Cloud-side collaborative intelligent autonomous monitoring method for dynamic industrial process

Abstract

The invention discloses a cloud edge collaborative intelligent autonomous monitoring method for a dynamic industrial process. The method comprises the steps that a cloud end establishes a dictionary model and trains and updates the model; the edge device obtains the latest dictionary, classifier and control limit from the cloud end, and then carries out abnormity monitoring and working condition identification on the current industrial process; wherein the cloud process comprises the following steps: establishing a dictionary model and training to obtain a dictionary, a classifier and a coding matrix; screening the historical samples based on the multi-azimuth hardness values of the historical samples, and forming a balanced data set with the new working condition samples to update the dictionary and the classifier; and taking the reconstructed data set and the reconstructed label as complete working condition knowledge, compressing and simplifying the updated over-complete dictionary and classifier, and taking the compressed and simplified dictionary and classifier as a dictionary and a classifier which are finally deployed to an edge layer. According to the invention, reliable and efficient monitoring of dynamic industrial occasions can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial process monitoring, and particularly relates to a cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes. Background Art

[0002] With the rise of the fourth industrial revolution represented by information technologies such as big data and the Internet of Things, the scale of industry has been continuously expanding. The internal structure of modern industrial systems is complex, and the components are closely related. If faults cannot be eliminated in time, it may cause huge economic losses or even casualties. Process monitoring detects abnormal behaviors such as product quality decline and casualties by monitoring the operating state of the process. It provides important working condition information for operators and is an important link to ensure the safety of the production process, which is of great significance for the efficient and healthy operation of industrial production.

[0003] Due to the deep integration of information technology and industrial systems, sensing devices can be connected to the industrial Internet of Things, and a large amount of industrial process data can be collected and stored. Data-driven industrial process monitoring methods can deeply mine the collected process data and establish effective monitoring models based on this. With the development of artificial intelligence, machine learning methods can accurately extract potential complex working condition characteristics, so machine learning methods have been continuously proposed and applied to process monitoring.

[0004] As an efficient machine learning method, dictionary learning finds a set of optimal basis vectors by learning process data. Using it for sparse representation of data can reduce the processing difficulty in the face of massive industrial process data. Therefore, it has been widely studied in the field of process monitoring. For example, Huang et al. proposed a dictionary learning method that combines characteristics and commonalities to extract data characteristics and commonalities. Peng et al. proposed a locally preserved sparse modeling method for multi-condition non-Gaussian processes, which effectively captures the local structure of data and improves the modeling ability and accuracy for complex industrial processes.

[0005] However, traditional monitoring methods often rely on the closed-world assumption, which is a static learning paradigm. In actual industrial processes, due to various factors such as raw material component fluctuations and production plan adjustments, the operating conditions are complex and variable. Therefore, it is difficult for the operating conditions included in historical training data to cover all possible situations in actual production. At the same time, it is difficult to estimate the operating condition switching time. When a new unknown operating condition arrives, it will be misjudged as a fault, which is called the mismatch between the static monitoring model and the dynamic production line requirements. Model adaptive update for dynamic industries is the most effective way to solve the model mismatch. However, considering the harsh industrial environment and limited computing power of edge devices, most industrial sites do not have the ability to update models. The development of cloud-edge collaboration technology provides a new idea to solve this problem. It makes up for the contradiction between computing power resources and real-time requirements in industrial process monitoring to meet the needs of large-scale data processing and applications in industrial processes.

[0006] However, in the cloud-edge collaboration framework, the integration between the cyber space and physical entities will inevitably lead to various cyber-physical problems. The open communication infrastructure results in limited data transmission bandwidth in the harsh industrial environment. And due to the transparency and untrustworthiness of the network, the deliberate destruction caused by network threats also makes the monitoring problem more complex. Specifically, attackers can obtain key data of industrial processes such as production processes and outputs through data theft. Moreover, attackers can carefully formulate attack strategies to tamper with industrial process data based on their understanding of the operating mechanism of the target process, deliberately avoiding triggering alarms of the monitoring system, making the monitoring system lose its monitoring ability, and thus causing serious safety accidents. Therefore, in the cloud-edge collaboration framework, how to implement an effective defense mechanism against illegal intrusion of network communication is a key problem that needs to be solved urgently. Summary of the Invention

[0007] The present invention provides a cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes, which can achieve reliable and efficient monitoring of dynamic industrial scenarios.

[0008] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0009] A cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes includes: establishing a dictionary model in the cloud and training and updating the model; the edge device obtains the latest dictionary, classifier, and control limits from the cloud, and uses the obtained dictionary and control limits to perform anomaly monitoring on the current industrial process, and uses the obtained dictionary and classifier to identify the operating conditions of the current industrial process;

[0010] Among them, establishing a dictionary model in the cloud and training and updating the model includes:

[0011] Establishing a dictionary model and using the historical monitoring data set Y of each operating conditionh and the corresponding operating condition category label H h , train the dictionary model to obtain the dictionary D h , classifier W h and the data set Y h Based on the dictionary D h encoding matrix X h ;

[0012] According to the dictionary D h , classifier W h , encoding matrix X h Calculate the hardness values of each sample in the historical monitoring data set Y h , and filter the samples from the historical monitoring data set Y based on the hardness values, and then jointly form the current balanced data set Y with the new operating condition data samples h ; b ;

[0013] Use the current balanced data set Y b to update the dictionary model to obtain the overcomplete dictionary D b , classifier W b and the data set Y b Based on the overcomplete dictionary D b encoding matrix X b ;

[0014] Use the overcomplete dictionary D b , classifier W b and the encoding matrix X b to calculate the reconstructed data set Y t and the reconstructed label H t ;

[0015] Take the reconstructed data set Y t and the reconstructed label H t as the complete operating condition knowledge, construct a dictionary compression model, and solve to obtain the simplified dictionary D s and the simplified classifier W s , as the finally updated dictionary and classifier.

[0016] Furthermore, the optimization function of the dictionary model is expressed as:

[0017]

[0018] Among them, D, A, W, X are the dictionary, transformation matrix, classifier, and encoding matrix in the optimization process, respectively. D h , A h , W h , X h are the optimized dictionary, transformation matrix, classifier, and encoding matrix; is the data set Y hThe sparse coding of the j-th sample data in the dictionary D h ; Q h is the discriminant coding matrix, and the element q at the k-th row and j-th column of it kj is a binary number. Only when the label of the k-th dictionary atom is the same as the class label of the j-th training condition data, q kj is a non-zero value; H h is the condition class label matrix corresponding to the data set Y h ; α and β are the coefficients that control the relative contributions between the discriminant sparse coding error and the classification error; || || F represents the Frobenius norm of the internal matrix, and T is the sparsity of the dictionary.

