Quality prediction method and device for multi-stage manufacturing system, and medium

Abstract

The invention relates to a quality prediction method and device for a multi-stage manufacturing system, and a medium, and the method comprises the following steps: obtaining an initial quality index value of a raw material and a device state parameter of each manufacturing stage, carrying out the preprocessing, carrying out the splicing through a broadcasting mechanism, inputting a trained quality prediction model for the multi-stage manufacturing system, and obtaining a quality prediction result. Obtaining a quality prediction result of each manufacturing stage; the model comprises a total attention module, a feedforward neural network, and an encoder and a decoder which are constructed based on a bidirectional long short-term memory network, and the encoder is used for generating an output vector of each manufacturing stage according to input data; the full attention module is used for calculating the attention weight of the output vector according to the hidden vector of the decoder; and the decoder is used for decoding the weighted output vector, and a decoding result is output as a quality sequence through the feedforward neural network. Compared with the prior art, the accuracy of quality prediction in the complex multi-stage manufacturing process is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial production quality control, and in particular relates to a quality prediction method, device and medium for a multi-stage manufacturing system. Background Art

[0002] Multistage Manufacturing Systems (MMS) are common systems in modern production processes. They consist of multiple continuous production stages. Products are processed or assembled in each stage in turn until the entire manufacturing process is completed. The production of complex products such as aviation panel manufacturing and automobile assembly are typical MMS examples. Such products often have varying degrees of quality deviations in the unit assembly and system assembly stages. The deviations in each stage will accumulate in the subsequent processing process and jointly affect the quality of the final product. In MMS, various factors such as the equipment status of the intermediate process and the quality of the initial raw materials have a significant impact on the quality of the final product. By predicting quality in advance, producers can adjust production plans in a timely manner to optimize the manufacturing process. Therefore, it is of great significance to study the quality transfer mechanism and its prediction method in MMS.

[0003] The Chinese invention patent with publication number CN118552092A proposes a multi-process quality prediction method based on the attention mechanism. The training set is used to train the feature encoding network based on the attention mechanism for multi-process quality prediction to obtain a quality prediction model, which realizes the acquisition of the mapping relationship between process parameters and quality indicators, provides support for the intelligent optimization of process parameters, and has good practicality. The Chinese patent application with publication number CN116757545A proposes a multi-stage manufacturing system quality prediction method based on multi-task deep learning, "constructing a multi-scale convolutional network with a control gate network based on the extraction of the transmission influence features of each stage, while considering the transmission influence of each stage, the prediction results of multiple quality indicators in each stage are output in parallel, the influence transmission of each stage is effective, and the results of the parallel output of multiple quality indicators are highly accurate. However, the above methods ignore the cumulative and interactive effects of different stages on the quality of the final product, resulting in a waste of resources and time, and most quality prediction models need to input all process deviations, which is difficult to meet the accuracy and generalization performance of parameter estimation in complex production scenarios. Therefore, it is necessary to redesign a quality prediction method for multi-stage manufacturing systems to further improve the accuracy of quality prediction in complex multi-stage manufacturing processes. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a quality prediction method, equipment and medium for multi-stage manufacturing systems with good generalization performance, taking into account the cumulative and interactive effects of different stages on the quality of the final product, and further improving the accuracy of quality prediction in complex multi-stage manufacturing processes.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] The present invention provides a quality prediction method for a multi-stage manufacturing system, comprising the following steps:

[0007] Obtain the initial quality index values ​​of raw materials and the equipment status parameters of each manufacturing stage, perform preprocessing and splice them through the broadcast mechanism, and then input the trained quality prediction model for multi-stage manufacturing systems to obtain the quality prediction results of each manufacturing stage;

[0008] Among them, the quality prediction model for multi-stage manufacturing systems includes an encoder, a full attention module, a decoder and a feedforward neural network connected in sequence. The encoder is constructed based on a bidirectional long short-term memory network, and is used to generate an output vector for each manufacturing stage according to input data; the full attention module is used to calculate the attention weight of the output vector of each manufacturing stage according to the hidden vector of the decoder, and input the weighted output vector into the decoder; the decoder is constructed based on a bidirectional long short-term memory network, and is used to decode the weighted output vector. The decoding results are integrated through the feedforward neural network and output as a quality prediction sequence, and the quality prediction results of each manufacturing stage are obtained based on the quality prediction sequence.

