A forecasting system.

Abstract

This invention relates to at least one training piece (2) which is generated with the characteristics specified by the user and performs predetermined functions, at least one training visual (3) which is the visual format of the training piece (2) and is provided by the user, at least one training matrix (4) in tabular form containing the information of the training visual (3), at least one training test schedule (5) which is predetermined by the user and stores the characteristics of the training piece (2) which includes the production time information, at least one training vector (6) which is a table consisting of at least one column, and at least one processor (7) which converts the training visual (3) into a training matrix (4), and converts the training matrix (4) into a training vector (6) by processing it through convolutional neural network methods specified by the user.

Description

1 TARIFF A PREDICTION SYSTEM This invention involves developing a model that estimates the production time of a part to be manufactured. It relates to a forecasting system that enables predictions to be made using a model. 5 The aviation industry involves high technological requirements and complex designs. In manufacturing processes, the information on how long it will take to produce a new part is usually provided by the part itself. It is not known until it is produced. Knowing the production time of the part allows us to understand the production processes. It plays an important role in planning. Furthermore, the expenditures made in production, 10 Analyzing inefficiencies and costs, simulations for different alternatives. It is extremely important that these processes are carried out and that these outputs are analyzed to make the right decisions. Production data is used to determine the production time, and the production hours are calculated accordingly. High accuracy cannot be achieved in prediction. 15 Chinese patent number CN102521442A, included in the prior art. The document describes the processing time of the neural network for the aircraft structure based on a characteristic example. A method for estimation is being discussed. The present invention relates to a characteristic sample. A method for estimating the processing time of a neural network based on an aircraft structure. and computer-aided design, computer-aided process planning, computer-aided manufacturing 20 It belongs to the technical field. Chinese patent number CN112884246, included in the prior art. in the document; working hour estimation method for aircraft structural element machining process It is mentioned that the variable amount of working hours in the invention is based on a numerical 25. A machining procedure for a workpiece using a controlled machine tool as the leading position. It is stated that working hours are influencing factors of a processing environment and feature complexity. It includes processing environment factors, macroscopic impact factors, and indices processing. This includes materials, processing precision, machine tool type, and similar factors. 30 Thanks to a prediction system developed with this invention, visual data of the parts to be produced is available. and the production of the part by running a predictive model using part specifications An estimated output with a high accuracy rate regarding the duration is obtained. 2 Another aim of this invention is to improve the production times, part characteristics, and visual appearance of manufactured parts. creating a model that allows for the estimation of production time using data It obtains a prediction system. The invention, developed to achieve its purpose, is defined in the initial claim and the requirements dependent on that claim. The prediction system generates data with user-defined specifications and performs specific operations. the training piece that is performed, which is the visualized version of the training piece, and The training visual, provided by the user, contains information in a table format. The training matrix, which includes the user-defined information and the production time information of the training piece, The training test schedule, which includes the characteristics of the training piece, is Table 10, consisting of at least one column. The training vector in this format converts the training visualization into a training matrix, and the training matrix... using convolutional neural network methods specified by the user It includes a processor that processes and converts it into a training vector. The prediction system described in the invention uses data from the processor's training schedule for training 15 The part vector obtained by processing it together with the vector, the part vector is processed by the processor training data in the part vector processed via artificial neural networks coefficients that enable finding the production time of the part, the coefficients that calculate the part The data stored in the vector is mathematically processed (the values ​​are processed using calculated coefficients). (by multiplying and summing the coefficients) determines the time it takes to produce the training piece, and 20 The predictive model generated by the processor, the training given as input to the predictive model. A test piece with different characteristics than the test piece allows for the estimation of information about the test piece. using the model to estimate how long it will take to produce the test piece. It includes the processor that makes it work. 25 In one application of the invention, the prediction system uses a drawing or photograph of the test piece as a reference. the image, the test matrix containing information about the test image, and the user's pre-defined data. and the test schedule containing the characteristics of the test piece, in tabular form with at least one column. The test vector converts the test matrix into a test vector, testing the data contained in the test schedule. The model vector obtained by adding to the vector, the model vector that makes up the prediction 30 using it as input to the model, running the predictive model, processing the model