A Multi-Condition Intelligent Monitoring Method and System for Industrial Processes Based on Element-Aware Dictionary Continuous Learning
By adopting a continuous learning method based on element-aware dictionaries, the problems of high model complexity and 'catastrophic forgetting' caused by multimodal changes in industrial production are solved, achieving efficient fault detection and model simplification, which is suitable for the stable and safe operation of large-scale industrial systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2023-10-08
- Publication Date
- 2026-06-30
AI Technical Summary
Existing technologies struggle to effectively address new modalities in industrial production, resulting in high model complexity, high training costs, and a tendency to produce 'catastrophic forgetting' problems.
We adopt an element-aware dictionary-based continuous learning approach. By constructing a global importance metric and an alternative loss function, we can update and constrain dictionary elements, establish a continuous learning model for new modalities, reduce model complexity, and maintain memory of historical modalities.
This method achieves efficient fault detection under multimodal conditions, improves the fault detection rate, reduces model complexity and training cost, and expands the applicability of the method.
Smart Images
Figure CN117349799B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial process monitoring, specifically relating to a multi-condition intelligent monitoring method and system for industrial processes based on continuous learning of an element-aware dictionary. Background Technology
[0002] Process monitoring is a crucial part of modern industrial production, helping companies improve production efficiency, reduce production costs, and ensure production line safety. With the continuous expansion of industrial scale and changes in production structures, industrial systems have become more complex. In modern complex industrial production processes, system failures can trigger a series of chain reactions, leading to decreased product quality, damage to production equipment, and even safety accidents causing operator injuries or fatalities. Due to the continuous development of modern industry and the rise of artificial intelligence, data-driven and AI-based process monitoring methods are widely used in industrial settings. However, changes in production environments and operating conditions inevitably lead to new modalities in industrial production processes, making traditional multimodal monitoring methods with a fixed number of modalities inadequate.
[0003] There are currently three common methods for addressing multimodal problems in industrial production:
[0004] 1) Multi-model approach: The multi-model approach treats each modality separately, using multiple different models to match each modality. However, as the number of modalities increases, the number of models also increases, placing a significant burden on management and training.
[0005] 2) Global modeling method: The global modeling method considers all modalities simultaneously, builds a model on data from all modalities, and uses this model for monitoring. When a new modality appears, the model needs to be retrained on all data, which is time-consuming.
[0006] 3) Online model update method; This method uses only one model to represent data from all modalities. This reduces the number of models and lowers training complexity. However, when modalities vary significantly, this method is prone to the "catastrophic forgetting" problem. Summary of the Invention
[0007] To reduce model complexity, decrease training costs, and overcome the "catastrophic forgetting" problem, this invention provides a multi-condition intelligent monitoring method and system for industrial processes based on continuous learning of element-aware dictionaries. It innovatively proposes a global importance measurement method for dictionary elements and then constructs a new alternative loss to constrain the updating of dictionary elements, thereby achieving continuous dictionary learning for new modalities. Finally, it establishes a model to complete online monitoring of multi-modal data of industrial processes.
[0008] To achieve the above technical objectives, the present invention adopts the following technical solution:
[0009] A multi-condition intelligent monitoring method for industrial processes based on continuous learning of element-aware dictionaries includes:
[0010] Collect monitoring datasets for the first N modes of an industrial process, and learn the optimal dictionary D for each mode sequentially. 1 D 2 ,...,D N The corresponding importance matrices are Ω. 1 ,Ω 2 ,...,Ω N and the comprehensive importance matrix Ω N,* Among them, the importance matrix Ω 1 ,Ω 2 ,...,Ω N The elements in the matrix refer to the importance of each element in the dictionary in its corresponding modality; the overall importance matrix Ω N,* The elements in the equation represent the importance of the first N modes.
[0011] The reconstruction error is obtained by reconstructing the monitoring dataset for the (N+1)th modality using a dictionary. An objective function is constructed based on the reconstruction error of the monitoring data for the (N+1)th modality and the loss of historical modalities. Solving the objective function yields the dictionary D for the (N+1)th modality. N+1 The loss for the historical modes is based on the dictionary to be determined in the (N+1)th mode and the dictionary D in the Nth mode. N and the importance matrix Ω N,* Calculated;
[0012] D using the dictionary for the (N+1)th modality N+1 Calculate the reconstruction error of each monitoring data point for each mode, fit a probability density function based on the reconstruction error of all monitoring data, and then calculate the control limit of the reconstruction error based on the preset confidence level.
