Real-time control system and method for fused deposition modeling process parameters

Abstract

A system and method for real-time control of fused deposition modeling process parameters. In the system, a printer is used for fused deposition modeling, a camera is directed toward the printer to capture an image of the fused deposition area, a local server is connected to the camera to obtain an input image, a multi-head convolutional network model predicts the printer process parameter state based on the input image and outputs a process parameter state prediction statement, a parameter state judge is connected to the multi-head convolutional network model, the process parameter state prediction statement is stored in the process parameter state prediction sequence in the parameter state judge, and an optimal process parameter state prediction statement is searched for, and then the optimal process parameter state prediction statement and the frequency p of occurrence of the optimal process parameter state prediction statement are input into a proportional corrector to determine the process parameter adjustment amount, a driver unit is connected to the local server and the printer, and the driver unit executes the adjustment command to update the printer process parameters.

Description

Technical Field

[0001] The present invention relates to the field of additive manufacturing technology, and in particular to a system and method for real-time control of fused deposition modeling process parameters. Background Art

[0002] Different from cutting processing characterized by subtractive forming, or forging and casting characterized by equal material forming, additive manufacturing technology characterized by accumulation forming has the characteristics of mold-free rapid free forming, full digitalization, high flexibility, and arbitrary multi-material composite manufacturing. It provides convenience for the manufacture of complex curved structural parts and has been widely used in the fields of automobiles, aerospace, industrial design, etc., and has been hailed by the media as "one of the twelve major technologies of the future economy."

[0003] Based on the characteristics of the process, additive manufacturing technology can be divided into processes such as stereolithography, laser selective melting, laser selective sintering, and fused deposition modeling. Fused deposition modeling is an additive manufacturing process that extrudes materials through a nozzle or orifice. It has the advantages of low cost and almost no post-processing. However, the fused deposition modeling process is affected by various uncertainties and is prone to quality defects such as low precision and poor strength in small parts, hindering the large-scale application of the fused deposition modeling process. How to control the process parameters of the molding process to ensure the quality of the molded parts is a difficult problem that the industry urgently needs to solve.

[0004] Existing control methods are mostly offline, focusing on suppressing a single process defect when a specific printer model is manufacturing a single part. These methods struggle to address the process parameter control requirements of multiple printer models producing parts with different structures in real-world production, and they cannot prevent the generation of multiple types of process defects.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] In response to the problems existing in the prior art, the present invention proposes a real-time control system and method for fused deposition modeling process parameters. Based on a multi-head convolutional network model, the system takes the image of the fused deposition area as input, outputs a process parameter state prediction statement, obtains the optimal process parameter state prediction statement through a parameter state judge, and adjusts the fused deposition modeling process parameters in real time through proportional control based on this statement and its frequency of occurrence.

[0007] The present invention is achieved through the following technical solutions:

[0008] A real-time control system for fused deposition modeling process parameters includes:

[0009] a printer for fused deposition modeling,

[0010] a camera, which is directed toward the printer to capture an image of the fused deposition area,

[0011] A local server is connected to the camera to crop and scale the fused deposition area image and normalize the RGB channels to obtain an input image.

[0012] A multi-head convolutional network model is embedded in the local server. The multi-head convolutional network model predicts the status of the printer process parameters based on the input image and outputs a process parameter status prediction statement.

[0013] A parameter state judger, which is connected to the multi-head convolutional network model to pass the process parameter state prediction statement to the process parameter state prediction sequence in the parameter state judger and find the best process parameter state prediction statement,

[0014] The proportional corrector is connected to the parameter state judger and the local server, and the optimal process parameter state prediction statement and the optimal process parameter state pre-occurrence frequency p are input to the proportional corrector to output the process parameter adjustment amount and send it to the local server.

[0015] The driver unit is connected to the local server and the printer. The local server generates an adjustment command based on the process parameter adjustment amount and sends it to the driver unit. The driver unit executes the adjustment command and updates the printer process parameters.

[0016] In the real-time control system for fused deposition modeling process parameters, when the updating of the printer process parameters is completed, the driver unit sends a completion instruction to the local server.

[0017] In the real-time control system for fused deposition molding process parameters, the printer process parameters include material extrusion rate, nozzle movement speed, nozzle temperature and Z-axis offset distance, the parameter state prediction sequence includes a material extrusion rate prediction sequence, a nozzle movement speed prediction sequence, a nozzle temperature prediction sequence, and a Z-axis offset distance prediction sequence, and the process parameter adjustment amount includes a material extrusion rate adjustment amount, a nozzle movement speed adjustment amount, a nozzle temperature adjustment amount, and a Z-axis offset distance adjustment amount.

