Workshop productivity prediction method based on data and physics dual drive

Abstract

The invention discloses a workshop productivity prediction method based on data and physics dual drive, and belongs to the field of productivity prediction. The method comprises the following steps: collecting operation data of a workshop, cleaning the operation data, performing feature selection in the cleaned operation data according to correlation with a target prediction variable and in combination with physical knowledge, and expanding the selected feature data by using a time sequence data generation method of a generative adversarial network to obtain expanded data; and constructing a data and physical dual-drive productivity prediction model based on a long short-term memory network, extracting physical knowledge related to the bottleneck according to the characteristics of the production line, designing a loss function, inputting the expanded data into the data and physical dual-drive productivity prediction model for training, and obtaining the relationship between the selected characteristics and the actual productivity. Therefore, guidance can be provided for production plan formulators, and the order on-time delivery rate is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of capacity prediction, and in particular to a workshop capacity prediction method based on data and physical dual drive. Background Art

[0002] Electronic information countermeasure equipment is an important foundation for defending the country's electromagnetic space security. It is equipped in various main combat equipment in the air, space, sea and land, and is the key to victory in modern non-contact warfare. In the context of preparing for war, the rapid, high-quality, and mass-produced delivery of equipment has become an important factor affecting the war situation. Among them, airborne electronic information countermeasure equipment has the characteristics of high density, cross-scale integration, long process chain, strong process flexibility, and unstable process state. Its manufacturing is a typical multi-variety variable batch production model, with R&D, trial production, and mass production in parallel.

[0003] A module product workshop is faced with difficulties such as difficulty in evaluating medium- and long-term manufacturing resource capabilities, difficulty in resource allocation and planning scheduling, and difficulty in coordinating resource allocation conflicts, which leads to slow response to external demand and long production cycle. It is urgent to study the workshop-level manufacturing capability evaluation method of complex electronic product manufacturing systems, study the medium- and long-term manufacturing planning methods, realize the coordinated allocation of electronic product production and manufacturing resources, study the resource allocation conflict coordination technology, realize conflict resolution and rescheduling under abnormal circumstances, and realize orderly, agile and intelligent management and control of the manufacturing process. Summary of the invention

[0004] The present invention provides a workshop capacity prediction method based on data and physical dual drive, which can provide guidance for production planners and improve the on-time delivery rate of orders.

[0005] The first aspect of the present invention provides a workshop capacity prediction method based on data and physical dual drive, comprising the following steps:

[0006] Collecting operation data of the workshop, cleaning the operation data, performing feature selection in the cleaned operation data according to the correlation with the target prediction variable and combining physical knowledge, and expanding the selected feature data by using a time series data generation method of a generative adversarial network to obtain expanded data;

[0007] A data- and physics-driven capacity forecasting model is constructed based on a long short-term memory network, wherein the data- and physics-driven capacity forecasting model is obtained by integrating bottleneck knowledge from constraint theory into the neural network forecasting model on the basis of capacity forecasting using a data-driven neural network forecasting model;

[0008] According to the characteristics of the production line, physical knowledge related to the bottleneck is extracted, a loss function is designed, and the expanded data is input into the capacity prediction model driven by both the data and physics for training to obtain the relationship between the selected features and the actual capacity.

[0009] Optionally, in one embodiment of the present invention, cleaning the operation data includes:

[0010] If the time windows for running data statistics are different, only the data in the overlapping part of the time windows will be used;

[0011] For repeated operation data, delete the redundant repeated operation data items;

[0012] For missing values ​​in running data, the mean or mode of similar data is used to fill them.

[0013] Optionally, in one embodiment of the present invention, expanding the selected feature data using the time series data generation method of the generative adversarial network includes:

[0014]

[0015] in, is the time series generated at (t a t a+1 ) production cycle data, is the value predicted based on previous data points of the time series, N t The end of the entire production cycle;

[0016] The loss functions of the generator and discriminator are:

[0017]

[0018] Among them, λ is the weight to balance the two losses, L D is the loss function of the discriminator of the generative adversarial network, L G is the loss function of the generator of the generative adversarial network.

