Molten pool monitoring method in SLM melting process

Abstract

The invention relates to the technical field of molten pool monitoring, and provides a molten pool monitoring method in an SLM (selective laser melting) process, which comprises the following steps: collecting molten pool image data and molten pool sound signal data in the SLM process in real time; calculating the vibration speed of the molten pool based on the molten pool acoustic signal data, and calculating the spectral irradiation of the molten pool based on the molten pool image data; constructing a deep belief network model based on the vibration velocity and spectral irradiation of the molten pool, and performing feature extraction on the molten pool image data according to the deep belief network model to obtain first molten pool image features and second molten pool image features; searching an optimal hyperplane through a support vector machine algorithm in combination with the first molten pool image features and the second molten pool image features to obtain a molten pool state discrimination function; and identifying the molten pool image data and the molten pool acoustic signal data based on the molten pool state discrimination function to obtain the real-time state of the molten pool. According to the method, the real-time monitoring capability of the molten pool state is improved, and the real-time state of the molten pool can be predicted more accurately through real-time updating of the multi-dimensional vector, so that the quality of an SLM forming part is predicted.

Description

Technical Field

[0001] The present invention relates to the technical field of molten pool monitoring, and in particular to a molten pool monitoring method in a SLM melting process. Background Art

[0002] Duplex stainless steel is a high-performance stainless steel in which a body-centered cubic ferrite phase and a face-centered cubic austenite phase coexist in the solid solution structure. It has the excellent properties of both ferritic stainless steel and austenitic stainless steel and is a structural material with high strength and excellent corrosion resistance.

[0003] At present, duplex stainless steel is additively manufactured using SLM (Selective Laser Melting), which uses a high-energy beam to melt fine metal powder layer by layer, and can produce fine parts with fine grains and high density. Due to the complex and rapid phase change process of the unique duplex structure of duplex stainless steel during the additive manufacturing process, it is very easy to cause structural cracking during the hot forming process. Therefore, it is necessary to monitor the depth and width of the molten pool during the SLM process to predict the quality of SLM formed parts.

[0004] The existing technology often uses a CNN model to identify the melt pool image and perform analysis based on the image. However, CNN is more suitable for static images, and has poor recognition ability for dynamic melt pool images, and the prediction accuracy is low. Summary of the invention

[0005] In view of this, the present invention proposes a molten pool monitoring method in the SLM melting process, which collects molten pool image data and molten pool acoustic signal data in the SLM process in real time, calculates the vibration speed of the molten pool based on the molten pool acoustic signal data, calculates the spectral irradiation of the molten pool based on the molten pool image data, constructs a deep belief network model, performs feature extraction on the molten pool image data, searches for the best hyperplane based on the support vector machine algorithm combined with the extracted image features to obtain the molten pool state discriminant function, improves the real-time monitoring capability of the molten pool state, and can more accurately predict the state parameters of the molten pool through real-time updating of the multi-dimensional vector, and can more accurately predict the real-time state of the molten pool, thereby predicting the quality of the SLM formed part, thereby solving the problem of poor recognition ability of the prior art for dynamic molten pool images.

[0006] Based on the molten pool state discriminant function, the data is identified to obtain the real-time state of the molten pool.

[0007] The technical solution of the present invention is implemented as follows: The present invention provides a molten pool monitoring method in a SLM melting process, comprising the following steps:

[0008] S1, real-time acquisition of molten pool image data and molten pool acoustic signal data during the SLM process;

[0009] S2, calculating the vibration velocity of the molten pool based on the molten pool acoustic signal data, and calculating the spectral irradiance of the molten pool based on the molten pool image data;

[0010] S3, constructing a deep belief network model based on the vibration velocity and spectral irradiation of the molten pool, and extracting features of the molten pool image data according to the deep belief network model to obtain a first molten pool image feature and a second molten pool image feature;

[0011] S4, combining the first molten pool image features and the second molten pool image features through a support vector machine algorithm, searching for the best hyperplane to obtain a molten pool state discriminant function;

[0012] S5, identifying the molten pool image data and the molten pool acoustic signal data based on the molten pool state discrimination function to obtain the real-time state of the molten pool, and predicting the quality of the SLM formed part based on the real-time state of the molten pool.

