CN-122020286-A - Abnormality detection method, abnormality detection system and abnormality detection storage medium for cigarette key index label noise

CN122020286ACN 122020286 ACN122020286 ACN 122020286ACN-122020286-A

Abstract

The invention relates to the technical field of industrial process anomaly detection, in particular to a method, a system and a storage medium for detecting the anomaly of a cigarette key index label noise in a space-time dual-channel mode, which comprise the steps of acquiring historical time sequence data of cigarettes and carrying out correlation analysis to acquire a correlation index data set; the method comprises the steps of preprocessing a correlated index data set to obtain a training sample set, constructing a space-time dual-channel detection model for obtaining the prediction probability of a time classifier and the prediction probability of a space classifier, training the space-time dual-channel detection model by adopting the training sample set, and detecting cigarette sample data acquired in real time by adopting the trained space-time dual-channel detection model to obtain a detection result. The space classifier combines the process sequence and the process coupling relation among modeling variables, thereby improving the anomaly identification capability from the space-time dual view angle.

Inventors

ZHU LIMING
ZHANG BO
LI MING
ZHANG QIANG
ZHANG LIHONG
PENG BIN

Assignees

浙江中烟工业有限责任公司

Dates

Publication Date: 20260512
Application Date: 20260113

Claims (10)

1. An anomaly detection method for cigarette key index label noise is characterized by comprising the following steps: Acquiring historical time sequence data of cigarettes and performing correlation analysis to acquire a correlation index data set; Preprocessing the associated index data set to obtain a training sample set; constructing a space-time dual-channel detection model for acquiring the prediction probability of the time classifier and the prediction probability of the space classifier; Training the space-time dual-channel detection model by adopting a training sample set; And detecting the cigarette sample data acquired in real time by adopting the trained space-time double-channel detection model so as to acquire a detection result.
2. The method of claim 1, wherein preprocessing the association index data set to obtain a training sample set comprises: Carrying out sliding window division on the associated index data set to obtain a sliding window sample set; Obtaining a sample overrun of each sliding window sample; and dividing the associated index data set into a clean sample set with high confidence and a label sample set with low confidence according to the sample overrun ratio, and taking the clean sample set and the label sample set with low confidence as training sample sets.
3. The method of claim 1, wherein constructing a spatio-temporal two-channel detection model for obtaining a temporal classifier prediction probability and a spatial classifier prediction probability comprises: constructing a time classifier by adopting a gating circulating unit to acquire hidden states of all time steps; inputting the hidden state to an attention mechanism to acquire an attention weight; calculating corresponding suction resistance priori attention weights according to the process values of the suction resistances corresponding to the key indexes; acquiring corrected attention weights according to the attention weights and the suction resistance priori attention weights; Adopting the corrected attention weight to weight and aggregate hidden states of all time steps so as to acquire process time sequence characterization; and obtaining the prediction probability of the time classifier according to the process time sequence representation.
4. The method of claim 3, wherein calculating the corresponding prior attention weight for the resistance according to the process value of the resistance corresponding to the key indicator comprises: calculating the distance between the process value and the standard value of each suction resistor; mapping the distance to a non-negative a priori score; And normalizing the prior score to obtain a suction resistance prior attention weight.
5. The method of claim 1, wherein constructing a spatio-temporal two-channel detection model for obtaining a temporal classifier prediction probability and a spatial classifier prediction probability comprises: Constructing a spatial classifier by adopting a graph attention network, wherein the spatial classifier comprises a graph structure learning layer and a graph information propagation layer; acquiring a learnable adjacency matrix by adopting the graph structure learning layer; acquiring strongly associated neighbor nodes in the adjacent matrix by adopting a sparsification method to obtain a simplified learned graph structure; embedding the rolling process sequence as process knowledge into the graph structure to obtain a learnable dynamic adjacency matrix; the initial characteristics of each node of the learnable dynamic adjacency matrix are updated in an aggregation mode through the graph information propagation layer so as to obtain spatial characterization; And obtaining the prediction probability of the spatial classifier according to the spatial characterization.
6. The method of claim 5, wherein embedding the splice process sequence as process knowledge into the graph structure to obtain a learnable dynamic adjacency matrix comprises: a process mask matrix is constructed based on the crimping process sequence according to formula (1), ,(1) Wherein, the For process mask matrix, representing variables only Belonging to the working procedures And the variables are Belonging to the next working procedure In the time-course of which the first and second contact surfaces, =1; A learnable dynamic adjacency matrix is obtained according to formula (2), ,(2) Wherein, the As a dynamic adjacency matrix that can be learned, In order to be able to construct the figure, Representing the hadamard product.
7. The method of claim 1, wherein training the spatio-temporal two-channel detection model with a training sample set comprises: Semi-supervised collaborative training is carried out on the space-time dual-channel detection model by adopting a low confidence coefficient sample set in the training sample set; And performing supervised training on the space-time dual-channel detection model by adopting a high confidence sample set in the training sample set.
8. The method of claim 7, wherein semi-supervised co-training the spatio-temporal two channel detection model with a low confidence sample set of the training sample sets comprises: acquiring a semi-supervised loss function according to the formulas (3) to (6), ,(3) ,(4) ,(5) ,(6) Wherein, the In order for the cross-entropy loss to occur, For the mean square error loss of the pseudo tag samples, In order for the loss to be regularized, As a semi-supervised loss function, Is the first Cross entropy loss of the individual samples, For the prediction of unlabeled exemplars in the low confidence exemplar set for the current channel, To use two-channel prediction of unlabeled exemplars in a low confidence exemplar set, In the case of a batch size of the product, As the total number of categories to be considered, In order to have a sample set of labels, As a set of pseudo tag samples, Representing the predicted probability of the current channel, Is that Is used for the weight coefficient of the (c), Is that Weight coefficient of (c) in the above-mentioned formula (c).
9. An anomaly detection system for cigarette key indicator tag noise, the system comprising a processor configured to perform the method of any one of claims 1 to 8.
10. A computer readable storage medium having instructions stored thereon which, when executed by a processor, implement the method of any of claims 1 to 8.