[0009] Furthermore, the preprocessing includes feature dimension reduction and data standardization.

[0010] Furthermore, feature dimensionality reduction is performed through principal component analysis.

[0011] Furthermore, the broadcast mechanism is specifically as follows:

[0012]

[0013] Concat means concatenating the input data in the second-to-last dimension, n is the total number of manufacturing stages, and x is the number of j is the equipment status parameter of the jth manufacturing stage, x j，m is the mth equipment status parameter in the jth manufacturing stage, n x，j represents the total amount of equipment status parameters in the jth manufacturing stage, and C is the common input.

[0014] Furthermore, the attention weight α of the output vector of the i-th manufacturing stage is iThe calculation formula is as follows:

[0015]

[0016] in, represents the unnormalized attention weight, n is the total number of manufacturing stages;

[0017] Calculate the attention weight α i and the encoder i-th output vector The dot product of Send to the decoder:

[0018]

[0019] Furthermore, the unnormalized attention weights Calculate using the following formula:

[0020]

[0021] Among them, v is the parameter vector, v T is the transpose of the vector, w is the weight matrix, represents the jth hidden vector of the decoder, Represents the i-th output vector of the encoder.

[0022] Furthermore, the specific expression of the quality sequence output by the feedforward neural network is as follows:

[0023] y=W2(σ(W1x+b1)+b2)

[0024] Among them, y is the quality sequence, x is the multidimensional output of the decoder, σ is the activation function, W1 and W2 are two linear transformation layers, and b1 and b2 are the bias parameters of the linear transformation layers respectively.

[0025] Furthermore, the activation function σ is ReLU.

[0026] The present invention also provides an electronic device, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements the above method when executing the program.

[0027] The present invention also provides a computer-readable storage medium on which a computer program is stored, and the program implements the above method when executed by a processor.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The present invention proposes a quality prediction method for a multi-stage manufacturing system. First, the initial quality index value of the raw materials and the equipment state parameters of each manufacturing stage are obtained, pre-processed and spliced ​​through a broadcast mechanism, and then the trained quality prediction model for the multi-stage manufacturing system is input to obtain the quality prediction result of each manufacturing stage; the sequence-to-sequence quality prediction model for the multi-stage manufacturing system specifically includes an encoder, a full attention module, a decoder and a feedforward neural network connected in sequence, wherein the encoder is constructed based on a bidirectional long short-term memory network, which can effectively process the dependency between features with a large number of input intervals, and the encoder generates an output vector for each manufacturing stage according to the input data; the full attention module is used to calculate the attention weight of the output vector of each manufacturing stage according to the hidden vector of the decoder, and input the weighted output vector into the decoder; the decoder is constructed based on a bidirectional long short-term memory network, which is used to decode the weighted output vector, and the decoding result is integrated through a feedforward neural network and output as a quality prediction sequence, thereby obtaining the quality prediction result of each manufacturing stage; the above method takes into account the cumulative and interactive effects of different stages on product quality, and can meet the accuracy and generalization performance of product quality estimation in complex production scenarios.

[0030] 2. The present invention uses a broadcast mechanism to splice the initial quality index values ​​of raw materials and the equipment status parameters of each manufacturing stage. The components take into account the performance and initial stage of the raw materials, and can cover complex situations such as returning workpieces and restarting processing, thereby improving the practicability of the model.

[0031] 3. The present invention dynamically pays attention to the weight of each stage through the full attention mechanism, thereby improving the prediction accuracy, realizing the weight control of the impact of key stages on the final quality output, and effectively improving the accuracy of quality estimation in complex multi-stage manufacturing processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the process of the present invention;

[0033] Figure 2 It is a schematic diagram of the structure of the quality prediction model for multi-stage manufacturing system;

[0034] Figure 3 This is a loss convergence diagram of a multi-stage prediction model obtained through simulation experiments in the embodiment;

[0035] Figure 4 It is a comparison chart of prediction accuracy indicators of various methods in the simulation experiment of the embodiment. DETAILED DESCRIPTION

[0036] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0037] Example:

[0038] This embodiment provides a quality prediction method for a multi-stage manufacturing system. Figure 1 As shown, the following steps are included:

[0039] S1. Obtain the initial quality indicators of raw materials in the historical processing stage and the equipment status parameters of each manufacturing stage, perform preprocessing, and then splice them through the broadcast mechanism to construct a training set.

