A process parameter analysis method, device and computer storage medium based on neural network

Abstract

The present invention discloses a process parameter analysis method, device and computer storage medium based on a neural network, which relates to the field of neural network technology to solve the technical problem that multiple process parameters of the equipment have a large correlation. The above-mentioned process parameter analysis method based on a neural network includes: creating a test data set based on the time series data of the process parameters generated by multiple process processes, wherein each process includes multiple process parameters. The test data set is input into a neural network classification model to extract the corresponding feature vectors of each process parameter, wherein the neural network classification model includes multiple gated recurrent networks, and each gated recurrent network is used to extract features from the time series data of a process parameter. According to the feature vector and the weight of the fully connected layer in the neural network classification model, the contribution vector of the corresponding process parameter is determined. According to the corresponding contribution vector of each process parameter, the influence of the corresponding process parameter on the process is determined.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to a process parameter analysis method, equipment and computer storage medium based on neural network. Background Art

[0002] In industrial manufacturing, a production line that seems to be operating well may not actually be operating at the optimal equipment process parameters. Therefore, the optimization of equipment process parameters is particularly important, as it determines the output and yield rate of the production line.

[0003] At present, the optimization of equipment process parameters mostly adopts the method of adjusting parameters one by one multiple times, but this method is time-consuming and labor-intensive, with high investment costs and low work efficiency. In addition, the correlation between multiple process parameters of the equipment is large, and it is difficult to optimize the work efficiency of the equipment by adjusting only a single process parameter. Summary of the invention

[0004] The purpose of the present invention is to provide a process parameter analysis method, device and computer storage medium based on a neural network to solve the technical problem that there is a large correlation between multiple process parameters of the equipment and it is difficult to optimize the working efficiency of the equipment by adjusting only a single process parameter.

[0005] In a first aspect, the present invention provides a method for process parameter analysis based on a neural network, comprising:

[0006] A test data set is created based on time series data of process parameters generated by multiple process steps, wherein each process step includes multiple process parameters.

[0007] The test data set is input into the neural network classification model to extract the feature vector corresponding to each process parameter; wherein the neural network classification model includes multiple gated recurrent networks, each gated recurrent network is used to extract features from the time series data of a process parameter.

[0008] According to the feature vector and the weight of the fully connected layer in the neural network classification model, the contribution vector of the corresponding process parameters is determined.

[0009] According to the corresponding contribution vector of each process parameter, the influence of the corresponding process parameter on the process is determined.

[0010] In the case of adopting the above scheme, the neural network-based process parameter analysis method provided by the present invention inputs a test data set created based on the time series data of the process parameters generated by multiple process steps into the neural network classification model to extract the feature vectors corresponding to each process parameter. The neural network classification model of the present invention includes multiple gated recurrent networks, each of which is used to extract features from the time series data of a process parameter. Based on this, since each process parameter corresponds to a gated recurrent network separately, the gated recurrent network parameters between the process parameters are not shared, which avoids mutual interference between the process parameters in the feature extraction stage, and helps to more accurately analyze the influence of each process parameter on the working efficiency of the equipment.

[0011] Furthermore, the present invention determines the contribution vector of the corresponding process parameter based on the above-mentioned feature vector and the weight of the fully connected layer in the neural network classification model. According to the corresponding contribution vector of each process parameter, the influence of the corresponding process parameter on the process is determined. Based on the influence of each process parameter on the equipment efficiency, it is possible to accurately analyze which process parameter has the greatest influence on the equipment efficiency, so that the working efficiency of the equipment can be improved by adjusting the process parameter. In other words, based on the influence of each process parameter on the equipment efficiency, the direction of improving the working efficiency of the equipment is provided.

[0012] Finally, compared with the prior art, the present invention avoids subjective human intervention, provides a direction for the optimization of process parameters, and solves the dilemma of technicians not knowing where to start. Compared with the method of manually searching for influencing factors one by one multiple times, this method has higher accuracy and efficiency, and this method can save manpower, material and financial resources, and help improve product yield and production line output.

[0013] In a possible implementation, the training process of the neural network classification model includes: creating a training data set, the training data set including a plurality of samples, and training a neural network based on the training data set to obtain the neural network classification model.

