Industrial process monitoring method and system based on cross-modal information fusion

Abstract

The invention discloses an industrial process monitoring method and system based on cross-modal information fusion, and the method comprises the steps: constructing a dictionary learning model for the refined decomposition of dictionary coefficients, and solving the dictionary learning model based on a training sample set which is constructed based on the data during the normal operation of a plurality of modals; and the definition of modal division is enhanced through separation of inter-class and intra-class information, so that different modal boundaries are clearer. Constructing a modal recognition and process monitoring model based on the dictionary learning model, solving an inter-class coefficient and an intra-class coefficient of the real-time sample, and obtaining a weighted error of the real-time sample; and judging the modal attribution of the real-time sample and whether a fault exists in the industrial process based on the weighted error. The original sparsity constraint is equivalently replaced by adopting a weighted quadratic form in the modal identification and process monitoring model, so that the model can adaptively adjust the weight of each element according to the characteristics of different data samples to realize more flexible and more accurate sparse representation.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial process monitoring, and particularly relates to an industrial process monitoring method and system based on cross-modal information fusion. Background Art

[0002] In recent years, the rapid development of data acquisition and processing technologies has greatly promoted the progress of data-driven process monitoring methods. Methods such as principal component analysis, partial least squares regression, and canonical correlation analysis have been widely applied in many fields. However, these traditional methods mainly rely on global statistical characteristics, such as covariance structure or correlation, and usually assume that samples come from a single-modal distribution. This assumption is often difficult to hold in actual industrial processes because, affected by various factors such as product specifications, manufacturing strategies, and environmental factors, the system often exhibits multi-modal characteristics, resulting in diverse data distributions. To address this challenge, some methods attempt to describe the operating states of each mode by establishing multiple models or hybrid models. However, this not only requires setting separate fault detection thresholds for each mode but also significantly increases the consumption of computing resources and the complexity of the model when the number of modes increases.

[0003] In contrast, inspired by the idea of sparse representation, the dictionary learning method provides a new solution. Traditional methods are difficult to directly associate with the local features of data, while dictionary learning "learns" a set of dictionary atoms from the data itself and represents the data as a sparse linear combination composed of these atoms. Each dictionary atom directly reflects the local structure and inherent characteristics of the data, enabling the model to more flexibly adapt to the distribution changes of multi-modal data. For normal samples, an over-complete dictionary can provide rich representation redundancy to accurately reconstruct the data through different atom combinations; for abnormal samples, due to the obvious deviation of their features from normal data, it is often impossible to find a suitable linear combination in the dictionary representation space, resulting in a high reconstruction error, which provides an effective basis for anomaly detection. It is this adaptive representation ability based on local features that makes dictionary learning show higher accuracy and reliability in multi-modal process monitoring and anomaly detection.

[0004] Although significant progress has been made in multi-modal industrial process monitoring technology based on dictionary learning, there are still the following limitations:

[0005] (1) Existing global dictionary models often assume that the data has a unified distribution structure. However, the data in industrial processes usually exhibits multi-modal characteristics. This makes it difficult for a global model to simultaneously adapt to the features of different modes, resulting in better reconstruction effects in some modes and higher false alarm rates and lower fault detection rates in other modes.

[0006] (2) To more precisely describe each modality, many methods choose to build a dictionary model for each modality separately. However, this approach not only significantly increases the computational and storage costs, but also easily overlooks the possible commonalities and interaction relationships between different modalities. In industrial processes, each modality often shares some configuration information. Therefore, separate modeling may not be able to fully utilize these shared characteristics to achieve cross-modal information fusion.

[0007] (3) Most existing methods focus on minimizing the data reconstruction error. However, within the same modality, the sample distribution is relatively loose, and the discrimination between different modalities is insufficient, thus affecting the stability and generalization ability of fault detection and pattern recognition.

[0008] (4) The current technology has deficiencies in distinguishing real modal features from process disturbances, resulting in difficult-to-clearly define boundaries between different modalities, and thus difficult to accurately capture the details of process changes. This not only affects the classification performance, but also restricts the accuracy of fault monitoring. Summary of the Invention

[0009] The object of the present invention is to provide an industrial process monitoring method and system based on cross-modal information fusion for multi-modal industrial process monitoring to achieve modal recognition and fault detection. This method provides a more accurate and efficient solution to the problems of insufficient modal recognition accuracy, lagging fault detection, and high false alarm rate and missed alarm rate in the existing technology.

