Flow monitoring method and system for wire feeding mechanism of 3D printer

Abstract

The invention relates to the technical field of 3D printing equipment monitoring, and discloses a flow monitoring method and system for a wire feeding mechanism of a 3D printer, and the method comprises the steps: collecting current signal data of a wire feeding motor drive unit, and carrying out the preprocessing of the current signal data; performing time-frequency analysis on the standardized current characteristic data by using a multi-scale signal analysis algorithm, outputting a high-dimensional current characteristic set, and establishing a characteristic fault mapping relation; a multi-model fault analysis system is adopted to process the high-dimensional current feature set, and fault analysis and positioning are carried out; a progressive three-level early warning mechanism and flow monitoring are realized, and a current flow mapping model is established; continuously optimizing the flow model and the fault identification model, and establishing a dynamically updated fault feature library; wire feeding flow monitoring can be achieved without an additional sensor, and system cost and complexity are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of 3D printing equipment monitoring, and more specifically, to a flow monitoring method and system for a wire feeding mechanism of a 3D printer. Background Art

[0002] In the field of 3D printing, the stability of the wire feed mechanism directly impacts print quality and efficiency. Existing monitoring technologies primarily rely on adding dedicated sensors to monitor wire feed status, which not only increases system complexity and cost but can also interfere with the normal wire feeding process. Furthermore, traditional monitoring methods suffer from delayed response, a lack of early warning, and difficulty accurately locating faulty components. During long printing processes, wire feed anomalies cannot be detected and located in a timely manner, leading to print failures, material waste, and equipment damage, severely impacting production efficiency.

[0003] At present, the common wire feeding monitoring solutions on the market mainly include the following categories: First, the displacement monitoring method based on the photoelectric encoder, which calculates the wire feeding amount by installing an encoder on the wire feeding wheel, but this method is easily affected by wire slippage and produces errors; the second is the quality monitoring method based on the weight sensor, which calculates the wire feeding amount by measuring the weight change of the consumables, but this method has a slow response speed and is difficult to monitor in real time; the third is the non-contact measurement method based on infrared or laser, which has high accuracy but is expensive and complicated to install.

[0004] These methods generally have problems such as low integration, high maintenance costs, and insufficient reliability. Especially in industrial-grade 3D printing application scenarios, printing time is usually as long as dozens of hours or even days. If minor anomalies in the wire feeding system are not discovered in time, the entire batch of products will be scrapped.

[0005] Therefore, developing a wire feeding flow monitoring method that does not require additional sensors and can provide early warning and accurately locate faults is of great significance to improving the reliability and production efficiency of 3D printing. Summary of the Invention

[0006] The present invention provides a flow monitoring method and system for a wire feeding mechanism of a 3D printer, which solves the technical problem in the related art that it is impossible to detect anomalies in advance and accurately locate faults without additional sensors.

[0007] The present invention provides a flow monitoring method for a wire feeding mechanism of a 3D printer, comprising: Collect the current signal data of the wire feeding motor drive unit, pre-process the current signal data, and generate standardized current characteristic data; Use a multi-scale signal analysis algorithm to perform time-frequency analysis on the standardized current characteristic data, extract wavelet transform features, output a high-dimensional current feature set, and establish a characteristic fault mapping relationship; Use a multi-model fault analysis system to process high-dimensional current feature sets for fault analysis and location; Based on the fault analysis results, a progressive three-level warning mechanism and flow monitoring are implemented, a multi-level mapping model of current and wire feeding status is established, a multi-level warning signal is generated, and a current flow mapping model is established; By utilizing standardized current feature data, high-dimensional current feature sets, fault analysis results, and current flow mapping models, we continuously optimize the flow model and fault identification model through incremental learning and transfer learning, establish a dynamically updated fault feature library, and adaptively adjust the detection threshold according to the degree of equipment aging.

[0008] Furthermore, the step of collecting current signal data of the wire feeding motor drive unit includes: Connecting the data acquisition module to the wire feeding motor drive unit; Set the sampling frequency to 20kHz; Continuously collect current signal data according to the preset time period.

[0009] Furthermore, the step of preprocessing the current signal data includes: Applying a bandpass filter to filter the collected current signal; Extract the amplitude, phase and harmonic characteristics of the current signal and construct the initial feature vector; Perform signal segmentation processing.

[0010] Furthermore, the step of performing time-frequency analysis on the pre-processed current signal data using a multi-scale signal analysis algorithm includes: Apply multi-scale wavelet packet decomposition algorithm to perform time-frequency analysis on current signal and extract features of different frequency bands; The wavelet scattering transform algorithm is used to extract high-order features of the current signal and construct a high-dimensional current fingerprint; Use recursive feature elimination algorithm to screen the most sensitive feature subset; Perform detailed analysis on the characteristic frequency bands of different fault modes and establish characteristic fault mapping relationships.