[0019] Furthermore, the method for solving the optimization function when training the dictionary model is as follows:

[0020] First, initialize the dictionary, transformation matrix, and classifier, and obtain the initial dictionary D (0) , the initial transformation matrix A (0) and the initial classifier W (0) respectively; among them, first use the data samples with the same working conditions in the historical monitoring data set Y h , and train the dictionary corresponding to the working conditions using the K-SVD algorithm, as shown in the following formula:

[0021]

[0022] where Y is the data sample with the same working conditions in the historical monitoring data set Y h , X is the coding matrix of Y based on the dictionary D, and T is the sparsity of the dictionary; combine the dictionary matrices D obtained for all working conditions to obtain the initial dictionary D (0) , and combine the coding matrices X obtained for all working conditions to obtain the initial dictionary D (0) the initial coding coefficient X h under the historical monitoring data set Y (0) ; then initialize A (0) and W (0) by solving the multiple ridge regression model:

[0023] A (0) = Q h X (0)T (X (0) X (0)T + λ 1 I) -1 (3)

[0024] W (0) = H h X (0)T (X (0) X (0)T+λ 2 I) -1 (4)

[0025] where λ 1 , λ 2 represents the regularization parameter, and I represents the identity matrix;

[0026] Then reconstruct the dataset and reconstruct the dictionary Convert Equation (2) to:

[0027]

[0028] Finally, solve Equation (5) by the K-SVD method to obtain D c , and then decompose D c to obtain D h , W h ; among them, first normalize the D c obtained by solving Equation (5) with the L 2 norm and then decompose it, and normalize the decomposed dictionary and classifier.

[0029] Furthermore, calculating the hardness values of each sample in the historical monitoring dataset Y h according to the dictionary D h , the classifier W h , and the encoding matrix X h includes:

[0030] First, use the dictionary D h to calculate the reconstruction error of each sample in the historical monitoring dataset Y h :

[0031]

[0032] where y j , and respectively represent the j-th data sample, its sparse coding, and the reconstruction error in the historical monitoring dataset;

[0033] Then, use the sparse coding and the classifier to calculate the reconstruction label of each sample, and then calculate the cross-entropy:

[0034]

[0035] where C j are the reconstruction label and cross-entropy of the sample y j respectively, and h ij , are the true labels of the sample y j respectively And the reconstruction label The corresponding label for the i-th working condition in

[0036] Then, the calculated reconstruction error and cross-entropy are respectively normalized as follows:

[0037]

[0038] In the formula, C j ' are the normalized reconstruction error and cross-entropy of the sample y j respectively, are the maximum and minimum values of the reconstruction errors of all samples in the dataset Y h under the same working condition as the sample y j respectively, C max , C min are the maximum and minimum values of the cross-entropies of all samples in the dataset Y h under the same working condition as the sample y j respectively;

[0039] Finally, by calculating the proximity of each sample to the optimal and worst samples under the same working condition, the hardness value of each sample is comprehensively obtained:

[0040]

[0041] In the formula, are the proximities of the sample y j to the optimal and worst samples respectively, ω r and ω q are the weights of the monitored hardness and the classified hardness respectively, are the maximum and minimum values of the normalized reconstruction errors of all samples in the same working condition as the sample y j respectively, C' max , C' min are the maximum and minimum values of the normalized cross-entropies of all samples in the same working condition as the sample y j respectively; M j is the hardness value of the sample y j ;

[0042] Furthermore, screening samples from the historical monitoring dataset Y h based on the hardness value specifically includes:

[0043] Dividing the hardness value range of all samples under each working condition in the historical monitoring dataset Y h into ε hardness levels on average;

[0044] Use the average comprehensive hardness value of each hardness grade as the sampling weight, and perform undersampling from all samples of the corresponding hardness grade under the current working condition; the undersampling quantity of each hardness grade is expressed as:

[0045]

[0046] In the formula, represents the average value of all sample hardness values in the d-th hardness grade under the current working condition, represents the average value of all sample hardness values in the l-th hardness grade under the current working condition, Num d represents the undersampling quantity of the d-th hardness grade under the current working condition, Num represents the total undersampling quantity from the dataset Y h under the current working condition.

[0047] Furthermore, the optimization function for constructing the dictionary compression model is:

[0048] The optimization function for constructing the dictionary compression model:

[0049]

[0050] In the formula, and are the distillation losses, and are the compressed dictionary and classifier, that is, the simplified dictionary and simplified classifier, and K s <K, K, K s are the atom numbers of the overcomplete dictionary D b and the simplified dictionary D s respectively; Q s is the discriminant coding matrix corresponding to the simplified dictionary D s ; γ 1 and γ 2 are used to control the relative contribution of the distillation term; A s are the compressed transformation matrices respectively, X s is the reconstructed dataset Y t based on the simplified dictionary D s of the simplified sparse matrix, is the j-th sparse coding in X s ; T is the sparsity of the dictionary; H b is the working condition category label matrix corresponding to the balanced dataset Y b respectively.

[0051] Furthermore, use the alternating optimization algorithm to optimize and solve formula (16), specifically:

[0052] First, perform the initialization operation to obtain and

[0053] Then, for the part of the fixed type (16) except X s the following optimization problem is obtained:

[0054]

[0055] In the formula, is the sparse coding matrix at the k-th iteration; through the OMP algorithm, the optimal at the k-th iteration is obtained. Among them, before each step of OMP solution, it is necessary to perform L 2 norm standardization;

[0056] Then fix X s and use the gradient descent method to optimize D s , W s , A s :

[0057]

[0058]

[0059] In the formula, respectively represent the simplified dictionary, simplified classifier, and simplified transformation matrix at the k-th iteration, and δ is the learning rate of the gradient descent method;

[0060] Continuously repeat the above process, alternately iterate and update until the convergence condition is met, and finally complete the dictionary compression to obtain the simplified dictionary D s and the simplified classifier W s .

