Printed circuit board function test item yield prediction method based on improved PatchTST model

Abstract

The invention relates to the technical field of function test item yield prediction, and discloses a printed circuit board function test item yield prediction method based on an improved PatchTST model, and the method comprises the steps: collecting the test item result data of a printed circuit board, and calculating the corresponding test item yield data; the method comprises the following steps of: decomposing test item yield data into a trend item and a residual item through moving average, respectively performing multi-scale time slice division, performing feature extraction on each divided branch time slice sequence through a Transform encoder, realizing intra-branch feature aggregation and inter-branch fusion through a dynamic weighting mechanism, and outputting prediction results of the trend item and the residual item. And linear fusion is carried out to generate a final test item yield prediction value. According to the invention, through time slice division of different scales, local fluctuation detail capture and long-term trend modeling are realized; and a weighted fusion module is utilized to balance trend and residual prediction results, so that the accuracy and reliability of model prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of functional test item yield prediction, and in particular to a printed circuit board functional test item yield prediction method based on an improved PatchTST model. Background Art

[0002] As an important indicator for measuring specific test items of printed circuit boards during the production process, the test yield rate plays an important role in optimizing the production process and reducing costs.

[0003] However, test yield data has strong trends, weak periodicity, and is prone to sudden changes. Existing prediction models have many shortcomings when dealing with these characteristics of test yield data.

[0004] Traditional time series models, such as the Autoregressive Integrated Moving Average (ARIMA) model, are based on the assumption of data stationarity. However, the presence of mutation points in test yield data makes it a non-stationary time series. Although data non-stationarity can be mitigated to a certain extent through methods such as differencing, this approach inevitably leads to the loss of key trend information, affecting the accuracy of predictions.

[0005] Even advanced deep learning-based models have limitations when processing test yield data. Deep learning models typically rely on large amounts of data for training to learn complex patterns within the data. However, local features near mutation points in test yield data are easily obscured by global trends, making it difficult for deep learning models to accurately capture and learn these key features. Furthermore, most existing deep learning models fail to fully account for variations at different time scales. For example, when capturing short-term mutations, they may overlook long-term trends. Conversely, when focusing on long-term trends, they struggle to respond promptly to subtle short-term fluctuations, limiting prediction accuracy. Some time series prediction models that use multi-scale modeling, such as HDMixer, are primarily targeted at multivariate time series prediction and are not suitable for univariate data like test yield. While Pathformer offers certain advantages in capturing multi-scale features, its training time is prohibitively long. This inefficiency makes it inefficient in scenarios requiring frequent retraining based on updated yield data, making it unsuitable for actual production needs. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a printed circuit board functional test item yield prediction method based on an improved PatchTST model.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] A printed circuit board functional test item yield prediction method based on an improved PatchTST model, comprising:

[0009] Collect printed circuit board test result data and calculate the corresponding test yield data;

[0010] The test item yield data is decomposed into a trend term and a residual term through moving average. The trend term is obtained by moving average calculation with a window size of N, and the residual term is obtained by subtracting the trend term from the test item yield data.

[0011] The trend term and residual term are divided into multi-scale time segments. The time segment sequences of each branch are extracted with the Transformer encoder, and the dynamic weighting mechanism is used to achieve intra-branch feature aggregation and inter-branch fusion. The prediction results of the trend term and residual term are output respectively. The Transformer encoder includes a multi-head attention mechanism, layer normalization and a feedforward network.

[0012] The prediction results of the trend term and the residual term are linearly fused according to the preset weighting coefficient to generate the final test item yield prediction value.

[0013] In one embodiment, collecting printed circuit board test result data and calculating corresponding test yield data includes:

[0014] Collect the result data of each test item in the production test process of printed circuit boards;

[0015] Calculate the yield rate of each test item based on the test result data. For a specific test item, the yield rate is calculated by dividing the number of tests passed for that test item by the total number of tests. Organize the test item yield rates into a time series in chronological order.

[0016] Normalize the time series of test item yield and map the data into a unified scale range.