[0019] Optionally, in one embodiment of the present invention, physical knowledge related to the bottleneck is extracted according to the characteristics of the production line, and the loss function is designed as:

[0020]

[0021] ReLU()=max(0,x)

[0022] in, is the standard loss function, and MSE is used to measure the inconsistency between the predicted value and the actual value in the data-driven model. is the physical loss function, is the predicted value, Y is the actual value, It is the prediction result after combining physical knowledge. Its value is obtained by combining the bottleneck capacity BN calculated based on physical knowledge with the original prediction value. Then, we use the commonly used activation function ReLU() to perform nonlinear transformation to obtain λ phy is the coefficient of the physical loss function, which is used to change the effect of the physical loss function, λ reg is the regularization coefficient, N s is the total number of samples, ω is the weight that converts the output of the previous layer into the input of the activation function of the current layer, ω PD are all weights in the proposed data and physical dual-drive model.

[0023] The workshop capacity prediction method based on data and physical dual drive of the embodiment of the present invention collects data of the workshop during operation through data collection equipment such as RFID and sensors deployed in the production and manufacturing workshop of an enterprise after digital transformation. The collected data is cleaned to improve the availability and quality of manufacturing data. The cleaned production data is analyzed to select usable features. In view of the problem that the collected data is small in scale, the data scale is expanded using data generation technology. According to the characteristics of the production line, the physical knowledge related to the bottleneck is extracted, the loss function is redesigned, and the selected feature data is input into the constructed data and physical dual drive capacity prediction model for training to find the relationship between the selected features and the actual capacity. Finally, the trained prediction model is applied to practice to provide guidance for production planners and improve the on-time delivery rate of orders.

[0024] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0026] Figure 1 A flowchart of a workshop capacity prediction method based on data and physical dual drive provided according to an embodiment of the present invention;

[0027] Figure 2 A schematic diagram of a data and physical dual-driven capacity prediction model according to an embodiment of the present invention;

[0028] Figure 3 A schematic diagram of the TimeGAN network structure according to an embodiment of the present invention;

[0029] Figure 4A schematic diagram of the long short-term memory network structure of an embodiment of the present invention;

[0030] Figure 5 This is a schematic diagram of the comparison between the TimeGAN generated data and the real data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0032] In the field of manufacturing, capacity forecasting is a crucial management activity. It aims to estimate the maximum production capacity that an enterprise or industry can achieve in a certain period of time in the future through scientific methods and models. This process is not only related to the strategic planning and resource allocation of the enterprise, but also directly affects the market response speed, cost control and profitability. The core of capacity forecasting lies in the comprehensive analysis of various internal and external factors to ensure the efficient operation and sustainable development of the production system.

[0033] A field survey of a certain enterprise module integration center revealed the following:

[0034] In terms of management process, the module integration center adopts the project management system. After receiving the project plan, the planning supervisor will break down the project according to personal experience and process data, and assign it to each line leader and working group in turn. The group will then evaluate whether its own production capacity matches the planned task volume and provide feedback. The supervisor will then adjust the plan.

[0035] In terms of production process, the digital production line belongs to the MTO (order-oriented) model. The material platform generates the design bill of materials EBOM, which is fine-tuned according to product characteristics and production environment before formulating the planned bill of materials PBOM. The material manager prepares the ingredients according to the PBOM, and after the sets are complete, they are packaged and sent to the production line buffer. After that, they are installed, tested for consistency, packaged, tested for basic performance, and tested in extreme environments. If any problems occur during the test, they need to be repaired and then retested. If everything goes well, they are packaged and put into storage after internal and external inspections, and finally delivered.

[0036] Capacity refers to the maximum number of products or the maximum capacity of services that a production system can produce under normal operating conditions within a certain period of time. Based on the above workshop processes and conditions, the representative product method is used to calculate capacity.

[0037] The representative product method is to select a representative product from the products produced by the production line. The so-called representative product refers to a product that reflects the professional aspects of the enterprise and has a large output and labor, or a product with a large output and representative structure and process. After the representative product is selected, the production capacity is calculated. The calculation method is divided into the following steps:

[0038] (1) Calculate the production capacity of the representative product using the following formula:

[0039]

[0040] Among them, M o represents the production capacity calculated in units of representative products; F represents the full effective working time of the unit equipment (taking into account the time for mold change and equipment failure); S represents the number of equipment in the bottleneck process; t o Indicates the standard working hours per unit of product.

[0041] (2) Calculate the product conversion factor using the following formula:

[0042]

[0043] Among them, K i represents the conversion factor of the i-th product; t i represents the unit standard working time of the i-th product; t o represents the standard working hours per unit of product; n represents the number of products.

[0044] (3) Calculate the production capacity of each specific product using the following formula:

[0045]

[0046] Among them, M i represents the production capacity of the i-th product; Q i represents the planned output of the i-th product; d i It represents the proportion of the i-th product in the total product output.

