An abnormal fluctuation attention time sequence channel-based quality prediction method and device

Abstract

The application discloses a quality prediction method and device based on an abnormal fluctuation attention timing channel, comprising the following steps: acquiring product production process data and preprocessing; the preprocessed data comprises non-timing data and timing data; timing features and non-timing features are extracted from the non-timing data and the timing data respectively; timing feature selection is performed based on mutual information; abnormal data is detected in the selected timing features, and the rest is normal data; an attention convergence value is taken as the mean value of the normal data, and an attention convergence operation is performed on the selected timing features; the non-timing features and the timing features after the convergence operation are fused, and the product quality probability is output through a full connection layer and a sigmod activation layer. The method improves the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of industrial artificial intelligence product quality prediction, and in particular to a quality prediction method and apparatus based on anomaly fluctuation attention time-series channel. Background Technology

[0002] For a long time, most discrete manufacturing industries have controlled product quality through random sampling. In some fields with even higher quality requirements, such as the semiconductor industry, each product needs to be inspected individually to ensure its high quality. Full inspection requires a significant amount of manpower and resources. Studies show that quality inspection accounts for 40% of labor costs. In recent years, with the development and maturation of the Industrial Internet, big data-driven quality prediction technology, due to its low cost and high accuracy, is gradually replacing inspection technology in some industries and has achieved certain results.

[0003] Predicting final product quality based on production process data is a typical time series forecasting problem. Typical methods for time series forecasting include Autoregressive Moving Average (ARMA), Autoregressive Integral Moving Average (ARIMA), exponential smoothing, Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Networks (GRU). Traditional statistical methods require assumptions about data distribution, correlation, and other characteristics, such as normality and linearity. In reality, these assumptions are not always valid, limiting the applicability of statistical methods. Furthermore, statistical methods are sensitive to outliers and missing values ​​and have limited ability to model nonlinear relationships. In contrast, machine learning and deep learning-based methods can automatically learn patterns and regularities in data without making strict assumptions about data distribution and characteristics. This makes machine learning methods more adaptable to complex and nonlinear relationships.

[0004] Among commonly used deep learning models, RNN-related networks perform best for modeling time-series data. For time-series data, the fluctuations at specific moments need to be emphasized, as this is crucial for quality prediction. While commonly used time-series feature extraction models such as LSTM and GRU can control the selection of inputs at certain moments through gating mechanisms to determine the model's focus, as a local mechanism, the gating structure means that the influence of historical information on subsequent moments may be lost, resulting in poor performance for long sequences. This paper proposes a quality prediction network based on anomaly fluctuation attention-based time-series channels. By adding a global operational attention mechanism based on feature analysis of anomalies, it improves both the contribution of indirect features from collected side-channel data and the memory performance for long sequences. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a quality prediction method and apparatus based on anomaly fluctuation attention time-series channel.

[0006] The objective of this invention is achieved through the following technical solution: a quality prediction method based on anomaly fluctuation attention time-series channels, comprising:

[0007] Acquire and preprocess product manufacturing process data; preprocessed data includes non-time-series data and time-series data.

[0008] Temporal features and non-temporal features are extracted from non-temporal data and temporal data, respectively.

[0009] Temporal feature selection based on mutual information;

[0010] If abnormal data is detected in the selected time-series features, the rest are considered normal data.

[0011] The attention convergence value is set to the mean of normal data, and attention convergence operation is performed on the selected temporal features.

[0012] The non-temporal features are fused with the temporal features after the convergence operation, and then passed through a fully connected layer and a sigmoid activation layer to output the product quality probability.

[0013] Furthermore, the product manufacturing process data shall include at least one of the following: manufacturing process parameters, the status of manufacturing equipment, and the manufacturing environment.

[0014] Furthermore, the preprocessing includes repetitive data encoding, data normalization, and discrete data continuity.

[0015] Furthermore, the temporal features in the preprocessed data are extracted using the GRU model.