[0014] Furthermore, before creating the training data set, the process parameter analysis method based on the neural network also includes: obtaining the time series data of the multiple samples. Based on the sample labels, the time series data of the multiple samples are grouped; wherein the time series data of the same sample corresponds to the same sample label. The time series data of the grouped multiple samples are normalized to obtain the normalized grouped sample time series data.

[0015] The creating of the training data set includes: creating a training data set based on the normalized grouped sample time series data.

[0016] When the above solution is adopted, the time series data of multiple samples are grouped according to the sample labels, and the time series data of the grouped multiple samples are normalized. Based on this, the convergence speed of the model processing data can be accelerated and the oscillation can be reduced.

[0017] In a possible implementation, the plurality of samples have a plurality of attributes. In the neural network, each of the attributes corresponds to a gated recurrent network.

[0018] Furthermore, the neural network is trained based on the training data set to obtain the neural network classification model, including:

[0019] The training data set is input into the neural network to obtain a prediction result.

[0020] The prediction result is input into the loss function, and the classification loss is minimized using the Adam optimizer to update the weights of the hidden layer and the fully connected layer in the neural network until the loss change of the neural network is less than a preset value, thereby obtaining the neural network classification model.

[0021] The training data set is input into the neural network to obtain prediction results including:

[0022] In the hidden layer, each of the gated recurrent networks is used to extract features of the corresponding attributes to obtain multiple attribute features; wherein the parameters of each of the gated recurrent networks are different;

[0023] The multiple attribute features are input into the fully connected layer and the classification layer of the neural network to obtain the prediction result of the neural network; wherein each of the attribute features corresponds to a different fully connected layer weight.

[0024] Each of the gated recurrent networks includes a plurality of hidden units, and the number of features of each of the attribute features is the same as the number of hidden units in the corresponding gated recurrent network.

[0025] Furthermore, each of the above gated recurrent networks includes 128 hidden units.

[0026] When the above scheme is adopted, each gated recurrent network of the present invention includes multiple hidden units, for example, each gated recurrent network includes 128 hidden units. In practice, the number of hidden units is directly related to the number of input and output units. If the number of hidden units is too small, the features between the data cannot be extracted, resulting in underfitting. If the number of hidden units is too large, the training time is increased, overfitting is prone to occur, and the generalization ability is poor. Through multiple experiments, the present invention sets the number of hidden units to 128, which can well balance the problems of overfitting and underfitting.

[0027] Furthermore, the present invention uses the Adam optimizer and the prediction results to update the hidden layer weights and the fully connected layer weights in the neural network classification model until the loss change of the neural network classification model is less than the preset value, and the trained neural network classification model is obtained. Since the Adam optimizer sets independent adaptive learning rates for different parameters, the learning rate is dynamically adjusted using the gradient first-order moment estimation and the second-order moment estimation, and each iterative training has a clear range of learning rates, so that the parameters change steadily, and the speed of model learning features is accelerated. Therefore, the neural network classification model trained in the present invention has the characteristics of fast learning speed.

[0028] In a possible implementation, determining the contribution vector of the corresponding process parameter according to the feature vector and the fully connected layer weight extracted by the neural network classification model includes:

[0029] After concatenating multiple feature vectors, they are multiplied by the fully connected layer weights extracted from the neural network classification model to obtain the total contribution vector;

[0030] According to the number of process parameters, the total contribution vector is evenly divided to obtain the contribution vector of each process parameter.

[0031] In a possible implementation, determining the influence of the corresponding process parameter according to the contribution vector corresponding to each of the process parameters includes:

[0032] Respectively obtaining the maximum value of the value in the contribution vector corresponding to each of the process parameters in each of the process processes, and obtaining a plurality of maximum values ​​corresponding to each process parameter;

[0033] An average value of the multiple maximum values ​​corresponding to each process parameter is obtained, and the average value is used as the influence degree of the corresponding process parameter.

[0034] In a possible implementation, after determining the influence of the corresponding process parameter according to the corresponding contribution vector of each process parameter, the process parameter analysis method based on the neural network further includes:

[0035] Arrange the multiple influence degrees in descending order according to their magnitude to obtain influence degree vectors of the multiple process parameters.