[0047] The production capacity of a production line is expressed by converting all products into corresponding working hours. In the manufacturing process, each production factor plays a certain role. Some production factors, such as the product process, will not change during the manufacturing process, while others, such as the number and type of products in the order being manufactured, will change as the manufacturing process progresses. These two types of production factors are interconnected and jointly affect the production capacity of the production line. In summary, the factors that affect production capacity can be summarized as follows:

[0048] (1) Order data. This includes different orders and the quantity requirements of each order for different types of products:

[0049]

[0050] In the formula, OD represents order data, It represents the demand quantity of the jth order for the ith product. N1 is the total order quantity, and N2 is the total number of product types.

[0051] (2) Working hours data. The time required for the product to be processed in each process:

[0052]

[0053] in, is the processing time of the i-th product in the p-th process, P i is the total number of processes for the ith product, is the total processing time of the ith product, so T represents the working time data.

[0054] (3) Product anomaly data. The frequency of product anomalies is calculated using the total number of product deliveries and the number of anomalies that occurred at each stage of the manufacturing process. Since the collection of product delivery data is relatively rudimentary, sufficient data has been accumulated and the frequency of product anomalies can be used as the probability of product anomalies.

[0055]

[0056] In the formula, Indicates the number of abnormalities that occurred in the history of product i. Represents the total historical delivery quantity of the product, so the ratio of the two is the frequency of historical anomalies of product i. Since a large amount of product delivery data has been accumulated, it is directly used As the probability of abnormality of product i P D Abnormal product data.

[0057] (4) Total machine difficulty. Different orders correspond to different product ratios, and different products have different processing difficulties, which may affect production capacity:

[0058]

[0059] Where N t To end the entire production cycle, The average working time per process of the product is used as the difficulty of the product and sorted. The difficulty of all products is divided into four levels according to the size relationship. The data of the central product of each level of difficulty is used to represent the product of that level as the model difficulty of the product of that level. The model difficulty is used to calculate the total model difficulty of the products to be produced in each production week Therefore, T P It is the collection of total difficulty of each production cycle.

[0060] (5) Start-up data. Including whether the work started on a certain date, and the total number of start-up days of the corresponding production cycle:

[0061]

[0062] D t ={0,1}

[0063]

[0064] Where D t It is a 0-1 variable. When its value is 1, it means that the work will start at time t, otherwise it will not start. Indicates the number of working days of the production cycle from date a to date b.

[0065] Based on the above analysis, using M PD To represent the total capacity, in the statistical production cycle, M PD The relationship between the factors affecting production capacity can be expressed as:

[0066] M PD =f PD (OD,T,P D ,T P ,D)

[0067] In the formula, use f PD To represent the relationship between the data used in production and manufacturing and the production capacity, this relationship is obtained through the training of the production capacity prediction model driven by both data and physics.

[0068] Specifically, Figure 1 A flow chart of a workshop capacity prediction method based on data and physical dual drive provided in an embodiment of the present invention.

[0069] like Figure 1 As shown, the workshop capacity prediction method based on data and physical dual drive includes the following steps:

[0070] Step S1, collect the operation data of the workshop, clean the operation data, perform feature selection in the cleaned operation data based on the correlation with the target prediction variable and combined with physical knowledge, and use the time series data generation method of the generative adversarial network to expand the selected feature data to obtain expanded data.

[0071] Step S2, constructing a data and physics dual-driven capacity prediction model based on the long short-term memory network, wherein the data and physics dual-driven capacity prediction model is obtained by integrating bottleneck knowledge from constraint theory into the neural network prediction model on the basis of capacity prediction using the data-driven neural network prediction model.

[0072] Step S3, extracting physical knowledge related to the bottleneck according to the characteristics of the production line, designing a loss function, inputting the expanded data into the data and physics dual-driven capacity prediction model for training, and obtaining the relationship between the selected features and the actual capacity.

[0073] The basic process of the workshop capacity prediction method based on data and physical dual drive in the embodiment of the present invention is as follows:

[0074] The first step is to collect data on the operation of the workshop through RFID, sensors and other data collection equipment deployed in the production and manufacturing workshop after the digital transformation of an enterprise.

[0075] The second step is to clean the collected data to improve the availability and quality of manufacturing data.

[0076] The third step is to analyze the cleaned production data and select the features that can be used.

[0077] The fourth step is to use data generation technology to expand the data size in order to solve the problem of small size of collected data.