[0040] The training set contains input variables (i.e., influencing factors) and output variables (i.e., quality indicators) at different stages. The input variable x is defined as {x1, ..., x J}, output variable index y = {y1, ..., y I},in is the input variable of stage j, x j，m is the mth equipment status parameter in the jth manufacturing stage, n x，j represents the total amount of equipment status parameters at the jth manufacturing stage, y i Indicates the quality indicator output of the i-th stage. Considering the multiple impacts of multiple stages, this embodiment designs an encoder-decoder structure based on BiLSTM, and designs an attention mechanism based on state weights according to the impact of different stages on the output variables.

[0041] Preprocessing includes feature dimensionality reduction and data standardization for data distribution. Specifically, PCA (principal component analysis) is used for dimensionality reduction to reduce redundant features, avoid overfitting, and improve the performance of the model on new data; standardization is performed on the data after dimensionality reduction to eliminate the impact of different feature dimensions on the prediction results. The features of each stage in the multi-stage manufacturing system are aligned, and the data of each stage is combined with the common input variables through the broadcast mechanism:

[0042]

[0043] Among them, Concat means concatenating the input data in the penultimate dimension, n is the total number of manufacturing stages, and C is the common input.

[0044] S2. Build and train a quality prediction model (Seq2Seq-FA) for multi-stage manufacturing systems.

[0045] like Figure 2As shown in the figure, the basic structure of the quality prediction model for multi-stage manufacturing systems consists of an encoder and a decoder. The encoder uses a bidirectional long short-term memory neural network to capture inter-sequence features, and then outputs a hidden vector through a feedforward neural network. A full attention module is added between the encoder and the decoder to capture the correlation between the stages, allocate more attention to the stage features that are more important for subsequent predictions, and enhance the model's attention to key features. The encoder output is weighted and summed with the attention mechanism and then sent to the decoder to be decoded into a quality output sequence. Finally, the bidirectional long short-term memory neural network in the decoder is input to restore the hidden vector to the quality indicator output of each stage, and the model's expression ability is increased through a feedforward neural network. The description of each part of the model is as follows:

[0046] (1) Encoder

[0047] The encoder BiLSTM has n time steps, each time step corresponds to a production stage in the model, and the latent vector of the jth stage can be expressed as h j , the output vector is y j , j = 1, 2, ..., n, latent vector h j Can be used to assist the next stage hidden vector h j+1 and the output vector y j+1 As the time step increases, j can be as large as n, that is, the latent vector of the last time step is h n , the output vector is y n .

[0048] The data spliced ​​by the broadcast mechanism is passed through multiple BiLSTM blocks to capture features, and the output sequence of the last BiLSTM is passed through a feedforward neural network module (FFN) to obtain the feature hidden state h n and the output sequence y n BiLSTM can effectively handle the dependencies between features with a large number of input intervals based on LSTM, where the forget gate, input gate, and output gate play a vital role in updating the information state of the storage unit at a given time. The calculation formula of BiLSTM is as follows:

[0049] Calculate the forward LSTM formula:

[0050] u j =σ(W uj x j +W uh h j-1 +b u )

[0051] f j =σ(W fx x j +W fh hj-1 +b f )

[0052]

[0053] o j =σ(W ox x j +W oh h j-1 +b o )

[0054] h j =o j ⊙tanh(C j )

[0055] Among them, u j 、f j , o j They are input gate, forget gate, and output gate respectively; C j represents the cell state of the forward LSTM at the jth time step, σ represents the sigmoid activation function; tanh represents the tanh activation function; W is the parameter matrix of the gate and storage unit; x j represents the representation of the j-th time step of the input sequence; b represents the bias, and ⊙ represents element-by-element multiplication.