[0036] In the case of adopting the above scheme, the present invention arranges multiple influences in descending order according to their size to obtain influence vectors of multiple process parameters. Therefore, the higher the influence order in the influence vector, the greater the influence of the corresponding process parameter on the equipment efficiency, thereby providing a direction for the optimization of the process parameters, which can not only improve the working efficiency of the equipment, but also help to improve the product yield and production line output.

[0037] In a second aspect, the present invention also discloses a process parameter analysis device based on a neural network, comprising: a processor and a communication interface, the communication interface and the processor are coupled, and the processor is used to run a computer program or instruction to implement the above-mentioned process parameter analysis method based on a neural network.

[0038] In a third aspect, the present invention further discloses a computer storage medium, in which instructions are stored. When the instructions are executed, the above-mentioned process parameter analysis method based on neural network is implemented.

[0039] The beneficial effects of the second aspect, the third aspect and their various implementations in the present invention are the same as the beneficial effects of the first aspect or any possible implementation of the first aspect, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0041] Figure 1 A flowchart of a process parameter analysis method based on a neural network provided by an embodiment of the present invention is shown;

[0042] Figure 2 A structural diagram of a neural network classification model provided by an embodiment of the present invention is shown;

[0043] Figure 3 The figure shows an internal structure diagram of a hidden layer unit provided by an embodiment of the present invention;

[0044] Figure 4 A flowchart showing another process parameter analysis method based on a neural network provided by an embodiment of the present invention is shown;

[0045] Figure 5 A schematic diagram showing the influence of various properties on a process provided by an embodiment of the present invention is shown;

[0046] Figure 6 A schematic diagram of the hardware structure of a process parameter analysis device based on a neural network provided in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0047] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0048] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. The meaning of "several" is one or more, unless otherwise clearly and specifically defined.

[0049] In industrial manufacturing, production lines that seem to be operating well are actually not operating in the best state. There is a lot of room for optimization of equipment parameters, which determine the output and yield rate of production lines. Taking the photovoltaic industry as an example, reducing the cost per kilowatt-hour has always been the direction of the photovoltaic industry's efforts and an important tool for enterprises to enhance their competitiveness. As an important link in the production of photovoltaic products, reducing its wire breakage rate will help improve the efficiency of crystal pulling, thereby reducing the cost per kilowatt-hour. At present, shoulder release and equal diameter are the key to determining the success of crystal pulling, and the setting of crystal pulling parameters determines the success or failure of shoulder release and equal diameter. Therefore, it is particularly important to optimize the crystal pulling process parameters. In terms of parameter optimization, on-site personnel modify the crystal pulling parameters through experience. Due to the large number of parameters, it is difficult to analyze the main factors affecting the failure of crystal pulling, which causes the dilemma of not knowing where to start. At present, the method of adjusting parameters one by one multiple times is often used. This method is time-consuming and labor-intensive, and the investment cost is large, the work efficiency is not high, and the correlation between parameters is large. It is not just adjusting a single parameter that can solve the problem of crystal pulling failure. Therefore, analyzing the factors affecting wire breakage has become a difficult problem for crystal pulling factories.

[0050] Based on this, Figure 1 It is shown that an embodiment of the present invention provides a method for process parameter analysis based on a neural network. The method includes:

[0051] S101, creating a test data set according to time series data of process parameters generated by multiple process steps, wherein each process step includes multiple process parameters.

[0052] The above-mentioned timing data of the process parameters are a set of timing data generated by the same process parameters in multiple process steps.

[0053] As a specific example, the crystal pulling process in the photovoltaic industry includes multiple process parameters such as pulling speed, heater heating power, crucible rotation, etc. Based on this, the time series data of process parameters generated by multiple process steps can be used to create a test data set, which can be to obtain the time series data of pulling speed, heater heating power, and crucible rotation in multiple process steps, and create a test data set based on the time series data of the multiple process parameters.

[0054] S102, inputting the test data set into the neural network classification model to extract the feature vector corresponding to each process parameter; wherein the neural network classification model includes a plurality of gated recurrent networks, each gated recurrent network is used to extract features from the time series data of a process parameter.

[0055] The created test data set is input into the trained neural network classification model to extract the feature vectors corresponding to each process parameter. In other words, each process parameter corresponds to a feature vector.