[0011] Furthermore, the step of using the recursive feature elimination algorithm to screen the most sensitive feature subset includes: First, train a linear support vector machine classifier to obtain the feature weight vector; Then recursively remove the feature with the smallest weight and retrain the classifier; Until the optimal feature subset is obtained, the expression is: ; in Indicates the The feature set of the iteration, Indicates the The feature set of the iteration, Indicates the Features, Indicates the The weight of the feature, Represents the minimum value in the weight vector.

[0012] Furthermore, the steps of performing fault analysis and location include: Build a deep network with a hierarchical attention module to separate the current features into two branches: electrical features and mechanical features for independent analysis; Apply the isolation forest algorithm to detect overall anomalies; Identify the specific fault type; Use causal inference algorithms to build a complete fault tree and analyze the propagation paths between faulty components; Combining current characteristics with a mapping model of the physical location, the fault can be accurately located to a specific mechanical or electrical component.

[0013] Furthermore, the steps of the progressive three-level warning mechanism and flow monitoring include: Establish a multi-level mapping model from microscopic current changes to macroscopic wire feeding status; Implement a three-level early warning module; Apply fast Fourier transform algorithm to process current signal and extract frequency domain features for flow monitoring; The incremental learning algorithm is used to establish a current flow mapping model to achieve real-time monitoring of wire feeding flow.

[0014] Furthermore, the incremental learning algorithm is used to establish a current flow mapping model, and the expression of the mapping model is: ; in Indicates the estimated flow value, represents the mapping function, represents the frequency domain characteristics, represents the time domain characteristics, Represents model parameters.

[0015] Furthermore, the parameter update formula for continuously optimizing the traffic model and the fault identification model through incremental learning and transfer learning is: ; in Indicates the The model parameters of times, Indicates the The model parameters of times, represents the learning rate, Indicates that based on new data Calculate the gradient of the loss function, Indicates newly collected data, represents the loss function, represents the gradient operator.

[0016] The present invention provides a flow monitoring system for a wire feeding mechanism of a 3D printer, which is used to execute the above-mentioned flow monitoring method for a wire feeding mechanism of a 3D printer, comprising: The signal acquisition and preprocessing module is used to obtain the current signal of the wire feeding motor drive unit and perform denoising and normalization processing to output standardized current characteristic data; Feature extraction and analysis module, used to perform multi-scale time-frequency decomposition of current signals, extract high-dimensional current features and establish the corresponding relationship between features and faults; A fault diagnosis and location module, which integrates multiple machine learning algorithms to analyze current characteristics, determine abnormal conditions, and accurately locate faulty components; The early warning and flow calculation module is used to implement a multi-level early warning mechanism based on fault analysis results, establish a current flow mapping model, and calculate the real-time wire feeding flow rate; The model adaptive optimization module is used to continuously update the fault feature library and traffic mapping model based on system operation data and dynamically adjust the detection threshold.

[0017] The beneficial effects of the present invention are as follows: wire feed flow monitoring can be achieved without the need for additional sensors, reducing system cost and complexity; anomalies can be detected in advance before a fault occurs, providing sufficient warning time; more than 10 specific fault types can be accurately identified and the fault can be located to a specific component; and the model can be continuously improved through incremental learning to adapt to changes in equipment aging and wear. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The present invention is a flow chart of a method for monitoring the flow of a wire feeding mechanism of a 3D printer. DETAILED DESCRIPTION

[0019] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a method for monitoring the flow rate of a wire feeding mechanism of a 3D printer. Figure 1 As shown, including: Step 1: collecting current signal data of the wire feeding motor drive unit, and preprocessing the current signal data to generate standardized current characteristic data; Step 1.1, using a high-frequency sampling circuit to collect the wire feeding motor current signal; The data acquisition module is connected to the wire feeding motor drive unit, the sampling frequency is set to 20kHz, and the current signal data is continuously collected according to the preset time period to obtain the current time domain raw data; a precision sampling resistor (0.01Ω) is connected in series in the power supply circuit of the wire feeding motor drive circuit, and the current signal is obtained by measuring the voltage across the resistor and passing it through the amplification circuit, and then digitally sampled through a 16-bit ADC (analog-to-digital converter).

[0021] Step 1.2, applying a bandpass filter to filter the collected current signal; The signal components in the 50Hz-5kHz frequency band are retained, and the environmental noise and power supply interference are eliminated; the filter transfer function is: ; in Indicates the frequency The filter transfer function when Indicates the signal frequency.

[0022] The filter is implemented using a digital FIR filter with an order of 64 to ensure filtering accuracy and phase linearity.

[0023] Step 1.3, extract the amplitude, phase and harmonic characteristics of the current signal and construct the initial feature vector; Specifically expressed as: ; in represents the initial eigenvector, 、 、 Represents the first, second, and Amplitude characteristics, 、 、 Represents the first, second, and Phase characteristics, 、 、 Respectively represent the first, second, and harmonic characteristics, Quantities representing amplitude and phase characteristics, Indicates the number of harmonic features.

