KR20230038839A - 3D printer monitoring system applying machine learning algorithm based on Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy - Google Patents

Abstract

According to the present invention, a 3D printer monitoring system applying a machine learning algorithm comprises: a 3D printer using a fused deposition modeling (FDM) method which generates a 3D output by discharging molten filament through a nozzle to form a plurality of layer; a thermocouple sensor unit which detects a temperature of the nozzle and outputs a detected temperature value as a digital value; a data collector which stores data output from the thermocouple sensor unit; and a control computer which receives data transmitted from the data collector and predicts the temperature value of the nozzle by applying a machine learning algorithm based on a temporal convolution neural network with two-stage sliding window strategy (TCN-TS-SW). Accordingly, deterioration in quality of a result can be reduced, and furthermore, a predicted temperature value can be provided accurately.

Description

3D printer monitoring system applying machine learning algorithm based on Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy}

The present invention relates to a 3D printer monitoring system, and more particularly, to a 3D printer monitoring system to which a machine learning algorithm based on TCN-TS-SW is applied.

In general, in order to manufacture a prototype having a three-dimensional three-dimensional shape, there are a wooden composite manufacturing method performed by hand depending on drawings and a manufacturing method by CNC milling. However, since the wooden composite manufacturing method is manual, precise numerical control is difficult and takes a lot of time, and the CNC milling manufacturing method enables precise numerical control, but there are many shapes that are difficult to process due to tool interference.

Therefore, recently, a so-called 3D printer method has appeared in which a prototype of a 3D solid shape is manufactured using a computer storing data generated from 3D modeling created by product designers and designers.

In this 3D printer method, SLA (StereoLithograhhic Apparatus) using the principle that a laser beam is injected into a photocurable resin and the scanned part is hardened, and a functional polymer or metal powder is used instead of a photocurable resin in SLA and laser beam is used. SLS (Selective Laser Sintering), which uses the principle of molding by scanning and solidifying, and LOM (Laminated Object Manufacturing), which cuts adhesive-coated paper into desired cross-sections using laser beams and laminates them one by one, , BPM (Ballistic Particle Manufacturing) using ink-jet printer technology, and FDM (Fused deposition modeling), which uses heated nozzles to melt molding materials and build them up in layers.

Filament, a raw material of FDM (Fused deposition modeling) method 3D printer, is used by processing heat-melting material into a thin thread form and winding it around a spool. ) is composed of a carrier that mounts a nozzle for melting and spraying and moves it to a printing position, and a bed that loads and moves the output, and the filament wound on the spool is continuously transported through a feeder and injected into the nozzle. The injected filament becomes a liquid state by the heat generated from the nozzle and is sprayed out of the nozzle and accumulated on the bed for loading the printed material. operated to form

That is, in the FDM (Fused Deposition Modeling) method, which is widely used in popular 3D printers, a filament, which is a thermoplastic resin, is injected into a heated nozzle and the molten filament is discharged at a certain amount. A small amount of the discharged filament is cured at room temperature, and a next layer is formed thereon to form a three-dimensional output product. In order to obtain a uniform discharge amount, a heater block is installed on the nozzle and heated to maintain a constant temperature.

In these FDM-type 3D printers, not only how precise the amount ejected from the nozzle is, but also how well the ejected filament is cured is an important factor in the quality of the printed product. Therefore, it is very important to keep the temperature of the nozzle discharge part constant.

KR 10-2017-0078396 A

The present invention is proposed to solve the above technical problems, and to improve the 3D printing technology of the FDM (Fused deposition modeling) method, TCN-TS-SW (Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy)-based Provides a 3D printer monitoring system to which machine learning algorithms are applied.

According to an embodiment of the present invention for solving the above problems, a 3D printer of a fused deposition modeling (FDM) method that generates a three-dimensional output while forming a plurality of layers by discharging a molten filament through a nozzle, and a nozzle A thermocouple sensor unit that detects temperature and outputs the detected temperature value as a digital value, a data collector that stores data output from the thermocouple sensor unit, and TCN-TS-SW (Temporal Convolution A 3D printer monitoring system applying a machine learning algorithm including a control computer for predicting the temperature value of the nozzle by applying a machine learning algorithm based on a Neural Network with Two-Stage Sliding Window Strategy is provided.

In addition, the control computer included in the present invention, in dividing the input data through the sliding window technique twice for data preprocessing, proceeds with the first sliding using the time window Δ and the sliding step δ, and the prediction interval h It is characterized in that the used second sliding is performed.

In addition, the δ and Δ values included in the present invention are characterized in that they are selected in consideration of the sampling capacity of the data collector.