[0010] The present invention provides an industrial process monitoring method based on cross-modal information fusion, which includes the following steps:

[0011] Collect multi-dimensional sensor signal data during the normal operation of multiple known modalities in the industrial process to construct a training sample set; construct a dictionary learning model; split the dictionary coefficients in the dictionary learning model into between-class coefficients and within-class coefficients; solve the dictionary learning model based on the training sample set to obtain class-specific dictionaries and within-class feature dictionaries;

[0012] Collect multi-dimensional sensor signal data for online monitoring to construct real-time samples of unknown modalities; construct a modal recognition and process monitoring model based on the dictionary learning model, solve the between-class coefficients and within-class coefficients of the real-time samples under each modality according to the class-specific dictionaries and within-class feature dictionaries, and obtain the weighted error of the real-time samples; judge the modal attribution of the real-time samples and whether there are faults in the industrial process based on the weighted error.

[0013] Preferably, in the dictionary learning model, the l 2,1 norm is used to constrain the between-class coefficients.

[0014] Preferably, in the modal recognition and process monitoring model, the l2,1 The norm constraint is equivalently replaced with a weighted quadratic form where s t is the between-class coefficient under the unknown mode t; Γ t = diag(γ1, γ2,..., γ k ); S kj represents the element in the k-th row and j-th column of the between-class coefficient S t ; N t is the number of columns of the between-class coefficient S t ; diag(·) represents a diagonal matrix.

[0015] Preferably, the method for obtaining the between-class coefficient of real-time samples is as follows: Obtain the gradient of the between-class coefficient according to the mode recognition and process monitoring model, and set the gradient equal to zero to obtain the closed-form solution of the between-class coefficient.

[0016] Preferably, the method for judging whether there is a fault in the industrial process based on the weighted error is as follows:

[0017] Obtain the control limit according to the training sample set, compare the weighted error of the real-time sample in the attributed mode with the control limit. If the weighted error in the attributed mode is greater than the control limit, it indicates that there is a fault.

[0018] Preferably, the method for obtaining the control limit is as follows: Obtain the weighted error of the samples in the training sample set according to the mode recognition and process monitoring model, and use the empirical quantile method to obtain the control limit.

[0019] Preferably, the method for judging the mode attribution of the real-time sample is as follows: Take the mode with the smallest weighted error as the mode attribution of the real-time sample.

[0020] Preferably, the class-specific dictionary and the within-class feature dictionary are obtained by solving using the online dictionary learning algorithm; the between-class coefficient set and the within-class coefficient set are obtained by solving using the fast iterative shrinkage threshold algorithm.

[0021] Preferably, the between-class coefficient of the real-time sample is obtained by solving using the l1-norm regularization method.

[0022] In a second aspect, the present invention provides an industrial process monitoring system based on cross-modal information fusion for performing the above-mentioned industrial process monitoring method; the industrial process monitoring system includes a data acquisition module, a data processing module, a mode determination module, and a fault judgment module; the data acquisition module is used to acquire industrial process data; the data processing module is used to process the acquired industrial process data, the mode determination module is used to determine the mode attribution of the acquired data according to the data processed by the data processing module, and input the data processed by the data processing module in combination to the fault judgment module for fault judgment.

[0023] The beneficial effects of the present invention are as follows:

[0024] 1. Aiming at the limitations of existing dictionary learning methods in multi-modal industrial process monitoring, the present invention splits the dictionary coefficients into two parts: inter-class coefficients (modal recognition) and intra-class coefficients (process variation). By characterizing local features with intra-class coefficients and capturing global features with inter-class coefficients to divide different modalities and describe the changes within modalities, it can effectively strengthen the boundaries between different modalities, make the modal division clearer, and at the same time accurately capture the dynamic changes in the industrial process, improve the sensitivity of fault monitoring, thus significantly enhancing the expression ability of samples and providing a more adaptable solution for the intelligent monitoring of complex industrial systems.

[0025] 2. By imposing a norm constraint on the inter-class part of the present invention to enhance the sparsity in the row direction, promoting samples under different modalities to share similar features and suppressing irrelevant information, it better conforms to the characteristic that there is partial configuration information sharing between different modalities in the industrial process and improves the reconstruction quality of samples. At the same time, the present invention uses a weighted quadratic form to equivalently replace the original sparsity constraint, enabling the model to adaptively adjust the weight of each element according to the characteristics of different data samples to achieve a more flexible and accurate sparse representation, which not only improves the expression ability of dictionary learning but also enhances the sensitivity to early faults.