[0078] The fifth step is to extract the physical knowledge related to the bottleneck according to the characteristics of the production line, redesign the loss function, input the selected feature data into the constructed data and physical dual-driven capacity prediction model for training, and find the relationship between the selected features and the actual capacity.

[0079] Finally, the trained prediction model will be applied in practice to provide guidance to production planners and improve the on-time delivery rate of orders.

[0080] Due to the complexity of the digital production line data of the enterprise module integration center, many factors need to be considered comprehensively (such as the machine configuration of the production line, the type of product, the composition of the personnel, etc.). The relationship between these influencing factors is relatively complex, and the dimension of the production data is high, resulting in the low accuracy of traditional physical-driven prediction methods when applied to complex production lines for capacity prediction. Machine learning is a type of method that can effectively solve high-dimensional and nonlinear feature problems, but its performance is also limited by the lack of representation of physical inconsistencies.

[0081] After the digital transformation, the company has collected some data from the operation of the workshop through the deployment of RFID, sensors and other data collection equipment in the production workshop. Since the order types received by the company are small batches of multiple varieties, and the processing time of products in the same process varies greatly, the data accumulation of each process in the current production process is insufficient, which makes it difficult to meet the data-driven model's requirements for the number of training samples, limiting the performance of the data-driven model.

[0082] Therefore, it is also necessary to use data generation methods and adopt the time series data generation method based on generative adversarial network (TimeGAN) to increase the sample size with relatively little impact on the original information of the data.

[0083] Finally, in response to the above pain points, on the basis of using data-driven neural network prediction models for capacity prediction, we consider the knowledge of bottlenecks from constraint theory and integrate it into the neural network prediction model. While mining the information in the data, we make it more consistent with physical laws to reduce physical inconsistency and obtain better performance. In this way, we can explore enough information from the data and physical fields to improve the generalization and robustness of the prediction model. Figure 2 shown.

[0084] In one embodiment of the present invention, before feature selection, the data needs to be cleaned. First, the data needs to be matched. Different data are generally not collected at the same time, so it is necessary to use product numbers, order numbers, machine and personnel numbers to link the various collected data; if the time windows of data statistics are different, only the data of the overlapping part of the time window can be intercepted and used; for repeated data, the redundant duplicates can be directly deleted; for missing values, the average or mode of similar data needs to be used to fill in. Through these methods, the quality and availability of data are improved.

[0085] When companies collect data during the manufacturing process, they often collect a wide variety of data due to the different levels of complexity of production lines and in order to solve various problems that may be encountered later in the manufacturing process. Therefore, when studying specific problems, the original data contains a lot of noise data that is irrelevant to the problem being studied, so it is necessary to use cleaned data and perform feature selection. When performing feature selection, the first thing to consider is the correlation between the selected features and the target prediction variables, and features with higher correlations are given priority; at the same time, physical knowledge needs to be considered. Some features are not statistically significant, but have important physical meanings, and these features also need to be retained. There are generally two types of selected features:

[0086] (1) Features obtained directly from manufacturing data.

[0087] (2) Artificial features added after certain processing of the original production and manufacturing data.

[0088] By combining physical knowledge to process the data in advance and generate some features to add to the model, the complexity of the model can be effectively reduced and the generalization ability of the model can be improved.

[0089] After feature selection is completed, a data set is formed and expressed using the following formula:

[0090]

[0091] Where X is the characteristic data, N c is the number of features.

[0092]

[0093] The predictor variable capacity M corresponds to the feature.

[0094] Data generation based on TimeGAN: When using data-driven models to solve problems, the good performance of the model requires the support of high-quality training sample data, where the number of data is one of the key indicators. If the training samples are insufficient, the data-driven model may not be able to learn the dependencies between various features in the training sample data, causing the data-driven model to fall into a local optimum, resulting in poor performance of the model.

[0095] The enterprise module centralized center has insufficient data accumulation due to the complexity of the production line and insufficient running time of the data collection equipment. Therefore, the data generation method can be used to expand the insufficient number of training samples. Although the data generation method reduces the credibility of the data, it greatly increases the scale of the data. GAN (Generative Adversarial Network) is a deep learning technology proposed by Ian Goodfellow et al. in 2014. It mainly consists of two parts: generator and discriminator.

[0096] Given a training dataset:

[0097] X Train =(X,M)

[0098] Establish the generator G and the discriminator D respectively:

[0099] X Train =G(z,,θ g )

[0100] Where z is random noise, θ g is the generator parameter, and G(z) is the fake data generated by the generator.