[0013] On the basis of the above technical solution, preferably, step S1 includes:

[0014] By arranging an optical sensor and an acoustic sensor around the molten pool, the molten pool image data in the SLM process is collected based on the optical sensor, and the molten pool acoustic signal data in the SLM process is collected based on the acoustic sensor.

[0015] On the basis of the above technical solution, preferably, step S1 further includes:

[0016] The molten pool image data includes video images, molten pool optical signal wavelength, and roughly measured temperature of the molten pool. The molten pool acoustic signal data includes acoustic impedance and molten pool acoustic wave velocity. The optical signal wavelength and the roughly measured temperature of the molten pool are used to calculate the spectral irradiation of the molten pool, and the acoustic impedance and the molten pool acoustic wave velocity are used to calculate the vibration velocity of the molten pool.

[0017] Based on the above technical solution, preferably, step S2 includes:

[0018] The vibration velocity of the molten pool is calculated based on the molten pool acoustic signal data, and the spectral irradiance of the molten pool is calculated based on the molten pool image data. The calculation formulas for the vibration velocity and spectral irradiance of the molten pool are respectively:

[0019]

[0020] Where v is the vibration speed of the molten pool, α is the laser absorption coefficient, β is the maximum energy distribution coefficient, γ is the gas adiabatic index, and v 声 is the acoustic wave velocity of the molten pool, Z is the acoustic impedance, f is the spectral irradiance of the molten pool, h and c are Planck constants, k is the Boltzmann constant, λ is the wavelength of the optical signal of the molten pool, T is the rough temperature of the molten pool, and σ is the error bias of the optical sensor.

[0021] On the basis of the above technical solution, preferably, step S3 includes:

[0022] S31, constructing an initial deep belief network model based on the vibration velocity and spectral irradiation of the molten pool, obtaining historical molten pool image data, and dividing the historical molten pool image data into a training set and a validation set;

[0023] S32, iteratively training the initial deep belief network model through the training set, configuring the loss function and hyperparameters, until the iteration stop condition is reached, and obtaining the trained deep belief network model;

[0024] S33, using the validation set to test the trained deep belief network model, and performing model evaluation on the model. When the model evaluation does not meet the requirements, return to step S32 to adjust the hyperparameters and re-train the model iteratively.

[0025] Based on the above technical solution, preferably, step S31 includes:

[0026] According to the vibration speed and spectral irradiation of the molten pool, the input layer structure of the initial deep belief network model is determined, the network weights and biases of the initial deep belief network model are initialized, the historical molten pool image data are collected and preprocessed, and the preprocessed historical molten pool image data are divided into a training set and a validation set with a ratio of 7:3. The hidden layer structure of the initial deep belief network model is determined, including the number of hidden layers and the number of neurons in each hidden layer.

[0027] Based on the above technical solution, preferably, step S32 includes:

[0028] The initial hyperparameters are set, including the number of pre-training iterations and the number of layers of the restricted Boltzmann machine. The initial deep belief network model uses the cross entropy loss function and the Adam optimization algorithm. For each training batch, the network output is calculated by forward propagation to check whether the iteration stopping condition is reached. The iteration stopping condition includes reaching the maximum number of iterations and the loss function convergence.

[0029] Based on the above technical solution, preferably, step S33 includes:

[0030] The mAP value of the trained deep belief network model is calculated on the validation set, and the current mAP value is compared with the preset mAP value. When the current mAP value is not less than the preset mAP value, the model evaluation is completed and the deep belief network model is obtained.