[0018] In the real-time control system for fused deposition modeling process parameters, the proportional corrector includes a linear mapping unit and a proportional amplification unit.

[0019] In the real-time control system for fused deposition modeling process parameters, the driver unit is a driver built into the printer.

[0020] The method for real-time control of fused deposition modeling process parameters of the real-time control system for fused deposition modeling process parameters comprises the following steps:

[0021] In the first step, the camera captures an image of the fused deposition area and sends it to a local server;

[0022] In the second step, the local server crops and scales the molten deposition area image with the extrusion nozzle as the center of the image, and normalizes the RGB channels to serve as the input image for the multi-head convolutional network model.

[0023] In the third step, the multi-head convolutional network model predicts the printer process parameter state based on the input image and outputs a process parameter state prediction statement, and passes the prediction statement to the process parameter state prediction sequence in the parameter state judger;

[0024] The multi-head convolutional network model is composed of a Resnet50 network and a fully connected network. The multi-head convolutional network model inputs the cropped original image. The network output layer includes 12 neurons. The neurons in the network output layer are divided into four equal parts, each corresponding to a different process parameter. The three neurons in each part correspond to the "high", "low" and "good" states of the process parameter.

[0025] Among them, the process of the multi-head convolutional network model predicting the process parameter state and outputting the process parameter state prediction statement is as follows: the multi-head convolutional network receives the cropped original image, first uses the Resnet50 network to process the image, extracts the key image features that notice the information associated with the process parameter state in the image, and the output of the Resnet50 network is flattened as the fully connected network input layer. The image information related to the process parameters extracted by the Resnet50 network will be linked to the printer process parameter state through the fully connected network to realize the mapping from image information to the printer process parameter state, that is, output the process parameter state prediction statement.

[0026] In the fourth step, when the process parameter state prediction sequence is completely filled, the parameter state judger searches for the best process parameter state prediction statement and sends the best process parameter state prediction statement and the best process parameter state prediction statement occurrence frequency p to the proportion corrector;

[0027] The process of the parameter state judger searching for the best process parameter state prediction statement is as follows: when a process parameter state prediction sequence is filled, the parameter state judger first calculates the frequency of occurrence of each process parameter state prediction statement in the state prediction sequence, that is, the percentage of the three state prediction statements of "high", "low" and "good" in the process parameter state prediction sequence. The parameter state judger compares the process parameter state prediction statement with the highest frequency in the process parameter state prediction sequence with the threshold θ of the process parameter state prediction sequence. When the frequency of the process parameter state prediction statement with the highest frequency is greater than the sequence threshold, the parameter state judger determines that the prediction statement is the best process parameter state prediction statement for the process parameter under the current state. Otherwise, the best process parameter state prediction statement is determined to be "good".

[0028] In the fifth step, the proportional corrector determines whether to adjust the process parameters based on the optimal process parameter state prediction statement. If it is determined to adjust, the relevant process parameter adjustment amount is output based on the percentage of the optimal process parameter prediction statement in the parameter state prediction sequence. The adjustment process consists of two stages: linear mapping and proportional amplification.

[0029] In the sixth step, after the local server receives the process parameter adjustment value output by the proportional corrector, it sends an adjustment command to the driver unit. The driver unit executes the adjustment command and updates the printer process parameters. When the update is completed, the driver unit sends a completion instruction to the local server.

[0030] In the method for real-time control of fused deposition modeling process parameters, the camera shooting frequency is 2 Hz.

[0031] In the method for real-time control of fused deposition modeling process parameters, during the image shearing, scaling, and normalization process, the image is first cropped to 320×320 ppi and then scaled to 224×224 ppi. The normalization method used is mean-variance normalization.

[0032] In the real-time control method of fused deposition modeling process parameters, the linear mapping process operation is as follows: linearly map the percentage of the optimal process parameter prediction statement in the interval [θ, 1] occupying the parameter state prediction sequence to the interval [I, 1]. The lower limits of the two sets of mapping intervals are both positive values. θ and I perform single process parameter state prediction and proportional control by cutting off the connection between other output heads in the multi-head neural network and the proportional controller.