vector, and It includes a processor that estimates the production time of the test piece. In one application of the invention, the prediction system converts the color information of the training image to grayscale. (Using shades of gray expressed with values ​​ranging from 0-255 in the matrix) will be in the range of 35 kept in the training matrix in this way, 3 -The first sub-matrix, which is the matrix containing the distinctive features included in the training visual, is the training The grayscale color information stored in the matrix is ​​processed by a convolutional neural network predefined by the user. Obtaining the first sub-matrix by operating the networks through a filter, -a version stored in the training image that contains less detail compared to the original version of the training image. The second sub-matrix, which contains features (including information such as edges, corners, and holes), is the second sub-matrix. The matrix processes the first sub-matrix through a user-defined convolutional neural network filter. obtaining the second sub-matrix, -Processing the data stored in the second sub-matrix using a convolutional neural network filter defined by the user. converting it into a training vector, -The part vector is obtained by adding the data stored in the training schedule to the training vector. It includes a processor that performs the steps of the process. In one application of the invention, the prediction system uses the part vector and the training vector. with numerous segment neurons containing data stored in the training schedule, segment the processing of neurons using user-defined artificial neural networks, 15 - Determining the coefficients required to obtain the production time of the training piece. -by applying mathematical operations using the identified coefficients in segment neurons The production time of the training part stored in the part vector is determined by processing the held values. the estimation model that includes the steps to obtain the value and the estimation It includes a processor that processes the part vector using the model. 20 In one application of the invention, the forecasting system inputs the model vector into the forecasting model. coefficients that transmit and enable the determination of the production time of the training piece By using mathematical operations, it processes the model vector and produces the test piece. a prediction model that estimates the duration, creating a part vector prediction model 25 It includes a processor that uses data to process and estimate the production time of the test piece. In one application of the invention, the prediction system uses the model vector and the test vector. The test schedule contains data from numerous model neurons, model neurons Using the coefficients determined in the prediction model, the user can select 30 It contains a processor that performs mathematical operations. In one application of the invention, the forecasting system processes the model with the forecasting model. Neurons are deployed in random or predetermined quantities according to user-defined rules. by deleting, the model neuron count is reduced to one, leaving a model 35. The processor determines the value stored in the neuron as the estimated production time result for the test piece. It includes. 4 In one application of the invention, the prediction system processes the training matrix using a kernel filter. creation of the first sub-matrix, -Creating the second sub-matrix by applying a pooling layer to the first sub-matrix, -The steps for applying a flattening layer to the second sub-matrix and creating the training vector are 5. It includes the processor that runs the prediction model. In one application of the invention, the prediction system uses the number of training pieces to determine the training matrix. the dimensions, the kernel filter that processes the training matrix, the pooling layer, and the smoothing The number of (flattening) layers and the number of model neurons to be deleted from the model neurons is 10 processor that uses a predictive model optimized by the user. It includes. In one application of the invention, the predictive system uses the training component, which is connected to the user's computer. It includes an educational visual, which is a photograph of a drawing and / or training piece prepared in the environment. 15 The prediction system implemented to achieve the purpose of this invention is shown in the attached figures. as shown, from these figures; Figure 1 – Schematic view of the forecasting system. 20 Figure 2 – Schematic view of the forecasting system. Figure 3 – Schematic view of the training matrix (4) and the training vector (6). Figure 4 – Schematic view of the steps of the prediction model (M). Figure 5 – Schematic of the prediction model (M) and the steps of the model neurons (141) It is the appearance. 25 The parts in the figures are individually numbered, and their corresponding numbers are shown below. It has been given. 1. Forecasting system 30 2. Training component 3. Educational visual 4. Training matrix 401. First sub-matrix 402. Second sub-matrix 35 5. Training test schedule 5 6. Training vector 7. Processor 8. Part vector 9. Test piece 10. Test image 5 11. Test matrix 12. Test schedule 13. Test vector 14. Model vector 141. Model neuron 10 (C) Coefficient (M) Prediction model 101. Training matrix (4) with color information of the training visual (3) in grayscale. storage on it 102. A first sub-matrix which is the matrix containing the distinctive features of the training visual (3) 15 (401) allows the user to pre-set the grayscale color information stored in the training matrix (4). Creating the first sub-matrix (401) by processing with the determined convolutional neural network filter 103. Matrix containing the features which are a reduced version of the details in the training visual (3). a second submatrix (402) which is pre-defined by the user as the first submatrix (401) Creating the second sub-matrix (402) by processing with the determined convolutional neural network filter 20 104. The data in the second sub-matrix (402) is determined by the user beforehand. converting