[0061] Furthermore, after the cloud obtains the simplified dictionary D s and the simplified classifier W s , it further calculates the reconstruction error threshold for anomaly monitoring, that is, the control limit:

[0062] First, calculate the reconstruction error of each sample in the current balanced data set Y b based on the simplified dictionary D s :

[0063]

[0064] In the formula, represents the j-th sample in the data set Y b , is the sparse coding of the sample based on the simplified dictionary D s , is the reconstruction error of the sample based on the simplified dictionary D s ;

[0065] Then, based on the kernel density estimation (KDE), establish the probability density function f(R) of the reconstruction error R:

[0066]

[0067] where h represents the bandwidth and K(x) represents the kernel function; N b is the number of samples in the dataset Y b ;

[0068] Finally, calculate the control limit R of the reconstruction error according to the given confidence level σ tr :

[0069]

[0070] Furthermore, the specific process of the edge device for anomaly monitoring and working condition identification is as follows:

[0071] If the real-time industrial process monitoring data x new based on the dictionary D s has a reconstruction error R new that does not exceed the control limit R tr , then the current industrial process is normal, and further use the simplified classifier W s to perform working condition discrimination according to x new : Calculate the label vector Take the working condition corresponding to the maximum value in the label vector as the current working condition of the industrial process;

[0072] If R new > R tr , then continue to obtain the industrial process data sequence within a continuous time period, calculate the average reconstruction error R ω of the sequence, and then compare the average reconstruction error R ω of the sequence with the control limit R tr : If R ω ≤ R tr , it means that the industrial process monitoring data x new is an outlier, and it is determined that the current industrial process is in an abnormal state; otherwise, it is determined that a new working condition appears in the current industrial process, and several monitoring data samples of this new working condition are continuously collected and uploaded to the cloud to update the dictionary model.

[0073] Furthermore, between the cloud and the edge device, a cloud-edge data security transmission protocol based on hybrid encryption is used for data transmission, including:

[0074] Use elliptic curve cryptography for key negotiation between the cloud and the edge device, and use the negotiated shared key for mutual authentication;

[0075] After mutual authentication, the AES-GCM algorithm and the shared key are used to symmetrically encrypt and decrypt the transmitted data.

[0076] Beneficial effects

[0077] The present invention provides a cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes. First, an effective monitoring model is established based on dictionary learning method of multi-task hardness evaluation in the cloud, and the historical working condition data with the best generalization performance is screened to update the dictionary model online after the emergence of new working conditions, ensuring the adaptability of the dictionary to new and old working conditions. Secondly, considering the resource limitations of edge devices, a dictionary distillation compression method is proposed to transfer the complete working condition knowledge extracted from the over-complete dictionary in the cloud to the reduced dictionary, minimizing the dictionary capacity while maintaining the original monitoring ability. In addition, the present invention realizes efficient and secure interaction between the cloud and the edge based on hybrid encryption for cloud-edge data transmission, ensuring the reliability of the monitoring model. Description of the drawings

[0078] Figure 1 is the overall framework diagram of the cloud-edge collaborative intelligent autonomous monitoring method described in the embodiments of the present application;

[0079] Figure 2 is the physical architecture of the experiment described in the embodiments of the present application. Detailed implementation manners

[0080] The embodiments of the present invention will be described in detail below. Based on the technical solutions of the present invention, detailed implementation manners and specific operation processes are given, and the technical solutions of the present invention are further explained and illustrated.

[0081] The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes proposed by the present invention has an overall framework as Figure 1 shown, mainly including model training and updating in the cloud, and edge layer devices using the dictionary and classifier updated in the cloud to perform anomaly monitoring and working condition identification on the current industrial process to meet the real-time requirements of the industrial process, and using encryption algorithms for secure data transmission between the cloud and edge devices.

[0082] 1. Initial training of the dictionary model.

[0083] For a complex industrial process with multiple working conditions, Y = [y 1 , y 2 , …, y N ∈ R m×N represents the collected training data samples, where m represents the dimension of the samples and N represents the number of samples. Dictionary learning learns a dictionary D = [d 1 , d 2 , …, d K∈R m×K to achieve the sparse representation of the training data, where d j represents the j-th atom of the dictionary (1 ≤ j ≤ K), and K represents the number of atoms in the dictionary. Utilizing sparsity can improve the representation ability of the sample data, and the dictionary is trained by minimizing the reconstruction error of the sparse representation of the sample data by the dictionary. Dictionary learning can be expressed as the following optimization problem:

[0084]

[0085] where, X = [x 1 , x 2 , …, x N ∈R K×N is called the coding matrix, and each column vector x j in the coding matrix is j the coding coefficient of the corresponding sample data y is the Frobenius norm of matrix A, is called the reconstruction error. T is called the sparsity of the dictionary. Applying the L 0 norm constraint to each column vector of the coding matrix enables each sample to be reconstructed using at most T dictionary atoms. Through these two norm constraints, it can be ensured that the dictionary has the minimum reconstruction error for the data and at the same time ensure the sparsity of the coding matrix.

[0086] For a complex industrial process with multiple working conditions, H h = [h 1 , h 2 , …, h N ∈R n×N represents the one-hot label corresponding to the data sample, where n represents the total number of existing working conditions. Based on Equation (1), the embodiments of the present invention hope that the atoms have discriminative class features, that is, the data of each working condition is represented by the corresponding atom. At the same time, a linear classifier is introduced to reconstruct the class label of the data so that the model can be used for working condition identification. Therefore, the embodiments of the present invention obtain the following optimization function:

[0087]

[0088] where D, A, W, X are the dictionary, transformation matrix, classifier, and coding matrix in the optimization process respectively, and D h , A h , W h , X h are the optimized dictionary, transformation matrix, classifier, and coding matrix; is the sparse coding of the j-th sample data in the historical monitoring data set Y h under the dictionary D, is called the reconstruction error. Q h= [q 1 , q 2 , …, q N ∈ R K×N is the discrimination coding matrix, Q h is a binary matrix, the matrix elements are composed of 0 and 1, and the element q kj at the k-th row and j-th column is a binary number. Only when the label of the k-th dictionary atom is the same as the label of the j-th training data, q kj is a non-zero value; A ∈ R K×K is a transformation matrix, represents the decision sparse coding error, which makes the coding matrix X approximate the discrimination coding matrix Q h , thereby promoting the discriminative class features of the dictionary atoms. H h is the label matrix corresponding to the data set Y h , W = [w 1 , w 2 , …, w K ∈ R n×K is a linear classifier, represents the classification error. By minimizing the error between the reconstructed label and the true label, the classification of sample data is achieved. α and β are the coefficients that control the relative contributions between the decision sparse coding error and the classification error; || || F represents the F-norm of the internal matrix, and T is the sparsity of the dictionary.