[0101] D(X Train ,θ d )

[0102] where θ d is the discriminator parameter, and the discriminator is used to discriminate the real data X Train When the false data G(z) is judged to be true, its value is 1; when it is judged to be false, its value is 0.

[0103] The loss function of the discriminator is:

[0104]

[0105]

[0106] in, P G(z) The real data X Train and the distribution of fake data G(z), x Train ,z s are the real data X Train and samples of fake data G(z), by minimizing L D To train the discriminator, it can better distinguish between real data and fake data.

[0107]

[0108] By minimizing L G To train the generator so that it can generate higher quality fake data, making it difficult for the discriminator to distinguish.

[0109] The two are trained alternately until a balance is reached, where the data generated by the generator is good enough and the discriminator cannot easily distinguish between real and fake data.

[0110] TimeGAN is a model for generating time series data, which combines the advantages of GAN's unsupervised learning and autoregressive model's supervised learning. TimeGAN was proposed by Jinsung Yoon et al. at the NeurIPS2019. The TimeGAN network structure is as follows: Figure 3 As shown,

[0111] TimeGAN adds a time series consistency loss to ensure that the generated data is temporally coherent:

[0112]

[0113] in is the time series generated at (t a t a+1) production cycle data, is the value predicted based on previous data points of the time series.

[0114] The modified generator and discriminator loss functions are:

[0115]

[0116] Here, λ is the weight to balance the two losses.

[0117] Output the data G(z) generated by the generator after training. At this time, the discriminator can no longer distinguish it from the original training data X Train , so it can be used as the expansion data X of the original training data GAN .

[0118] Data-driven prediction based on LSTM:

[0119] After using TimeGAN for data generation, expand the data X GAN It can already meet the requirements of data-driven model prediction. GAN Divide into training sets according to a certain ratio and test set The training set data The input is used in the RNN-based LSTM model to train the capacity prediction model, and then the trained model is used in the test set data. Tests are performed on the model to determine its performance.

[0120] Recurrent Neural Networks (RNN) is a type of neural network suitable for processing sequence data. Unlike traditional feedforward neural networks (FNN), RNN can transfer information between different time steps of a sequence, which enables it to process time-dependent data.

[0121] The working principle of RNN is to use the output of the previous time step as one of the inputs of the current time step, so that the network can consider the information of the previous input while processing the current input. The basic structure of RNN can be expressed as:

[0122] h t =f(x t ,h t-1 )

[0123] y t =g(h t )

[0124] Among them, h tis the hidden state at time step t, x t is the input at time step t, y t is the output at time step t, and f and g are the transition functions of the hidden state and the output state respectively.

[0125] However, RNNs encounter the problem of gradient vanishing or gradient exploding when processing long sequence data, which makes it difficult for the network to learn the dependencies between data.

[0126] In order to solve these problems, LSTM (Long Short-Term Memory) network was proposed, such as Figure 4 As shown in Figure 1, LSTM is a variant of RNN, which controls the flow of information by introducing three gating mechanisms. These three gates are:

[0127] Forget Gate: Determines which information to discard from the current hidden state.

[0128] Input Gate: Determines what new information is added to the current hidden state.

[0129] Output Gate: Determines how much of the current hidden state information the hidden state of the next time step should contain.

[0130] The gating mechanism of LSTM can be expressed as the following formula:

[0131] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0132] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0133] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0134]

[0135] h t =o t *tanh(C t )

[0136] Among them, f t,i t ,o t are the activation values ​​of the forget gate, input gate, and output gate, respectively, and C t is the cell state, is the candidate unit state, W and b are weight and bias parameters, σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function, and * indicates element-wise multiplication.

[0137] The design of LSTM enables it to effectively capture the dependencies between data and has a good prediction effect on sequence data.

[0138] In order to verify the effect of the data-driven model, the mean absolute error (MAE), mean percentage absolute error (MPAE), and root mean square error (RMSE) indicators are used to verify the effect of the trained model on the test set. The definitions are as follows:

[0139]

[0140]

[0141] Among them, n is the number of test samples when testing on the test set, and y i is the actual observation value of the ith is the i-th predicted value. The smaller the value calculated by the three indicators, the higher the accuracy.

[0142] Physical knowledge based on bottlenecks: When using data-driven models such as LSTM to make predictions, since the information obtained comes entirely from the data, there are physical inconsistencies that are inconsistent with the actual situation, resulting in poor prediction results when the data-driven model is making predictions, or although the prediction results are good, when applied to practice, the results are inconsistent with the prediction results.