[0016] Furthermore, the temporal feature selection based on mutual information specifically involves: calculating the mutual information between temporal features, setting a threshold, and if the mutual information is greater than or equal to the threshold, then only one of the relevant features is retained; mutual information I H The calculation formula is as follows:

[0017]

[0018] Where k is the nearest neighbor number; N is the number of sample points; D k Let z be the time series data point at time i. i The distance to the k-th nearest neighbor. The horizontal direction to point z i The distance is less than D k The number of points, The vertical direction to point z i The distance is less than D kThe number of points; the function Conforms to the formula:

[0019] Furthermore, the detection of outlier data in the selected time-series features specifically involves using the Tukey test method to detect outliers in the selected time-series feature sequence. This method uses upper and lower quartiles as a basis to form an outlier range; the formula is as follows:

[0020] z>Q3+k(IQR)∪z <Q1-k(IQR) (6)

[0021] IQR = (Q3 - Q1), k ≥ 0

[0022] Where z is the time series data point, Q1 is the lower quartile, Q3 is the upper quartile, IRQ is the interquartile range, and k is the coefficient.

[0023] Furthermore, the specific steps of using the mean of normal data as the attention convergence value and performing attention convergence operation on the selected temporal features are as follows:

[0024] Remove the abnormal datasets S of each time series feature obtained E Calculate the remaining time-series features of the normal dataset S c The mean value is the attention convergence value Q:

[0025] Q = Avg(S) c )#(7)

[0026] The attention scoring function S is calculated using Euclidean distance. t :

[0027]

[0028] Where X t Here, N represents the time series data input at time t, and N is the number of sample points.

[0029] Furthermore, the product quality probability is:

[0030] O pred =sigmod(O A W A +X N W N )#(10)

[0031] Among them, O pred For product quality probability, O A To achieve the desired attention outcome, W A The weights for the attention convergence results; X N As a non-temporal feature, W N Weights for non-time-series features.

[0032] The present invention also provides a quality prediction device based on anomaly fluctuation attention time-series channel, including one or more processors, for implementing the above-described quality prediction method based on anomaly fluctuation attention time-series channel.

[0033] The present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, is used to implement the above-described quality prediction method based on anomaly fluctuation attention time-series channels.

[0034] The beneficial effects of this invention are: Based on the extraction of time-series channel features, this invention takes the abnormal fluctuation moments of bypass data as the focus of multi-head attention, thus overcoming the problem of poor memory effect of traditional gating mechanisms on ultra-long sequences and passive focusing. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A schematic diagram of an AFAC-GRU model structure provided in an embodiment of the present invention;

[0037] Figure 2 A comparison chart of ROC curves;

[0038] Figure 3 This is a hardware structure diagram provided for an embodiment of the present invention. Detailed Implementation

[0039] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0040] The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0041] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0042] During normal production, the fluctuations in bypass data often follow a certain trend, but abnormal data fluctuations can occur at certain times. These abnormal fluctuations are often related to quality issues, so abnormal fluctuations in bypass data are a key feature that should be analyzed. Considering the local characteristics and context of the abnormal fluctuation period, the abnormal fluctuations in bypass data are analyzed from five aspects: the significance of abnormal features, the stability of abnormal changes, the duration of abnormalities, the linkage of abnormal changes, and the integrity of abnormal fluctuations.

[0043] The significance of anomalies refers to the magnitude of abnormal fluctuations relative to normal fluctuations. This is the essential characteristic of anomalies, representing the degree of influence of external factors on the final quality. The stability of abnormal changes refers to the frequency of abnormal fluctuations throughout the production process, representing the stability of the influence of external factors. The duration of anomalies refers to the temporal range of abnormal fluctuations, representing the persistence of the influence of external factors. Anomalous linkage changes refer to the correlation of abnormal changes in different bypass data at the same time, representing the correlation of the influence of external factors. The completeness of abnormal fluctuations refers to the completeness of abnormal changes in the actually acquired production data. Actual data collection may suffer from the loss of information at the time of fluctuations. To address this issue, this method simulates this data loss by randomly discarding some fluctuation time data during training, thus enabling the model to better handle information loss situations.