[0024] In actual implementation, the amplitude feature takes the first 10 main components, the phase feature takes the corresponding 10 components, and the harmonic feature takes the first 5 main harmonics.

[0025] Step 1.4, perform signal segmentation processing; The continuous current signal is segmented according to the wire feeding cycle, with each segment corresponding to a complete wire feeding process, facilitating subsequent analysis. During the wire feeding process of a 3D printer, the current is segmented by detecting the peak points of the cyclical current change. Each wire feeding cycle typically corresponds to a signal segment of 0.5-2 seconds.

[0026] Step 2: Use a multi-scale signal analysis algorithm to perform time-frequency analysis on the standardized current feature data, extract wavelet transform features, output a high-dimensional current feature set, and establish a feature fault mapping relationship; The current signal preprocessed in step 1 is processed using a multi-scale signal analysis algorithm to generate a multi-scale current feature set and a high-dimensional current fingerprint. Specifically, the algorithm includes: Step 2.1: Apply the multi-scale wavelet packet decomposition algorithm to perform time-frequency analysis on the current signal and extract the features of different frequency bands; The mathematical expression of wavelet packet decomposition is: ; in represents the wavelet packet transform coefficient, is the current signal function, represents the time variable, is the wavelet packet basis function, represents the decomposition scale (determines the frequency resolution), Indicates the time position parameter (determines time positioning), represents the oscillation parameter (determines the frequency bandwidth), Represents the integral from negative infinity to positive infinity.

[0027] For the wire feeding motor current signal, the decomposition scale If it is set to 4, a feature subspace of 16 frequency bands can be obtained; Step 2.2, applying the wavelet scattering transform algorithm to extract the high-order features of the current signal and construct a high-dimensional current fingerprint; The mathematical expression of wavelet scattering transform is: ; in represents the scattering transformation coefficient, is the input current signal, is the wavelet function, is the wavelet scale parameter, represents the scattering path, represents the convolution operation, Indicates taking the absolute value.

[0028] By calculating the first- and second-order scattering coefficients, a feature representation that is sensitive to small changes and stable to deformation can be obtained.

[0029] In this embodiment, Morlet wavelet is used as the wavelet function, and the scale parameter It is set to 8 discrete values, covering the frequency band range of 50Hz-5kHz, the number of first-order scattering paths is 8, the number of second-order scattering paths is 28, and a total of 36 scattering features.

[0030] Step 2.3, use recursive feature elimination algorithm to select the most sensitive feature subset; First, a linear support vector machine classifier is trained to obtain the feature weight vector. Then, the feature with the smallest weight is recursively removed and the classifier is retrained until the optimal feature subset is obtained. The algorithm expression is: ; in Indicates the The feature set of the iteration, Indicates the The feature set of the iteration, Indicates the Features, Indicates the The weight of the feature, Represents the minimum value in the weight vector.

[0031] In practice, the initial feature set includes over 100 features, including wavelet packet features and scattering transform features. Recursive feature elimination is used to select the 30-40 most sensitive features for subsequent analysis. Feature selection uses cross-validation to evaluate classification performance at each iteration, stopping the feature removal process when performance begins to degrade.

[0032] Step 2.4: Perform detailed analysis on the characteristic frequency bands of different fault modes and establish characteristic fault mapping relationships; The mapping records the characteristic patterns corresponding to different fault types: For example, gear failures are mainly manifested in energy changes in the 200-500Hz frequency band, bearing failures are mainly manifested in characteristic peaks in the 1-3kHz frequency band, and motor winding problems are manifested as distortion of the fundamental frequency and its harmonic components.

[0033] Through multi-dimensional feature extraction and optimization, the raw data of step 1 is converted into a high-dimensional current feature set. These high-dimensional features provide a rich information basis for fault analysis and location, and can sensitively capture tiny current changes and associate them with specific fault modes.

[0034] Step 3: Use a multi-model fault analysis system to process the high-dimensional current feature set to perform fault analysis and location. The multi-model fault analysis system is used to process the current characteristics extracted in step 2 to generate fault type determination results and fault location information; specifically, the system includes: Step 3.1: Build a deep network with a hierarchical attention module to separate the current features into two branches: electrical features and mechanical features for independent analysis. The network structure includes feature extraction layer, attention calculation layer and classification layer, where the calculation formula of the attention module is: ; in For the The attention weight of each feature, and Respectively Features and The importance score of a feature, and Indicates the Features and The hidden representation of features, represents the exponential function, Represents the summation symbol.

[0035] This deep network uses a multi-layer perceptron architecture, consisting of an input layer (feature dimension), two hidden layers (64 nodes and 32 nodes), an attention layer, and an output layer (number of fault types). The attention module uses a feedforward neural network to calculate feature importance scores, and attention weights are normalized using a softmax function. The network is trained using the Adam optimizer with a learning rate of 0.001, 200 training epochs, and a batch size of 32.