[0089] After establishing the optimization function of the dictionary model, it is necessary to optimize and solve it. First, consider D (0) , A (0) and W (0) for initialization. For D (0) , for the data under the same working conditions, the dictionaries are trained using the K-SVD algorithm respectively, and then the dictionaries corresponding to the training of each working condition are combined together to obtain the initial dictionary D (0) . The label of each dictionary atom corresponds to the corresponding working condition label. By solving the multivariate ridge regression model, A (0) and W (0) are initialized as follows:

[0090] A (0) = Q h X T (XX T + λ 1 I) -1 (3)

[0091] W (0) = H h X T (XX T + λ 2 I)-1 (4)

[0092] Among them, X is for the sparse coding with the initial dictionary D (0) . After initialization, Equation (2) can be rewritten as:

[0093]

[0094] Among them, Solve Equation (5) by the K-SVD method to obtain D c . Substitute D c After decomposition, D h and W h are obtained. Among them, in order to reduce the scale difference between dictionary atoms, it is necessary to perform L c norm normalization on D 2 , so renormalize after decomposition:

[0095]

[0096] So far, the cloud has trained to obtain the dictionary D h and the classifier W h .

[0097] 2. Comprehensive multi-hardness value sampling.

[0098] By learning the historical data Y h and H h , D h , W h , and X h can be obtained for anomaly detection and working condition identification. Since the dictionary and the classifier are jointly optimized during training, insufficient information in any aspect will lead to a decline in the overall performance of the model. In addition, sample imbalance does not have a decisive impact on the task difficulty itself, but other difficulties embedded in the data nature, such as noise and class overlap, will significantly deteriorate the model performance. Therefore, the embodiments of the present invention introduce the concepts of detection hardness and classification hardness to integrate potential difficulties. Detection hardness reflects the difficulty of reconstructing sample data using the dictionary, and classification hardness represents the difficulty of correctly classifying sample data. Through in-depth analysis of the model performance, the concept of hardness can help identify whether a sample poses a challenge to the model's ability and reveal its potential weaknesses. This enables more targeted improvement of the model during the optimization process, ultimately enhancing the robustness of the model. First, the model can reconstruct the historical data and labels, and then calculate the detection hardness and classification hardness.

[0099] The ability to represent data represented by the monitoring hardness can be measured by the reconstruction error. Generally, data samples located at the decision boundary are more difficult to be represented by the dictionary, and their reconstruction errors are larger. The reconstruction error of historical working condition samples is calculated as follows:

[0100]

[0101] Among them, y j , and represent the j-th training data sample, its sparse coding, and the reconstruction error. Similarly, using sparse coding and a classifier, the labels of historical operating condition data can be reconstructed as follows:

[0102]

[0103] The vector is the reconstructed label of the j-th historical operating condition data. After normalizing it by Softmax, the output of the reconstructed label can be transformed into a probability distribution, and thus the cross-entropy of the j-th operating condition sample data can be calculated:

[0104]

[0105] The cross-entropy of historical data is used to evaluate the classification performance of the classifier. The lower the value of the cross-entropy, the closer the predicted classification result is to the actual result. In summary, detecting hardness and classifying hardness can be represented by the reconstruction error and the cross-entropy. From the perspective of hardness, the greater the hardness value corresponding to the sample, the more information it contains.

[0106] Next, a comprehensive index that simultaneously considers detecting hardness and classifying hardness needs to be calculated. First, unify the dimensions:

[0107]

[0108] Among them, and represent the maximum and minimum reconstruction errors under the current operating condition, C max and C min are the same. The dimensions are unified within [0,1] through Equation (10). Different from common evaluation methods, if a certain sample has a high hardness for only a single task, the embodiments of the present invention also consider such samples to be contributory. Because it can help the model improve the performance in other fields in a specific scenario. Therefore, the embodiments of the present invention compare the degree of closeness of each scheme to the best scheme and the worst scheme as the evaluation score. The comprehensive evaluation method based on distance can effectively prevent key information from being masked and ensure the comprehensiveness of the evaluation method. The specific calculation is as follows:

[0109]

[0110] Among them, and are the degrees of closeness of the sample y j to the optimal sample and the worst sample under the same operating condition, respectively, ωr and ω q is the weight relationship for monitoring hardness and classifying hardness. In the embodiments of the present invention, the weight relationship is estimated from the data itself by the entropy weight method. M j is the comprehensive hardness score of the j-th sample under this working condition. Since M j combines the detected hardness and the classified hardness through a weighting method, so M j is selected. Updating the samples with higher values can improve the overall monitoring performance. At the same time, to prevent overfitting, a small number of low-hardness samples are retained to represent the skeletons of their respective distributions. Therefore, by sorting according to the comprehensive hardness score, the historical working condition data can be evenly divided into ε hardness levels. The number of undersampled samples from each hardness level is as follows:

[0111]

[0112] In this embodiment, the average comprehensive hardness score of the d-th hardness level is used as the sampling weight. Num represents the number of undersampled samples in each category of historical data. High-hardness samples provide in-depth optimization for specific tasks, while low-hardness samples ensure the wide adaptability of the model in more common situations. Reasonable proportion of resource allocation is carried out based on the complexity of the samples, so as to expand the influence of high-hardness samples while retaining a reasonable proportion of low-hardness samples as skeletons. Therefore, the sampling method proposed in this paper can cover the distribution of the entire working condition data with the fewest samples, ensure the optimization ability of the model for specific tasks, and prevent overfitting at the same time.

[0113] Finally, the most representative historical data set obtained by screening can be represented

[0114] 3. Merge new working condition samples and update the model.

[0115] When a new working condition appears in the current industrial process, the edge device can upload a part of the data in real time, denoted as Y new . A balanced data set can be established and the corresponding number of categories is also updated to n = n h + n new . Use the latest balanced data set to update the dictionary model to obtain an overcomplete dictionary D b , classifier W b and the corresponding coding matrix X b , ensuring the adaptability of the obtained dictionary and classifier to new and old working conditions.