[0013] When monitoring data of any mode of an industrial process is received online, dictionary D is used. N+1 Calculate the reconstruction error of the monitoring data, and compare the obtained reconstruction error with the control limit to determine whether the industrial process is currently malfunctioning.
[0014] Furthermore, an objective function is constructed based on the reconstruction error of the monitoring dataset at the (N+1)th modality and the loss of historical modalities, expressed as:
[0015]
[0016] In the formula, Y N+1Y is the monitoring dataset for the (N+1)th modality, D is the dictionary to be determined for the (N+1)th modality, and X is the monitoring dataset Y. N+1 The sparse coding matrix obtained by sparse representation using dictionary D; λ1 and λ2 are manually set hyperparameters; d a,b It is the element in row a and column b of dictionary D, and it needs to be optimized. It is dictionary D N The element in row a and column b is a definite value; The importance of the first N modalities is summarized for the element in row a, column b of dictionary D.
[0017] Furthermore, the element in row a, column b of dictionary D represents the importance of the first N modalities. The calculation formula is:
[0018]
[0019] In the formula, k is the modal index; For dictionary D k The element in row a and column b, For the initial dictionary D 0 The element in row a, column b; For elements The importance weight; ξ is the introduced damping coefficient.
[0020] Furthermore, using an alternating and iterative optimization approach, the variables D and X in the objective function are alternately optimized in each iteration:
[0021] When optimizing X, fix D, retain the terms related to X in the formula, solve directly using the OMP algorithm, and obtain the sparse coding matrix X;
[0022] When optimizing D, X is fixed, and the terms related to D are retained in the formula, resulting in the following optimization problem:
[0023]
[0024] Optimizing the above formula element-wise, i.e., taking the derivative of each element and setting it to zero, yields the following formula for updating the elements:
[0025]
[0026] In the formula, Y a Y represents N+1 The data in row a, X a This represents the data in row a of X, X b This represents the data in row b of X.
[0027] Furthermore, using dictionary D N+1 The reconstruction error of the monitoring data is calculated using the following formula:
[0028]
[0029] Among them, R i Represents the i-th monitoring data y i The reconstruction error, x i It is monitoring data y i The sparse coding of can be obtained using the OMP algorithm;
[0030] After calculating the reconstruction error of all monitoring data, a probability density function is fitted based on the reconstruction error of all data:
[0031]
[0032] Where K(·) is the kernel density function, h represents the bandwidth, and M represents the number of monitoring data samples used for training;
[0033] After obtaining the probability density function of the reconstruction error, the control limit R of the reconstruction error is calculated based on the preset confidence level α. tr .
[0034] Furthermore, online monitoring data y new Reconstruction error R new With control limit R tr Comparison, if R new Greater than R tr Determine the online monitoring data y new For fault data, it indicates that a fault has occurred in the current industrial process; conversely, for normal data, it indicates that an online fault has occurred. new This is normal data, meaning the current industrial process is not malfunctioning.
[0035] A system based on the element-aware dictionary-based continuous learning intelligent monitoring method for multi-condition industrial processes as described above, comprising:
[0036] The dictionary learning module is used to: collect monitoring datasets for the first N modes of an industrial process, and sequentially learn the optimal dictionary D for each mode. 1 D 2 ,...,D N The corresponding importance matrices are Ω. 1 ,Ω 2 ,...,Ω N and the comprehensive importance matrix Ω N,* Among them, the importance matrix Ω 1 ,Ω 2 ,...,Ω N The elements in the matrix refer to the importance of each element in the dictionary in its corresponding modality; the overall importance matrix Ω N,*The elements in the dictionary represent the importance of the first N modes. The dictionary learning module is further used to: reconstruct the monitoring dataset for the (N+1)th mode using the dictionary to obtain the reconstruction error; construct an objective function based on the reconstruction error of the monitoring data for the (N+1)th mode and the loss of historical modes; and solve the objective function to obtain the dictionary D for the (N+1)th mode. N+1 The loss for the historical modes is based on the dictionary to be determined in the (N+1)th mode and the dictionary D in the Nth mode. N and the importance matrix Ω N,* Calculated;
[0037] The control limit calculation module is used to calculate the dictionary D when using the (N+1)th mode. N+1 Calculate the reconstruction error of each monitoring data point for each mode, fit a probability density function based on the reconstruction error of all monitoring data, and then calculate the control limit of the reconstruction error based on the preset confidence level.