[0017] In one embodiment, the method of decomposing the test item yield data into a trend term and a residual term by moving average, wherein the trend term is obtained by moving average calculation with a window size of N, and the residual term is obtained by subtracting the trend term from the test item yield data, specifically includes:

[0018] Determine the moving average window size N based on the fluctuation and trend changes of the data;

[0019] Calculate the moving average of the test item yield data according to the window size N as the trend item T t : Yield t-i Indicates the test item yield at time ti;

[0020] The residual term R is obtained by subtracting the trend term from the yield data of the test item Yield. t :R t =Yield-T t .

[0021] In one embodiment, dividing the trend term and the residual term into multi-scale time segments specifically includes:

[0022] Set multiple different time segment lengths and divide the trend term and residual term into multi-scale time segments respectively; let the time segment length of the kth branch be P (k) , divide the trend term and residual term into non-overlapping time segment series x p :

[0023]

[0024] in, N (k) is the number of time segments in the time segment sequence of the kth branch, and L is the length of the time segment sequence.

[0025] In one embodiment, the divided branch time segment sequences are subjected to a Transformer encoder to extract features, specifically including:

[0026] Through the trainable linear projection W p The time segment sequence x p Mapped to the Transformer latent space of dimension D, and applying a learnable additive positional encoding W pos , get the encoding result x d :x d =W p x p +W pos ;

[0027] Each head h in the multi-head attention mechanism will encode the result x d Perform linear transformation respectively and transform into query matrix Q h , key matrix K h Sum matrix V h : They are used to generate Q h , K h and V h The trainable weight matrix based on Q h , K h and V h Calculate the attention output O h ;

[0028] Output O to the attention hAfter layer normalization, the input is fed into a feedforward network with residual connections, and the output of the feedforward network is FFN(O h ):FFN(O h )=max(0,O h W1+b1)W1+b2; where W1 and W2 are the weight matrices of the linear layer, b1 and b2 are bias vectors, and max(0,·) is the ReLU activation function; the output of the feedforward network is the feature extracted by the Transformer encoder for the time segment sequence of the kth branch.

[0029] In one embodiment, the Q-based h , K h and V h Calculate the attention output O h , specifically including:

[0030]

[0031] d k is the dimension of the key vector; Softmax is the Softmax function.

[0032] In one embodiment, the dynamic weighting mechanism is used to achieve intra-branch feature aggregation and inter-branch fusion, and output prediction results of trend terms and residual terms respectively, specifically including:

[0033] Through the learnable weight matrix W intra To calculate the importance score of each feature in the current branch:

[0034]

[0035] is the i-th feature vector extracted by the Transformer encoder for the time segment sequence of the k-th branch, for Importance score;

[0036] Use the Softmax function to normalize the importance score and obtain the dynamic fusion coefficient of the i-th feature vector in the k-th branch

[0037] Using dynamic fusion coefficient Perform weighted summation on the feature vectors within the branch to obtain the aggregate feature h of the kth branch (k) :

[0038]

[0039] Through the learnable weight matrix W interTo calculate the importance score t of the aggregated features of the kth branch (k) :

[0040] t (k) =h (k) W inter ;

[0041] Use the Softmax function to aggregate the importance score t of the features (k) Perform normalization processing to obtain the branch dynamic fusion coefficient β of the kth branch (k) ;

[0042] Using the branch dynamic fusion coefficient β (k) , perform weighted summation on the aggregated features of all branches to obtain the fusion feature z fused ∈R D :

[0043] The fused feature z is transformed into fused Map to the prediction space to obtain the prediction results of the trend term or residual term The weight matrix of the linear projection layer is W p , the bias vector is b p .

[0044] In one embodiment, the importance score is normalized using the Softmax function to obtain the dynamic fusion coefficient of the i-th feature vector in the k-th branch. Specifically include:

[0045]

[0046] N (k) is the number of time segments in the time segment sequence of the kth branch.