[0116] 4. Cloud dictionary distillation and compression.

[0117] The dictionary trained in the cloud is over-complete to enhance its representation ability. Although it ensures effective monitoring performance, it also leads to a large number of redundant atoms. Deploying the redundant dictionary on edge devices will increase the data transmission load. At the same time, online monitoring will also increase the computational amount. Therefore, it is necessary to compress the model before deployment. Traditional methods achieve compression by deleting redundant dictionary atoms according to their contribution degrees, but directly removing atoms will weaken the model's representation ability for key features and cause a decline in model performance. Therefore, the embodiment of the present invention proposes a dictionary distillation compression method to extract key information from the original model and transfer it to its streamlined version, so that the expressiveness of the original model can be retained during the compression process.

[0118] First, using the complex over-complete dictionary trained in the cloud as a guide, extract its reconstruction result as complete operating condition knowledge, and its mathematical representation is as follows:

[0119]

[0120] Among them, Y t and H t are the reconstruction data and reconstruction labels of the uncompressed over-complete dictionary. By introducing a distillation term, distill the extracted complete operating condition knowledge into a streamlined dictionary model to achieve model compression. The optimization function of the dictionary compression model is obtained:

[0121]

[0122] Among them, and are the distillation losses, and are the compressed dictionary and classifier (K s < K), γ 1 and γ 2 control the relative contribution of the distillation term. The purpose of formula (16) is to make the streamlined dictionary approximate the reconstruction results of the uncompressed dictionary and classifier during the process of learning training data, so as to minimize the dictionary capacity while retaining the original monitoring ability.

[0123] To obtain the compressed simplified dictionary D s and the simplified classifier W s , the embodiment of the present invention innovatively designs an alternating optimization algorithm to optimize and solve formula (16).

[0124] First, perform the initialization operation in the same way as above to obtain and Then perform iterative solution. For the kth iteration, fix the part except X s to obtain the following optimization problem:

[0125]

[0126] Among them, is the sparse coding matrix at the k-th iteration. Through the OMP algorithm, the optimal It should be noted that before each step of OMP solution, it is necessary to perform L 2 norm standardization.

[0127] Then fix X s , and the gradient descent method can be used to optimize D s , W s , A s , as shown in the following formula:

[0128]

[0129] Among them, represents the compressed dictionary matrix at the k-th iteration, and the other symbols are the same. δ is the learning rate of the gradient descent method. Continuously repeat the above process, alternately iterate and update until the convergence condition is met, and finally complete the dictionary compression to obtain the simplified dictionary D s and the simplified classifier W s .

[0130] 5. Cloud computing control limit.

[0131] Since the edge side will perform online monitoring according to the compressed model, it is necessary to use this model to evaluate the reconstruction error threshold for anomaly detection, that is, the control limit, before sending the model. First, calculate the reconstruction error of each training data under the compressed dictionary, as shown in Equation (7). According to the kernel density estimation KDE, the probability density function of the reconstruction error can be calculated as:

[0132]

[0133] Among them, h represents the bandwidth, and K(x) represents the kernel function. In the embodiments of the present invention, the Gaussian kernel function is selected. After obtaining the probability density function f(R), given the confidence level of σ, the control limit R tr :

[0134]

[0135] Finally, the cloud will deploy the compressed dictionary D s , the classifier W s and the control limit R tr to the edge side for online monitoring.

[0136] 6. Edge side online monitoring and update trigger.

[0137] The edge device will utilize D s , W s and R tr to conduct online monitoring of real-time working condition data, including anomaly detection and working condition identification. In addition, when new unforeseen working conditions occur, the edge device should be able to provide an update signal to minimize the duration of model mismatch. Therefore, this embodiment also proposes a dynamic error triggering strategy to trigger updates in a timely manner when new unforeseen working conditions occur. Therefore, this section will be described from two parts: online monitoring and dynamic error triggering strategy.

[0138] 6.1 Online monitoring.

[0139] When the sensors in the industrial field collect new process data y new , it is sent to the edge device for online monitoring. The edge device utilizes the dictionary model D s sent by the cloud to reconstruct the process data. First, sparse representation is performed:

[0140]

[0141] x new can be calculated using the OMP method. Then, the reconstruction error R new can be calculated. The normal data learned by the dictionary has a smaller reconstruction error when reconstructed using this dictionary. Therefore, R new is compared with R tr . If R new ≤R tr , the online data belongs to normal data. Subsequently, W s can be used to further discriminate the working conditions of the normal data. The label vector can be calculated:

[0142]

[0143] The index of the largest element in the label vector is the working condition to which the data belongs. If the i-th element in the vector is the largest, the data belongs to the i-th working condition.

[0144] 6.2 Dynamic error triggering strategy.

[0145] If the online monitoring determines that R new >R tr, it indicates the emergence of brand-new features not covered by the dictionary model. This presents two possible scenarios: a fault or new unforeseen operating conditions. Since the operating conditions in a large industrial site occur sequentially and the new operating conditions will persist for a relatively long time after their emergence, the reconstruction error of the new operating condition data will stably exceed the control limit within a certain period; while the faults generated in the actual industrial process will cause data fluctuations in the early stage, resulting in the continuous fluctuation of its reconstruction error between the control limits. Therefore, the present invention proposes a dynamic error-triggering strategy that can detect the switching of new operating conditions in a timely manner without affecting fault detection. Specifically, a moving window with a length of ω is used to form a sample sequence. This sequence includes the current sampling sample and the previous ω - 1 samples. The edge device will calculate and store the average reconstruction error R in this sample sequence. ω . When brand-new features appear, R ω is compared with R tr . If R ω ≤ R tr , it indicates that this sample is an outlier and is judged as a fault; otherwise, it is judged as a new operating condition. Subsequently, the edge device collects a small amount of process data of the new operating condition and uploads it to the cloud for online model update.

[0146] 7. Secure data transmission between the cloud and the edge.

[0147] The challenges brought by dynamic industrial processes are addressed through self-updating dictionary learning within the cloud-edge collaboration framework. Therefore, data transmission is required between the cloud and the edge during the update. However, due to the transparency and untrustworthiness of the network, there are security risks during data transmission. Resource constraints and real-time requirements also complicate this issue. This section will introduce a secure data transmission protocol for cloud-edge data based on hybrid encryption, enabling data to be transmitted securely and quickly. This protocol consists of a shared key protocol and symmetric encryption. Specifically, ECC is used to synchronize the shared key and verify the identity between the cloud and the edge. Subsequently, this shared key is used for symmetric encryption to securely transmit data.