[0036] Step 3.2, apply the isolation forest algorithm to detect overall anomalies; This algorithm isolates data points by constructing random trees, and outliers are usually easier to isolate. The calculation formula for the anomaly score is: ; in Representation sample The anomaly score, represents the input sample, Indicates the size of the dataset, It is a sample The path length, is the expected path length, The amount of data is Normalization factor when .

[0037] For the 3D printer wire feeding system, the number of trees in IsolationForest is set to 100, the subsampling size is set to 256, and the anomaly threshold is set to 0.6, that is, samples with an anomaly score greater than 0.6 are judged as anomalies.

[0038] Step 3.3, identify the specific fault type; The one-class support vector machine (One-ClassSVM) algorithm is used to identify the specific fault type. The algorithm separates normal data from abnormal data by finding an optimal hyperplane. The decision function is: ; in represents the decision function, represents the input sample, represents the symbolic function, Indicates from 1 to The sum of represents the number of support vectors, For the Lagrange multipliers, is the kernel function, Indicates the support vectors, is the bias term.

[0039] For the fault identification of wire feeding system, radial basis function is used as kernel function, and the kernel parameter Set to 0.1, abnormal ratio It is set to 0.05. In order to identify multiple fault types, a one-to-many strategy is adopted to train a One-ClassSVM model for each fault type.

[0040] Step 3.4: Use the causal inference algorithm to build a complete fault tree and analyze the propagation path between faulty components; The inference of causality is based on the probability of intervention, which is expressed as: ; in Indicates that the intervention variable The value is Under the condition that The probability distribution of Represents the intervention variable Perform human intervention and set its value to , represents the outcome variable, represents the cause variable, represents the intervention value.

[0041] In the actual implementation, the structural equation model is used to model the causal relationship, and the causal structure is learned from the current characteristics and fault data through the Bayesian network learning algorithm.

[0042] For the 3D printer wire feeding system, a causal network containing 10 key component nodes, such as motor, controller, gear, bearing, conduit, nozzle, etc., is established to describe the propagation relationship of the fault.

[0043] Step 3.5: Combine the current characteristics with the physical location mapping model to accurately locate the fault to a specific mechanical or electrical component; The positioning accuracy can reach the component level, and the position mapping function is: ; in represents the coordinates of the fault location, is the mapping function, Indicates electrical characteristics, Represents mechanical characteristics.

[0044] Mapping Function This is achieved using a gradient boosting decision tree model. The input is electrical and mechanical features (a total of 30 features), and the output is the faulty component number and location coordinates. The GBDT model has 200 trees, a maximum depth of 5, and a learning rate of 0.1. It is trained using historical fault sample data.

[0045] The step is to convert the current characteristics into specific fault information through a multi-model collaborative analysis system, providing a basis for subsequent early warning and monitoring, enabling the system to quickly locate the fault type and location when an anomaly is detected, and support precise maintenance.

[0046] Step 4: Implement a progressive three-level warning mechanism and flow monitoring based on the fault analysis results, establish a multi-level mapping model of current and wire feeding status, generate multi-level warning signals, and establish a current flow mapping model; The progressive warning algorithm is used to process the abnormality detection results in step 3 to generate multi-level warning signals and flow status information. Step 4 is based on the fault analysis results output in step 3 and combines the current signal features extracted in steps 1 and 2 to build a warning system and flow monitoring model; specifically, it includes: Step 4.1, establish a multi-level mapping model from microscopic current changes to macroscopic wire feeding states; The model maps the tiny current change pattern to the working state of the wire feeding mechanism to form a state transfer matrix.

[0047] Specifically, the hidden Markov model is used to implement it, and five hidden states (normal, slightly abnormal, moderately abnormal, severely abnormal, and fault) are defined. The state transition probability and emission probability are learned from the current feature sequence through the Baum-Welch algorithm, and then the most likely state sequence is determined through the Viterbi algorithm.

[0048] Step 4.2, implement the three-level warning module; Early warning: When a small current anomaly is detected, an early warning is issued, and potential problems are discovered 100-300 wire feeding cycles (about 10-30 minutes) in advance. The early warning signal is ; Fault type identification: When the abnormality intensifies, the specific fault type is determined and the warning signal is ; Precise positioning: Identify the specific fault components and locations, and the early warning signal is ; The warning level determination formula is: ; in Indicates the The warning level of each sample, is the anomaly score, Represents the input sample, with a value range of [0,1].

[0049] In actual applications, early warning signals are transmitted to operators through sound and light indications, system push and maintenance interface display, and recorded in the system log.

[0050] Each warning level corresponds to a different processing strategy: Levels recommend increasing the frequency of monitoring; Level recommendations are to be checked after the current print task is completed; The level recommends pausing printing immediately for maintenance.