In addition, the TCN-TS-SW (Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy) included in the present invention uses a one-dimensional causal convolution, but uses a residual block so that the predicted value depends only on the history value of the series It is characterized in that it is configured to.

The 3D printer monitoring system to which the machine learning algorithm according to the embodiment of the present invention is applied is based on TCN-TS-SW (Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy) to improve the 3D printing technology of the FDM (Fused deposition modeling) method. A 3D printer monitoring system applying a machine learning algorithm based on this was proposed.

That is, using the proposed machine learning algorithm, it is possible to reduce the quality degradation of the result due to the decrease in the output temperature of the 3D printer, and furthermore, it is possible to accurately provide the predicted temperature value.

1 is a block diagram of a 3D printer monitoring system 1 to which a machine learning algorithm is applied.
Figure 2 is a diagram showing the process of the monitoring program of the 3D printer monitoring system 1 to which a machine learning algorithm is applied
3 is a flow chart of a monitoring program
Figure 4 is a LabVIEW block diagram of a program for data collection and storage
5 is an exemplary view of a dataset stored as a program for data collection and storage;
6 is an exemplary diagram of a user interface of a data collection and storage program;
7 is an exemplary diagram of a sliding window technique of steps
8 is a structural diagram of TCN used in the present invention (Rb = 10, rf = 8, db = 2, k = 3)
9 is a diagram showing actual data (black) and predicted data (blue)
10 is a diagram showing the accuracy of two models,
11 is a diagram showing loss rates of two models

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings in order to describe in detail enough for those skilled in the art to easily implement the technical idea of the present invention.

1 is a configuration diagram of a 3D printer monitoring system 1 to which a machine learning algorithm is applied.

Referring to FIG. 1, a monitoring system to which a machine learning algorithm based on TCN-TS-SW (Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy) is applied to improve 3D printing technology of FDM (Fused deposition modeling) method is shown in FIG. is composed as

The 3D printer monitoring system 1 to which a machine learning algorithm is applied includes a 3D printer 100 of a fused deposition modeling (FDM) method, a thermocouple sensor unit 200, a data collector 300, and a control computer 400. do.

That is, the hardware of the 3D printer monitoring system (1) to which the machine learning algorithm is applied includes the FDM printer (100), K-thermocouple to Digital converter (MAX6675) - thermocouple sensor unit (200) -, data collector (Arduino Uno, 300), It is composed of a control computer 400.

The thermocouple sensor mounted on the nozzle of the 3D printer 100 can measure up to 350 degrees (Celsius) and can withstand the high temperature of the nozzle. The temperature value output from the thermocouple sensor unit 200 is converted using a module (MAX6675) that generates data in 12-bit, 0.25 degree (Celsius) units, and is transmitted to the data collector 300. Data transmitted to the data collector 300 is stored in a (.csv) format in the control computer 400.

Figure 2 is a diagram showing the process of the monitoring program of the 3D printer monitoring system 1 to which the machine learning algorithm is applied, Figure 3 is a flow chart of the monitoring program, Figure 4 is a LabVIEW block diagram of the program for data collection and storage.

- Data collection and storage

The nozzle temperature of a 3D printer has a great influence on the quality of the output. Therefore, the proposed monitoring system collects and stores the temperature value of the nozzle. To this end, the LabVIEW program was used in this embodiment. When the temperature value of the nozzle is entered, the data is processed and saved through this value.

5 is an exemplary view of a dataset stored as a program for data collection and storage. In the dataset, time, temperature, humidity, and distance are expressed as voltage.

- Data display (display)

6 is an exemplary view of a user interface of a data collection and storage program.

The input data is displayed to the user, and through this, the user can easily check the nozzle temperature value while 3D printing is in progress. In addition, when an abnormality is detected in the temperature value predicted through machine learning, the result is displayed so that the user can proceed with a response to maintain or improve quality. 6 shows a user interface (UI) of a data collection and storage program. The measured temperature, humidity and distance are displayed (displayed) through the UI (User Interface).

- Proposed Machine Learning Algorithm (TCN-TS-SW)

For the 3D printer monitoring system of the present invention, TCN-TS-SW (Temporal Convolutional Neural Network with a Two-Stage Sling Window strategy) is proposed. The proposed technique can be processed through the control computer 400.

The proposed method is based on TCN (Temporal Convolution Neural Network), but existing TCN uses a lot of memory for prediction because it requires the entire input data as input to the next layer compared to RNN. In the present invention, a sliding window technology capable of data pre-processing is applied to reduce memory usage and prediction time.