[0148] The shared key negotiation based on ECC is established on the discrete logarithm problem under the elliptic curve. The communication parties share the domain parameters E{e p , e a , e b , G, e n , e h} of the elliptic curve used before the negotiation. In the embodiment of the present invention, the elliptic curve is selected as secp256k1, which is constructed in a special, non-random manner and has a faster calculation speed compared to other curves. To achieve simple and efficient identity authentication, both parties have a pre-shared secret password PW. The cloud and the edge calculate an integer t according to this password PW and the following formula:

[0149] t = hash(PW) mod e n (25)

[0150] Where hash(PW) is the hash result of the password PW, and SHA-256 is selected as the hash function here. mod represents the modulo operation. The key agreement process based on the password PW is as follows:

[0151] (1) Edge side: Randomly generate r e ∈[1, e n -1], calculate R e = [r e + t]G. The edge side transmits R e ;

[0152] (2) Cloud side: Randomly generate r c ∈[1, e n -1], calculate R c = [r c + t]G. The cloud side transmits R c ;

[0153] (3) Edge side: After receiving R c transmitted by the cloud side, calculate S ec = [r e (R c + [-t]G);

[0154] (4) Cloud side: After receiving R e transmitted by the edge side, calculate S ce = [r c (R e + [-t]G).

[0155] Where -t and t are additive inverses modulo e n . Based on the discrete logarithm problem of elliptic curves, an attacker cannot obtain the randomly generated private keys r e and r c from the public key. And since both the cloud and the edge share the integer t generated by the password, a man-in-the-middle cannot launch an effective attack. Therefore, after the key exchange is completed, since the ECC operation satisfies the Abelian group, it can be known that S = S ec = [r e ([r c G) = [r c ([r e G) = S ce . If both parties can negotiate a consistent shared key, the authenticated key agreement is completed.

[0156] After negotiating a consistent shared key, data transmission between the cloud and the edge can be processed using symmetric encryption. In the transmission protocol of the embodiments of the present invention, AES-GCM is used, which is a mode of the Advanced Encryption Standard (AES). As an AEAD-type algorithm, AES-GCM has confidentiality, integrity, and authenticity at the same time. In this encryption mode, the plaintext does not require any padding, and the MAC value can be compared to ensure the integrity of the message. It provides three different key lengths. To ensure data security to the greatest extent, the key length of AES in this protocol is selected to be 256 bits. Take the abscissa of S as the symmetric encryption key, denoted as S.x. The process of encrypting and decrypting the plaintext m using AES-GCM is as follows:

[0157] The encryption by the sender is represented by the function p = encrypt(m, S):

[0158] (1) Generate a 16-byte random number nonce;

[0159] (2) Use the random number noice and the key S.x to perform AES encryption on the plaintext m to generate a ciphertext c of the same length;

[0160] (3) Use the GMAC algorithm to generate the message authentication code mac of nonce and ciphertext c to verify the integrity;

[0161] (4) Combine nonce, c, and mac together as p = nonce||c||mac and send it to the receiver.

[0162] The decryption by the receiver is represented by the function m = decrypt(p, S):

[0163] (1) Split nonce, c, and mac from the received p;

[0164] (2) Use the GMAC algorithm to generate the message authentication code mac' of nonce and ciphertext c again;

[0165] (3) Verify whether mac is equal to mac'. If they are not equal, it means that the message has been tampered with or the shared key S used is inconsistent, and a tampering attack has occurred during data transmission;

[0166] (4) If the verification passes, use the random number noice and the key S.x to decrypt the ciphertext c to restore the plaintext m.

[0167] Through the two functions p = encrypt(m, S) and m = decrypt(p, S), data transmission can be carried out using symmetric encryption after negotiating the shared key. After triggering the model update, the data transmission between the cloud and the edge through the secure data transmission protocol can be divided into a key negotiation phase, a data collection phase, and a model update phase. First, key negotiation and identity authentication are carried out through the key negotiation phase; subsequently, in the data collection phase, the edge device encrypts the uploaded real-time data using the shared key and the symmetric encryption algorithm. After reaching the specified upload volume, the model update phase is started. At this time, the cloud uses the uploaded new working condition data and the extracted representative old working condition data for online update and model compression, and deploys the obtained models D s , W s , and R tr after encryption to the edge device.

[0168] 8. Validity verification.

[0169] To verify the effectiveness of the present invention in actual industrial applications, this example takes the production process of a certain aluminum alloy wheel manufacturing factory as an example. There are more than 200 types of wheels that can be produced on the production line, but in actual production, the daily production plan of the production line only includes a few types. The production plan of the wheel production line will be adjusted due to reasons such as order changes. Therefore, the wheel production process is a dynamic industrial process. The embodiment of the present invention monitors the wheel production and conducts fault detection on the production process to avoid the mixing of non-production plan wheel models into the production line, and the working condition identification can guide the grasping process of the manipulator program on the production line. When the production plan is adjusted, a safe and reliable model update is carried out between the cloud and the edge to ensure that the monitoring model matches the dynamic production line.

[0170] In this experiment, the wheel production system, Internet of Things sensors, edge devices, and cloud servers together constitute the physical architecture of cloud-edge collaboration. As Figure 2 shown, the edge devices are deployed on-site in the wheel production system to process the real-time data collected by the sensors and upload the data to the cloud server for training and updating. The physical layer communicates with the edge layer through the OPC UA protocol, and the edge layer communicates with the cloud server through the TCP protocol. The cloud server is responsible for processing large-scale data and communicating with the model. To simulate the real production process, 3 types of wheel models are randomly selected as normal working conditions and 1 type of wheel model is selected as a fault in the experiment. Among them, the data of wheel models M1 and M2 are collected in advance and stored in the cloud. Wheel model M3 is the newly added production plan on the day, and wheel model M4 is an abnormal wheel model that is not produced. There are 150 historical wheel model images stored in the cloud. The experiment is verified from three aspects: model compression effect, online monitoring update and security effect analysis, and model monitoring performance.