[0051] Step 4.3, apply the fast Fourier transform algorithm to process the current signal and extract the frequency domain features for flow monitoring; The transformation formula is: ; in represents the transformed frequency domain signal, represents the frequency index, is the time domain current signal, Represents the time index, is the number of sampling points, is the imaginary unit, represents the complex exponential function, Indicates from 0 to The sum of .

[0052] In order to improve the computational efficiency, the Radix-2 Fast Fourier Transform (Radix-2, Radix-2-FFT) algorithm is used, and the number of sampling points is Set to 2048 and use Hanning window to reduce spectrum leakage. For flow monitoring, focus on the power spectrum density in the 2-20 Hz frequency band, which is directly related to the wire feed rate.

[0053] Step 4.4, using incremental learning algorithm to establish current flow mapping model to achieve real-time monitoring of wire feeding flow; The expression of the mapping model is: ; in Indicates the estimated flow value, represents the mapping function, represents the frequency domain characteristics, represents the time domain characteristics, Represents model parameters.

[0054] The mapping model is implemented using a random forest regressor with 100 trees and a maximum depth of 10. The input features include 10 frequency domain features and 5 time domain features. The output is the volume flow value (unit: mm 3 Grid search and cross-validation are used to determine the optimal hyperparameters for model training.

[0055] This step integrates the results of all the previous steps to form a complete monitoring and early warning system, providing operation and maintenance personnel with graded early warning and flow monitoring information, enabling the system to take measures in advance before problems occur, improving printing success rates and reducing material waste.

[0056] Step 5: Using standardized current feature data, high-dimensional current feature sets, fault analysis results, and current flow mapping models, we continuously optimize the flow model and fault identification model through incremental learning and transfer learning, establish a dynamically updated fault feature library, and adaptively adjust the detection threshold based on the degree of equipment aging. Adaptive optimization algorithms are used to update model parameters to improve system adaptability and accuracy. Model parameters in each step are continuously optimized using the new current signal data collected in step 1, the new features extracted in step 2, the fault analysis results in step 3, and the early warning feedback in step 4. Step 5 ensures that the system can adapt to equipment aging and the emergence of new failure modes during long-term operation. Specifically, this includes: Step 5.1: As the equipment accumulates data during operation, the current flow mapping model and fault identification model are continuously optimized through incremental learning methods to improve identification accuracy. The parameter update formula for incremental learning is: ; in Indicates the The model parameters of times, Indicates the The model parameters of times, represents the learning rate, Indicates that based on new data Calculate the gradient of the loss function, Indicates newly collected data, represents the loss function, represents the gradient operator.

[0057] In practical applications, the model is updated every time 50 new samples are accumulated, and the learning rate Decays over time, with an initial value of 0.01 and a decay factor of 0.95.

[0058] For the current flow mapping model, the partial fitting method is used to achieve online update; for the fault identification model, the incremental learning forest algorithm is used to dynamically add new decision trees.

[0059] Step 5.2: Establish a fault feature library to record the current fingerprints of various faults and continuously update them as the equipment is used. The feature library structure is: ; in represents the fault signature database, 、 、 Respectively represent the first, second, and The characteristic vector of the fault, 、 、 Respectively represent the first, second, and Type identification of the fault, 、 、 Respectively represent the first, second, and The location information of the fault, Indicates the total number of fault types.

[0060] The signature database is implemented as a key-value store, supporting fast retrieval and updates. It initially contains 10 common failure modes, with at least 10 samples for each. This database is continuously expanded as the system is used and faults are discovered. The signature database is updated using an incremental clustering algorithm, automatically discovering new failure modes and adding them to the database.

[0061] In step 5.3, for the newly emerged fault mode, a transfer learning algorithm is used to quickly build a new fault recognition model using existing knowledge. The objective function of transfer learning is: ; in Indicates the target parameters The minimization operation, represents the loss function of the target domain, represents the target domain data, represents the balance coefficient, represents the parameter regularization term, Represents the source model parameters.

[0062] This regularization term ensures that the new parameters do not deviate too far from the source domain. In practical implementation, a parameter transfer approach is used, using the parameters of the existing fault model as the initial parameters for the new fault model, which is then fine-tuned using a small number of new fault samples. For deep network models, the parameters of the first few layers are retained unchanged, and only the parameters of the later layers are updated. For tree models, a model fusion strategy is used, combining the prediction results of the existing and newly trained models in a weighted manner.

[0063] Step 5.4: Automatically adjust the detection threshold according to the aging degree of the equipment to maintain the accuracy of the detection. The threshold adjustment formula is: ; in represents the adjusted threshold, represents the basic threshold, represents the aging coefficient, Indicates the device usage time.

[0064] Aging coefficient Typically set to 0.02-0.05. In practice, the system records the device's usage time, number of print jobs, and total wire feed volume as indicators of aging, and automatically adjusts the detection threshold every three months.

[0065] The steps are to continuously improve the performance of the entire monitoring system through adaptive optimization algorithms. The optimization mechanism ensures that the system maintains high accuracy and adaptability during long-term use, can identify new failure modes and adapt to changes caused by equipment aging.