[0047] In one embodiment, the importance score t of the aggregated features is calculated using the Softmax function. (k) Perform normalization processing to obtain the branch dynamic fusion coefficient β of the kth branch (k) , specifically including:

[0048]

[0049] K represents the total number of branches in a time segment.

[0050] In one embodiment, the linear fusion of the prediction results of the trend term and the residual term according to the preset weighting coefficient to generate the final test item yield prediction value specifically includes:

[0051]

[0052] Y p is the predicted value of the test item yield, α and β are the predicted results of the trend item respectively And the prediction results of the residual term The weighting coefficient of .

[0053] Compared with the prior art, the beneficial technical effects of the present invention are:

[0054] This paper, based on an in-depth analysis of the strong trend, weak periodicity, and susceptibility to mutation in test item yield data, designs a printed circuit board functional test item yield prediction model based on an improved PatchTST. The model's moving average-based yield decomposition module decouples yield data into trend and residual terms, effectively reducing the interference of mutations on prediction and highlighting the trend term as the primary component. An improved PatchTST module, based on a multi-scale patch partitioning strategy, captures local fluctuation details and models long-term trends through patch partitioning at different scales. A weighted fusion module balances trend and residual prediction results, thereby improving the accuracy and reliability of the model's predictions. This proposed method is expected to provide a more effective solution for test item yield prediction, meeting the practical needs of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 4 is a flow chart of a method in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] A preferred embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0057] like Figure 1 As shown, the present invention provides a printed circuit board functional test item yield prediction method based on the improved PatchTST, comprising the following steps:

[0058] S1, collect printed circuit board test result data and calculate the corresponding test item yield data;

[0059] S2, decompose the test item yield data into a trend term and a residual term through moving average, where the trend term is obtained by moving average calculation with a window size of N, and the residual term is obtained by subtracting the trend term from the test item yield data;

[0060] S3, divides the trend term and residual term into multi-scale time segments, extracts features from each branch time segment sequence after division, and realizes intra-branch feature aggregation and inter-branch fusion through a dynamic weighting mechanism, outputting the prediction results of the trend term and residual term respectively; the Transformer encoder includes a multi-head attention mechanism, layer normalization and a feedforward network;

[0061] S4, linearly fuse the prediction results of the trend term and the residual term according to the preset weighting coefficient to generate the final test item yield prediction value.

[0062] In one embodiment, collecting printed circuit board test result data and calculating corresponding test yield data in step S1 includes:

[0063] Collect the result data of each test item in the production test process of printed circuit boards;

[0064] Calculate the yield rate of each test item based on the test result data. For a specific test item, the yield rate is calculated by dividing the number of tests passed for that test item by the total number of tests. Organize the test item yield rates into a time series in chronological order.

[0065] Normalize the time series of test item yield and map the data into a unified scale range.

[0066] Specifically, the test result data should cover the test conditions of each test item in different production batches and different time periods, including whether the test passed, test time, test equipment and other information, to ensure the comprehensiveness and accuracy of the data and provide a reliable basis for subsequent yield calculations; common normalization methods such as minimum-maximum normalization or Z-score normalization can be used for normalization processing to map the data to a unified scale range, thereby improving the effect and stability of model training.

[0067] In one embodiment, step S2 decomposes the test item yield data into a trend term and a residual term by moving average, wherein the trend term is obtained by moving average calculation with a window size of N, and the residual term is obtained by subtracting the trend term from the test item yield data, specifically including:

[0068] Determine the moving average window size N based on the fluctuation and trend changes of the data;

[0069] Calculate the moving average of the test item yield data according to the window size N as the trend item T t : Yield t-i Indicates the test item yield at time ti;

[0070] The residual term R is obtained by subtracting the trend term from the yield data of the test item Yield. t :R t =Yield-T t .

[0071] Specifically, the selection of the moving average window size N requires comprehensive consideration of the data fluctuations and trend changes, and can be determined through experiments or experience. Generally speaking, the window size should include the current point and a certain number of previous historical points to accurately capture the long-term trend of the data; decomposing the test item yield data into trend terms and residual terms can separate long-term trends from mutation characteristics.