[0143] Therefore, it is necessary to extract some physical knowledge from the actual situation and add it to the data-driven prediction model to improve the performance of the data-driven model and improve the effect of the data-driven model in actual application. For manufacturing companies, combined with the knowledge of TOC constraint theory, it is generally necessary to find the bottleneck of production. When identifying the bottleneck, it is generally necessary to analyze the data of various resources and conduct field investigations.

[0144] Due to the complex production line structure of the enterprise module integration center, it is impossible to directly identify the bottleneck of the production line through simple steps. We mainly rely on the collected production data combined with domain knowledge to determine the bottleneck of the current production line. For the bottleneck of the production line, its output limits the output of the entire production line, so the load on this resource accounts for a high proportion and has a small fluctuation.

[0145] Decompose historical capacity data by process:

[0146]

[0147] In the formula is the pth process at t a t b The total working hours of the time period, the function Mean() and the function σ() are used to calculate the mean and standard deviation of the data respectively. p Sort and observe the historical working time distribution of the process to determine the bottleneck process.

[0148] Add physical features: Bottlenecks represent resources that are relatively important in the production line, so when using data-driven models for prediction, bottlenecks need to be given a larger proportion. For example, in the feature selection stage, according to the characteristics of the bottleneck, additional bottleneck-related features are selected and added to the model, which can enhance the model's ability to capture key constraints in the manufacturing process and improve the model's interpretability.

[0149] Loss function with added physics knowledge: The standard loss function is used to evaluate the inconsistency between the output value and the actual value in the data-driven model, as shown below:

[0150]

[0151] in is the predicted value, and Y is the actual value. However, the performance of the data-driven model is limited by the limited size of the training data set and the lack of information consistent with the real physical mechanisms of the real world. Therefore, in order to solve this problem, a loss function with added physical knowledge is proposed. By adding the physical laws of the real world to the loss function of the original data-driven model, the prediction results of the data-driven model are made more physically consistent. The modified loss function is shown below:

[0152]

[0153] In regression prediction problems, the loss function generally uses the mean square error (MSE) method. The smaller the MSE, the smaller the difference between the model's predicted value and the actual value, and the better the model performance.

[0154]

[0155] λ phy It is the coefficient of the physical loss function, which is used to change the effect of the physical loss function.

[0156]

[0157] ReLU()=max(0,x)

[0158] In the formula, It is the prediction result after combining physical knowledge. Its value is obtained by combining the bottleneck capacity BN calculated based on physical knowledge with the original prediction value. The meaning is that if the predicted capacity exceeds the bottleneck capacity, that is, it violates the actual physical laws, then its value must be discarded, and the predicted value that does not exceed the bottleneck capacity needs to be retained.

[0159] When discussing the generalization ability of a model in machine learning, two indicators are usually involved: bias and variance. A high bias is manifested as underfitting, which means that the model has poor performance and cannot complete the intended target task well. High variance is manifested as overfitting, that is, the model performs well on the training samples, but performs poorly when replaced with test samples or other data sets. An effective way to solve the high variance problem is to remodel and adjust the structure of the model, but modeling is a complex system engineering, especially when the model is complex, the cost of remodeling is very high. Another way to solve this problem in simpler cases is regularization, which adds prior probabilities to the model from a Bayesian perspective. The general practice of regularization methods is to add additional terms to the loss function, called L1 norm and L2 norm. The L1 norm produces a sparse matrix by resetting the interference values ​​to zero during the feature selection process. The L2 norm adds a regularization term to prevent overfitting, which is a widely used method in deep learning. The mechanism of the L2 norm is as follows:

[0160]

[0161] Where ω is the weight that transforms the output of the previous layer into the input of the activation function of the current layer. At the same time, ω is the solution to this equation to minimize the loss J(ω) with the L2 norm. E in express It represents the training loss when the L2 norm is not added. ω (x i ) is the deep learning model using x i The result calculated according to the parameter ω as input. i is the actual label. is the L2 norm. m is the number of training samples, and λ is the regularization coefficient. The principle of the L2 norm is to assume that the value of λ is reasonable to simplify

[0162] Therefore, the final loss function used is:

[0163]

[0164] ReLU()=max(0,x)

[0165] in, is the standard loss function, and MSE is used to measure the inconsistency between the predicted value and the actual value in the data-driven model. is the physical loss function, is the predicted value, Y is the actual value, It is the prediction result after combining physical knowledge. Its value is obtained by combining the bottleneck capacity BN calculated based on physical knowledge with the original prediction value. Then, we use the commonly used activation function ReLU() to perform nonlinear transformation to obtain λ phy is the coefficient of the physical loss function, which is used to change the effect of the physical loss function, λ reg is the regularization coefficient, N s is the total number of samples, ω is the weight that converts the output of the previous layer into the input of the activation function of the current layer, ω PD are all weights in the proposed data and physical dual-drive model.