[0066] The five main steps of this implementation form a complete closed-loop system, with each step closely linked and mutually supportive: Step 1 provides basic data for the entire system; Step 2 extracts high-dimensional features based on step 1; Step 3 uses the features of step 2 to identify and locate the fault; Step 4 generates warning signals and traffic monitoring based on the analysis results of step 3; Step 5 continuously optimizes the model parameters in the first four steps through new data and feedback to form a complete closed loop.

[0067] This closely linked step design ensures the efficiency, accuracy and adaptability of the system, enabling high-precision flow monitoring and fault warning of the 3D printer wire feeding mechanism without the need for additional sensors.

[0068] A flow monitoring system for a wire feeding mechanism of a 3D printer, used to execute the above-mentioned flow monitoring method for a wire feeding mechanism of a 3D printer, comprising: The signal acquisition and preprocessing module is used to obtain the current signal of the wire feeding motor drive unit and perform denoising and normalization processing to output standardized current characteristic data; Feature extraction and analysis module, used to perform multi-scale time-frequency decomposition of current signals, extract high-dimensional current features and establish the corresponding relationship between features and faults; A fault diagnosis and location module, which integrates multiple machine learning algorithms to analyze current characteristics, determine abnormal conditions, and accurately locate faulty components; The early warning and flow calculation module is used to implement a multi-level early warning mechanism based on fault analysis results, establish a current flow mapping model, and calculate the real-time wire feeding flow rate; The model adaptive optimization module is used to continuously update the fault feature library and traffic mapping model based on system operation data and dynamically adjust the detection threshold.

[0069] Here, the present invention provides an implementation example: In order to improve printing efficiency and quality, the wire feeder flow monitoring method of the present invention was implemented on the production line of a large-scale 3D printing factory. The specific implementation process and effects of this method in actual application are described in detail below.

[0070] The plant primarily produces high-precision metal parts for aerospace and medical devices, using 20 industrial-grade DMLS (Direct Metal Laser Sintering) 3D printers, each equipped with a dual-motor wire feeder. Printing materials include titanium alloy (Ti6Al4V), stainless steel (316L), and nickel-based superalloy (Inconel718). Before implementing this invention, the plant faced the following challenges: High material waste rate: Due to unstable wire feeding, excessive or insufficient material feeding results in an average material waste cost of approximately RMB 12,000 per machine per month; High print failure rate: The print failure rate due to wire feeding failure is as high as 8.6%. The economic losses caused by the failure of printing large parts are particularly serious. Long equipment downtime: Wire feed mechanism failure usually requires 4-6 hours of downtime for maintenance, with an average annual unplanned downtime of 460 hours; Weak fault prediction capabilities: Using traditional regular maintenance methods, potential faults cannot be predicted in a targeted manner.

[0071] The factory decided to implement the method of the present invention on all its equipment. The system deployment architecture is as follows: the motor drive unit of the wire feeding mechanism of each 3D printer is connected to the data acquisition module, the acquisition module is connected to the edge computing unit via industrial Ethernet, and the edge computing unit is then connected to the central monitoring server to form a distributed computing architecture.

[0072] Phase 1: Initial data collection and model building (implementation time: January-February 2023); Data collection plan design: Determine the sampling frequency: According to the working characteristics of the wire feeding mechanism, the sampling frequency is set to 200Hz; Set data acquisition parameters: motor current, voltage, speed, wire feed tension sensor data, equipment ambient temperature; Deploy data acquisition hardware: Each device is equipped with an NI-9215 data acquisition card and a self-developed data pre-processing unit based on the STM32F407; Motor current change feature extraction: Continuous data collection was performed on a normally operating printer for 15 days, collecting 400GB of motor current data. By using the Fourier transform and wavelet analysis in the method of the present invention to extract features, it is found that the current of the normally operating motor exhibits the following characteristics: The main frequency components are concentrated in the range of 30-45Hz; The standard deviation of the current amplitude change is controlled within 0.08A; The current change response time is less than 50ms when the load fluctuates; Traffic model establishment: Determine the mapping relationship between motor current and wire feeding flow rate through experiments; Multiple regression analysis was used to determine the weights of the influencing factors: current change accounted for 65%, speed accounted for 25%, and temperature compensation accounted for 10%; Build a mathematical model: ; in is the flow value, is the current value, is the rotation speed, is the ambient temperature, 、 、 Represent the weight coefficients of current value, speed and ambient temperature respectively, is the compensation constant; Phase II: Abnormal pattern recognition and real-time monitoring deployment (implementation time: March-April 2023); Abnormal pattern database establishment: Simulate 6 common fault conditions: wire feed wheel wear, motor overheating, motor bearing wear, wire feed pipe blockage, uneven material quality, and power supply fluctuation; Conduct 50 data collection experiments for each fault and record the changing characteristics of the motor current data; Establish an abnormal mode current feature database, including 300 sets of sample data; The abnormal mode current feature database is shown in Table 1: Table 1: Abnormal mode current feature database (partial sample)