[0033] In the real-time control method for fused deposition modeling process parameters, the proportional amplification operation is: multiplying the mapping value on the [I, 1] interval by the amplification coefficient ±K to obtain the adjustment amount of the fused deposition modeling process parameters.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] The present invention can realize real-time correction of various process parameters in the molding processes of different printing systems and parts with different structures, reduce problems such as unfused parts and low geometric accuracy of small-sized parts, and ensure the quality of products in the fused deposition molding process. At the same time, the deployment of cameras is low-cost and the control process does not require human intervention, which extends the universal applicability of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are intended only to illustrate preferred embodiments and are not to be construed as limiting the present invention. It should be understood that the drawings described below are merely examples of the present invention, and that those skilled in the art will be able to derive other drawings from these drawings without inventive effort. Throughout the drawings, identical reference numerals are used to denote identical components.

[0037] In the attached figure:

[0038] Figure 1 A flow chart of the method for real-time control of fused deposition modeling process parameters provided by the present invention;

[0039] Figure 2 is a schematic diagram of a printer in this embodiment;

[0040] Figure 3 is a schematic diagram of image adjustment in this embodiment;

[0041] Figure 4 Schematic diagram of a multi-head convolutional network model in this embodiment;

[0042] Figure 5 Schematic diagram of the parameter state judgement device in this embodiment;

[0043] Figure 6 This is a schematic diagram of the scale corrector in this real-time example;

[0044] Figure 7 Schematic diagram of the residual block structure;

[0045] Figure 8 The multi-head convolutional neural network training process.

[0046] The present invention will be further explained below with reference to the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0047] Specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although specific embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0048] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the invention. The scope of protection of the present invention shall be as defined in the attached claims.

[0049] To facilitate understanding of the embodiments of the present invention, further explanation will be given below using specific embodiments as examples in conjunction with the accompanying drawings, and the accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0050] like Figures 1 to 8 As shown, the real-time control system of fused deposition modeling process parameters includes:

[0051] a printer for fused deposition modeling,

[0052] a camera, which is directed toward the printer to capture an image of the fused deposition area,

[0053] A local server is connected to the camera to crop and scale the fused deposition area image and normalize the RGB channels to obtain an input image.

[0054] A multi-head convolutional network model is embedded in the local server. The multi-head convolutional network model predicts the status of the printer process parameters based on the input image and outputs a prediction statement.

[0055] A parameter state judger, which is connected to the multi-head convolutional network model to pass the process parameter state prediction statement to the process parameter state prediction sequence in the parameter state judger and find the best process parameter state prediction statement,

[0056] The proportional corrector is connected to the parameter state judger and the local server, and the optimal process parameter state prediction statement and the optimal process parameter prediction statement occurrence frequency p are input to the proportional corrector to output the process parameter adjustment amount and send it to the local server.

[0057] The driver unit is connected to the local server and the printer. The local server generates an adjustment command based on the process parameter adjustment amount and sends it to the driver unit. The driver unit executes the adjustment command and updates the printer process parameters.

[0058] In a preferred embodiment of the real-time control system for fused deposition modeling process parameters, when the updating of the printer process parameters is completed, the driver unit sends a completion instruction to the local server.

[0059] In a preferred embodiment of the real-time control system for fused deposition molding process parameters, the printer process parameters include material extrusion rate, nozzle movement speed, nozzle temperature and Z-axis offset distance, the parameter state prediction sequence includes a material extrusion rate prediction sequence, a nozzle movement speed prediction sequence, a nozzle temperature prediction sequence, and a Z-axis offset distance prediction sequence, and the process parameter adjustment amount includes a material extrusion rate adjustment amount, a nozzle movement speed adjustment amount, a nozzle temperature adjustment amount, and a Z-axis offset distance adjustment amount.

[0060] In a preferred embodiment of the real-time control system for fused deposition modeling process parameters, the proportional corrector includes a linear mapping unit and a proportional amplification unit.

[0061] In a preferred embodiment of the real-time control system for fused deposition modeling process parameters, the driver unit is a driver built into the printer.

[0062] The method for real-time control of fused deposition modeling process parameters of the real-time control system for fused deposition modeling process parameters comprises the following steps:

[0063] In the first step, the camera captures an image of the fused deposition area and sends it to a local server;

[0064] In the second step, the local server crops and scales the molten deposition area image with the extrusion nozzle as the center of the image, and normalizes the RGB channels to serve as the input image for the multi-head convolutional network model.