[0038] The online data monitoring module is used to: when receiving monitoring data of any mode of an industrial process online, use dictionary D N+1 Calculate the reconstruction error of the monitoring data, and compare the obtained reconstruction error with the control limit to determine whether the industrial process is currently malfunctioning.
[0039] Beneficial effects
[0040] This invention presents a multi-condition intelligent monitoring method and system for industrial processes based on continuous learning using an element-aware dictionary. It can be applied to large-scale industrial systems with various operating conditions. This method continuously learns new modes while maintaining a memory of historical modes, overcoming the "catastrophic forgetting" problem inherent in traditional process monitoring methods under multi-modal conditions. Simultaneously, it achieves a high level of monitoring accuracy, ensuring the stable and safe operation of industrial processes.
[0041] Specifically, compared to existing methods, this invention improves the fault detection rate by more than 2 times while maintaining the same false alarm rate. Furthermore, this invention reduces model complexity and the amount of data required for modeling, significantly expanding the applicability of the method. Attached Figure Description
[0042] Figure 1 This is the framework for continuously learning new modalities in the embodiments of this application;
[0043] Figure 2 This is the framework for online monitoring in the embodiments of this application;
[0044] Figure 3 These are visualizations of the dimensionality reduction of the four modal data in the embodiments of this application;
[0045] Figure 4These are experimental results comparing the method of this application embodiment with two other advanced methods; (a) shows the reconstruction error and control limit of the DL method on the test set, (b) shows the reconstruction error and control limit of the JMSDL method on the test set, and (c) shows the reconstruction error and control limit of the EaLDL method of this invention on the test set. Detailed Implementation
[0046] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.
[0047] This invention provides a multi-condition intelligent monitoring method for industrial processes based on element-aware lifelong dictionary learning (EaLDL), mainly comprising two parts: continuous learning of new modalities and online monitoring. Continuous learning of new modalities, such as... Figure 1 As shown, firstly, the optimal dictionary and importance matrix corresponding to the first N modes are learned sequentially; based on this, an objective function is constructed based on the reconstruction error of the N+1th mode monitoring data and the historical mode loss, and the optimal dictionary for the N+1th mode is learned; then, the optimal dictionary for the latest mode is used to establish a monitoring model based on control limits. Online monitoring, such as... Figure 2 As shown, control limits are used to monitor new online data to determine whether the industrial process is currently malfunctioning.
[0048] The technical approach of this invention is explained below: First, dictionary learning is analyzed from a Bayesian perspective to identify the causes of catastrophic forgetting. Then, the true posterior is transformed using the concept of substitution loss, while a quadratic loss is introduced as a substitution loss, with its coefficient representing the importance of the elements. Next, the importance of elements is derived for both unimodal and multimodal cases, thus proposing an element-aware dictionary continuous learning method. Finally, a model is built based on the dictionary obtained from the continuous learning of the element-aware dictionary to monitor multimodal data online.
[0049] I. Dictionary Learning from a Bayesian Perspective
[0050] In dictionary learning, assuming the model parameters are D and the training data is Y, according to the conditional probability formula:
[0051]
[0052] When data Y is divided into two parts Y A Y BWhen the former represents data from mode A and the latter represents data from mode B, in order to facilitate the derivation of the probability distribution under continuous learning from a Bayesian perspective, this paper makes the following assumption: Y A Y B The parameters D are independent. Therefore, at this point we have:
[0053] p(Y A ,Y B |D)=p(Y A |D)p(Y B |D) (2)
[0054] From the perspective of maximum a posteriori, the trained model needs to maximize the posterior p(D|Y), that is, maximize p(D|Y). A ,Y B Through the derivation of Bayes' theorem, it can be transformed into maximizing the following formula.
[0055]
[0056] Taking the logarithm of the above expression, we get:
[0057] logp(D|Y A ,Y B ) = logp(Y B |D)+logp(D|Y A )+C (4)
[0058] Where C = -logp(Y) B |Y A The maximum a posteriori (MAP) estimate remains unchanged during training, so the maximum a posteriori estimate can be transformed into the following optimization objective:
[0059]
[0060] Referring to the Bayesian optimization model under two modalities, a Bayesian model under multimodalities can be derived. Assume there are N modalities to be learned, denoted as Y1, Y2, ..., Y... N Similar to the formula, the optimization objective under N modes can be obtained, as shown in the following equation:
[0061]
[0062] The optimization objectives derived from Equations (5) and (6) are analyzed. The former is the maximum likelihood of the new mode, which can be represented by the loss function when learning the new mode; the latter is the posterior estimate of the historical mode, and all information about the historical mode should be incorporated into this posterior.