[0044] Based on the aforementioned feature analysis of bypass data, this method proposes the AFAC-GRU (Abnormal Fluctuation Attention Channel GRU) model. The overall structure of this model consists of three parts: a temporal channel feature extraction layer, an abnormal fluctuation attention convergence layer, and a multimodal feature fusion layer. The temporal feature extraction layer, based on GRU, utilizes its temporal feature extraction capabilities to obtain temporal features such as the stability and duration of abnormal changes, and embeds a Dropout layer to enhance the model's compatibility with incomplete abnormal data. The abnormal fluctuation attention convergence layer selects different feature combinations as focus points based on the abnormal linkage changes, giving more attention to moments of significant abnormal fluctuations. The feature fusion layer combines the temporal feature extraction results from the attention convergence layer with non-temporal data such as raw material features to ultimately output a prediction of product quality.

[0045] Figure 1 This invention provides a schematic diagram of an overall model structure; the quality prediction method includes the following steps:

[0046] (1) Acquire product manufacturing process data and preprocess it; the preprocessed data includes non-time series data and time series data;

[0047] The product manufacturing process data includes at least one of the following: manufacturing process parameters, manufacturing equipment status, and manufacturing environment. The preprocessing includes repetitive data encoding, data normalization, and discrete data continuation.

[0048] (2) Extract time series features from time series data;

[0049] like Figure 1 As shown, the temporal channel feature extraction layer is used to extract temporal features of the stability and duration of abnormal changes from temporal data;

[0050] RNN is a type of neural network widely used for modeling sequential data. It introduces hidden variables H... t This allows the network to capture and retain historical information about the sequence up to the current time step, giving it the ability to remember and learn the temporal characteristics of the input data. Suppose that at time step t, there is an input X... t The hidden state of the previous time step is stored in H. t-1 In the middle, use H t Let H represent the hidden state at time step t. t Input X from the current time step t The hidden state H at the previous time step t-1 Joint decision:

[0051]

[0052] The meanings of each parameter are shown in Table 1:

[0053] Table 1 Dimensions of Variables in Temporal Feature Extraction Layer

[0054]

[0055] Where n is the number of temporal features and h is the size of the hidden layer.

[0056] The introduction of hidden state structures means that the input at the current time step will influence the output at that time step to some extent, but this influence gradually decreases as the time step increases. RNN variants such as LSTM and GRU can alleviate this problem to some extent through gating mechanisms; therefore, this model uses GRU as the base model for the temporal feature extraction layer. However, the gating structure means that the influence of historical information on subsequent time steps may be lost, so GRU / LSTM still doesn't perform well in remembering very long sequences. The next step is to use attention mechanisms to consider the hidden states at all time steps from a global perspective, which can better extract temporal features from longer production sequences.

[0057] Preferably, embedding a Dropout layer makes the model more compatible with incomplete or outlier data.

[0058] (3) Select time series features based on mutual information; detect abnormal data in the selected time series features, and the rest are normal data; use the mean of normal data as the attention convergence value, and perform attention convergence operation on the selected time series features;

[0059] like Figure 1 As shown, the abnormal fluctuation attention convergence layer is used to select temporal features based on mutual information; abnormal data is detected in the selected temporal features, and the rest are normal data; the average value of the normal data is used as the attention convergence value to perform attention convergence operation on the selected temporal features.

[0060] (3.1) Selection of temporal features based on mutual information;

[0061] If a multi-head attention mechanism is directly based on all temporal features, it will result in many redundant information combinations. To remove these redundant feature combinations, the mutual information between temporal features is calculated, and a threshold is set. If the mutual information is greater than or equal to the threshold, only one of the relevant features is retained; mutual information I H The calculation formula is as follows:

[0062]

[0063] Where k is the nearest neighbor number; N is the number of sample points; D k Let z be the time series data point at time i. i The distance to the k-th nearest neighbor. The horizontal direction to point z i The distance is less than D k The number of points, The vertical direction to point z i The distance is less than D k The number of points; the function Conforms to the formula: (3.2) Detect abnormal data in the selected time-series features;

[0064] To ensure the model's applicability, the Tukey test method is used to detect outliers in each time-series feature sequence. This method is based on the upper and lower quartiles, with an added margin of error to define an outlier range. The detailed formula is as follows:

[0065] z>Q3+k(IQR)∪z <Q1-k(IQR)#(6)

[0066] IQR = (Q3 - Q1), k ≥ 0

[0067] Where z represents the time-series data point, Q1 is the lower quartile, Q3 is the upper quartile, IRQ is the interquartile range, and k is a coefficient. Based on this, the anomaly dataset S for each time-series feature is obtained. E .

