calculate the feature representation in this mode. and The Hotelling statistic measures the degree to which a sample deviates from its distribution in a low-dimensional space, while the squared prediction error reflects the error in reconstructing the sample in the original space.

[0043] In one embodiment of the present invention, the process for evaluating the performance of monitoring may include: In step S9, the Hotelling statistic and squared prediction error of the test sample are obtained.

[0044] In step S10, if either the Hotelling statistic or the squared prediction error of the test sample is greater than the corresponding preset threshold, it is determined that the corresponding test sample may be faulty.

[0045] In step S11, the number of faults in the test samples and the total number of samples are counted, and the corresponding fault detection rate and false alarm rate are obtained according to formula (8): Formula (8) In step S12, the monitoring performance is evaluated based on the calculated fault detection rate and false alarm rate.

[0046] In this invention, when evaluating detection performance, the Hotelling statistic and squared prediction error of the test sample can be obtained. When either the Hotelling statistic or the squared prediction error of the test sample exceeds a corresponding preset threshold, it is determined that the corresponding test sample may be faulty. Then, the number of faults in the test samples and the total number of samples can be counted, and the corresponding fault detection rate and false alarm rate can be obtained according to formula (8). Based on the calculated fault detection rate and false alarm rate, the monitoring performance can be evaluated.

[0047] like Figure 3 As shown, through the calculation and judgment of the two statistics, the fault detection rates are as high as 85.78% and 99.77%, respectively, while the false alarm rates are both as low as 1.02%. This result fully demonstrates the high accuracy and reliability of the method of this invention in fault detection. In contrast, traditional hybrid principal component analysis methods, such as... Figure 4 As shown, through the calculation and judgment of the two statistics, although the false alarm rate was 0% during the detection process, the fault detection rate was only 5.84% and 37.85%.

[0048] Low fault detection rates in industrial fragrance replenishment processes can lead to missed faults, resulting in reduced production efficiency, equipment damage, and even safety accidents. Missed faults not only prevent timely preventative measures but can also cause prolonged abnormal operation, leading to resource waste and economic losses, seriously threatening the safety and stability of industrial production. The high fault detection rate of this invention is primarily due to the introduction of the Fisher discriminant criterion and the orthogonality penalty term of the projection matrix. This further enhances the model's sensitivity to fault samples, allowing abnormal states to significantly deviate from the normal data distribution in low-dimensional space. This ensures the timely detection and handling of potential faults in practical applications, effectively improving the safety and stability of industrial processes.

[0049] On the other hand, the present invention can also provide a multimodal fault monitoring system for the tobacco flavoring process, the multimodal fault monitoring system comprising: Multidimensional sensors are used to collect multidimensional sensor signal data during the normal operation of multiple modes in the tobacco flavoring process; The PLC module is connected to the multidimensional sensor and is used to execute a multimodal fault monitoring method for the tobacco flavoring process as described above.

[0050] Through the above technical solution, this invention provides a multimodal fault monitoring method and system for the tobacco flavoring process. It collects multidimensional sensor signal data during the normal operation of multiple modes in the tobacco flavoring process, and then constructs a training sample dataset based on this data. After obtaining the training sample dataset, an adaptive discriminant principal component analysis (PCA) model can be constructed, and the projection matrix and projection coefficients can be solved using an offline iterative algorithm. After obtaining the projection matrix and projection coefficients, they can be input into a PLC. Online monitoring of multidimensional sensor signal data can be collected, and a test dataset can be constructed. Based on the PCA model constructed using the projection matrix and projection coefficients in the PLC, the test dataset is input into the PCA model to calculate the low-dimensional projection of the test dataset, and then modal identification can be performed. Based on the results of the modal identification, sample anomaly analysis can be performed on the test dataset, and the fault detection rate and false alarm rate can be calculated based on the anomaly calculation results, thereby evaluating the detection performance. This multimodal fault monitoring method enables real-time intelligent monitoring of multimodal data in the flavoring process on a PLC with highly limited computing resources.

[0051] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0052] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0056] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0057] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0058] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0059] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A multimodal fault monitoring method for tobacco flavoring process, characterized in that, The multimodal fault monitoring method includes: Collect multidimensional sensor signal data during normal operation of multiple modes in the tobacco flavoring process, and construct a training sample dataset; An adaptive discriminant principal component analysis model is constructed based on the training sample dataset, and the projection matrix and projection coefficients are solved by an offline iterative algorithm. The projection matrix and projection coefficients are then input into the PLC; Collect multi-dimensional sensor signal data for online monitoring and construct a test dataset; Based on the principal component analysis model constructed from the projection matrix and projection coefficients in the PLC, the low-dimensional projection of the test dataset is calculated, and then modal recognition is performed. Based on the results of the modality recognition, sample anomaly calculation is performed on the test dataset, and the fault detection rate and false alarm rate are calculated based on the sample anomaly calculation results to evaluate the monitoring performance.

2. The multimodal fault monitoring method according to claim 1, characterized in that, An adaptive discriminative principal component analysis model is constructed based on the training sample dataset, and the projection matrix and projection coefficients are solved using an offline iterative algorithm, including: Construct an adaptive discriminant principal component analysis model based on formula (1): Official (1) in, This indicates that the training sample dataset belongs to a certain modality. A subset of samples, each subset containing Sample , It is modal The adaptive transformation matrix, Indices representing other modalities , for modality The corresponding projection matrix, and It is derived from the empirical covariance matrix of all training data through singular value decomposition, and then the top contributors are selected based on the cumulative variance contribution rate. A shared fundamental projection matrix consisting of principal component vectors, where , Indicates that for the first For a specific mode, the number of principal components required to ensure that the main information of that mode is not lost is estimated separately using the cumulative variance contribution rate method. It is modal Projection coefficients in the principal component space. and They are modal projection coefficient The mean of and the global mean of all modes, and For model parameters, It is the identity matrix. Denotes the square of the Frobenius norm. Representing modes The transpose of the adaptive transformation matrix, the same below. Representing modes The adaptive transformation matrix, The projection coefficient matrix of all modes in the principal component space is also known as the global projection coefficient matrix. This represents the Frobenius norm.