[0166] Data- and physics-driven capacity forecasting: The original loss function of the data-driven model is modified into a physical loss function that combines physical knowledge. The training set data is re-entered. For training, use the test set data The predicted results are compared with the original predicted results.

[0167] The workshop capacity prediction method based on data and physical dual drive of the present invention is described in detail below through a specific embodiment.

[0168] Feature Selection:

[0169] The enterprise module integration center needs to control the production capacity M of the entire production line. PD Forecasting is done by referring to the factors affecting production capacity in the mathematical model established above, and according to the various production and manufacturing data collected that are related to production capacity M PD Based on the correlation between them and physical knowledge, the following features are selected and input into the model:

[0170] Features directly obtained from manufacturing data:

[0171] (1) Number of product types. The total number of types of orders currently received.

[0172] (2) The average number of products produced per day of planned production.

[0173] (3) Average working hours per planned working day. These two features represent the number of products in the order and the average value of the corresponding product working hours distributed on each working day after the specific production cycle.

[0174] Artificial features added after some processing of the original manufacturing data:

[0175] (4) The percentage of unfinished products in the previous production cycle.

[0176] (5) The proportion of unfinished product hours in the previous production cycle. Similarly, the number of unfinished products in the previous production cycle and the corresponding product hours account for the proportion of the entire production cycle. Since the company's current order arrangement is unreasonable, resulting in a large number of orders being postponed and the production and manufacturing of the next production cycle, this feature is selected to consider the impact of these orders.

[0177] (6) The average expected number of abnormal products per planned production day.

[0178] (7) The average expected abnormal product working hours per day of planned operation. The probability of abnormality occurring for each product Calculate the number of products that may be abnormal in each production cycle and the corresponding product working hours that will be wasted.

[0179] (8) Total machine difficulty. The difficulty of producing products in each production cycle is calculated by clustering method to measure the impact of different production process difficulties on production capacity M. PD impact.

[0180] Since the data accumulation of the enterprise module integration center is less than one year, the week is selected as the production cycle, and the production capacity in the selected eight features and historical data is used as label data to obtain the original data set with a dimension of (34, 9).

[0181] Data Generation:

[0182] Data of this dimension is obviously not enough for the data-driven model, so the original data set is used as X Train Input it into TimeGAN for training to generate data, and generate data with distribution almost the same as the original data. The parameter setting refers to other papers and problem characteristics, and finally the dimension X is (240, 9) GAN . Comparison between TimeGAN generated data and real data Figure 5 shown.

[0183] Data-driven predictions:

[0184] For the augmented data X after data generation GAN Divide into training sets with dimensions of (210, 9) and (30, 9) and test set And the training set data The data was input into the LSTM data-driven model and trained five times, with 1000 iterations per training. The results of the last five training times were averaged to obtain the results in Table 1:

[0185] Table 1 LSTM prediction results using the original loss function

[0186]

[0187] Physics knowledge:

[0188] The historical working time data M of the enterprise module integration center that has not been generated is decomposed according to the process, totaling 23 types of processes. Their discrete coefficients are calculated and sorted respectively, and the "debugging" and "testing" processes are found to be production bottlenecks. The working time distribution graphs of these two types of processes are observed over time, and their bottleneck capacities are determined to be 1955.8 and 4522.6 respectively. Assuming that the relationship between the bottleneck capacity and the final capacity is linear, the linear regression method is used to obtain the corresponding reasonable total capacity of 23496.12.

[0189] Add physical features:

[0190] Bottlenecks represent resources that are relatively important in a production line. Therefore, when using data-driven models for prediction, bottlenecks need to be given a larger weight. For example, in the feature selection phase, based on the characteristics of the bottleneck, additional features related to the bottleneck are selected and added to the model.

[0191] Based on the bottlenecks identified in the previous article based on physical knowledge, consider adding them to the data-driven model. The first method is to add them as features to the data-driven model. The features that can be added are:

[0192] (1) Consume bottleneck hours. Select the bottleneck corresponding to the time used when identifying the bottleneck. Add to feature set X.