[0026] 3. By introducing an improved Fisher discriminant criterion, the present invention effectively constrains the distribution of samples within a class, making the distribution of samples within the same modality more compact, and at the same time increasing the discrimination between different modalities, thus significantly improving the stability of modal classification and the sensitivity to faults, enabling the monitoring system to more accurately identify abnormal states. Description of the Drawings

[0027] Figure 1 is the overall flowchart in the present invention.

[0028] Figure 2 is the process flowchart of the air flow inside the blast furnace.

[0029] Figure 3 is a schematic diagram comparing the modal recognition results of the present invention and the Fisher discriminant sparse dictionary learning method; among them, (a) is the schematic diagram of the modal recognition result of the present invention; (b) is the schematic diagram of the modal recognition result of the Fisher discriminant sparse dictionary learning method.

[0030] Figure 4 is a schematic diagram comparing the fault monitoring results of the present invention and the Fisher discriminant sparse dictionary learning method; among them, (a) is the schematic diagram of the fault monitoring result of the present invention; (b) is the schematic diagram of the fault monitoring result of the Fisher discriminant sparse dictionary learning method. Detailed Implementation Modes

[0031] The present invention will be further described below with reference to the accompanying drawings.

[0032] As Figure 1 shown, an industrial process monitoring method based on cross-modal information fusion uses an industrial process monitoring system including a data acquisition module, a data processing module, a modality determination module, and a fault judgment module; the data acquisition module is used to acquire industrial process data; the data processing module is used to process the acquired industrial process data, and the modality determination module is used to determine the modality attribution of the acquired data according to the data processed by the data processing module, and input the data processed by the data processing module into the fault judgment module for fault judgment.

[0033] The industrial process monitoring method includes the following steps:

[0034] Step S1: Collect multi-dimensional sensor signal data during the normal operation of multiple known modalities in the industrial process, and construct a training sample set where n is the dimension of the sensor signal data; N i is the number of samples of modality i; C is the number of modalities. The dictionary learning model is constructed as follows:

[0035]

[0036] where D ∈ R n×k and are the class-specific dictionary and the intra-class feature dictionary respectively; k and k0 are the number of atoms of the class-specific dictionary and the intra-class feature dictionary respectively; S is the set of inter-class coefficients, S = {S1, S2,..., S C}; S0 is the set of intra-class coefficients, S0 = {S 0,1 , S 0,2 ,..., S 0,C}; is the sample subset belonging to modality i in the training sample set; is the inter-class coefficient of the class-specific dictionary, used to characterize modality information; is the intra-class coefficient of the intra-class feature dictionary, used to describe the intra-modal variation; M i and M 0,i are the mean matrices of the inter-class coefficient and the intra-class coefficient respectively; d k represents the k-th atom of the dictionary; λ1, λ2, and λ3 are model parameters, and λ1 is dynamically adjusted according to the data scale to ensure that on datasets of different scales, the weight of the l 2,1 norm constraint can maintain a reasonable balance; ||·||1, and ||·|| 2,1 represent the l1 norm, the square of the Frobenius norm, and e 2,1Norm.

[0037] The online dictionary learning (ODL) algorithm is used to solve the class-specific dictionary D and the intra-class feature dictionary D0 in the dictionary learning model. It adopts a column-by-column update strategy, effectively reducing the computational complexity. Meanwhile, it avoids matrix inversion, improves numerical stability, and makes dictionary learning more efficient and robust. The class-specific dictionary D and the intra-class feature dictionary D0 are used to represent modal discrimination information and intra-modal variation features respectively. In the dictionary learning model, represents the reconstruction error of the sample, As the inter-class reconstruction error, it aims to prevent the intra-class reconstruction part D0S 0,i from over-approximating the training sample set X i and causing the inter-class coefficient S i to tend to 0. ||S i || 2,1 is to impose an l 2,1 norm constraint on the inter-class coefficient, which will promote its row sparsity, that is, many elements in the same row tend to zero, enabling samples under the same modality to share similar features and suppressing irrelevant information. ||S 0,i ||1 means imposing an l1 norm constraint on the intra-class coefficient to make it as sparse as possible, that is, most elements tend to zero, and only a small number of key information is retained to more accurately describe the intra-modal changes. The improved Fisher discriminant criterion further optimizes the feature extraction and classification capabilities. The first two terms are used to reduce the intra-class scatter and increase the inter-class scatter of the encoded samples respectively to enhance the discriminant ability of the model. Since the intra-class coefficients contribute differently to each category, an additional third term is introduced to reduce their common influence. The introduction of the last term helps to enhance the convexity of the model. Constraint conditions means that the product of the k-th atom of the class-specific dictionary and the intra-class feature dictionary and its transpose is less than or equal to 1, aiming to normalize the dictionary atoms to prevent the infinite amplification of the dictionary vectors.