[0065] In the third step, the multi-head convolutional network model predicts the printer process parameter state based on the input image and outputs a process parameter state prediction statement, and passes the prediction statement to the process parameter state prediction sequence in the parameter state judger;

[0066] The multi-head convolutional network model is composed of a Resnet50 network and a fully connected network. The multi-head convolutional network model inputs the cropped original image. The network output layer includes 12 neurons. The neurons in the network output layer are divided into four equal parts, each corresponding to a different process parameter. The three neurons in each part correspond to the "high", "low" and "good" states of the process parameter.

[0067] The multi-head convolutional network model predicts process parameter states and outputs a process parameter state prediction statement as follows: The multi-head convolutional network receives the cropped original image and first processes it using the ResNet50 network to extract key image features that indicate information related to the process parameter state. The output of the ResNet50 network is then flattened and used as the input layer of the fully connected network. The image information related to the process parameters extracted by the ResNet50 network is then linked to the printer's process parameter state through the fully connected network, mapping the image information to the printer's process parameter state and outputting a process parameter state prediction statement. The multi-head convolutional network outputs a binary code of 100010001100. When the multi-head convolutional network outputs the process parameter state prediction statement, it first divides the output code into four equal parts, with each sub-code corresponding to a specific process parameter state prediction statement. For example, "100" corresponds to the prediction statement "high," "010" corresponds to the prediction statement "low," and "001" corresponds to the prediction statement "good." Therefore, when the multi-head convolutional network outputs the code 100010001100, the corresponding process parameter state prediction statement is: high material extrusion rate, low nozzle movement speed, good nozzle temperature, and high Z-axis offset distance.

[0068] In the fourth step, when the process parameter state prediction sequence is completely filled, the parameter state judger searches for the best process parameter state prediction statement and sends the best process parameter state prediction statement and the best process parameter state prediction statement occurrence frequency p to the proportion corrector;

[0069] The process of the parameter state judger searching for the best process parameter state prediction statement is as follows: when a process parameter state prediction sequence is filled, the parameter state judger first calculates the frequency of occurrence of each process parameter state prediction statement in the state prediction sequence, that is, the percentage of the three state prediction statements of "high", "low" and "good" in the process parameter state prediction sequence. The parameter state judger compares the process parameter state prediction statement with the highest frequency in the process parameter state prediction sequence with the threshold θ of the process parameter state prediction sequence. When the frequency of the process parameter state prediction statement with the highest frequency is greater than the sequence threshold, the parameter state judger determines that the prediction statement is the best process parameter state prediction statement for the process parameter under the current state. Otherwise, the best process parameter state prediction statement is determined to be "good".

[0070] In the fifth step, the proportional corrector determines whether to adjust the process parameters based on the optimal process parameter state prediction statement. If it is determined to adjust, the relevant process parameter adjustment amount is output based on the percentage of the optimal process parameter prediction statement in the parameter state prediction sequence. The adjustment process consists of two stages: linear mapping and proportional amplification.

[0071] In the sixth step, after the local server receives the process parameter adjustment value output by the proportional corrector, it sends an adjustment command to the driver unit. The driver unit executes the adjustment command and updates the printer process parameters. When the update is completed, the driver unit sends a completion instruction to the local server.

[0072] In a preferred embodiment of the method for real-time control of fused deposition modeling process parameters, the camera shooting frequency is 2 Hz.

[0073] In a preferred embodiment of the method for real-time control of fused deposition modeling process parameters, during the image cropping, scaling, and normalization process, the image is first cropped to 320×320 ppi and then scaled to 224×224 ppi, and the normalization method used is mean-variance normalization.

[0074] In a preferred embodiment of the method for real-time control of fused deposition modeling process parameters, the linear mapping process operates by linearly mapping the percentage of the optimal process parameter prediction statement in the interval [θ, 1] occupying the parameter state prediction sequence to the interval [I, 1]. The lower limits of both mapping intervals are positive. During the control of a single process parameter, θ and I are disconnected from the proportional controller in the multi-head neural network, performing single process parameter state prediction and proportional control, thereby balancing the molding effect and control effect. The molding effect is considered because the printing process contains randomness that does not affect the molding effect. When the sequence threshold θ is small, the parameter state determiner is sensitive to this randomness and may frequently output incorrect statements, disrupting the normal printing process. When the sequence threshold θ is large, the parameter state determiner tends to assume that the process parameters are in good condition and do not require adjustment. This bias can lead to molding failure. The parameter I is selected with full consideration of the control process response. When the proportional coefficient K is determined, a large I value results in a smaller process parameter adjustment range, while a small I value results in a larger process parameter adjustment range. When the frequency of the optimal process parameter state prediction statement changes significantly, an excessively large I value will result in a small adjustment of the process parameters, a weak control effect, and difficulty in meeting timely response requirements; an excessively small I value will result in a large adjustment of the process parameters, causing a large overshoot in the response process, and even causing control system oscillation, leading to molding failure.