[0171] First, test the impact of model compression on the monitoring effect. In this experiment, a model trained with historical operating condition data stored in the cloud is used, and 100 data are generated for operating condition 1, operating condition 2, and abnormal operating condition in sequence as the test set. Different model compression ratios are set to make the compressed dictionary capacities different, and the monitoring effects of the dictionary on the test data after each compression are recorded, including FAR, FDR, recognition rate, and monitoring time consumption. The experimental results are shown in the following table. According to the experimental results, since the dictionary model is compressed, the time spent on matching atoms during sparse representation is correspondingly reduced, ensuring the real-time nature of monitoring. Moreover, since the model fully learns the knowledge of the original complex dictionary during the compression process and can maintain the original monitoring ability as much as possible during compression, although the model is compressed, it can still maintain a high monitoring accuracy and recognition rate.

[0172]

[0173] Due to the dynamic changes in the production process, when new operating conditions appear, the model needs to be updated online between the cloud and the edge. This experiment demonstrates the online update in the face of a dynamic industrial process. When the dynamic error strategy determines the generation of a new operating condition, the edge device uploads 30 new operating condition data in real time for online model update. Moreover, through ARP spoofing for man-in-the-middle attack, 5 new operating condition data uploaded by the edge device are maliciously tampered with as faulty data, affecting the fault detection performance of the monitoring model. The overall monitoring performances with and without the cloud-edge secure data transmission protocol are compared, as shown in Table 4 below. Without using the encryption scheme, it is impossible to determine whether the data has been tampered with, and using the data tampered with by the attacker to update the monitoring model results in the loss of the monitoring ability for faults.

[0174]

[0175] Finally, through comparative experiments, explore the monitoring performance of the model after update, and compare the proposed method with PCA-EWC and ETDL that also samples the cloud-edge collaboration method. The monitoring results are shown in the following table. ETDL ignores the imbalance problem between new and old samples during model update, and PCA-EWC is difficult to learn knowledge from a small amount of data and has a catastrophic forgetting problem in the face of significant operating condition changes. The present invention makes full use of the advantages of cloud-edge collaboration for model update to provide fault detection and operating condition recognition results, and can obtain the best monitoring performance in the face of a dynamic industrial process.

[0176]

[0177] The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes of the present invention focuses on the cyber-physical problems existing in dynamic industrial processes and modern industrial Internet of Things. First, an effective monitoring model is established through a dictionary learning method based on multi-task hardness evaluation, and the dictionary is dynamically updated by combining the most influential historical data and new working condition data, ensuring its adaptability to working conditions. Then, by distilling and compressing the over-complete dictionary, a simplified dictionary and a classifier are obtained and deployed to the edge side for monitoring and working condition identification, thus saving the computing resources at the edge side to ensure the real-time performance of monitoring. Finally, a cloud-edge data transmission protocol based on hybrid encryption is introduced to achieve efficient and reliable interaction of data and models between the cloud and the edge.

[0178] The above embodiments are the preferred embodiments of the present application. Those of ordinary skill in the art can also make various transformations or improvements based on this. Without departing from the general concept of the present application, these transformations or improvements should fall within the scope of protection required by the present application.

Claims

1. A cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes, characterized in that: include: The cloud builds a dictionary model and trains and updates the model; The edge device obtains the latest dictionary, classifier, and control limit from the cloud, and uses the obtained dictionary and control limit to monitor the abnormality of the current industrial process, and uses the obtained dictionary and classifier to identify the working condition of the current industrial process; The cloud establishes a dictionary model and trains and updates the model, including: Establish a dictionary model and use the historical monitoring data set Y of each working condition h And the corresponding working condition category label H h , train the dictionary model to obtain the dictionary D h , classifier W h And the dataset Y h Based on dictionary D h The encoding matrix X h ; According to the dictionary D h , classifier W h , encoding matrix X h Calculate the historical monitoring data set Y h The hardness value of each sample in the hardness value is obtained from the historical monitoring data set Y h The samples are selected and then combined with the new working condition data samples to form the current balanced data set Y b ; Using the current balanced dataset Y b Update the dictionary model to obtain an overcomplete dictionary D b , classifier W b And the dataset Y b Based on the overcomplete dictionary D b The encoding matrix X b ; Using an overcomplete dictionary D b , classifier W b and the encoding matrix X b , calculate the reconstructed data set Y t and refactoring label H t ; To reconstruct the dataset Y t and refactoring label H t As the complete working condition knowledge, the dictionary compression model is constructed and the simplified dictionary D is obtained by solving s and the simplified classifier W s , as the final updated dictionary and classifier.

2. The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes according to claim 1 is characterized in that: The optimization function of the dictionary model is expressed as: Among them, D, A, W, and X are the dictionary, transformation matrix, classifier, and encoding matrix in the optimization process respectively. h ,A h ,W h ,X h The dictionary, transformation matrix, classifier and encoding matrix obtained for optimization; For data set Y h The jth sample data in the dictionary D h Sparse coding under Q h is the discriminant coding matrix, whose element q in the kth row and jth column is kj is a binary number. Only when the label of the kth dictionary atom is the same as the category label of the jth training condition data, q kj Only non-zero value; H h For data set Y h The corresponding working condition category label matrix; α and β are the coefficients of the relative contribution between the control decision sparse coding error and the classification error; || || F represents the F norm of the internal matrix, and T is the sparsity of the dictionary.

3. The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes according to claim 2 is characterized in that: The method for solving the optimization function when training the dictionary model is: First, initialize the dictionary, transformation matrix, and classifier to obtain the initial dictionary D (0) , initial transformation matrix A (0) and the initial classifier W (0) ; Among them, the historical monitoring data set Y is used first h The data samples of the same working conditions in the training set are trained using the K-SVD algorithm to obtain the dictionary under the corresponding working conditions, as shown in the following formula: Among them, Y is the historical monitoring data set Y h The data samples have the same working conditions, X is the encoding matrix of Y based on the dictionary D, and T is the sparsity of the dictionary; the dictionary matrix D obtained from all working conditions is combined to obtain the initial dictionary D (0) , combine the encoding matrices X obtained from all working conditions to obtain the initial dictionary D (0) In the historical monitoring dataset Y h The initial coding coefficient X under (0) ; Then, by solving the multivariate ridge regression model, A (0) and W (0) To initialize: A (0) =Q h X (0)T (X (0) X (0)T +λ1I) -1 (3) W (0) =H h X (0)T (X (0) X (0)T +λ2I) -1 (4) In the formula, λ1,λ2 represent regularization parameters, and I represents the unit matrix; Then reconstruct the dataset And reconstruct the dictionary Convert equation (2) to: Finally, the K-SVD method is used to solve equation (5) to obtain D c , and then D c Decompose to get D h , W h ; Among them, first solve the D obtained by formula (5) c After L2 norm normalization, decomposition is performed, and the decomposed dictionary and classifier are standardized.