[0073] Machine Learning Model Training: A hybrid model combining the support vector machine and random forest of the present invention is adopted; Use 80% of the data as a training set and 20% as a test set; Model accuracy: fault detection accuracy is 94.7%, fault type classification accuracy is 91.3%; Deploy the model to the edge computing unit to achieve real-time response within 40ms; Data real-time monitoring system deployment: Deploy monitoring systems on all 20 devices; Configure alarm thresholds: slight abnormality (current fluctuation exceeds 3 times the normal standard deviation), moderate abnormality (5 times), severe abnormality (8 times); Establish a three-level early warning mechanism: screen prompt, sound and light alarm, and automatic shutdown; Phase 3: Self-learning and optimization (implementation time: May-July 2023); Incremental learning system implementation: Establish a closed feedback loop: After each fault event, the technicians will confirm the fault type and enter it into the system; Implementing the stochastic gradient descent-based incremental learning algorithm of the present invention; Perform global optimization of the model once a month to continuously improve recognition accuracy; Adaptive Optimization: Automatically adjust monitoring parameters for different material types: titanium alloy wire feeding monitoring frequency increased to 250Hz, stainless steel maintained at 200Hz; Establish differentiated detection models for different batches of materials to improve adaptability; Implement the dynamic threshold adjustment algorithm of the present invention to automatically adjust the alarm threshold according to the equipment operation time; System integration and data analysis platform: Integrate the monitoring system with the factory MES (Manufacturing Execution System); Develop a visual analysis platform to provide equipment health scores and predictive maintenance recommendations; Realize remote monitoring function, technicians can receive alarms and conduct remote diagnosis through mobile devices; After the method of the present invention was deployed on all 20 3D printing devices in the factory and operated for 6 months, it was specifically demonstrated in the following aspects: By implementing the wire feeder flow monitoring method of the present invention, the factory achieved the following quantitative results: Wire feeding failure early warning time: On average, potential failures can be predicted 4.2 hours in advance, providing maintenance personnel with ample response time; Fault detection rate: The average detection rate for six common fault types reached 94.7%, including: The detection rate of wire feed wheel wear is 96.3%; the detection rate of motor overheating is 98.5%; the detection rate of motor bearing wear is 93.1%; the detection rate of wire feed pipe blockage is 97.2%; the detection rate of uneven material quality is 91.8%; and the detection rate of power supply fluctuation is 91.4%. Fault classification accuracy: The accuracy of correctly identifying fault types reaches 91.3%, which is 35% higher than traditional methods; False Alarm Rate: The system had a false alarm rate of 3.2%, a decrease from the plant’s previous threshold-based monitoring method (false alarm rate of 15.7%).

[0074] The comparison between the method of the present invention and the traditional monitoring method is shown in Table 2: Table 2: Comparison between the present invention and traditional monitoring methods

[0075] After implementing the method of the present invention, the production efficiency and product quality of the factory are improved: Equipment downtime: reduced from an average of 72 hours per month to 18.5 hours, a 74.3% reduction; The number of unplanned maintenance operations decreased from an average of 38 per month to 11 per month, a decrease of 71.1%. Material waste rate: reduced from 8.7% to 2.1%, saving a large amount of high-value metal powder materials; Print success rate: increased from 83.5% to 96.8%, reducing the waste rate; Overall equipment efficiency: increased from 72.4% to 91.6%, an increase of 19.2 percentage points.

[0076] Through the self-learning function of the present invention, the system is continuously optimized during actual application: Improved model accuracy: The fault detection accuracy was 89.5% during initial deployment, increasing to 94.7% after six months of self-learning. Adaptability to different materials: The system successfully adaptively adjusted the wire feed flow monitoring parameters for six different metal powder materials; Prediction time extended: The initial warning time averaged 2.8 hours, which was extended to 4.2 hours after optimization; Based on the successful application of the method of this invention, the factory plans to further expand the system's functions and extend similar monitoring methods to other key components of 3D printing equipment, such as laser systems and printing platforms, to form a more comprehensive equipment health monitoring system.

[0077] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for monitoring the flow rate of a wire feeding mechanism of a 3D printer, characterized in that: include: Collect the current signal data of the wire feeding motor drive unit, pre-process the current signal data, and generate standardized current characteristic data; Use a multi-scale signal analysis algorithm to perform time-frequency analysis on the standardized current characteristic data, extract wavelet transform features, output a high-dimensional current feature set, and establish a characteristic fault mapping relationship; Use a multi-model fault analysis system to process high-dimensional current feature sets for fault analysis and location; Based on the fault analysis results, a progressive three-level warning mechanism and flow monitoring are implemented, a multi-level mapping model of current and wire feeding status is established, a multi-level warning signal is generated, and a current flow mapping model is established; By utilizing standardized current feature data, high-dimensional current feature sets, fault analysis results, and current flow mapping models, we continuously optimize the flow model and fault identification model through incremental learning and transfer learning, establish a dynamically updated fault feature library, and adaptively adjust the detection threshold according to the degree of equipment aging.