[0068] (3.3) Perform attention convergence operations;

[0069] Remove the outlier datasets S obtained in step (3.2) for each time series feature. E Calculate the remaining time-series features of the normal dataset S c The mean value is the attention convergence value Q:

[0070] Q = Avg(S) c (7)

[0071] The attention scoring function S is calculated using Euclidean distance. t :

[0072]

[0073] Where X t The input time series data is at time t, and N is the number of sample points (i.e., the sequence length).

[0074] The abnormal fluctuation attention convergence layer processing procedure can be obtained based on the attention score:

[0075] O A =H t S t +b m (9)

[0076] The meanings of each parameter are shown in Table 2.

[0077] Table 2: Variable Dimensions of Multi-Head Attention Convergence Layer

[0078]

[0079] Where m is the number of heads in the multi-head attention mechanism, and h is the size of the hidden layer.

[0080] (4) Obtain non-time-series features from non-time-series data, fuse the non-time-series features with the time-series features after the aggregation operation, and output the product quality probability through a fully connected layer and a sigmoid activation layer.

[0081] like Figure 1 As shown, the multimodal feature fusion layer is used to fuse non-temporal features with temporal features after convergence. After passing through a fully connected layer and a sigmoid activation layer, it outputs the product quality probability. The calculation process is as follows:

[0082] O pred =sigmod(O A W A +X N W N )#(10)

[0083] Among them Opred For product quality probability, O A To achieve the desired attention outcome, W A The weights for the attention convergence results; X N As a non-temporal feature, W N Weights for non-time-series features.

[0084] Table 3 Summary of Information Prediction Layer Variable Dimensions

[0085]

[0086] Where k is the number of non-time-series features.

[0087] Example 1:

[0088] 3.1 Dataset and Preprocessing

[0089] The dataset for this experiment comes from a cable manufacturing company whose main products are silver-plated monofilaments and steel strands. Steel strands refer to reinforced conductors consisting of single or multiple layers of aluminum strands twisted around a galvanized steel core. They are mainly used in the power and transmission line industries, and their quality directly affects the safety of power transmission. Many factors influence the quality of steel strands, such as the matching degree between the monofilament diameter and the conduit, the pitch ratio, and the tension. The experiment selected 3009 production process data points. Based on the factors affecting product quality, two non-time-series features (copper type, wire diameter) were selected from raw material information, and four time-series features (drawing speed, rotation speed, tension, machine speed difference) and one non-time-series feature (wire diameter difference) were selected from equipment status and process parameters. The overall results are shown in Table 4.

[0090] Table 4: Summary of Predicted Features

[0091]

[0092] Data preprocessing mainly includes encoding repetitive data, data normalization, and converting discrete data to continuous data. Under normal production conditions, the process data contains a large amount of repetition, increasing training time. Therefore, data with the same sequence are merged, and a duration field is added. Data normalization ensures that different features use the same units, reducing the influence of outliers and accelerating gradient descent. In the dataset, both wire drawing speed and tension are set to 10. 2 The magnitude is on the order of magnitude, while the rotational speed reaches 10. 3To avoid the excessive influence of rotational speed variations on the prediction results, normalization is required. Here, the z-score standardization method is used to unify the dimensions of each feature. To input discrete variables (such as copper material type and quality) into the network, they need to be transformed into continuous variables. To avoid a surge in input dimensionality after encoding, the LeaveOneOutEncoder method is used to achieve continuous transformation.