it into a training vector (6) by processing it with convolutional neural network filter 105. By combining the training vector (6) with the data held in the training schedule (5), the part obtaining the vector (8) 106. The part vector (8) is formed and the training vector (6) is kept in the training schedule (5) 25 Multiple segment neurons (801) containing data, segment neurons (801) by the user processing using predefined artificial neural networks 107. Coefficients required to obtain the production time information of the training piece (2) (C) determination 108. Through mathematical operations using the specified coefficients (C), part 30 Production time of the training piece (2) held in the part vector (8) by processing its neurons (801) obtaining the value (108) 109. Obtaining the second sub-matrix (402) by applying a pooling layer to the first sub-matrix (401). to be done 110. Obtain the training vector (6) by applying a smoothing layer to the second sub-matrix (402) 35 to be done 6 The prediction system (1) is produced with the characteristics specified by the user and in advance At least one training piece (2) that performs the specified functions, the visual of the training piece (2) at least one training visual (3) provided by the user, in the format of training At least one training matrix (4) in tabular form containing information of the visual (3) by the user 5 which stores the properties of the training piece (2) which are predetermined and contain information about the production time. at least one training vector which is a table consisting of at least one column, at least one training test schedule (5) (6), converts the training visual (3) into a training matrix (4), the training matrix (4) by the user by processing the training vector through predetermined convolutional neural network methods (6) It contains at least one processor (7) that converts. 10 The subject of the invention is the prediction system (1), which holds data in the training schedule (5) of the processor (7). by processing the training vector (6), it creates at least one part vector (8), part vector (8) Production of the training part (2) that holds the data in the part vector (8) by processing with artificial neural networks at least one coefficient (C) that enables obtaining the duration, determining the coefficients (C) and the part By processing the data held in the vector (8) with the determined coefficients (C), the production of the training piece (2) 15 at least one created by the processor (7) that enables the obtainment of the duration The prediction model (M) is provided with a training piece (2) as input to the prediction model (M). At least one test piece (9) that is different, with the model (M) predicting information about the test piece (9). It includes a processor (7) that determines the production time of the test piece (9) by processing it. 20 Having predefined characteristics and predetermined features by the user. The training piece (2) that performs the functions is produced. The visual data of the training piece (2) The training visual (3) is provided as input by the user in a computer environment. The information in the training visual (3) is stored in the training matrix (4) in tabular format. The user, Production time, part number, material information, dimension information of the training piece (2) 25 It creates an educational test schedule (5) that includes the features. The table consists of at least one column. The training vector (6) in the format is created. The processor (7) creates the training visual (3) training The processor (7) converts the training matrix (4) into a convolutional matrix (4) as defined by the user. It converts the training vector (6) using neural network methods. In this way, training The training visual (2) of part (3) is converted to matrix format and stored. Training 30 The matrix (4) is processed and converted into a training vector (6) consisting of a single column. (Figure-1, Figure 2) The processor (7) combines the data from the training schedule (5) with the training vector (6) to form the part vector. (8) obtains. The numerical data (5) in the training schedule is directly transferred to the part vector (8) 35 non-numeric data are written, converted into numeric values ​​and part vector (8) It is written. The processor (7) processes the part vector (8) with artificial neural networks and the part 7 obtaining the production time of the training piece (2) stored in the data in the vector (8) The processor (7) determines the coefficients (C) and the part The vector (8) processes the data with the coefficients (C) it determined and the training part (2) training Prediction model (M) that enables obtaining the production time held in the vector (6) It consists of a processor (7) which has not yet been manufactured and has different features from the training part (2). By processing the information of the test piece (9) with the prediction model (M), the test piece (9) It generates an estimate of the production time. In this way, the training vector (6) is in the training schedule. (5) Part vector formed by combining visual data and tabular data by adding the retained data (8) is obtained. The coefficients (C) that determine the production time of the training piece (2) are determined. The prediction model (M) processes the data of the test part (9), which is the new part to be produced, and test 10 It estimates the production time of the part (9). In this way, the visual details of the test part (9) and part features are combined and provided as input to a prediction model (M) and An estimated value for production hours is obtained. (Figure-1) In one application of the invention, the prediction system (1) is based on a drawing or photograph of the test piece (9). test image (10), at least one test matrix (11) containing information of test image (10), user at least one test schedule determined in advance by and including the characteristics of the test piece (9) (12), at least one test vector (13), which is a table consisting of at least one column, test the test matrix (11) converting the data held in the test schedule (12) into a test vector (13) At least one