[0056] Backward LSTM calculation formula:

[0057] u′ j =σ(W′ ux x j +W′ uh h′ j-1 +b′ u )

[0058] f′ j =σ(W′ fx x j +W′ fh h′ j-1 +b′ f )

[0059]

[0060] o′ j =σ(W′ ox x j +W′ oh h j-1 +b′ o )

[0061] h′ j =o′ j ⊙tanh(C′ j )

[0062] Similar to the forward LSTM, u′ j , f′ j , o′ j are the states of input gate, forget gate and output gate respectively, C′ j is the cell variable of the jth time step of the backward LSTM; h j is the hidden state variable of the jth time step of the forward LSTM; h′ j is the hidden state variable of the jth time step of the backward LSTM, σ represents the sigmoid activation function; tanh represents the tanh activation function; W is the parameter matrix of the gate and storage unit; x j represents the representation of the j-th time step of the input sequence; b represents the bias, and ⊙ represents element-by-element multiplication.

[0063] (2) Full-Attention Module (FA)

[0064] The decoder outputs the hidden state of the feature The output characteristics of the encoder After a full-attention (FA) module, the sequence attention score is obtained

[0065] The attention mechanism is used to capture the correlation between stages, allocate more attention to the stage features that are more important for subsequent predictions, and enhance the model's attention to key features. The encoder output is weighted and summed with the attention mechanism and then sent to the decoder to be decoded into a quality output sequence. i and Multiply element by element to get the hidden state with attention weights Attention weight α i is generated by the decoder hidden state and all encoder outputs. The relevant equation is as follows:

[0066]

[0067] Among them, v is the parameter vector, v T is the transpose of the vector, which is used to map the calculation result to an attention weight to generate an attention score. w is the weight matrix used to transform the hidden state and output of the decoder. The parameter represents the mutual dependency between input and output, and uses the softmax function to ensure that the sum of all attention weights is 1. represents the i-th output vector of encoder e(encoder), Represents the jth hidden vector of the decoder d(decoder).

[0068] Then, we calculate the dot product between the weight parameters and the encoder output Send to decoder:

[0069]

[0070] (3) Decoder

[0071] As the input of the first BiLSTM in the decoder, it passes through n BiLSTM modules at a time, and then passes through the FFN layer to integrate all features and finally output the quality prediction sequence. Specifically, the FFN layer is composed of two linear transformation layers and an activation function. The activation function uses ReLU:

[0072] y=W2(σ(W1x+b1)+b2)

[0073] Wherein, y is the quality sequence, σ is the activation function, ReLU is used in this embodiment, W1 and W2 are two linear transformation layers, b1 and b2 are bias parameters of the linear transformation layers respectively.

[0074] S3. Use the trained quality prediction model to predict the quality of the target data.

[0075] Compared with the prior art, the above method has the following beneficial effects:

[0076] 1. The cumulative and interactive effects of different stages on the quality of the final product are taken into account to ensure the accuracy and generalization performance of parameter estimation in complex production scenarios.

[0077] 2. The performance and initial stage of raw materials are considered through the broadcast mechanism, and the process deviation is predicted using global state variables. This can cover complex situations such as returning workpieces to restart processing, thereby improving the practicality of the model.

[0078] 3. By establishing a sequence-to-sequence quality prediction model, multi-level and multi-dimensional data can be effectively encoded, multi-stage quality prediction of different lengths can be achieved, and the quality transfer relationship between stages can be effectively fitted.

[0079] 4. Through the full attention mechanism, the weight mechanism of each stage is dynamically focused, which improves the prediction accuracy, realizes the weight control of the impact of key stages on the final quality output, and effectively improves the accuracy of quality estimation in complex multi-level manufacturing processes.