[0056] The neural network classification model provided by the embodiment of the present invention includes multiple gated recurrent networks, each of which is used to extract features from the time series data of a process parameter. Based on this, since each process parameter corresponds to a gated recurrent network separately, the network parameters between the process parameters are not shared, which avoids mutual interference between the process parameters in the feature extraction stage, and helps to more accurately analyze the influence of each process parameter on the working efficiency of the equipment.

[0057] The training process of the neural network classification model in the embodiment of the present invention includes:

[0058] A training data set is created, wherein the training data set includes a plurality of samples.

[0059] In an embodiment of the present invention, multiple samples may be created from historical data generated during a process.

[0060] For example, in the crystal pulling process, each sample includes historical data generated by pulling a single crystal silicon rod.

[0061] Furthermore, before creating the training data set, the process parameter analysis method based on the neural network also includes:

[0062] Acquire time series data of the plurality of samples.

[0063] Based on the sample labels, the time series data of the multiple samples are grouped, wherein the time series data of the same sample corresponds to the same sample label. Each sample has multiple attributes.

[0064] Specifically, when the time series data of multiple samples are grouped, an index field is added to group the time series data of multiple samples. The time series data of the same sample represents the time series data generated by the same process and corresponds to the same sample label.

[0065] For example, the same crystal pulling process is used to produce a crystal rod, and the timing data of the pulling speed, the timing data of the heater heating power, and the timing data of the crucible rotation generated in the process of pulling the same crystal rod all correspond to the same label.

[0066] For the crystal pulling process, the properties of the above samples may include pulling speed, heating power, and crucible rotation.

[0067] Furthermore, in an embodiment of the present invention, in order to increase the convergence speed of the model processing data and reduce oscillation, the time series data of the plurality of grouped samples are normalized to obtain normalized grouped sample time series data.

[0068] Specifically, the normalized expression is:

[0069]

[0070] Among them, x new is the normalized time series data, x is the original time series data, min(x) is the minimum value in the time series data corresponding to each attribute, and max(x) is the maximum value in the time series data corresponding to each attribute.

[0071] In a specific implementation, taking the crystal pulling process in the photovoltaic industry as an example, the original time series data is a set of crystal pulling time series data formed by the time series data of the pulling speed, the time series data of the heater heating power, and the time series data of the crucible rotation generated in the process of pulling multiple crystal rods. min(x) can be the minimum value in the time series data of the pulling speed, or the minimum value in the time series data of the heater heating power, or the minimum value in the time series data of the crucible rotation. max(x) can be the maximum value in the time series data of the pulling speed, or the maximum value in the time series data of the heater heating power, or the maximum value in the time series data of the crucible rotation. new It is the normalized crystal pulling timing data.

[0072] The creating of the training data set includes: creating a training data set based on the normalized grouped sample time series data.

[0073] Based on the training data set, the neural network is trained to obtain the neural network classification model.

[0074] In the training process of the neural network in the embodiment of the present invention, each attribute corresponds to a gated recurrent network. Specifically, the neural network is trained based on the training data set to obtain a neural network classification model including:

[0075] The training data set is input into the neural network to obtain the prediction results.

[0076] Each gated recurrent network includes multiple hidden units, and the multiple hidden units process the data in the input training data set in turn.

[0077] Before inputting the above training data set into the neural network classification model, in order to improve the accuracy of the prediction results of the neural network classification model, random downsampling is used to balance the positive and negative samples in the training data set. In order to make the training set and test set data representative and avoid overfitting of the neural network classification model, the full amount of data is shuffled according to the label ratio.

[0078] Furthermore, the above-mentioned prediction results of the neural network classification model obtained by inputting the training data set into the neural network classification model include:

[0079] In the hidden layer, each gated recurrent network is used to extract features of the corresponding attributes to obtain multiple attribute features; wherein the parameters of each gated recurrent network are different.

[0080] Based on the multiple attribute features and the fully connected layer weights of the neural network classification model, the prediction result of the neural network classification model is calculated, wherein each attribute feature corresponds to a different fully connected layer weight.

[0081] Each attribute includes a set of time series data, and the number of features of each attribute feature is the same as the number of hidden units in the corresponding gated recurrent network.

[0082] Specifically, each of the above gated recurrent networks includes 128 hidden units.