[0031] Based on the above technical solution, preferably, step S4 includes:

[0032] The calculation formula of the molten pool state discriminant function is:

[0033]

[0034] Among them, f(x) is the molten pool state discriminant function, sign(·) is the signal function, and x i is the i-th sample of the best hyperplane, D is the number of samples of the best hyperplane, η i is the Lagrange multiplier of the i-th sample of the optimal hyperplane, y i For sample x i The observed value, G(x i ,x) is the kernel function, b1 and b2 are the biases of the first molten pool image feature and the second molten pool image feature, respectively.

[0035] Based on the above technical solution, preferably, step S5 includes:

[0036] Based on the molten pool state discriminant function, the molten pool image data and the molten pool acoustic signal data are identified to obtain molten pool state identification parameters, wherein the molten pool state identification parameters include the predicted molten pool depth, the predicted molten pool width, the predicted molten pool temperature, the predicted molten pool vibration frequency and the molten pool radiation intensity. A multidimensional vector is constructed based on the predicted molten pool depth, the predicted molten pool width, the predicted molten pool temperature, the predicted molten pool vibration frequency and the molten pool radiation intensity. The molten pool image data and the molten pool acoustic signal data are updated in real time to update the multidimensional vector in real time to obtain the real-time state of the molten pool, and the quality of the SLM formed part is predicted based on the real-time state of the molten pool.

[0037] The molten pool monitoring method in the SLM melting process of the present invention has the following beneficial effects compared with the prior art:

[0038] (1) By collecting the molten pool image data and molten pool acoustic signal data in real time during the SLM process, the vibration speed of the molten pool is calculated based on the molten pool acoustic signal data, and the spectral irradiation of the molten pool is calculated based on the molten pool image data. A deep belief network model is constructed to extract features from the molten pool image data. According to the support vector machine algorithm and the extracted image features, the optimal hyperplane is searched to obtain the molten pool state discriminant function, which improves the real-time monitoring capability of the molten pool state. Through the real-time update of the multi-dimensional vector, the real-time state of the molten pool can be more accurately predicted, thereby predicting the quality of the SLM formed parts.

[0039] (2) By combining the deep belief network model and the support vector machine algorithm, comprehensive and accurate identification of the molten pool state is achieved, improving the accuracy and reliability of molten pool state monitoring;

[0040] (3) By updating the molten pool image data and acoustic signal data in real time, a multidimensional vector is constructed and updated, thus realizing dynamic monitoring and real-time adjustment of the molten pool status. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 This is a flow chart of a method for monitoring a molten pool during an SLM melting process according to the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] See also Figure 1 , this embodiment provides a method for monitoring a molten pool during a SLM melting process, comprising the following steps:

[0045] S1, real-time acquisition of molten pool image data and molten pool acoustic signal data during the SLM process;

[0046] S2, calculating the vibration velocity of the molten pool based on the molten pool acoustic signal data, and calculating the spectral irradiance of the molten pool based on the molten pool image data;

[0047] S3, constructing a deep belief network model based on the vibration velocity and spectral irradiation of the molten pool, and extracting features of the molten pool image data according to the deep belief network model to obtain a first molten pool image feature and a second molten pool image feature;

[0048] S4, combining the first molten pool image features and the second molten pool image features through a support vector machine algorithm, searching for the best hyperplane to obtain a molten pool state discriminant function;

[0049] S5, identifying the molten pool image data and the molten pool acoustic signal data based on the molten pool state discrimination function to obtain the real-time state of the molten pool, and predicting the quality of the SLM formed part based on the real-time state of the molten pool.

[0050] Specifically, a molten pool monitoring method in an SLM melting process of the present embodiment collects molten pool image data and molten pool acoustic signal data in real time during the SLM process, calculates the vibration speed of the molten pool based on the molten pool acoustic signal data, calculates the spectral irradiation of the molten pool based on the molten pool image data, constructs a deep belief network model, performs feature extraction on the molten pool image data, searches for the optimal hyperplane based on the support vector machine algorithm combined with the extracted image features to obtain a molten pool state discriminant function, thereby improving the real-time monitoring capability of the molten pool state, and can more accurately predict the state parameters of the molten pool through real-time updating of multi-dimensional vectors.