[0038] Since the dictionary learning model is composed of a smooth convex function and a non-smooth sparse regularization term, the fast iterative shrinkage threshold algorithm (FISTA) is used to solve the inter-class coefficient set S and the intra-class coefficient set S0. The fast iterative shrinkage threshold algorithm adopts a momentum acceleration technique, and the convergence speed reaches the time complexity O(1 / k 2 ), so it only needs to calculate the gradients of the inter-class coefficient set S and the intra-class coefficient set S0 respectively and substitute them into this algorithm to solve efficiently. The gradients of the inter-class coefficient set S and the intra-class coefficient set S0 and are calculated as follows:

[0039]

[0040] Among them, M and M0 are the mean matrices of the between-class coefficient set S and the within-class coefficient set S0, respectively.

[0041] Step S2: Collect multi-dimensional sensor signal data from on-line monitoring to construct a real-time sample x of the unknown mode t t ∈R n ×1 ; On the premise of the known class-specific dictionary D and the within-class feature dictionary D0, based on the dictionary learning model of the present invention, construct a mode recognition and process monitoring model to solve the between-class coefficients and within-class coefficients of the real-time sample under each mode, so as to obtain its mode attribution and process change characteristics. The constructed mode recognition and process monitoring model is as follows:

[0042]

[0043] Among them, s t ∈R k×1 and are the between-class coefficient and the within-class coefficient under the mode t, respectively; m t ∈R k×1 is the mean vector of the between-class coefficient s t ; is the mean vector of the within-class coefficient s 0,t ; Γ t = diag(γ1, γ2,..., γ k ); S kj represents the element in the k-th row and j-th column of the between-class coefficient S t ; N t is the number of columns of the between-class coefficient S t ; diag(·) represents a diagonal matrix.

[0044] In order to improve the sensitivity of the mode recognition and process monitoring model to early faults, compared with the dictionary learning model, the l 2,1 norm constraint imposed on the between-class coefficient is equivalently replaced by a weighted quadratic form: The advantage of this equivalent replacement is that it is convenient for solution and can adaptively adjust the weight of each element to achieve a sparse effect. If a certain element of the between-class coefficient S t is large, the corresponding weight is small; on the contrary, if the element is small, the corresponding weight is large.

[0045] For the solution of the between-class coefficient s t in the mode recognition and process monitoring model, since there are only quadratic form and l2 norm constraints, the closed-form solution can be obtained by calculating its gradient and setting the gradient equal to zero, thus avoiding the problem of unstable optimization. The between-class coefficient st The calculation of the closed-form solution is as follows:

[0046] s t =(2D T D + λ1Γ t + λ3I) -1 [D T (2x t - D0s 0,t ) + λ3m t )

[0047] where I ∈ R k×k is the identity matrix.

[0048] For the solution of s 0,t , due to the l1-norm constraint, the LASSO regularization (l1 regularization) method can be used to ensure the sparsity of the solution.

[0049] Step S3: For the real-time sample x t ∈ R n×1 , the weighted error E t of the real-time sample x t is jointly measured by the reconstruction error and the within-class deviation error. The calculation formula is as follows:

[0050]

[0051] where w ∈ [0, 1] is a preset weight parameter, and the preferred value is 0.9, which is used to balance the contributions of the two error terms.

[0052] Finally, the mode with the minimum weighted error E t is used as the mode attribution of the unknown mode t, that is:

[0053]

[0054] where identity(x t ) is the mode attribution of the real-time sample x t ; argmin represents the variable value when the objective function takes the minimum value.