2. A method for monitoring the flow rate of a wire feeding mechanism of a 3D printer according to claim 1, characterized in that: The step of collecting current signal data of the wire feeding motor drive unit includes: Connecting the data acquisition module to the wire feeding motor drive unit; Set the sampling frequency to 20kHz; Continuously collect current signal data according to the preset time period.

3. A method for monitoring flow of a wire feeding mechanism of a 3D printer according to claim 1, characterized in that: The step of preprocessing the current signal data includes: Applying a bandpass filter to filter the collected current signal; Extract the amplitude, phase and harmonic characteristics of the current signal and construct the initial feature vector; Perform signal segmentation processing.

4. A method for monitoring flow of a wire feeding mechanism of a 3D printer according to claim 1, characterized in that: The step of performing time-frequency analysis on the pre-processed current signal data using a multi-scale signal analysis algorithm comprises: Apply multi-scale wavelet packet decomposition algorithm to perform time-frequency analysis on current signal and extract features of different frequency bands; The wavelet scattering transform algorithm is used to extract high-order features of the current signal and construct a high-dimensional current fingerprint; Use recursive feature elimination algorithm to screen the most sensitive feature subset; Perform detailed analysis on the characteristic frequency bands of different fault modes and establish characteristic fault mapping relationships.

5. A method for monitoring flow of a wire feeding mechanism of a 3D printer according to claim 4, characterized in that: The step of using the recursive feature elimination algorithm to screen the most sensitive feature subset comprises: First, train a linear support vector machine classifier to obtain the feature weight vector; Then recursively remove the feature with the smallest weight and retrain the classifier; Until the optimal feature subset is obtained, the expression is: ； in Indicates the The feature set of the iteration, Indicates the The feature set of the iteration, Indicates the Features, Indicates the The weight of the feature, Represents the minimum value in the weight vector.

6. A method for monitoring flow of a wire feeding mechanism of a 3D printer according to claim 5, characterized in that: The steps of performing fault analysis and location include: Build a deep network with a hierarchical attention module to separate the current features into two branches: electrical features and mechanical features for independent analysis; Apply the isolation forest algorithm to detect overall anomalies; Identify the specific fault type; Use causal inference algorithms to build a complete fault tree and analyze the propagation paths between faulty components; Combining current characteristics with a mapping model of the physical location, the fault can be accurately located to a specific mechanical or electrical component.

7. The method for monitoring flow of a wire feeding mechanism of a 3D printer according to claim 1, wherein: The steps of the progressive three-level early warning mechanism and flow monitoring include: Establish a multi-level mapping model from microscopic current changes to macroscopic wire feeding status; Implement a three-level early warning module; Apply fast Fourier transform algorithm to process current signal and extract frequency domain features for flow monitoring; The incremental learning algorithm is used to establish a current flow mapping model to achieve real-time monitoring of wire feeding flow.

8. A method for monitoring flow of a wire feeding mechanism of a 3D printer according to claim 7, characterized in that: The incremental learning algorithm is used to establish a current flow mapping model, and the expression of the mapping model is: ； in Indicates the estimated flow value, represents the mapping function, represents the frequency domain characteristics, represents the time domain characteristics, Represents model parameters.

9. The method for monitoring the flow rate of a wire feeding mechanism of a 3D printer according to claim 1, wherein: The parameter update formula for continuously optimizing the traffic model and the fault identification model through incremental learning and transfer learning is: ； in Indicates the The model parameters of times, Indicates the The model parameters of times, represents the learning rate, Indicates that based on new data Calculate the gradient of the loss function, Indicates newly collected data, represents the loss function, represents the gradient operator.

10. A flow monitoring system for a 3D printer wire feeding mechanism, characterized in that: A method for monitoring the flow rate of a wire feeding mechanism of a 3D printer, used for executing any one of claims 1 to 9, comprising: The signal acquisition and preprocessing module is used to obtain the current signal of the wire feeding motor drive unit and perform denoising and normalization processing to output standardized current characteristic data; Feature extraction and analysis module, used to perform multi-scale time-frequency decomposition of current signals, extract high-dimensional current features and establish the corresponding relationship between features and faults; A fault diagnosis and location module, which integrates multiple machine learning algorithms to analyze current characteristics, determine abnormal conditions, and accurately locate faulty components; The early warning and flow calculation module is used to implement a multi-level early warning mechanism based on fault analysis results, establish a current flow mapping model, and calculate the real-time wire feeding flow rate; The model adaptive optimization module is used to continuously update the fault feature library and traffic mapping model based on system operation data and dynamically adjust the detection threshold.

Images