- Data pre-processing using sliding technique

7 is an exemplary diagram of a sliding window technique of steps <a) first sliding using time window Δ and sliding step δ, (b) second sliding using prediction interval h>.

For data preprocessing, input data is divided through two sliding window techniques. Through the first sliding window technique, as shown in FIG. At this time, the values of δ and Δ are selected in consideration of the sampling capacity and printing speed of the data collection module.

The split sequence transforms x into a 2D vector consistent with a predefined level of unfolding.

The second sliding can be seen in Figure 7(b). The second sliding is performed after the data is divided into training and validation sets. At this time, in order to define the prediction parameter h calculated through the movement difference between the input series of length m and the target series, the previously slid data is additionally divided as shown in FIG. 7(b). The prediction parameter h reduces errors and enables better prediction performance. After the second sliding window is performed, segment m is processed in the predictive model. The processed data is used for predictive models.

- TCN-based prediction model

8 is a structural diagram of TCN used in the present invention (Rb = 10, rf = 8, db = 2, k = 3).

TCN-TS-SW consists of a residual block composed of stacked dilated convolution layers, non-linearity functions, and dropout layers used in TCN. The structure of the TCN used is shown in FIG. 8.

Instead of using a one-dimensional causal convolution, we use a residual block so that the predicted value depends only on the historical value x of the series. In order to keep the same length between input and output, zero padding is added to the input to ensure length uniformity in each layer, and this is also applied to successive layers of dilated convolutional layers in the residual block.

Since dilated layers increase the receptive field rf of the input while keeping the number of layers relatively small, dilated convolution layers replace the existing convolutional layers. The output ζ of the causal convolution layer within the residual block is normalized using weight regularization and further transformed to learn a nonlinear representation of the data using the Rectified Linear Unit (ReLU) function. A dropout layer is introduced at the end of the convolution to prevent the model from overfitting. In the final output of the model, ReLU is disabled so that y takes negative values. Finally, after the network obtains a set of predicted values, the output y, the 'predicted value', is displayed on the graphical user interface along with the actual data as shown in FIG. 9 . - Figure 9 shows actual data (black) and predicted data (blue) -

- Performance of TCN-TS-SW

10 is a diagram showing the accuracy of two models, and FIG. 11 is a diagram showing loss rates of the two models.

It can be seen that TCN-TS-SW (Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy) introducing a two-stage sliding technique has higher accuracy and lower loss rate than conventional TCN.

The 3D printer monitoring system to which the machine learning algorithm according to the embodiment of the present invention is applied is based on TCN-TS-SW (Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy) to improve the 3D printing technology of the FDM (Fused deposition modeling) method. A 3D printer monitoring system applying a machine learning algorithm based on this was proposed.

That is, using the proposed machine learning algorithm, it is possible to reduce the quality degradation of the result due to the decrease in the output temperature of the 3D printer, and furthermore, it is possible to accurately provide the predicted temperature value.

As such, those skilled in the art to which the present invention pertains will be able to understand that the present invention may be embodied in other specific forms without changing its technical spirit or essential features. Therefore, the embodiments described above should be understood as illustrative in all respects and not limiting. The scope of the present invention is indicated by the following claims rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

100: 3D printer
200: thermocouple sensor unit
300: data collector
400: control computer

Claims (4)

A 3D printer of a fused deposition modeling (FDM) method that generates a three-dimensional output while forming a plurality of layers by discharging molten filament through a nozzle;
a thermocouple sensor unit that senses the temperature of the nozzle and outputs the detected temperature value as a digital value;
a data collector to store data output from the thermocouple sensor unit; and
a control computer that receives the data transmitted from the data collector and predicts the temperature value of the nozzle by applying a machine learning algorithm based on TCN-TS-SW (Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy);
3D printer monitoring system to which machine learning algorithms are applied.
According to claim 1,
The control computer,
In dividing input data through two sliding window techniques for data preprocessing, first sliding is performed using a time window △ and a sliding step δ, and second sliding is performed using a prediction interval h. Characterized in that 3D printer monitoring system applying machine learning algorithm.
According to claim 2,
The δ and Δ values are 3D printer monitoring systems applying a machine learning algorithm, characterized in that selected in consideration of the sampling capacity of the data collector.
According to claim 1,
TCN-TS-SW (Temporal Convolution Neural Network with Two-Stage Sliding Window Strategy),
A 3D printer monitoring system applying a machine learning algorithm, characterized in that the prediction value is configured to depend only on the history value of the series using a one-dimensional causal convolution, but using a residual block.

Images