[0075] In a preferred embodiment of the method for real-time control of fused deposition modeling process parameters, the proportional amplification operation is: multiplying the mapping value on the [I, 1] interval by the amplification coefficient ±K to obtain the adjustment amount of the fused deposition modeling process parameters.

[0076] In one embodiment, Figure 1As shown, a real-time control system for fused deposition modeling process parameters includes a printer, a camera, a local server, a multi-head convolutional network model, a parameter state judger, a scale corrector and a driver unit. The printer, such as Figure 2 As shown in FIG, the controlled object in this embodiment, the targets to be adjusted for the controlled object are four process parameters: material extrusion rate, nozzle movement speed, nozzle temperature, and Z-axis offset distance; the camera is connected to the above-mentioned printer to obtain images of the printer's fused deposition molding area, and the shooting frequency is 2 Hz; the local server is connected to the above-mentioned camera, and its tasks include receiving images taken by the camera, and performing cropping, scaling, and RGB channel normalization on the images to generate multi-head convolutional network model input, as shown in FIG. Figure 3 In addition, the local server also takes on the role of converting the output of the scale corrector into adjustment instructions and sending them to the driver unit. In the above image cropping, scaling, and normalization process, the image is first cropped to 320×320 ppi and then scaled to 224×224 ppi. The normalization method used is mean-variance normalization.

[0077] The multi-head convolutional network model is implanted in the above local server, such as Figure 4 As shown, the multi-head convolutional network model input is a cut and scaled image of the fused deposition modeling area centered on the nozzle. Its function is to predict the current process parameter state based on the model input graph and output the process parameter state prediction statement; the parameter state judger, such as Figure 5 As shown, the parameter state judger is connected to the above-mentioned multi-head convolutional network model, the input is the process parameter state prediction statement, and the output is the optimal process parameter prediction statement and the frequency p of the optimal process parameter prediction statement; the proportion corrector, such as Figure 6 As shown, it is connected to the above-mentioned state parameter judge, the input is the optimal process parameter state prediction statement and the frequency p of the optimal process parameter prediction statement, and the output is the process parameter adjustment amount. The conversion process between the two includes linear mapping and proportional amplification; the driver unit is connected to the local server, and its function is to execute the adjustment instructions issued by the local server and adjust the fused deposition molding process parameters in real time.

[0078] The parameter state judger includes process parameter state prediction sequences. The number of process parameter state prediction sequences is the same as the number of process parameters to be adjusted. The length and threshold of the process parameter state prediction sequences are obtained through repeated debugging, taking into full consideration the requirements of real-time control and judgment confidence. In this embodiment, the process parameter state prediction sequences included in the parameter state judger are of four types, namely: material extrusion rate prediction sequence, nozzle movement speed prediction sequence, nozzle temperature prediction sequence, and Z-axis offset distance prediction sequence.

[0079] The material extrusion rate prediction state prediction sequence is 16, and the sequence threshold is 0.720;

[0080] The nozzle movement speed prediction state prediction sequence 13, the sequence threshold is 0.700;

[0081] The nozzle temperature prediction state prediction sequence is 10, and the sequence threshold is 0.850;

[0082] The Z-axis offset distance prediction state prediction sequence is 14, and the sequence threshold is 0.700;

[0083] The linear mapping process is operated as follows: the percentage of the optimal process parameter prediction statement in the interval [θ, 1] occupying the parameter state prediction sequence is linearly mapped to the interval [I, 1]. The lower limits of the two sets of mapping intervals are both positive values. θ and I are obtained by cutting off the connection between other output heads in the multi-head neural network and the proportional controller, performing single process parameter state prediction and proportional control, and balancing the molding effect and control effect.

[0084] The proportional amplification process is operated as follows: multiplying the mapping value on the [I, 1] interval by the amplification coefficient ±K to obtain the adjustment amount of the fused deposition modeling process parameter. The positive or negative amplification coefficient is determined according to the parameter state judgment result. When the parameter state judgment result is "low", the amplification coefficient is positive, otherwise it is negative.