[0072] In one embodiment, the step S3 of dividing the trend term and the residual term into multi-scale time segments specifically includes:

[0073] Set multiple different time segment lengths and divide the trend term and residual term into multi-scale time segments (Patch) respectively; let the time segment length of the kth branch be P (k) , divide the trend term and residual term into non-overlapping time segment series x p :

[0074]

[0075] in, N (k) is the number of time segments in the time segment sequence of the kth branch, and L is the length of the time segment sequence.

[0076] In the present invention, smaller-scale patches are used to capture subtle fluctuations in data, while larger-scale patches are used to grasp long-term trends.

[0077] In one embodiment, the features of each branch time segment sequence after the division in step S3 are extracted by a Transformer encoder, specifically including:

[0078] Through the trainable linear projection W p The time segment sequence x p Mapped to the Transformer latent space of dimension D, and applying a learnable additive positional encoding W pos , so that the model can perceive the temporal order information of the Patch and obtain the encoding result x d :x d =W p x p +W pos ;

[0079] Each head h=1,2,…,H in the multi-head attention mechanism will encode the result x d Perform linear transformation respectively and transform into query matrix Q h , key matrix K h Sum matrix V h : They are used to generate Q h , K h and V h The trainable weight matrix based on Q h , K h and V h Calculate the attention output O h ;

[0080] Output O to the attention h After layer normalization, the input is fed into a feedforward network with residual connections, and the output of the feedforward network is FFN(O h ):FFN(O h )=max(0,O h W1+b1)W2+b2; where W1 and W2 are the weight matrices of the linear layer, b1 and b2 are bias vectors, and max(0,·) is the ReLU activation function; the output of the feedforward network is the feature extracted by the Transformer encoder for the time segment sequence of the kth branch.

[0081] The role of layer normalization is to normalize each sample in the feature dimension so that it has a stable mean and variance, which helps the training and convergence of the model. After layer normalization, a feedforward network with residual connections is connected. The feedforward network usually consists of two linear layers and an activation function, which can further transform and extract features from the data.

[0082] In one embodiment, the Q-based h , K h and V h Calculate the attention output O h , specifically including:

[0083]

[0084] d k is the dimension of the key vector; Softmax is the Softmax function. The Softmax function is used here to normalize the attention scores so that the sum of all attention scores is 1, thereby obtaining the attention weight of each position and then adding it to the value matrix V h Multiply them together to get the attention output of each head.

[0085] In one embodiment, step S3 implements intra-branch feature aggregation and inter-branch fusion through a dynamic weighting mechanism, and outputs prediction results of trend terms and residual terms respectively, specifically including:

[0086] In order to determine the importance of each feature in the branch, a learnable weight matrix W is used intra To calculate the importance score of each feature in the current branch:

[0087]

[0088] is the i-th feature vector extracted by the Transformer encoder for the time segment sequence of the k-th branch, for The importance score of reflects the relative importance of the feature vector in the current branch;

[0089] In order to convert the feature importance score into a coefficient for feature aggregation, the importance score is normalized using the Softmax function to obtain the dynamic fusion coefficient of the i-th feature vector in the k-th branch. Dynamic fusion coefficient satisfy It represents the weight of each feature vector when it is aggregated within a branch.

[0090] Using dynamic fusion coefficient Perform weighted summation on the feature vectors within the branch to obtain the aggregate feature h of the kth branch (k) :

[0091]

[0092] Get the aggregated feature h of each branch (k) After that, k=1,…,K, K is the total number of branches, and these branch features need to be fused to obtain the final multi-scale feature representation; in order to determine the importance of each branch in the whole, the learnable weight matrix W inter To calculate the importance score t of the aggregated features of the kth branch (k) :

[0093] t (k) =h (k) W inter ;

[0094] The score here is t (k) It reflects the relative importance of this branch among all branches.