[0080] To prove the effectiveness of the above method, a verification experiment was carried out in this embodiment. The verification experiment was carried out on a computer with a CPUDELL R740Gold 5220R, a memory of 32G 3200A, and a graphics card Nvidia TeslaV100 32G. The compiler used Pycharm and the compilation language used Python 3.11. The data set used in the verification experiment comes from the public data set https: / / www.kaggle.com / datasets / podsyp / production-quality / data. The sensor data of the discrete production process (RMDT) of multiple chambers of a roasting machine is used as historical data, with a total of 12,000 manufacturing data, which has sequential continuity in time. It is now divided into three stages according to time, where the initial state includes the initial original quality and the initial process. Each stage contains 17 sensor parameters P1, T x , ..., including pressure, humidity, water content and other characterization quantities. The quality index of stage i is S i-Quality , the data is randomly divided into 80% and 20% for training and testing, and the results are shown in Table 1.

[0081] Table 1 Multi-stage manufacturing raw dataset RMDT

[0082]

[0083] In data preprocessing, PCA was used to reduce the dimension of data and remove the variables with correlation > 0.9. The number of device state variables in the first, second and third stages was 15, 17 and 14 respectively. In order to align the sequence tensors when inputting the Seq2Seq model, the vectors of different stages were padded with zeros to make them all 17 dimensions. Considering the impact of different initial distributions of different features on model performance, the remaining features were standardized by removing the mean and scaling to unit variance. The data set was randomly divided into three data sets, namely training set (80%), validation set (10%), and test set (10%).

[0084] The characteristics of each stage in the multi-stage manufacturing system are aligned, and the data of each stage is combined with the common input variables to capture the common characteristics of all stages. In this verification experiment, there are two common input variables, namely initial pressure and temperature. Finally, the processed data is input into the model for training and verification.

[0085] This example uses eight popular regression algorithms for comparison, including state-of-the-art methods that can be reproduced on this dataset:

[0086] ① Polynomial regression (PR): prediction is performed through multi-dimensional combination of features, and the maximum degree of polynomial features is set to 2.

[0087] ②Support Vector Regression (SVR): One of the most robust prediction methods based on kernel learning and Vapnik-Chervonenkis theory (VC theory), the kernel type is selected as rbf radial basis function, and the regularization parameter C = 1.0. The width of the ε-insensitive loss function in support vector regression is epsilon = 0.1.

[0088] ③ Lasso regression: Introduce L1 regularization (Lasso penalty) to perform feature selection and prevent overfitting. The regularization strength alpha = 1.0.

[0089] ④ Random Forest (RF): An efficient gradient boosting method that uses squared error as the loss function, n_estimators = 100, max_depth = 5, random_state = 42.

[0090] ⑤ Extreme Gradient Boosting (Xgboost): The number of trees in the forest is assigned to 100, the maximum depth is 3, and the learning rate is 0.15.

[0091] ⑥K Nearest Neighbor (KNN): An instance-based learning algorithm that makes predictions by finding the nearest k data points and averaging their target values, with k being 5.

[0092] ⑦Long Short-Term Memory Network (LSTM): It can remember information in long time series and is suitable for processing data with long-term dependencies.

[0093] ⑧Transformer: A model architecture based on the self-attention mechanism. It does not rely on the sequential processing of sequences, but processes the input data through the self-attention mechanism task, which can capture global dependencies.

[0094] This example selects four commonly used indicators in regression tasks (mean square error RMSE, mean absolute error MAE, mean absolute percentage MAPE and determination coefficient R 2 ), where RMSE, MAE, and MAPE are as small as possible, R 2 The bigger the better.

[0095] After multiple experiments, the experimental hyperparameters were adjusted, as shown in Table 2:

[0096] Table 2 Model hyperparameter settings

[0097] Batch size Learning Rate Optimizer Number of training sessions Encoder hidden layer size 128 0.001 Adam 150 128

[0098] To avoid overfitting, early stopping was used and the MSE of the validation set was evaluated every 30 steps. The parameters in Seq2Seq-FA were optimized using the Adam optimizer, with the training batch size batch_size set to 128. Figure 3 As shown in the figure, the loss tends to be stable after about 40 iterations, so the epoch is set to 150. Compared with the baseline method, the overall performance of Seq2Seq-FA proposed in this embodiment on the RMDT dataset is shown in Table 3. It can be observed that the proposed Seq2Seq-FA substantially outperforms the existing best methods, and the best indicators and methods are indicated in bold.