model vector obtained by adding (14), model vector (14) which forms model 20 The vector (14) provides input to the prediction model (M) and the model (M) makes predictions. By processing the model vector (14) via a test piece (9) regarding the production time It contains at least one processor (7) that generates the prediction. In this way, the test image of the test piece (9) By combining the data stored in (10) and the test schedule (12), the test vector (13) is obtained. is being done. 25 In one application of the invention, the prediction system (1) uses grayscale color information of the training visual (3). Storing on the training matrix (4) in a tonal way (101), - a first sub-matrix which is the matrix containing the distinctive features of the training visual (3). (401) grayscale color information held in the training matrix (4) is pre-set by the user 30 Creating the first sub-matrix (401) by processing with the determined convolutional neural network filter (102), - matrix containing the features which are a reduced version of the details in the training visual (3). a second sub-matrix (402) which is the first sub-matrix (401) determined by the user Creating the second sub-matrix (402) by processing with convolutional neural network filter (103), 35 - data in the second sub-matrix (402) is determined by the user beforehand. (104), by processing with convolutional neural network filter and converting into training vector (6), 8 - by combining the data held in the training schedule (5) with the training vector (6) The processor (7) is configured to process the steps (105) of obtaining the vector (8). It includes. In this way, the data kept in the training visual (3) is made less detailed. and data size is saved. 5 In one application of the invention, the prediction system (1) generates the -segment vector (8) and training Multiple segment neurons (801) containing data held in the training schedule (5) with vector (6), using user-defined artificial neural networks of part neurons (801) processing (106), - Coefficients (C) 10 required to obtain the production time information of the training piece (2). determination (107), -part neurons through mathematical operations using the specified coefficients (C) (801) by processing the production time value of the training part (2) held in the part vector (8) is obtained creating the forecasting model (M) which includes the steps (108) and forecasting model (M) It includes a processor (7) configured to process the part vector (8). In this way, training 15 Coefficients (C) and estimation that enable the calculation of the production time information of part (2). The model (M) is determined. The training piece (2) has been produced beforehand and the production time is known. and is stored in the training vector (6). In one application of the invention, the prediction system (1) estimates the model vector (14) 20 providing input to the model (M) and obtaining the production time of the training piece (2). model through mathematical operations using the coefficients (C) that enable By processing its vector (14), it creates an estimate of the production time of the test piece (9). by creating the prediction model (M) and processing the part vector (8) with the prediction model (M) test piece (9) configured to allow the creation of a production time estimate 25 It includes a processor (7). In this way, the test piece (9) can be used by using the prediction model (M). An estimate of the production time is being created. The production time estimate of the test piece (9) is the test Part (9) is determined before it is manufactured. In one application of the invention, the prediction system (1) forms the model vector (14) and test 30 Multiple model neurons (141) containing data held in the test schedule (12) with vector (13), Using the coefficients (C) determined in the prediction model (M) for model neurons (141), the user a processor configured to perform mathematical operations predetermined by the system (7) contains the model vector (14) in which the information of the test piece (9) is stored. It is created and processed with a prediction model (M). 35 9 In one application of the invention, the prediction system (1) is the model applied to the prediction model (M). neurons (141) are set by the user at random or predetermined frequencies reducing the number of model neurons (141) to one by deleting in accordance with the determined rule estimation of production time of test piece (9) of data held in the remaining model neuron (141) It includes a processor (7) that enables the determination of the prediction model (M). This ensures that overshooting is prevented during the development phase. In one application of the invention, the prediction system (1) adds a kernel filter to the training matrix (4). Obtaining the first sub-matrix (401) by applying, - Obtaining the second sub-matrix (402) by applying a pooling layer to the first sub-matrix (401) 10 (109), - Obtaining the training vector (6) by applying a smoothing layer to the second sub-matrix (402) It includes the processor (7) that runs the prediction model (M) which includes steps (110). In this way training matrix (4) will have the details determined by the user in advance It is being simplified. 15 In one application of the invention, the prediction system (1) determines the number of training pieces (2) in the training matrix. (3) dimensions, kernel filter, pooling layer and (4) applied to the training matrix the number of flattening layers and the model neurons to be deleted among the model neurons (141) (141) Optimize the prediction model (M) by predetermining the number by the user. It includes a processor (7) that allows the user to do this beforehand. the optimum number of filters that will allow simplification to have the specified level of detail It is determined. In one application of the invention, the prediction system (1) uses the training piece (2) by the user 25 (2) the computer-assisted visualized version and / or training piece It includes the training visual (3) with a photograph. In this way, the user can draw the training piece (2). or can use its photograph as input in the prediction model (M).