[0083] As a specific example, see Figure 2 In the embodiment of the present invention, a neural network classification model for analyzing factors affecting crystal pulling is constructed based on a gated recurrent network, wherein a hidden layer is used to extract the features of each attribute, the extracted features are summarized through a fully connected layer (FC), and then a classification layer (CL) is used to classify the time series data of the training data set. In the hidden layer, Figure 2 Medium x1-x n Represents input data, h represents each hidden unit, the upper right subscript of h represents the attribute number, and the lower right subscript represents the hidden unit code. Each attribute corresponds to a gated recurrent network, and the network parameters between attributes are not shared. This structure avoids mutual interference between attributes and helps to analyze the impact of each attribute on product results. In the fully connected layer, the attribute features extracted by the hidden layer are concatenated and input into the fully connected layer. In the figure, f represents the concatenated features, and c represents the output category of the neural network classification model.

[0084] For example, the model contains 128 hidden units. represents the 128th hidden unit of attribute n. For the crystal pulling process, x n Indicates the time series data of the nth crystal pulling parameter, f128 It represents the features corresponding to the time series data of the training data set extracted by the 128th hidden unit, and c1, c2, and c3 represent different categories respectively.

[0085] Figure 3 for Figure 2 The structure of the hidden unit, refer to Figure 3 , x t 、h t-1 、h t , They represent the input at time t, the output of the hidden unit at time t-1, the output of the hidden unit at time t, the memory content, etc., u t Indicates h t-1 Pass to h t The amount of information, its function expression is as follows:

[0086] u t =σ(ω xu x t +ω hu h t-1 +b u );

[0087] Where σ is the sigmoid activation function, ω xu The input value x t The weight of hu is the weight of the hidden unit output at the previous moment, b u For bias.

[0088] r t Indicates h t-1 Deliver to The degree of forgetting is expressed as follows:

[0089] r t =σ(ω xr x t +ω hr h t-1 +b r );

[0090] Where σ is the sigmoid activation function, ω xr The input value x t The weight of hr is the weight of the hidden unit output at the previous moment, b r For bias.

[0091] h t-1 The information retained after rt is the same as x t Splice and then use the tanh activation function to get the new memory content The function expression is as follows:

[0092]

[0093] In the formula, The input value x t The weight of is the weight of the hidden unit output at the previous moment, b h For bias.

[0094] The final memory content at time t is h t , its function expression is as follows:

[0095]

[0096] In the formula, (1-u t )×h t-1 Indicates the retained t-1 time information, represents the retained information at time t, and the sum of the two is the final retained information at time t.

[0097] The prediction results are input into the loss function, and the Adam optimizer is used to minimize the classification loss to update the weights of the hidden layer and the fully connected layer in the neural network until the change in the loss of the neural network is less than the preset value, thus obtaining the neural network classification model.

[0098] Reference Figure 4 First, the time series data of each parameter in the process is obtained, and the obtained time series data of each parameter is grouped according to the label. Then the grouped time series data is preprocessed, specifically, the grouped time series data is normalized. Secondly, the preprocessed data is input into the neural network split model for GRU feature learning. Finally, in the training process of the neural network classification model, the Adam optimizer is used to minimize the classification loss so that the weight ω in the hidden layer h and the weights ω in the fully connected layer f Keep updating until the change in the loss of the neural network classification model is less than 0.001, save the neural network classification model, and stop training. The Adam optimizer sets independent adaptive learning rates for different parameters, and uses the gradient first-order moment estimation and second-order moment estimation to dynamically adjust the learning rate. For each iterative training, the learning rate has a clear range, which makes the parameter change smooth and speeds up the model learning features. Its function expression is as follows:

[0099]

[0100] In the formula, m t and v t are the first-order momentum term and the second-order momentum term, β1 and β2 are the magnitudes of the dynamic values, β1 = 0.9, β2 = 0.999; is the correction value; wt is the parameter of the neural network classification model at the tth iteration, g t =ΔJ(w t ) represents the gradient of the loss function with respect to w at the tth iteration; η is the learning rate, ∈ is a very small number. In order to prevent the denominator from being 0, its value is 1e-8.

[0101] In the embodiment of the present invention, cross entropy is used as the loss function of the disconnection factor analysis neural network classification model to evaluate the difference between the probability distribution obtained by the current training and the true distribution. That is to say, the smaller the entropy value, the higher the accuracy of the neural network classification model prediction. Its function expression is as follows:

[0102]

[0103] In the formula, ω represents the weight between neurons, b represents the bias of neurons, N represents the total number of samples in the sample data, i represents the i-th sample, and y (i) represents the prediction result, f(x (i) ) represents actual historical results.