[0085] The amplification factor K is limited by the printer itself and is determined by comprehensively considering the applicable range of each process parameter of the printer and the I value in the linear mapping process.

[0086] In one embodiment, a method for real-time control of fused deposition modeling process parameters includes the following steps:

[0087] In the first step, an image of the molten deposition area is captured by a camera and sent to a local server;

[0088] In the second step, the image of the molten deposition area is cropped and scaled with the extrusion nozzle as the image center, and the RGB channels are normalized to obtain the input of the multi-head convolutional network model;

[0089] In the third step, the multi-head convolutional network model predicts the printer process parameter state based on the input image, outputs a process parameter state prediction statement, and passes the process parameter state prediction statement to the process parameter state prediction sequence in the parameter state judger;

[0090] In the fourth step, when the process parameter state prediction sequence is filled, the parameter state judge will find the best process parameter state prediction statement, and send the best process parameter state prediction statement and the best process parameter state statement occurrence frequency p to the proportion corrector;

[0091] In the fifth step, the proportional corrector determines whether to adjust the process parameters based on the optimal process parameter state prediction statement. If it is determined to adjust, it outputs the relevant process parameter adjustment amount based on the occurrence frequency p of the optimal process parameter state prediction statement. The adjustment process consists of two stages: linear mapping and proportional amplification.

[0092] In the sixth step, after the local server receives the process parameter adjustment value output by the proportional corrector, it sends an adjustment command to the driver unit. The driver unit executes the adjustment command and updates the printer process parameters. When the update command is executed, the driver unit sends a completion instruction to the server.

[0093] In the method described, steps one, two, three, four, five, and six constitute a complete real-time control process.

[0094] In the method, in the third step, the prediction statement includes "high", "low", and "good".

[0095] In the method, in the third step, the clarity of the captured image is positively correlated with the process parameter judgment ability of the multi-head convolutional network model.

[0096] In the method, in the third step, the number of heads of the multi-head convolutional network model is the same as the number of controlled process parameters.

[0097] In the method, in the third step, the number of process parameter state prediction sequences in the parameter state determiner is equal to the number of process parameters to be adjusted.

[0098] In the method, in the third step, the length of the parameter state prediction sequence needs to comprehensively consider the response time settings of the printer firmware and hardware. A shorter state prediction sequence may cause the parameter adjustment frequency to exceed the response time of the printer firmware and hardware, causing oscillation of the printer process parameters, which is not conducive to the smooth progress of the fused deposition modeling process; an excessively long state prediction sequence affects the real-time adjustment of the process parameters.

[0099] In the method, in the fourth step, the optimal process parameter state prediction statement is the prediction statement that occupies the largest percentage of the process parameter state prediction sequence length and is greater than the sequence threshold, and the frequency of occurrence is the percentage of the prediction statement that occupies the process parameter state prediction sequence length. If there is no prediction statement that occupies a percentage of the parameter state prediction sequence length that is greater than the sequence threshold, the optimal process parameter prediction statement is considered to be "good".

[0100] In the above method, in the fourth step, the process parameter state prediction sequence threshold is obtained by debugging.

[0101] In the method, in the fifth step, if the optimal process parameter state prediction statement received by the proportional corrector is "good", no adjustment is applied to the process parameters, and the process parameters are maintained to continue molding.

[0102] In the method, in the fifth step, if the optimal process parameter state prediction statement received by the proportional corrector is "high" or "low", the process parameters are adjusted.

[0103] In the method, in the fifth step, the linear mapping stage operation in the adjustment process is: linearly map the frequency p of the optimal process parameter prediction statement located in the interval [θ, 1] to the interval [I, 1], and the lower limits of the two sets of mapping intervals are both positive values. In the process of single process parameter control, the connection between other output heads in the multi-head neural network and the proportional controller is cut off to perform single process parameter state prediction and proportional control, and the molding effect and control effect are weighed.

[0104] In the method, in the fifth step, the proportional amplification stage operation in the adjustment process is: multiply the mapping value on the [I,1] interval by the amplification coefficient ±K to obtain the adjustment amount of the fused deposition modeling process parameter. The positive or negative amplification coefficient is determined according to the parameter state judgment result. When the optimal process parameter state prediction statement is "low", the amplification coefficient is positive, otherwise it is negative.