[0099] Table 3 Overall performance on the RMDT dataset

[0100]

[0101]

[0102] Draw the above results into a bar graph, such as Figure 4 As shown, it can be clearly seen that the loss values ​​of Seq2Seq-FA in the three stages are lower than those of other methods. In addition, LSTM, Transformer and Seq2Seq-FA are methods based on sequence prediction. Simple LSTM is not as good as Seq2Seq with an encoder structure. This embodiment considers that the structure of Seq2Seq-FA takes into account the attention of each element in the sequence and the output under variable length. Transformer is an attention-based sequence-to-sequence prediction, which is more advantageous in long-distance sequences. The current sequence length of 3 is prone to overfitting. Some machine learning-based methods PR, SVR, KNN, and XGBoost also show good performance on the current data set. These models extract patterns in the data through feature workers, which provide good supplements in flexibility and interpretability, but still lack certain fitting capabilities in the latent vector transmission of the sequence. Therefore, the structural advantages of Seq2Seq-FA enable it to have higher robustness and accuracy in complex sequence prediction tasks.

[0103] If the above method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0104] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A quality prediction method for a multi-stage manufacturing system, characterized in that: The following steps are involved: Obtain the initial quality index values ​​of raw materials and the equipment status parameters of each manufacturing stage, perform preprocessing and splice them through the broadcast mechanism, and then input the trained quality prediction model for multi-stage manufacturing systems to obtain the quality prediction results of each manufacturing stage; Among them, the quality prediction model for multi-stage manufacturing systems includes an encoder, a full attention module, a decoder and a feedforward neural network connected in sequence. The encoder is constructed based on a bidirectional long short-term memory network, and is used to generate an output vector for each manufacturing stage according to input data; the full attention module is used to calculate the attention weight of the output vector of each manufacturing stage according to the hidden vector of the decoder, and input the weighted output vector into the decoder; the decoder is constructed based on a bidirectional long short-term memory network, and is used to decode the weighted output vector. The decoding results are integrated through the feedforward neural network and output as a quality prediction sequence, and the quality prediction results of each manufacturing stage are obtained based on the quality prediction sequence.

2. A quality prediction method for a multi-stage manufacturing system according to claim 1, characterized in that: The preprocessing includes feature dimension reduction and data normalization.

3. A quality prediction method for a multi-stage manufacturing system according to claim 2, characterized in that: Feature dimensionality reduction was performed using principal component analysis.

4. The quality prediction method for a multi-stage manufacturing system according to claim 1, characterized in that: The broadcast mechanism is specifically: Concat means concatenating the input data in the second-to-last dimension, n is the total number of manufacturing stages, and x is the number of j is the equipment status parameter of the jth manufacturing stage, x j,m is the mth equipment status parameter in the jth manufacturing stage, n x,j represents the total amount of equipment status parameters in the jth manufacturing stage, and C is the common input.

5. The quality prediction method for a multi-stage manufacturing system according to claim 1, characterized in that: The attention weight α of the output vector of the i-th production stage i The calculation formula is as follows: in, represents the unnormalized attention weight, n is the total number of manufacturing stages; Calculate the attention weight α i and the encoder i-th output vector The dot product of Send to the decoder:

6. A quality prediction method for a multi-stage manufacturing system according to claim 5, characterized in that: Unnormalized attention weights Calculate using the following formula: Among them, v is the parameter vector, v T is the transpose of the vector, w is the weight matrix, represents the i-th output vector of the encoder, represents the jth hidden vector of the decoder.

7. The quality prediction method for a multi-stage manufacturing system according to claim 1, characterized in that: The specific expression of the quality prediction sequence output by the feedforward neural network is as follows: y=W2(σ(W1x+b1)+b2) Among them, y is the quality sequence, x is the multidimensional output of the decoder, σ is the activation function, W1 and W2 are two linear transformation layers, and b1 and b2 are the bias parameters of the linear transformation layers respectively.

8. A quality prediction method for a multi-stage manufacturing system according to claim 7, characterized in that: The activation function σ is ReLU.

9. An electronic device comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images