4. The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes according to claim 1 is characterized in that: According to the dictionary D h , classifier W h , encoding matrix X h Calculate the historical monitoring data set Y h The hardness values ​​of each sample in the test include: First, use the dictionary D h Calculate the historical monitoring data set Y h The reconstruction error of each sample in is: In the formula, y j , and Respectively represent the j-th data sample in the historical monitoring data set and its sparse coding and reconstruction error; Then, sparse coding and classifier are used to calculate the reconstructed labels of each sample, and then the cross entropy is calculated: In the formula, C j are samples y j The reconstructed labels and cross entropy, h ij , are samples y j The real label and refactor tags The corresponding label of the i-th working condition in ; The calculated reconstruction error and cross entropy are then standardized: In the formula, C j ' are sample y j The standardized reconstruction error and cross entropy of The data sets Y h In the sample y j The maximum and minimum values ​​of the reconstruction error of all samples under the same working condition, C max ,C min The data sets Y h In the sample y j The maximum and minimum cross entropy values ​​of all samples under the same working condition; Finally, by calculating the closeness of each sample to the best sample and the worst sample under the same working condition, the hardness value of each sample is obtained comprehensively: In the formula, are samples y j The closeness of the best sample and the worst sample, ω r and ω q are the weights of monitoring hardness and classification hardness, are respectively j The maximum and minimum values ​​of the normalized reconstruction error of all samples under the same working condition, C' max ,C' min are respectively j The maximum and minimum values ​​of the cross entropy normalized for all samples under the same working condition; M j For sample y j Hardness value.

5. The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes according to claim 1 is characterized in that: Based on the hardness value from the historical monitoring data set Y h Screening samples include: The historical monitoring data set Y h The hardness value range of all samples under each working condition is evenly divided into ε hardness grades; The average comprehensive hardness value of each hardness grade is used as the sampling weight, and under-sampling is performed from all samples of the corresponding hardness grade under the current working condition; the number of under-sampling for each hardness grade is expressed as: In the formula, It represents the average hardness value of all samples in the dth hardness grade under the current working condition. Indicates the average hardness value of all samples in the first hardness grade under the current working condition, Num d Indicates the number of undersampling of the dth hardness grade under the current working condition, Num represents the number of undersampling from the data set Y h The total number of undersamplings for the current condition.

6. The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes according to claim 1 is characterized in that: The optimization function for constructing the dictionary compression model is: Construct an optimized function for the dictionary compression model: In the formula, and is the distillation loss, and is the compressed dictionary and classifier, that is, the simplified dictionary and simplified classifier, and K s <K, K, K s They are respectively the overcomplete dictionary D b and the simplified dictionary D s The number of atoms; Q s To simplify the dictionary D s The corresponding discriminant coding matrix; γ1 and γ2 are used to control the relative contribution of the distillation terms; A s are the transformation matrices after compression, X s To reconstruct the dataset Y t Based on the simplified dictionary D s The simplified sparse matrix of For X s The jth sparse code in ; T is the sparsity of the dictionary; H b To balance the dataset Y b The corresponding working condition category label matrix.

7. The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes according to claim 6 is characterized in that: The alternating optimization algorithm is used to optimize and solve equation (16), specifically: First, the initialization operation gets and Then, fix the equation (16) by dividing by X s The following optimization problem is obtained: In the formula, is the sparse coding matrix at the kth iteration; through the OMP algorithm, the optimal kth iteration is obtained Among them, before each step of OMP solution, Perform L2 norm normalization; Fix X again s , using the gradient descent method to optimize D s , W s , A s : In the formula, They represent the simplified dictionary, simplified classifier and simplified transformation matrix of the kth iteration respectively, and δ is the learning rate of the gradient descent method; Repeat the above process and iterate and update alternately until the convergence condition is met, and finally complete the dictionary compression to obtain the simplified dictionary D s and the simplified classifier W s .

8. The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes according to claim 1 is characterized in that: Simplified dictionary D is available in the cloud s and the simplified classifier W s After that, the reconstruction error threshold for abnormal monitoring, i.e., the control limit, is further calculated: First, calculate the current balanced data set Y b Each sample in is based on the simplified dictionary D s The reconstruction error is: In the formula, Represents the data set Y b The jth sample in For sample Based on the simplified dictionary D s The sparse coding of For sample Based on the simplified dictionary D s The reconstruction error of Then, based on the kernel density estimation KDE, the probability density function f(R) of the reconstruction error R is established: In the formula, h represents bandwidth, K(x) represents kernel function; N b For data set Y b The number of samples in ; Finally, the control limit R of the reconstruction error is calculated according to the given confidence level σ tr :

9. The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes according to claim 1 is characterized in that: Edge devices perform abnormal monitoring and condition identification as follows: If real-time industrial process monitoring data x new Based on dictionary D s The reconstruction error R new The control limit R is not exceeded tr , then the current industrial process is normal, and the simplified classifier W is further used s According to x new Perform working condition identification: Calculate label vector Get the label vector The operating condition corresponding to the maximum value is the current operating condition of the industrial process; If R new >R tr , then continue to obtain the industrial process data sequence within the continuous period and calculate the average reconstruction error R of the sequence ω , and then the average reconstruction error R of the sequence ω and control limit R tr Comparison: If R ω ≤R tr , then it means industrial process monitoring data x new is an abnormal value, and the industrial process is judged to be in an abnormal state; otherwise, it is judged that a new operating condition has occurred in the industrial process, and several monitoring data samples of the new operating condition are continuously collected and uploaded to the cloud to update the dictionary model.

10. The cloud-edge collaborative intelligent autonomous monitoring method for dynamic industrial processes according to claim 1 is characterized in that: The cloud-edge data security transmission protocol based on hybrid encryption is used for data transmission between the cloud and edge devices, including: Elliptic curve cryptography is used for key negotiation between the cloud and edge devices, and the negotiated shared key is used for mutual authentication; After mutual authentication is passed, the AES-GCM algorithm and shared key are used to symmetrically encrypt and decrypt the transmitted data.

Images