[0104] Based on this, each gated recurrent network of the embodiment of the present invention includes multiple hidden units, for example, each gated recurrent network includes 128 hidden units. In practice, the number of hidden units is directly related to the requirements of the problem, the number of input and output units, and if the number is too small, the information obtained is too little, resulting in underfitting. If the number is too large, the training time is increased, overfitting is prone to occur, and the generalization ability is poor. The present invention is implemented multiple times, and the number of hidden units is set to 128, which can well balance the problems of overfitting and underfitting.

[0105] Furthermore, the embodiment of the present invention uses the Adam optimizer to update the hidden layer weights and the fully connected layer weights in the neural network classification model until the loss change of the neural network classification model is less than a preset value, thereby obtaining a trained neural network classification model.

[0106] The Adam optimizer sets independent adaptive learning rates for different parameters, uses the gradient first-order moment estimation and the second-order moment estimation to dynamically adjust the learning rate, and each iterative training has a clear range of learning rates, so that the parameters change smoothly and the speed of model learning features is accelerated. Therefore, the neural network classification model trained in the embodiment of the present invention has the characteristic of fast learning speed.

[0107] S103, determining the contribution vector of the corresponding process parameters according to the feature vector and the fully connected layer weights extracted by the neural network classification model.

[0108] Specifically, the weights of the fully connected layer are extracted through the above-trained neural network classification model, that is, Figure 2The weights of the FC layer.

[0109] In the fully connected layer, multiple feature vectors are concatenated and multiplied with the fully connected layer weights extracted by the neural network classification model to obtain the total contribution vector. According to the number of process parameters, the total contribution vector is evenly divided to obtain the contribution vector of each process parameter.

[0110] S104, determining the influence of the corresponding process parameter on the process according to the contribution vector corresponding to each process parameter.

[0111] Specifically, the maximum value of the value in the contribution vector corresponding to each process parameter in each process is obtained respectively, and multiple maximum values ​​corresponding to each process parameter are obtained;

[0112] An average value of multiple maximum values ​​corresponding to each process parameter is calculated, and the average value is used as the influence degree of the corresponding process parameter.

[0113] In practice, multiple influences can be arranged in descending order to obtain influence vectors of multiple process parameters. It can be understood that in the influence vector, the higher the influence, the greater the influence of the corresponding process parameter on the process.

[0114] As a specific example, if the sample data to be analyzed has N samples, then an attribute has N contribution vectors and N maximum values. The N maximum values ​​are averaged and their means are sorted in descending order. Suppose the weights of the FC layer [w1w2w3···w n ] is denoted as w, the weight of each contribution vector is denoted as w′, and the features of the FC layer [f1,f2,f3,···,f n ] is recorded as f, and the characteristic of each contribution vector is recorded as f′, then the influence of each contribution vector is shown as follows:

[0115]

[0116] In the formula, C′ represents the influence, and N represents that there are N samples in the sample data. In the embodiment of the present invention, 128 hidden units are constructed in the neural network classification model, and each sample of each attribute outputs 128 feature values. Then, there are 128 weights in the FC layer corresponding to the above features, and the corresponding features are multiplied by the corresponding weights to obtain a 1*128-dimensional contribution vector; then the maximum value of the contribution vector is calculated; finally, the maximum value of 1300 samples is averaged to obtain the influence C′ of the attribute.

[0117] The embodiment of the present invention inputs 1300 samples into the neural network model to obtain a 1300*768-dimensional contribution vector, which is visualized by the plt.plot() method. Figure 5,In the figure, the horizontal axis represents the number of feature nodes, and the vertical axis represents the feature.,It can be seen that the process parameters with contribution values ​​fluctuating around 0 have very little impact on the process, and vice versa. Figure 5 The six parameters are represented as f1, f2, f3, f4, f5, and f6. The contribution values ​​of f4 and f6 are near 0, which means that the corresponding process parameters of f4 and f6 have very little effect on the process. The contribution values ​​of f2, f3, and f5 are between 0.5 and 0.2. Therefore, the corresponding process parameters of f2, f3, and f5 have a greater impact on the process than f4 and f6. The contribution value of f1 is 0.6. Therefore, the corresponding process parameters of f1 have a greater impact on the process than f2, f3, and f5.