3. The multimodal fault monitoring method according to claim 2, characterized in that, An adaptive discriminative principal component analysis model is constructed based on the training sample dataset, and the projection matrix and projection coefficients are solved using an offline iterative algorithm, including: Set learning rate , This represents the number of iterations. Calculate the information matrix: ,in For the objective function with respect to gradient: ,in Indicates the t-th The adaptive transformation matrix in the first iteration. This indicates the state before mode i is updated, i.e., the t-th time. The adaptive transformation matrix in the first iteration. express transpose, express transpose, express transpose, express transpose, express Transpose of; The adaptive transformation matrix is ​​updated according to the information matrix using formula (2): Official (2) in, This represents the updated adaptive transformation matrix. The adaptive transformation matrix before the update Indicates the learning rate; If the updated adaptive transformation matrix satisfies formula (3), stop iterating on the adaptive transformation matrix: Official (3) in, Representing scientific notation; After the iteration is complete, based on the iterated adaptive transformation matrix, the formula is used... Calculate the corresponding projection matrix.

4. The multimodal fault monitoring method according to claim 2, characterized in that, An adaptive discriminative principal component analysis model is constructed based on the training sample dataset, and the projection matrix and projection coefficients are solved using an offline iterative algorithm, including: Set learning rate , For the number of iterations, As auxiliary parameters, This indicates the initialization of the auxiliary matrix; The objective function is calculated using formula (4) with respect to... gradient: Official (4) in, This represents the global mean matrix after expansion and concatenation. Representing the A column vector of the mean values ​​of each modality. Indicates mode 1 to... The transpose of the adaptive transformation matrix. Indicates the corresponding mode arrive a subset of samples It is a row vector consisting entirely of 1s. Indicates the first The global projection coefficient matrix at the next iteration Representing modes The transpose of the corresponding projection matrix, where 1 represents a column vector of all 1s with appropriate dimensions; Initial update : ,in, This represents the temporary update matrix generated by the regular gradient descent step; Calculate parameters and update again ,in, This represents the acceleration parameters calculated in the current iteration step. This indicates the acceleration parameters of the previous iteration step; Update the remaining parameters: ; Based on updated parameters and ,judge Convergence: If convergence reaches a set threshold, iteration stops, and the corresponding value is obtained. ,according to It can obtain the corresponding projection coefficients.

5. The multimodal fault monitoring method according to claim 1, characterized in that, Based on the principal component analysis model constructed using the projection matrix and projection coefficients in the PLC, the low-dimensional projection of the test dataset is calculated, and then modal recognition is performed, including: The test dataset is obtained and fed into the principal component analysis model constructed based on the projection matrix and projection coefficients in the PLC to obtain the corresponding parameters; Collect test samples and corresponding parameters from the test dataset, and obtain the corresponding low-dimensional representation based on formula (5): Official (5) in, For the first The mean of the data type. Representing modes The transpose of the adaptive transformation matrix. Indicates the first One test sample.

6. The multimodal fault monitoring method according to claim 5, characterized in that, Based on the principal component analysis model constructed from the projection matrix and projection coefficients in the PLC, the low-dimensional projection of the test dataset is calculated, and then modal identification is performed, including: obtaining the modality corresponding to the test sample through formula (6): Official (6) in, Preset weighting parameters are used to balance the contributions of the two error terms. This represents the total number of modes.

7. The multimodal fault monitoring method according to claim 6, characterized in that, Based on the results of the modality recognition, sample anomaly calculation is performed on the test dataset, and the fault detection rate and false alarm rate are calculated based on the sample anomaly calculation results to evaluate the monitoring performance, including: obtaining the Hotelling statistic and squared prediction error of the test samples through formula (7): Official (7) in, This represents the Hotelling statistic. Indicates the squared prediction error. Indicates test sample When identified as a modality Low-dimensional projection of time.

8. The multimodal fault monitoring method according to claim 7, characterized in that, Based on the modality recognition results, sample anomaly calculation is performed on the test dataset, and the fault detection rate and false alarm rate are calculated based on the sample anomaly calculation results to evaluate the monitoring performance, including: Obtain the Hotelling statistic and squared prediction error for the test sample; If either the Hotling statistic or the squared prediction error of the test sample exceeds the corresponding preset threshold, it is determined that the corresponding test sample may be faulty. The number of faults in the test samples and the total number of samples were statistically analyzed, and the corresponding fault detection rate and false alarm rate were obtained according to formula (8): Official (8) The performance of the monitoring is evaluated based on the calculated fault detection rate and false alarm rate.

9. A multimodal fault monitoring system for tobacco flavoring process, characterized in that, The multimodal fault monitoring system includes: Multidimensional sensors are used to collect multidimensional sensor signal data during the normal operation of multiple modes in the tobacco flavoring process; A PLC module, connected to the multidimensional sensor, is used to execute a multimodal fault monitoring method for tobacco flavoring process as described in any one of claims 1-8.

Images