[0105] In the method, in step 6, considering the response time of the printer firmware and hardware when executing commands, the parameter update must be performed after the driver has completed executing the last parameter adjustment command, otherwise it will cause oscillation of the printer process parameters.

[0106] The residual neural network for detecting the forming process consists of a shared backbone network and four output heads, which correspond to four process parameters, namely material extrusion rate, nozzle movement speed, nozzle temperature, and Z-axis offset distance. The network structure diagram is shown in the figure. Figure 7 shown.

[0107] The shared backbone network used to extract image features consists of six residual blocks and three attention models. Each residual block includes a convolutional layer, batch normalization, activation function activation, and pooling layer. The residual block structure diagram is shown in the figure below. Figure 8 As shown in Figure 2, the ReLu function is used to activate the output of the convolutional layer after batch normalization, and the activation value is sent to the next layer.

[0108] After extracting feature information from the forming process image using a shared backbone network consisting of six serially connected residual blocks, its output is flattened into a column vector. A fully connected layer then maps the image features to the four process parameter states. This means that the four output heads share the feature information extracted by the shared backbone. During training, the weights of the shared backbone network are updated using the total loss of the four output heads. This means that the shared backbone network can learn the interactions between process parameters and the relationship between each process parameter and the forming state image, allowing it to propose multiple process parameter adjustment methods for a single undesirable forming state.

[0109] In order to obtain an effective and stable multi-head convolutional network to predict accurate process parameter states, the training process of the multi-head convolutional network is divided into three parts, such as Figure 8 As shown in the figure, the network is first trained using images from the training set that display distinct process parameter state features. Due to these distinct state features, the multi-head convolutional neural network can effectively and quickly capture image information and detect important features. Next, the pre-trained multi-head convolutional neural network model is fine-tuned on the full training set to improve its state recognition capabilities for images with less distinct features. Finally, considering the imbalance of the training set samples—that is, due to the influence of forming characteristics, excessively low Z-axis offsets can lead to nozzle-heatbed collisions and forming failures—the number of samples with excessively high Z-axis offsets far exceeds those with excessively low offsets. To mitigate the impact of sample imbalance on the predictive performance of the multi-head convolutional neural network, a transfer learning approach is used in the final stage of training to freeze the shared backbone network weights and fine-tune the fully connected layer weights using balanced samples, resulting in improved process parameter state prediction results. Finally, the accuracy of the trained model is verified on the test set, where the classification accuracy for each parameter is 87.1% for flow rate, 86.4% for lateral velocity, 85.5% for Z-offset, and 78.3% for hot-end temperature.

[0110] Although the embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments and application fields. The above-mentioned specific embodiments are merely illustrative and instructive, and are not restrictive. A person skilled in the art, guided by this specification and without departing from the scope of protection of the claims of the present invention, may also devise various forms, all of which fall within the scope of protection of the present invention.

Claims

1. A real-time control system for fused deposition modeling process parameters, characterized in that: These include, a printer for fused deposition modeling, a camera, which is directed toward the printer to capture an image of the fused deposition area, A local server is connected to the camera to crop and scale the fused deposition area image and normalize the RGB channels to obtain an input image. A multi-head convolutional network model is implanted in the local server. The multi-head convolutional network model predicts the process parameter state of the printer based on the input image and outputs a process parameter state prediction statement. A parameter state judger is connected to the multi-head convolutional network model to store the process parameter state prediction statement into the process parameter state prediction sequence in the parameter state judger, and find the best process parameter state prediction statement. The proportional corrector is connected to the parameter state judger and the local server, and inputs the optimal process parameter state prediction statement and the frequency of occurrence of the optimal process parameter state prediction statement. p , output the process parameter adjustment amount and send it to the local server, A driver unit is connected to the local server and the printer, the local server generates an adjustment command based on the process parameter adjustment amount and sends it to the driver unit, and the driver unit executes the adjustment command to update the printer process parameters; Among them, the printer process parameters include material extrusion rate, nozzle movement speed, nozzle temperature and Z-axis offset distance, the parameter state prediction sequence includes material extrusion rate prediction sequence, nozzle movement speed prediction sequence, nozzle temperature prediction sequence, Z-axis offset distance prediction sequence, and the process parameter adjustment amount includes material extrusion rate adjustment amount, nozzle movement speed adjustment amount, nozzle temperature adjustment amount, and Z-axis offset distance adjustment amount.

2. A real-time control system for fused deposition modeling process parameters according to claim 1, characterized in that: When the updating of the printer process parameters is completed, the driver unit sends a completion instruction to the local server.