[0118] Based on the above description, the embodiment of the present invention determines the contribution vector of the corresponding process parameter according to the above feature vector and the weight of the fully connected layer extracted by the neural network classification model. According to the corresponding contribution vector of each process parameter, the influence of the corresponding process parameter on the process is determined. Based on the influence of each process parameter on the equipment efficiency, it is possible to accurately analyze which process parameter has the greatest influence on the equipment efficiency, so that the working efficiency of the equipment can be improved by adjusting the process parameter. In other words, based on the influence of each process parameter on the equipment efficiency, the direction of improving the working efficiency of the equipment is provided.

[0119] Finally, compared with the prior art, the embodiment of the present invention avoids subjective human intervention, provides a direction for the optimization of process parameters, and solves the dilemma of technicians not knowing where to start. Compared with the method of manually searching for influencing factors one by one multiple times, this method has higher accuracy and efficiency, and this method can save manpower, material and financial resources, and help improve product yield and production line output.

[0120] Reference Figure 6 The actions performed by the above-mentioned neural network-based process parameter analysis device can be stored as computer instructions in the memory 220 of the neural network-based process parameter analysis device, and the computer instructions stored in the memory 220 are executed by the processor 210.

[0121] The process parameter analysis device 200 based on a neural network includes: a processor 210 and a communication interface 230, the communication interface 230 is coupled to the processor 210, and the processor 210 is used to run a computer program or instruction. The process parameter analysis device 200 based on a neural network can communicate with a slicing device, a degumming device, a sheet sorting device, and a driving component of the sorting device through the communication interface 230.

[0122] like Figure 6As shown, the processor 210 may be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of the program of the present invention. The communication interface 230 may be one or more. The communication interface 230 may use any transceiver-like device for communicating with other devices or communication networks.

[0123] like Figure 6 As shown, the neural network-based process parameter analysis device 200 may further include a communication line 240. The communication line 240 may include a path to transmit information between the above components.

[0124] Optional, such as Figure 6 As shown, the process parameter analysis device 200 based on neural network may further include a memory 220. The memory 220 is used to store computer instructions for executing the solution of the present invention, and is controlled to execute by the processor 210. The processor 210 is used to execute the computer instructions stored in the memory 220.

[0125] like Figure 6 As shown, the memory 220 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 220 may exist independently and be connected to the processor 210 via the communication line 240. The memory 220 may also be integrated with the processor 210.

[0126] Optionally, the computer instructions in the embodiment of the present invention may also be referred to as application program codes, which is not specifically limited in the embodiment of the present invention.

[0127] In a specific implementation, as an example, Figure 6As shown, the processor 210 may include one or more CPUs, such as Figure 6 CPU0 and CPU1 in.

[0128] In a specific implementation, as an example, Figure 6 As shown, the neural network-based process parameter analysis device 200 may include multiple processors 210, such as Figure 6 The processor 210 and the processor 250 in the embodiment of the present invention are shown in FIG. 1 . Each of these processors may be a single-core processor or a multi-core processor.

[0129] The embodiment of the present invention further provides a computer-readable storage medium. The computer-readable storage medium stores instructions, and when the instructions are executed, the functions performed by the process parameter analysis device based on the neural network in the above embodiment are realized.

[0130] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instruction is loaded and executed on the computer, the process or function of the embodiment of the present invention is executed in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a terminal, a user device or other programmable device. The computer program or instruction can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer program or instruction can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless means. The computer-readable storage medium can be any available medium that can be accessed by the computer or a data storage device such as a server, data center, etc. that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, a hard disk, or a tape; it can also be an optical medium, such as a digital video disc (DVD); it can also be a semiconductor medium, such as a solid state drive (SSD).

[0131] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art may understand and implement other variations of the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in the claims. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0132] Although the present invention has been described in conjunction with specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present invention. Accordingly, this specification and the accompanying drawings are merely exemplary illustrations of the present invention as defined by the appended claims and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present invention. Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, the present invention is intended to include such modifications and variations if they fall within the scope of the claims of the present invention and their equivalents.