[0063] In practice, the true posterior in the formula is very tricky and requires special handling. Here, we take equation (5) as an example and expand it using the conditional probability formula as follows:
[0064] -logp(D|Y A ) = -logp(Y A |D)-logp(D)+logp(Y A (7)
[0065] Ignoring the constant term, we can find that this posterior can be approximated as the likelihood of the parameter in mode A. Therefore, we can learn the parameter in the training data Y. A The loss function is used to represent the current task's loss. However, in reality, historical data and historical losses are not continuously retained, and even if they are, they are difficult to optimize simultaneously with the current task's loss, and will increase additional computational complexity. Therefore, in this paper, we use the idea of alternative loss to find a function that can approximate the original loss function -logp(Y) to the greatest extent. A The loss function of |D) is used to approximate the intractable posterior -logp(D|Y) to some extent. A ).
[0066] II. Derivation of Element Importance
[0067] 1. Derivation of the importance of single-modal elements
[0068] First, make the following assumptions:
[0069] a. The model parameters for dictionary learning are D, and the parameter update trajectory is D. 0 D 1 ,...,D m That is, it has undergone m iterations;
[0070] b. For element d in row a and column b of dictionary D a,b Its update trajectory is The element's importance is Ω a,b Subsequent analyses are all based on d a,b superior;
[0071] c. During the training process of historical modes, any element d a,b The contribution to the actual loss is f. a,b (d) Actual loss Where A and B are the total number of rows and columns of the dictionary, respectively;
[0072] d. The form of agency loss used is Ω a,b It can be uniquely determined;
[0073] e. Considering the substitution loss function requires precise matching of the descent dynamics with the original loss function in three aspects: the substitution loss and the original loss decrease by the same amount, the total net motion in the parameter space is equal, and both reach a minimum at the endpoint.
[0074] Integrating the gradient along the update path of any element, based on the previous assumptions, we can obtain the gradient for any element d. a,b Contribution f to the actual loss a,b (d), as shown in the following formula:
[0075]
[0076] For the alternative loss, at the end and start points of the training process, we have:
[0077]
[0078]
[0079] Since the total change in the substitution loss needs to be the same as the original loss, the following equation must be satisfied:
[0080]
[0081] Substituting the relevant parameters yields:
[0082]
[0083] Solving the above equation and introducing the damping coefficient ξ, we can obtain the formula for calculating the importance of elements in a single mode:
[0084]
[0085] 2. Derivation of the Importance of Multimodal Elements
[0086] First, make the following assumptions:
[0087] a. The total number of modes is N+1. For any element d in the a-th row and b-th column of the matrix... a,b Its update trajectory is represented as
[0088] b. When training the k-th modality, element d a,b The contribution to the actual loss is The importance weight of the elements is Based on this, its expression can be obtained as follows:
[0089]
[0090] d. The form of agency loss used is in This represents the element importance weights for the combined N modalities, which are the element importance weights used when learning the (N+1)th modality.
[0091] For element d a,b The changes in actual loss and vicarious loss between the training endpoint and the starting point are as follows:
[0092]
[0093]
[0094] in, It can be calculated by the formula, that is, the numerator of formula (13) represents the path integral of the gradient.
[0095] By combining the equations, we can obtain the following formula for calculating the element importance of the first N modes:
[0096]
[0097] III. Element-Aware Dictionary Continuous Learning
[0098] The relevant symbols are defined as follows: using D 0 Let D represent the initial dictionary. After learning N modes sequentially, the optimal dictionaries obtained are respectively D. 1 D 2 ,...,D N Furthermore, the importance matrix Ω of any element in each mode can be obtained using the formula. 1 ,Ω 2 ,...,Ω N The comprehensive importance matrix Ω of the first N modes is obtained according to the formula. N,* This importance is the element importance used when learning the (N+1)th modality. Based on this, the objective function for continuously learning new modalities can be obtained as follows:
[0099]
[0100] Where λ1 and λ2 are manually defined hyperparameters. a,b This is the element in row a and column b of dictionary D, and it needs to be optimized. It is dictionary D N The element in row a and column b is a definite value. The importance of the first N modalities is summarized for the element in row a, column b of the dictionary.
[0101] In the formula, The terms are derived from traditional dictionary learning. By minimizing this reconstruction error, the dictionary's representation ability on the new modality N+1 is improved, thereby learning the new modality. The term, based on the formula, introduces an alternative loss that minimizes the model's ability to represent historical modalities, thereby avoiding "catastrophic forgetting."