[0055] This method can more accurately characterize the mode attribution of the sample and improve the reliability of anomaly detection and classification.

[0056] Step S4: Based on Steps S2 and S3, obtain the weighted error of each training sample in the training sample set, and use the empirical quantile method to determine the control limit E0. Compare the weighted error of the real-time sample with the control limit E0. If , it indicates that there is a fault.

[0057] Step S5: Collect the real industrial data set from the blast furnace. The present invention is applicable to a multi-dimensional sensor network deployed in an industrial field. These sensors can monitor key variables in the blast furnace ironmaking process in real time and generate multi-dimensional signal data. A blast furnace is a shaft furnace for producing hot metal. During its operation, raw materials such as iron ore and coke slowly fall from the top of the furnace under the action of gravity and come into contact with hot air and pulverized coal blown in from the bottom of the tuyere. A series of complex chemical changes occur during the combustion reaction in the furnace, and finally hot metal and slag are produced and intermittently discharged from the bottom of the furnace, while the waste gas is discharged from the top of the furnace. The process flow of the gas flow inside the blast furnace is as Figure 2 shown. The data set is sampled at intervals of 2 minutes, covering 10 key process variables related to gas flow, and finally forms a training data set X∈R 10×1500 .

[0058] Using the dictionary learning model constructed according to Step S1 with this training data set, solve the class-specific dictionary and the within-class feature dictionary, which are used to represent modal discrimination information and within-modal variation characteristics respectively; the model parameters λ1, λ2, and λ3 in the dictionary learning model are set to 0.1 and 1.5 respectively; the class-specific dictionary D∈R 10×9 ; the within-class feature dictionary D0∈R 10×20 . Collect the multi-dimensional sensor signal data of online monitoring as test samples to construct a test data set; the single test sample collected is x t ∈R 10×1 . Based on the modal recognition and process monitoring model constructed in Step S2, on the premise of knowing the class-specific dictionary and the within-class feature dictionary, solve the between-class coefficient and the within-class coefficient of the test sample under each mode to characterize its modal attribution and process variation characteristics.

[0059] As Figure 3 shown, for the traditional Fisher discriminant sparse dictionary learning method, its basic idea is to force the sparse coefficients to have small within-class dispersion but large between-class dispersion. Its correct recognition rate for modes is only 84.24%, which is significantly lower than 99.46% proposed by the present invention (structured joint sparse dictionary learning method). Especially in the challenging sample region (i.e., after the 800th sample), the traditional method misidentifies all samples, while the method of the present invention still maintains a high accuracy rate with only a few individual errors. This significant performance difference is mainly attributed to the l 2,1 -norm constraint imposed in the optimization process of the present invention. This constraint can further sparsify the between-class coefficients, prompting samples of different modes to share the same dictionary atoms, thereby better reconstructing the samples. At the same time, aiming at the problem that the within-class coefficients in the traditional Fisher discriminant criterion have different contributions to each category, is additionally introduced to reduce its commonality influence, thereby improving the modal discrimination ability and the generalization performance of the model.

[0060] Based on the training data set, steps S2 and S3 are repeatedly executed to calculate the weighted error of each training sample, and the control limit is determined by using the empirical quantile method. Then, according to whether the weighted error of the test sample exceeds the control limit, normal and abnormal samples are distinguished, and the fault detection rate and false alarm rate are calculated. The calculation methods of the fault detection rate and false alarm rate are as follows:

[0061]

[0062] Among them, FDR is the fault detection rate; F P is the number of samples correctly detected as faults; P is the total number of actual fault samples; FPR is the false alarm rate; F P is the number of normal samples misjudged as faults; N is the total number of actual normal samples.

[0063] As Figure 4 shown, starting from the 492nd sample, due to the pulverized coal injection amount exceeding the normal range, abnormal fluctuations occur in the CO concentration in the flue gas. The method proposed by the present invention can accurately and timely identify this fault, ensuring that abnormal situations can be quickly detected and processed. Through the determination of the weighted error, the fault detection rate of the method proposed by the present invention is as high as 96.74%, while the false alarm rate is as low as 1.02%. In contrast, the fault detection rate of the traditional Fisher discriminant sparse dictionary learning method is only 81.59%, and the false alarm rate is as high as 1.22%, and its performance in the early detection of faults is poor, which may lead to the neglect or delayed identification of fault symptoms, thus missing the best intervention opportunity and affecting the safety and stability of the system.