[0093] 3.2 Attention Feature Selection Based on Mutual Information

[0094] The fluctuations in different combinations of side-channel data features contain different information, and the introduction of multi-head attention mechanisms is precisely to make full use of this information. However, due to the multi-feature nature of industrial big data, direct combination will result in a large number of feature combinations, increasing the complexity of the model. This experiment calculates the correlation between features using mutual information and selects only features with weak correlation to reduce the number of side-channel feature combinations. Since mutual information is not a metric and cannot be normalized, the results on different datasets cannot be compared. Therefore, this experiment chose the maximum information coefficient (MIC) based on mutual information to measure the correlation between side-channel features. The specific implementation uses the MINE class in the minepy library, and the MICs between the four time-series features are shown in Table 5.

[0095] Table 5 Maximum information coefficients among time-series features

[0096]

[0097] In this implementation, a threshold of 0.9 is set. If the value is greater than or equal to the threshold, the two features are considered to be strongly correlated; otherwise, they are considered to be weakly correlated. Analysis of Table 5 shows that the rotational speed S... r and wire drawing speed S d The correlation is too high, so only one of the two should be selected as the feature of interest. Tension F and rotational speed S r And wire drawing speed S d While there is some correlation, it is not a strong correlation, and both can be selected simultaneously. The duration T and other features have relatively low correlations, so they can be selected. In summary, the rotational speed S is selected. r Tension F and duration T are the attention feature set of the attention mechanism. The feature set is decomposed to obtain [S] r ]、[S r 、F、[S r [T], [S] r There are 7 attention points, namely the 7-head attention mechanism.

[0098] 3.3 Outlier Detection in Bypass Data

[0099] In the previous section, we determined the temporal features that the attention mechanism should focus on by calculating the maximum information coefficient. Next, in order to obtain the query value corresponding to the attention mechanism (i.e., the average value of the features under normal conditions), we need to first detect abnormal moments in the time-series data under attention.

[0100] Statistical anomaly detection methods can be broadly categorized into two types based on whether they conform to a normal distribution: one type uses methods that follow a normal distribution, including the 3σ criterion and Dixon test; the other type uses methods that do not follow a normal distribution, including the ESD test and Tukey test.

[0101] In order to determine whether the bypass data conforms to a normal distribution and to select the correct anomaly detection method, the KS method is used to determine the data distribution. The Kolmogorov-Smirnov test

[18] is a commonly used method to determine whether the frequency distribution and the theoretical distribution are consistent. It compares the maximum absolute difference D between the cumulative frequency of the sample and the cumulative probability of the theoretical distribution. n =max(|f(x)-g(x)|) and critical value D na The size is used to determine the result. Here, the kstest function from the scipy toolkit is used directly for testing, and the results are shown in Table 6.

[0102] Table 6. Results of KS test for key features

[0103]

[0104] The test results show that all features of interest should be detected using anomaly detection methods based on the non-normal distribution assumption. This experiment used the Tukey test method. By defining the normal range of time-series data within the product lifecycle using the Tukey test method, the abnormal moments in the data can be identified.

[0105] 3.4 Analysis of Experimental Results

[0106] In neural networks, parameter updates are typically performed using gradient descent. The magnitude of these updates depends on the chosen loss function and learning rate. Ideally, the magnitude of the parameter updates should be smaller as it approaches the target value; therefore, this experiment chose cross-entropy as the loss function. To fully utilize the data in the dataset, a 10-fold cross-validation method was used to divide the dataset into training and testing sets. 10-fold cross-validation involves dividing the data into 10 equal parts, selecting 9 parts as the training set and 1 part as the testing set for each training iteration, allowing for 10 training and testing iterations. To maintain consistency in the data distribution between the training and testing sets, the data was proportionally obtained from both training and testing datasets during the partitioning process. Furthermore, to reduce discrepancies caused by sample partitioning, the dataset was randomly partitioned 10 times, and finally, 10 iterations of 10-fold cross-validation were performed.

[0107] This experiment selected RNN and its gated variant GRU for comparison. The performance was directly compared in five aspects: accuracy, precision, recall, F1 score, and AUC (Area Under ROC Curve). The results are shown in Table 7, and the ROC curves are shown below. Figure 2 As shown.