[0095] Use the Softmax function to aggregate the importance score t of the features (k) Perform normalization processing to obtain the branch dynamic fusion coefficient β of the kth branch (k) ; Branch dynamic fusion coefficient β (k) satisfy It represents the weight of each branch when fusion occurs between branches.

[0096] Using the branch dynamic fusion coefficient β (k), perform weighted summation on the aggregated features of all branches to obtain the fusion feature z fused ∈R D :

[0097] The fused feature z is transformed into fused Map to the prediction space to obtain the prediction results of the trend term or residual term The weight matrix of the linear projection layer is W p , the bias vector is b p .

[0098] In one embodiment, the importance score is normalized using the Softmax function to obtain the dynamic fusion coefficient of the i-th feature vector in the k-th branch. Specifically include:

[0099]

[0100] In one embodiment, the importance score t of the aggregated features is calculated using the Softmax function. (k) Perform normalization processing to obtain the branch dynamic fusion coefficient β of the kth branch (k) , specifically including:

[0101]

[0102] K represents the total number of branches in a time segment.

[0103] In one embodiment, the linear fusion of the prediction results of the trend term and the residual term according to the preset weighting coefficient in step S4 to generate the final test item yield prediction value specifically includes:

[0104]

[0105] Y p is the predicted value of the test item yield, α and β are the predicted results of the trend item respectively And the prediction results of the residual term The weighting coefficient of .

[0106] Specifically, the present invention determines the weighted coefficients α and β of the trend term and residual term prediction results based on the training set data through experiments or optimization algorithms. These coefficients are used to balance the contribution of the trend term and residual term in the final prediction results, making the prediction results more accurate and reliable.

[0107] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times. The order of execution of these steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0108] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. It is intended that all variations within the meaning and range of equivalents of the claims be embraced herein, and any reference signs in the claims should not be construed as limiting the claims to which they relate.

[0110] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A printed circuit board functional test item yield prediction method based on an improved PatchTST model, characterized in that: include: Collect printed circuit board test result data and calculate the corresponding test yield data; The test item yield data is decomposed into a trend term and a residual term through moving average. The trend term is obtained by moving average calculation with a window size of N, and the residual term is obtained by subtracting the trend term from the test item yield data. The trend term and residual term are divided into multi-scale time segments. The time segment sequences of each branch are extracted with the Transformer encoder, and the dynamic weighting mechanism is used to achieve intra-branch feature aggregation and inter-branch fusion. The prediction results of the trend term and residual term are output respectively. The Transformer encoder includes a multi-head attention mechanism, layer normalization and a feedforward network. The prediction results of the trend term and the residual term are linearly fused according to the preset weighting coefficient to generate the final test item yield prediction value.

2. The method for predicting the yield of printed circuit board functional test items based on the improved PatchTST model according to claim 1, characterized in that: The collecting of printed circuit board test result data and calculating corresponding test yield data includes: Collect the result data of each test item in the production test process of printed circuit boards; Calculate the yield rate of each test item based on the test result data. For a specific test item, the yield rate is calculated by dividing the number of tests passed for that test item by the total number of tests. Organize the test item yield rates into a time series in chronological order. Normalize the time series of test item yield and map the data into a unified scale range.

3. The method for predicting the yield of printed circuit board functional test items based on the improved PatchTST model according to claim 1, characterized in that: The method of decomposing the test item yield data into a trend term and a residual term by moving average, wherein the trend term is obtained by moving average calculation with a window size of N, and the residual term is obtained by subtracting the trend term from the test item yield data, specifically includes: Determine the moving average window size N based on the fluctuation and trend changes of the data; Calculate the moving average of the test item yield data according to the window size N as the trend item T t : Yield t-i Indicates the test item yield at time ti; The residual term R is obtained by subtracting the trend term from the yield data of the test item Yield. t :R t =Yield-T t .