[0064] The traditional method has a weak identification ability in the initial stage of fault occurrence. The main reason is that by only focusing on the reconstruction error to improve the detection sensitivity, it may have a negative impact on the ability of the dictionary coefficients to distinguish these initial faults. The method proposed by the present invention not only improves the data reconstruction quality by imposing the e 2,1 norm constraint, but more importantly, introduces a weighted quadratic penalty constraint, making the part with smaller dictionary coefficients receive less penalty, so that weak fault features can be more sensitively captured in the initial stage of the fault, and will not be ignored due to the weak fault signal.

Claims

1. An industrial process monitoring method based on cross-modal information fusion, characterized in that: Including the following steps: Collect multi-dimensional sensor signal data during normal operation in multiple known modes of the industrial process to construct a training sample set; Construct a dictionary learning model; the dictionary coefficients in the dictionary learning model are split into between-class coefficients and within-class coefficients; Solve the dictionary learning model based on the training sample set to obtain class-specific dictionaries and within-class feature dictionaries; Collect multi-dimensional sensor signal data for on-line monitoring to construct real-time samples of unknown modes; Construct a mode recognition and process monitoring model based on the dictionary learning model, solve the between-class coefficients and within-class coefficients of the real-time samples in each mode according to the class-specific dictionaries and within-class feature dictionaries, and obtain the weighted error of the real-time samples; judge the mode attribution of the real-time samples and whether there are faults in the industrial process based on the weighted error.

2. The industrial process monitoring method based on cross-modal information fusion according to claim 1, wherein: In the dictionary learning model, the l 2,1 norm is used to constrain the between-class coefficients.

3. The industrial process monitoring method based on cross-modal information fusion according to claim 2, wherein: In the modal recognition and process monitoring model described above, the l 2,1 -norm constraint imposed on the between-class coefficient is equivalently replaced with a weighted quadratic form where s t is the between-class coefficient under the unknown mode t; Γ t = diag(γ1, γ2,..., γ k ); S kj represents the element in the k-th row and j-th column of the between-class coefficient S t ; N t is the number of columns of the between-class coefficient S t ; diag(·) represents a diagonal matrix.

4. The industrial process monitoring method based on cross-modal information fusion according to claim 3, wherein: The method for obtaining the between-class coefficients of the real-time samples is as follows: obtain the gradient of the between-class coefficients according to the mode recognition and process monitoring model, and set the gradient equal to zero to obtain the closed-form solution of the between-class coefficients.

5. The industrial process monitoring method based on cross-modal information fusion according to claim 1, characterized in that: The method for judging whether there are faults in the industrial process based on the weighted error is as follows: Obtain the control limit according to the training sample set, compare the weighted error of the real-time samples in the attributed mode with the control limit. If the weighted error in the attributed mode is greater than the control limit, it indicates that there are faults.

6. The industrial process monitoring method based on cross-modal information fusion according to claim 5, wherein: The method for obtaining the control limit is as follows: obtain the weighted error of the samples in the training sample set according to the mode recognition and process monitoring model, and use the empirical quantile method to obtain the control limit.

7. A method for industrial process monitoring based on cross-modal information fusion according to claim 1, characterized in that: The method for judging the mode attribution of the real-time samples is: take the mode with the smallest weighted error as the mode attribution of the real-time samples.

8. The industrial process monitoring method based on cross-modal information fusion according to claim 1, characterized in that: The class-specific dictionaries and within-class feature dictionaries are obtained by solving using the online dictionary learning algorithm; the between-class coefficient set and within-class coefficient set are obtained by solving using the fast iterative shrinkage threshold algorithm.

9. The industrial process monitoring method based on cross-modal information fusion according to claim 1, characterized in that: The between-class coefficients of the real-time samples are obtained by solving using the l1-norm regularization method.

10. An industrial process monitoring system based on cross-modal information fusion, including a data acquisition module, a data processing module, and a fault judgment module; the data acquisition module is used to acquire industrial process data; the data processing module is used to process the acquired industrial process data and input it to the fault judgment module for fault judgment; It is characterized in that: It is used to execute an industrial process monitoring method according to claim 1; this industrial process monitoring system further includes a mode determination module; the mode determination module is used to determine the mode attribution of the acquired data according to the data processed by the data processing module and cooperate with the fault judgment module for fault judgment.

Images