[0108] Table 7. Performance Comparison Results

[0109]

[0110] From basic RNNs to GRUs with added gating structures, and then to AFAC-GRUs with added anomaly attention channels, improvements in both accuracy and generalization performance can be observed. For GRUs, the gating structure passively increases attention at certain times, further improving prediction accuracy. However, for bypass data collected due to cost factors, the indirect data itself contributes little to the final quality, limiting the effectiveness of passive attention. Therefore, analyzing various characteristics of anomalies and adding active attention based on the specific anomalies is more effective.

[0111] Corresponding to the aforementioned embodiment of a quality prediction method based on anomaly fluctuation attention time-series channel, the present invention also provides an embodiment of a quality prediction device based on anomaly fluctuation attention time-series channel.

[0112] See Figure 3 The present invention provides a quality prediction device based on anomaly fluctuation attention time-series channel, comprising one or more processors, for implementing a quality prediction method based on anomaly fluctuation attention time-series channel as described in the above embodiments.

[0113] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0114] An embodiment of the quality prediction device based on anomaly fluctuation attention time-series channel of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 3 The diagram shown is a hardware structure diagram of any data processing-capable device, including the quality prediction device based on anomaly fluctuation attention temporal channel according to the present invention. Except for... Figure 3 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0115] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0116] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0117] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a quality prediction method based on an abnormal fluctuation attention time-series channel as described in the above embodiments.

[0118] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0119] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A quality prediction method based on anomaly fluctuation attention time-series channels, characterized in that, include: Acquire and preprocess product manufacturing process data; Preprocessed data includes both non-time-series and time-series data; Temporal features and non-temporal features are extracted from non-temporal data and temporal data, respectively. Temporal feature selection based on mutual information; Specifically, the process involves calculating the mutual information between time-series features, setting a threshold, and retaining only one of the relevant features if the mutual information is greater than or equal to the threshold. The calculation formula is as follows: in The nearest neighbor number; The number of sample points; For the first Time series data points at each moment To the The distance of a neighbor, To point in the horizontal direction The distance is less than The number of points, The vertical direction to the point The distance is less than The number of points; the function Conforms to the formula: ; If outliers are detected in the selected time-series features, the rest are considered normal data. Specifically, the Tukey test method is used to detect outliers in the selected time-series feature sequences. This method uses upper and lower quartiles as a basis to form an outlier range; the formula is as follows: in For time series data points, The lower quartile, The upper quartile, Interquartile range, For coefficients; The attention convergence value is set to the mean of normal data, and attention convergence operation is performed on the selected temporal features. The non-temporal features are fused with the temporal features after the convergence operation, and then passed through a fully connected layer and a sigmoid activation layer to output the product quality probability.

2. The quality prediction method based on anomaly fluctuation attention time-series channel according to claim 1, characterized in that, The product manufacturing process data shall include at least one of the following: manufacturing process parameters, status of manufacturing equipment, and manufacturing environment.

3. The quality prediction method based on anomaly fluctuation attention time-series channel according to claim 1, characterized in that, The preprocessing includes repetitive data encoding, data normalization, and discrete data continuity.

4. The quality prediction method based on anomaly fluctuation attention time-series channel according to claim 1, characterized in that, Temporal features are extracted from preprocessed data using the GRU model.

5. The quality prediction method based on anomaly fluctuation attention time-series channel according to claim 1, characterized in that, The specific steps of performing attention convergence operation on the selected temporal features using the mean of normal data as the attention convergence value are as follows: Remove the abnormal datasets of each time series feature Calculate the remaining time-series features of the normal dataset. The mean is the attention convergence value. : Calculate the attention scoring function using Euclidean distance. : in The time series data input at time t, This represents the number of sample points.

6. The quality prediction method based on anomaly fluctuation attention time-series channel according to claim 1, characterized in that, The probability of product quality: in, For product quality probability, To focus attention on results Weights for the attention convergence results; It is a non-temporal feature. Weights for non-time-series features.

7. A quality prediction device based on anomaly fluctuation attention time-series channel, characterized in that, It includes one or more processors for implementing the quality prediction method based on the abnormal fluctuation attention time-series channel according to any one of claims 1-6.

8. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program is used to implement the quality prediction method based on the abnormal fluctuation attention time-series channel as described in any one of claims 1-6.

Images