[0102] The optimization formula is obtained using an alternating iterative approach. When optimizing X, D is fixed, and the terms related to X are retained. The OMP algorithm is used to directly solve the problem and obtain the sparse coding matrix X. When optimizing D, X is fixed, and the terms related to D are retained, leading to the following optimization problem:
[0103]
[0104] Optimizing the above formula element-wise, i.e., taking the derivative of each element and setting it to zero, we can derive the following formula for updating the elements:
[0105]
[0106] In the formula, Y a Y represents N+1 The data in row a, X a This represents the data in row a of X, X b This represents the data in row b of X.
[0107] IV. Online Monitoring Process Based on EaLDL
[0108] 1. Calculate control error and control limits
[0109] Assuming that incremental learning is performed on all modalities, the resulting dictionary is D. N+1 Then for any sample y in the training set i The reconstruction error is calculated as follows:
[0110]
[0111] Among them, R i Let x represent the reconstruction error of the i-th sample. i It is sample y i The sparse coding can be obtained using the OMP algorithm. After calculating the reconstruction error of all samples, the KDE algorithm can be used to obtain the control limits, as shown in the following expression:
[0112]
[0113] Where K(·) is the kernel density function, h represents the bandwidth, which can be automatically obtained using Silverman's rule of thumb, and R i Let f(R) represent the reconstruction error of the i-th data point. After estimating the reconstruction errors of all samples using the KDE algorithm, a suitable normal distribution model f(R) can be obtained. Based on this distribution model, by selecting an appropriate confidence level α, the control limit R can be obtained.tr The expression:
[0114]
[0115] Since the final dictionary has a good representation ability for all modalities, the control limits obtained from the above steps can be used to determine whether a sample is normal or not with a certain degree of confidence.
[0116] 2. Online monitoring process
[0117] When new data y new Upon arrival, the OMP algorithm is first used to obtain the sparse code x of the new data. new The reconstruction error of the new data is calculated based on this sparse coding as follows:
[0118]
[0119] At this point, you can use R new and R tr The size of R is used to determine the state of the data: if R new >R tr If the condition is met, the new data is considered faulty; otherwise, the new data is considered normal.
[0120] In summary, to address the multimodal problem in industrial processes, this paper first uses the EaLDL method for modeling to obtain a dictionary that can represent all modes. Then, the KDE algorithm is used to estimate the control limits based on the reconstruction error of all training data. Finally, the control limits are used to monitor the online data.
[0121] To verify the effectiveness of the method of this invention, a 20-dimensional numerical simulation experiment was constructed. The experiment included four different modes, and the data was visualized after dimensionality reduction as follows. Figure 3 As shown.
[0122] This invention, based on continuous learning of an element-aware dictionary for intelligent monitoring of multi-condition industrial processes, is compared with two other advanced methods while maintaining the same initialization conditions. Experimental results are as follows: Figure 4 As shown, (a) represents the reconstruction error and control limits of the DL method on the test set, (b) represents the reconstruction error and control limits of the JMSDL method on the test set, and (c) represents the reconstruction error and control limits of the EaLDL method based on this invention on the test set. Test results show that traditional DL methods are affected by "catastrophic forgetting"; the JMSDL method can memorize multiple modalities but cannot guarantee that the reconstruction error is at the same level; while the EaLDL-based method in this paper can effectively solve the multimodal problem.
[0123] The above experiments verified the superiority of the method of the present invention over traditional methods in monitoring multimodal industrial production processes.
[0124] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.