4. The method for predicting the yield of printed circuit board functional test items based on the improved PatchTST model according to claim 1, characterized in that: The multi-scale time segment division of the trend term and the residual term specifically includes: Set multiple different time segment lengths and divide the trend term and residual term into multi-scale time segments respectively; let the time segment length of the kth branch be P (k) , divide the trend term and residual term into non-overlapping time segment series x p : in, N (k) is the number of time segments in the time segment sequence of the kth branch, and L is the length of the time segment sequence.

5. The method for predicting yield of printed circuit board functional test items based on the improved PatchTST model according to claim 1, characterized in that: The features of each branch time segment sequence after the division are extracted by the Transformer encoder, specifically including: Through the trainable linear projection W p The time segment sequence x p Mapped to the Transformer latent space of dimension D, and applying a learnable additive positional encoding W pos , get the encoding result x d :x d =W p x p +W pos ; Each head h in the multi-head attention mechanism will encode the result x d Perform linear transformation respectively and transform into query matrix Q h , key matrix K h Sum matrix V h : They are used to generate Q j , K h and V h The trainable weight matrix based on Q h , K h and V h Calculate the attention output O h ; Output O to the attention h After layer normalization, the input is fed into a feedforward network with residual connections, and the output of the feedforward network is FFN(O h ):FFN(O h )=max(0,O h W1+b1)W2+b2; where W1 and W2 are the weight matrices of the linear layer, b1 and b2 are bias vectors, and max(0,·) is the ReLU activation function; the output of the feedforward network is the feature extracted by the Transformer encoder for the time segment sequence of the kth branch.

6. The method for predicting the yield of printed circuit board functional test items based on the improved PatchTST model according to claim 5, characterized in that: The Q-based h , K h and V h Calculate the attention output O h , specifically including: d k is the dimension of the key vector; Softmax is the Softmax function.

7. The method for predicting the yield of printed circuit board functional test items based on the improved PatchTST model according to claim 1, characterized in that: The dynamic weighting mechanism is used to achieve intra-branch feature aggregation and inter-branch fusion, and output the prediction results of the trend term and the residual term respectively, specifically including: Through the learnable weight matrix W intra To calculate the importance score of each feature in the current branch: is the i-th feature vector extracted by the Transformer encoder for the time segment sequence of the k-th branch, for Importance score; Use the Softmax function to normalize the importance score and obtain the dynamic fusion coefficient of the i-th feature vector in the k-th branch Using dynamic fusion coefficient Perform weighted summation on the feature vectors within the branch to obtain the aggregate feature h of the kth branch (k) : Through the learnable weight matrix W inter To calculate the importance score t of the aggregated features of the kth branch (k) : t (k) =h (k) W inter ; Use the Softmax function to aggregate the importance score t of the features (k) Perform normalization processing to obtain the branch dynamic fusion coefficient β of the kth branch (k) ; Using the branch dynamic fusion coefficient β (k) , perform weighted summation on the aggregated features of all branches to obtain the fusion feature z fused ∈R D : The fused feature z is transformed into fused Map to the prediction space to obtain the prediction results of the trend term or residual term The weight matrix of the linear projection layer is W p , the bias vector is b p .

8. The method for predicting yield of printed circuit board functional test items based on the improved PatchTST model according to claim 7, characterized in that: The importance score is normalized using the Softmax function to obtain the dynamic fusion coefficient of the i-th feature vector in the k-th branch. Specifically include: N (k) is the number of time segments in the time segment sequence of the kth branch.

9. The method for predicting the yield of printed circuit board functional test items based on the improved PatchTST model according to claim 7, characterized in that: The importance score t of the aggregated feature is calculated using the Softmax function. (k) Perform normalization processing to obtain the branch dynamic fusion coefficient β of the kth branch (k) , specifically including: K represents the total number of branches in a time segment.

10. The method for predicting the yield of printed circuit board functional test items based on the improved PatchTST model according to claim 1, characterized in that: The linear fusion of the prediction results of the trend term and the residual term according to the preset weighting coefficient to generate the final test item yield prediction value specifically includes: Y p is the predicted value of the test item yield, α and β are the predicted results of the trend item respectively And the prediction results of the residual term The weighting coefficient of .

Images