[0193] (2) The proportion of product quantity and working hours that need to pass through the bottleneck. After determining the bottleneck, determine whether each product has passed through the bottleneck, and calculate the proportion of the number of products that pass through the bottleneck in each production cycle to the total number of products and the corresponding proportion of working hours. These two types of data are also added as features to the feature set X.

[0194] Physics-guided loss function:

[0195] According to the previous calculation, the final reasonable total capacity calculated using the bottleneck capacity is 23496.12, so the hyperparameters in the loss function are assigned, and BN=23496.12, the physical loss function coefficient λ phy Set to 1, and the regularization coefficient to 0.1. Change the loss function of the original data-driven model to a physical-guided loss function and re-enter Conduct training.

[0196] Capacity forecast driven by both data and physics:

[0197] The prediction results using the physical guidance loss function are compared with those using the classic loss function. The comparison of various performance evaluation indicators is shown in Table 2:

[0198] Table 2 LSTM prediction results and comparison using physical loss function

[0199]

[0200] According to the workshop capacity prediction method based on data and physical dual drive proposed in the embodiment of the present invention, data of the workshop during operation is collected through data collection equipment such as RFID and sensors deployed in the production and manufacturing workshop of an enterprise after digital transformation. The collected data is cleaned to improve the availability and quality of manufacturing data. The cleaned production data is analyzed to select usable features. In view of the problem that the collected data is small in scale, the data scale is expanded using data generation technology. According to the characteristics of the production line, the physical knowledge related to the bottleneck is extracted, the loss function is redesigned, and the selected feature data is input into the constructed data and physical dual drive capacity prediction model for training to find the relationship between the selected features and the actual capacity. Finally, the trained prediction model is applied to practice to provide guidance for production planners and improve the on-time delivery rate of orders.

[0201] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0202] 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 at least one of the features. In the description of the present invention, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0203] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

Claims

1. A workshop capacity prediction method based on data and physical dual drive, characterized in that: The following steps are involved: Collecting operation data of the workshop, cleaning the operation data, performing feature selection in the cleaned operation data according to the correlation with the target prediction variable and combining physical knowledge, and expanding the selected feature data by using a time series data generation method of a generative adversarial network to obtain expanded data; A data- and physics-driven capacity forecasting model is constructed based on a long short-term memory network, wherein the data- and physics-driven capacity forecasting model is obtained by integrating bottleneck knowledge from constraint theory into the neural network forecasting model on the basis of capacity forecasting using a data-driven neural network forecasting model; According to the characteristics of the production line, physical knowledge related to the bottleneck is extracted, a loss function is designed, and the expanded data is input into the capacity prediction model driven by both the data and physics for training to obtain the relationship between the selected features and the actual capacity.

2. The method according to claim 1, characterized in that Cleaning the operation data includes: If the time windows for running data statistics are different, only the data in the overlapping part of the time windows will be used; For repeated operation data, delete the redundant repeated operation data items; For missing values ​​in running data, the mean or mode of similar data is used to fill them.

3. The method according to claim 1, characterized in that The formula for expanding the selected feature data using the time series data generation method of the generative adversarial network is: in, is the time series generated at (t a t a+1 ) production cycle data, is the value predicted based on previous data points of the time series, N t The end of the entire production cycle; The loss functions of the generator and discriminator are: Among them, λ is the weight to balance the two losses, L D is the loss function of the discriminator of the generative adversarial network, L G is the loss function of the generator of the generative adversarial network.

4. The method according to claim 1, characterized in that According to the characteristics of the production line, the physical knowledge related to the bottleneck is extracted and the loss function is designed as: ReLU()=max(0,x) in, is the standard loss function, and MSE is used to measure the inconsistency between the predicted value and the actual value in the data-driven model. is the physical loss function, is the predicted value, T is the actual value, It is the prediction result after combining physical knowledge. Its value is obtained by combining the bottleneck capacity BN calculated based on physical knowledge with the original prediction value. Then, we use the commonly used activation function ReLU() to perform nonlinear transformation to obtain λ phy is the coefficient of the physical loss function, which is used to change the effect of the physical loss function, λ reg is the regularization coefficient, N s is the total number of samples, ω is the weight that converts the output of the previous layer into the input of the activation function of the current layer, ω PD are all weights in the proposed data and physical dual-drive model.

Images