Claims

10 REQUESTS 1. Produced with the specifications determined by the user and pre-defined. At least one training piece (2) that performs the functions, the visual of the training piece (2) at least one training visual (3) provided by the user, in the format of training At least one training matrix (4) in tabular form containing information of the visual (3), user 5 information on the production time of the training piece (2) determined in advance by a training test schedule (5) that stores the features of at least one column table at least one training vector (6) which converts the training image (3) into a training matrix (4), training matrix (4) convolutional neural networks predetermined by the user At least one processor (7) that converts the training vector (6) by processing it through methods, 10 data held in the training schedule (5) of the processor (7) with the training vector (6) at least one part vector (8) created by processing the part vector (8) artificial neural networks by processing the production time of the training part (2) whose data is kept in the part vector (8) at least one coefficient (C) that enables obtaining, determining the coefficients (C) and part By processing the data held in the vector (8) with the determined coefficients (C), the training piece 15 (2) created by the processor (7) which enables the production time to be obtained. at least one forecasting model (M), training provided as input to forecasting model (M) At least one test piece (9) that differs from piece (2), information about test piece (9) The processor determines the production time of the test piece (9) by processing with the estimation model (M). An estimation system characterized by (7) (1). 20 2. The test image (10), which is a drawing or photograph of the test piece (9), is the test image (10) At least one test matrix (11) containing information, predetermined by the user and At least one test chart (12) containing the characteristics of the test piece (9), in at least one column The resulting table is at least one test vector (13), test matrix (11) to test vector (13) 25 The transformation is obtained by adding the data kept in the test schedule (12) to the test vector (13). At least one model vector (14) is obtained, which constitutes the model vector (14), the model vector (14) providing input to the forecasting model (M) and forecasting through the model (M) By processing the model vector (14), a production time for the test piece (9) is determined. a 30 like in request 1 characterized by at least one processor (7) that generates an estimate forecasting system (1).