3. A real-time control system for fused deposition modeling process parameters according to claim 1, characterized in that: The scale corrector includes a linear mapping unit and a scale amplification unit.

4. A real-time control system for fused deposition modeling process parameters according to claim 1, characterized in that: The driver unit is a driver built into the printer.

5. A method for real-time control of fused deposition modeling process parameters according to the real-time control system for fused deposition modeling process parameters according to any one of claims 1 to 4, characterized in that: The steps include: In the first step, the camera captures an image of the fused deposition area and sends it to a local server; In the second step, the local server crops and scales the molten deposition area image with the extrusion nozzle as the center of the image, and normalizes the RGB channels to serve as the input image for the multi-head convolutional network model. In the third step, the multi-head convolutional network model predicts the printer process parameter state based on the input image and outputs a process parameter state prediction statement, and passes the prediction statement to the process parameter state prediction sequence in the parameter state judger; The multi-head convolutional network model consists of a Resnet50 network and a fully connected network. The multi-head convolutional network model inputs the cropped original image. The network output layer includes 12 neurons. The neurons in the network output layer are divided into four equal parts, each corresponding to a different process parameter. The three neurons in each part correspond to the "high", "low" and "good" states of the process parameter. Among them, the process of the multi-head convolutional network model predicting the process parameter state and outputting the process parameter state prediction statement is as follows: the multi-head convolutional network receives the cropped original image, first uses the Resnet50 network to process the image, extracts the key image features that notice the information associated with the process parameter state in the image, and the output of the Resnet50 network is flattened as the fully connected network input layer. The image information related to the process parameters extracted by the Resnet50 network will be linked to the printer process parameter state through the fully connected network to realize the mapping from image information to the printer process parameter state, that is, output the process parameter state prediction statement. In the fourth step, when the process parameter state prediction sequence is fully filled, the parameter state judger searches for the best process parameter state prediction statement and compares the best process parameter state prediction statement with the best process parameter state prediction statement occurrence frequency. p Send to the proportional corrector; Among them, the process of the parameter state judger searching for the best process parameter state prediction statement is as follows: when a process parameter state prediction sequence is filled, the parameter state judger first calculates the frequency of occurrence of each process parameter state prediction statement in the state prediction sequence, that is, the percentage of the three state prediction statements of "high", "low" and "good" in the process parameter state prediction sequence, and the parameter state judger compares the process parameter state prediction statement with the highest frequency in the process parameter state prediction sequence with the threshold value of this process parameter state prediction sequence. When the frequency of the process parameter state prediction statement with the highest frequency is greater than the sequence threshold, the parameter state judger determines that the prediction statement is the best process parameter state prediction statement for the process parameter under the current state. Otherwise, the best process parameter state prediction statement is judged as "good"; In the fifth step, the proportional corrector determines whether to adjust the process parameters based on the optimal process parameter state prediction statement. If it is determined to adjust, the relevant process parameter adjustment amount is output based on the percentage of the optimal process parameter prediction statement in the parameter state prediction sequence. The adjustment process consists of two stages: linear mapping and proportional amplification. In the sixth step, after the local server receives the process parameter adjustment value output by the proportional corrector, it sends an adjustment command to the driver unit. The driver unit executes the adjustment command and updates the printer process parameters. When the update is completed, the driver unit sends a completion instruction to the local server.

6. The method for real-time control of fused deposition modeling process parameters according to claim 5, characterized in that: The camera capture frequency is 2 Hz.

7. The method for real-time control of fused deposition modeling process parameters according to claim 5, characterized in that: During the image cropping, scaling, and normalization process, the image was first cropped to 320×320 ppi and then scaled to 224×224 ppi. The normalization method used was mean-variance normalization.

8. The method for real-time control of fused deposition modeling process parameters according to claim 5, characterized in that: The linear mapping process is as follows: The percentage of the best process parameter prediction statement in the interval that occupies the parameter state prediction sequence is linearly mapped to the interval , the lower limits of the two mapping intervals are both positive values, and I By cutting off the connection between other output heads in the multi-head neural network and the proportional controller, single process parameter state prediction and proportional control are performed.

9. The method for real-time control of fused deposition modeling process parameters according to claim 8, characterized in that: The proportional amplification operation is: The mapped value on the interval is multiplied by the magnification factor , and obtain the adjustment amount of the fused deposition modeling process parameters.

Images