Claims
1. A multi-condition intelligent monitoring method for industrial processes based on continuous learning of an element-aware dictionary, characterized in that, include: Collect monitoring datasets for the first N modes of the industrial process, and learn the optimal dictionaries for each mode sequentially. The corresponding importance matrices are as follows: and the comprehensive importance matrix Among them, the importance matrix The elements in the importance matrix refer to the importance of each element in the dictionary in the corresponding modality; The elements in the equation represent the importance of the first N modes. The reconstruction error is obtained by reconstructing the monitoring dataset for the (N+1)th modality using a dictionary. An objective function is constructed based on the reconstruction error of the monitoring data for the (N+1)th modality and the loss of historical modalities. Solving the objective function yields the dictionary for the (N+1)th modality. The loss for the historical modes is based on the dictionary of the unknown at the (N+1)th mode and the dictionary at the Nth mode. and importance matrix Calculated; Dictionary when using the (N+1)th modality Calculate the reconstruction error of each monitoring data point for each mode, fit a probability density function based on the reconstruction error of all monitoring data, and then calculate the control limit of the reconstruction error based on the preset confidence level. When monitoring data of any mode of an industrial process is received online, a dictionary is used. Calculate the reconstruction error of the monitoring data, and compare the obtained reconstruction error with the control limit to determine whether the industrial process is currently malfunctioning. The objective function is constructed based on the reconstruction error of the monitoring dataset at the (N+1)th modality and the loss of historical modalities, and is expressed as follows: ； In the formula, Let X be the monitoring dataset for the (N+1)th modality, D be the dictionary to be determined for the (N+1)th modality, and X be the monitoring dataset. The sparse coding matrix obtained by sparse representation of dictionary D; and These are hyperparameters that are set manually; It is a dictionary The Middle line, number The elements in the column need to be optimized; It is a dictionary The Middle line, number The elements of a column are definite values; For dictionary The Middle Line 1 Before column element synthesis The importance of each mode is calculated using the following formula: ； In the formula, k is the modal index; For dictionary The Middle Line 1 Column elements, For the initial dictionary The Middle Line 1 Column elements; For elements Importance weights; The damping coefficient is introduced.
2. The intelligent monitoring method for multi-condition industrial processes based on continuous learning of an element-aware dictionary according to claim 1, characterized in that, Using an alternating and iterative optimization approach, the variables D and X in the objective function are alternately optimized in each iteration: optimization Time fixed In the reserved form The relevant terms are solved directly using the OMP algorithm to obtain the sparse coding matrix. ; optimization Time fixed In the reserved form The relevant terms lead to the following optimization problem: ； Optimizing the above formula element-wise, i.e., taking the derivative of each element and setting it to zero, yields the following formula for updating the elements: ； In the formula, express The Middle Row data, express The Middle Data of rows express The Middle The data in the row.
3. The intelligent monitoring method for multi-condition industrial processes based on continuous learning of an element-aware dictionary according to claim 1, characterized in that, Use a dictionary The reconstruction error of the monitoring data is calculated using the following formula: ； in, Indicates the first One monitoring data Reconstruction error, It is monitoring data The sparse coding of can be obtained using the OMP algorithm; After calculating the reconstruction error of all monitoring data, a probability density function is fitted based on the reconstruction error of all data: ； in, For kernel density function, This represents the bandwidth, and M represents the number of monitoring data samples used for training. After obtaining the probability density function of the reconstruction error, the control limit of the reconstruction error is calculated based on the preset confidence level α. .
4. The intelligent monitoring method for multi-condition industrial processes based on continuous learning of an element-aware dictionary according to claim 1, characterized in that, Online monitoring data y new Reconstruction error R new With control limit R tr Comparison, if R new Greater than R tr Determine the online monitoring data y new For fault data, it indicates that a fault has occurred in the current industrial process; conversely, for normal data, it indicates that an online fault has occurred. new This is normal data, meaning the current industrial process is not malfunctioning.
5. A system based on the multi-condition industrial process intelligent monitoring method based on continuous learning of an element-aware dictionary as described in any one of claims 1-4, characterized in that, include: The dictionary learning module is used to: collect monitoring datasets for the first N modes of an industrial process, and sequentially learn the optimal dictionary for each mode. The corresponding importance matrices are as follows: and the comprehensive importance matrix Among them, the importance matrix The elements in the importance matrix refer to the importance of each element in the dictionary in the corresponding modality; The elements in the dictionary represent the importance of the first N modes. The dictionary learning module is further used to: reconstruct the monitoring dataset for the (N+1)th mode using the dictionary to obtain the reconstruction error; construct an objective function based on the reconstruction error of the monitoring data for the (N+1)th mode and the loss of historical modes; and solve the objective function to obtain the dictionary for the (N+1)th mode. The loss for the historical modes is based on the dictionary of the unknown at the (N+1)th mode and the dictionary at the Nth mode. and importance matrix Calculated; The control limit calculation module is used to: use the dictionary when using the (N+1)th mode. Calculate the reconstruction error of each monitoring data point for each mode, fit a probability density function based on the reconstruction error of all monitoring data, and then calculate the control limit of the reconstruction error based on the preset confidence level. The online data monitoring module is used to: when receiving monitoring data of any mode of an industrial process online, use a dictionary... Calculate the reconstruction error of the monitoring data, and compare the obtained reconstruction error with the control limit to determine whether the industrial process is currently malfunctioning.