Claims

1. A process parameter analysis method based on neural network, characterized in that: The method for process parameter analysis based on neural network includes: Creating a test data set based on time series data of process parameters generated by a plurality of process steps, wherein each of the process steps includes a plurality of process parameters; Inputting the test data set into a neural network classification model to extract feature vectors corresponding to each of the process parameters; wherein the neural network classification model includes a plurality of gated recurrent networks, each of which is used to extract features from time series data of one of the process parameters; Determining the contribution vector of the corresponding process parameter according to the feature vector and the weight of the fully connected layer in the neural network classification model, including: concatenating a plurality of the feature vectors and multiplying the concatenated vectors by the weight of the fully connected layer in the neural network classification model to obtain a total contribution vector; dividing the total contribution vector equally according to the number of the process parameters to obtain a contribution vector of each process parameter; According to the contribution vector corresponding to each of the process parameters, the influence of the corresponding process parameter is determined.

2. The process parameter analysis method based on neural network according to claim 1, characterized in that: The training process of the neural network classification model includes: Creating a training data set, wherein the training data set includes a plurality of samples; Based on the training data set, the neural network is trained to obtain the neural network classification model.

3. The process parameter analysis method based on neural network according to claim 2 is characterized in that: Before creating the training data set, the process parameter analysis method based on a neural network further includes: Acquire time series data of the plurality of samples; Based on the sample labels, the time series data of the multiple samples are grouped; wherein the time series data of the same sample corresponds to the same sample label; Normalizing the time series data of the grouped samples to obtain normalized grouped sample time series data; The creating of the training data set comprises: A training data set is created based on the normalized grouped sample time series data.

4. The process parameter analysis method based on neural network according to claim 2 is characterized in that: Each of the samples has a plurality of attributes; In the neural network, each of the attributes corresponds to a gated recurrent network.

5. The process parameter analysis method based on neural network according to claim 2, characterized in that: The step of training the neural network based on the training data set to obtain the neural network classification model comprises: Inputting the training data set into a neural network to obtain a prediction result; The prediction result is input into the loss function, and the classification loss is minimized using the Adam optimizer to update the weights of the hidden layer and the fully connected layer in the neural network until the loss change of the neural network is less than a preset value, thereby obtaining the neural network classification model.

6. The process parameter analysis method based on neural network according to claim 5, characterized in that: The step of inputting the training data set into the neural network to obtain a prediction result comprises: In the hidden layer, each of the gated recurrent networks is used to extract features of the corresponding attributes to obtain multiple attribute features; wherein the parameters of each of the gated recurrent networks are different; The multiple attribute features are input into the fully connected layer and the classification layer of the neural network to obtain the prediction result of the neural network; wherein each of the attribute features corresponds to a different fully connected layer weight.

7. The process parameter analysis method based on neural network according to claim 6, characterized in that: Each of the gated recurrent networks includes a plurality of hidden units, and the number of features of each of the attribute features is the same as the number of hidden units of the hidden units in the corresponding gated recurrent network.

8. The process parameter analysis method based on neural network according to claim 7, characterized in that: Each of the gated recurrent networks includes 128 hidden units.

9. The process parameter analysis method based on neural network according to any one of claims 1 to 8, characterized in that: Determining the influence of the corresponding process parameter according to the contribution vector corresponding to each process parameter includes: Respectively obtaining the maximum value of the value in the contribution vector corresponding to each of the process parameters in each of the process processes, and obtaining a plurality of maximum values ​​corresponding to each process parameter; An average value of the multiple maximum values ​​corresponding to each of the process parameters is calculated, and the average value is used as the influence degree of the corresponding process parameter.

10. The process parameter analysis method based on neural network according to any one of claims 1 to 8, characterized in that: After determining the influence of the corresponding process parameter according to the contribution vector corresponding to each process parameter, the process parameter analysis method based on neural network further includes: Arrange the multiple influence degrees in descending order according to their magnitude to obtain influence degree vectors of the multiple process parameters.

11. A process parameter analysis device based on a neural network, characterized in that: include: A processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run a computer program or instruction to implement the process parameter analysis method based on a neural network according to any one of claims 1 to 10.

12. A computer storage medium, characterized in that: The computer storage medium stores instructions, and when the instructions are executed, the process parameter analysis method based on a neural network according to any one of claims 1 to 10 is implemented.

Images