3. –The training matrix (4) with the color information of the training visual (3) in grayscale. storage on (101), - a first sub-matrix which is the matrix containing the distinctive features of the training visual (3) 35 (401) grayscale color information held in the training matrix (4) by the user 11 by processing the first sub-matrix (401) with a predetermined convolutional neural network filter creation (102), - matrix containing the features which are a reduced version of the details in the training visual (3). a second submatrix (402) which is pre-defined by the user as the first submatrix (401) By processing the second sub-matrix (402) with the determined convolutional neural network filter, 5 creation (103), - data in the second sub-matrix (402) is determined by the user beforehand. (104), by processing with convolutional neural network filter and converting into training vector (6), - by combining the data held in the training schedule (5) with the training vector (6) The processor (7) configured to process the steps (105) of obtaining the vector (8) 10 an estimation system such as in request 1 or 2 characterized by (1).

4. The part vector (8) is formed and kept in the training schedule (5) with the training vector (6). Multiple segment neurons (801) containing data, user of segment neurons (801) processing by using artificial neural networks predetermined by (106), 15 - Coefficients (C) required to obtain the production time information of the training piece (2) determination (107), -part through mathematical operations using the specified coefficients (C) by processing its neurons (801) the production of the training piece (2) held in the piece vector (8). 20 forming the estimation model (M) which includes the steps of obtaining the duration value (108). and processor configured to process the part vector (8) with the prediction model (M) An estimate like any of the above-mentioned requests characterized by (7). system (1).

5. The model vector (14) is provided as input to the prediction model (M) and training 25 coefficients (C) that enable the production time of part (2) to be obtained by processing the model vector (14) through the mathematical operations used, the test the forecasting model (M) which creates an estimate of the production time of part (9) by creating and processing the part vector (8) with the prediction model (M) test part (9) Processor 30 configured to allow for the creation of production time estimates. (7) like any of the claims numbered 2-4 characterized by forecasting system (1).

6. The model vector (14) is formed and kept in the test schedule (12) with the test vector (13). Multiple model neurons containing data (141), predicting model neurons (141) 35 using the coefficients (C) determined in the model (M) by the user beforehand with the processor (7) configured to perform the specified mathematical operations 12 an estimation like in any of the characterized claims 2-5. system (1).

7. The model neurons (141) to which the prediction model (M) is applied are random or pre-selected 5 according to the rules defined by the user, at specified frequencies deleting reduces the number of model neurons (141) to one and the remaining model The estimated production time of the test piece (9) of the data held in the neuron (141) is a request like the one in 6 characterized by the processor (7) that enables its determination forecasting system (1). 10 8. Obtaining the first sub-matrix (401) by applying a kernel filter to the training matrix (4), - Obtaining the second sub-matrix (402) by applying a pooling layer to the first sub-matrix (401) (109), - Obtain the training vector (6) by applying a flattening layer to the second sub-matrix (402). The processor (7) that runs the prediction model (M) which includes the steps (110) and 15 An estimate like any of the characterized claims 3-7 system (1).

9. The number of training pieces (2), the dimensions of the training matrix (4), the training matrix (4) the number of kernel filters, pooling layers and smoothing layers applied and 20 The number of model neurons (141) to be deleted among the model neurons (141) is determined by the user This allows the prediction model (M) to be optimized by being predetermined. in any of the requests numbered 6-8 characterized by the processor (7) providing such a prediction system (1). 25 10. The training piece (2) is prepared by the user using computer assistance. training visual (3) which is a visualized version and / or a photograph of the training piece (2) an estimate like any of the above-mentioned claims characterized by system (1). 30

Images