Description

Abnormality detection method, abnormality detection system and abnormality detection storage medium for cigarette key index label noise Technical Field The invention relates to the technical field of industrial process anomaly detection, in particular to a method, a system and a storage medium for detecting space-time double-channel anomaly of cigarette key index label noise. Background As industrial production continues to advance to the large-scale, intelligent and green directions, the complexity of process control and optimization continues to deepen. In the background of increasingly competitive cigarette markets, cigarette quality has become an important component of enterprise core competitiveness. The rolling connection link in the cigarette manufacturing process is used as the last processing procedure before the product leaves the factory, and has decisive influence on the quality of the finished cigarette. The suction resistance and the ventilation degree are two key indexes for measuring whether the cigarettes are qualified or not, wherein the suction experience is weakened due to the fact that the suction resistance is too large or the ventilation degree is insufficient, and the fragrance is insufficient due to the fact that the suction resistance is smaller or the ventilation degree is too high. Therefore, the two indexes are not only important basis for product quality evaluation, but also key monitoring signals for identifying production process abnormality. At present, a Cigarette Inspection System (CIS) is commonly adopted in the industry, and the overrun cigarettes are removed through manually set suction resistance and ventilation threshold values. However, the detection mode based on the experience threshold has obvious limitations, including high equipment dependence, single monitoring index, incapability of analyzing abnormal root causes and the like. More importantly, the manual threshold has strong subjectivity, and is easy to cause missed detection or false detection, so that the label inevitably contains noise, and the stability of quality control is further damaged. In recent years, with rapid advancement of digital transformation of factories, data-driven abnormality detection methods have received attention. The development of the neural network and the deep learning technology enables the model to accurately predict or reconstruct normal working condition data, autonomously learn a normal mode from massive process data, do not need to rely on prior knowledge in the field, and are applied to cigarette abnormality detection tasks. However, such methods typically rely on reliable normal/abnormal labels. In actual cigarette production, in order to achieve both energy consumption and product quality, abnormal thresholds of suction resistance and ventilation degree are often set manually, which inevitably introduces label noise, so that application effects of the models in real industrial scenes are limited. On the other hand, the rolling process is a typical space-time coupling production system essentially, wherein in the time dimension, a rolling machine performs continuous mass production, obvious time sequence dependence exists among process parameters, and in the space dimension, a complex nonlinear coupling structure is formed among different process units due to a production mechanism and a linkage relation and is closely related to a key quality index. Aiming at the problem of label noise, a space-time anomaly detection method facing to the noise label is needed to be constructed so as to improve the detection performance and the robustness of the model under complex production conditions. Disclosure of Invention The embodiment of the invention aims to provide a space-time dual-channel anomaly detection method, a system and a storage medium for cigarette key index label noise, which are used for solving the technical problems that label noise and a data driving method are limited due to the fact that cigarette detection depends on a manual threshold value in the prior art, and the space-time coupling characteristic of a cigarette making process is not considered so as to restrict anomaly detection precision and root cause diagnosis. In order to achieve the above object, an embodiment of the present invention provides an anomaly detection method for label noise of key indicators of cigarettes, including: Acquiring historical time sequence data of cigarettes and performing correlation analysis to acquire a correlation index data set; Preprocessing the associated index data set to obtain a training sample set; constructing a space-time dual-channel detection model for acquiring the prediction probability of the time classifier and the prediction probability of the space classifier; Training the space-time dual-channel detection model by adopting a training sample set; And detecting the cigarette sample data acquired in real time by adopting the trained space-time double-channe