[0051] Step S1 includes:

[0052] By arranging an optical sensor and an acoustic sensor around the molten pool, the molten pool image data in the SLM process is collected based on the optical sensor, and the molten pool acoustic signal data in the SLM process is collected based on the acoustic sensor.

[0053] Specifically, this embodiment realizes that by using optical sensors and acoustic sensors at the same time, image data and acoustic signal data of the molten pool can be obtained, thereby realizing multi-dimensional and multi-angle information collection.

[0054] Optical sensors can be used to capture the visual characteristics of the molten pool, while acoustic sensors can be used to capture the acoustic characteristics of the molten pool. The combination of the two sensors makes data collection more comprehensive and accurate.

[0055] By arranging sensors around the molten pool, the molten pool data during the SLM process can be collected in real time, providing a basis for real-time analysis and monitoring.

[0056] Different types of sensors can capture different types of information. This combination enables the present embodiment to adapt to more process parameters and material types. The combination of molten pool image data and molten pool acoustic signal data provides a richer source of information for deep learning and feature extraction.

[0057] Step S1 also includes:

[0058] The molten pool image data includes video images, molten pool optical signal wavelength, and roughly measured temperature of the molten pool. The molten pool acoustic signal data includes acoustic impedance and molten pool acoustic wave velocity. The optical signal wavelength and the roughly measured temperature of the molten pool are used to calculate the spectral irradiation of the molten pool, and the acoustic impedance and the molten pool acoustic wave velocity are used to calculate the vibration velocity of the molten pool.

[0059] Specifically, this embodiment obtains the original data required for calculating key physical quantities by collecting video images, optical signal wavelength, rough temperature, acoustic impedance and sound wave velocity of the molten pool. The collected optical signal wavelength and rough temperature can be directly used to calculate the spectral irradiation of the molten pool, and the acoustic impedance and sound wave velocity can be directly used to calculate the vibration velocity of the molten pool.

[0060] By simultaneously acquiring visual, spectral and acoustic data, the state of the molten pool can be characterized from multiple dimensions, enhancing the ability to describe the characteristics of the molten pool. By collecting the wavelength of optical signals and the speed of sound waves, the monitoring system's sensitivity to tiny changes in the molten pool is improved.

[0061] Step S2 includes:

[0062] The vibration velocity of the molten pool is calculated based on the molten pool acoustic signal data, and the spectral irradiance of the molten pool is calculated based on the molten pool image data. The calculation formulas for the vibration velocity and spectral irradiance of the molten pool are respectively:

[0063]

[0064] Where v is the vibration speed of the molten pool, α is the laser absorption coefficient, β is the maximum energy distribution coefficient, γ is the gas adiabatic index, and v 声 is the acoustic wave velocity of the molten pool, Z is the acoustic impedance, f is the spectral irradiance of the molten pool, h and c are Planck constants, k is the Boltzmann constant, λ is the wavelength of the optical signal of the molten pool, T is the rough temperature of the molten pool, and σ is the error bias of the optical sensor.

[0065] Specifically, this embodiment uses the molten pool acoustic signal data, combined with the sound wave velocity and acoustic impedance, to calculate the vibration velocity of the molten pool, thereby accurately reflecting the dynamic characteristics of the molten pool; by using the molten pool image data, combined with the optical signal wavelength and roughly measured temperature, the spectral irradiation of the molten pool is calculated, thereby accurately reflecting the thermal radiation characteristics of the molten pool.

[0066] This embodiment takes into account other influencing factors such as laser absorption coefficient, maximum energy distribution coefficient, gas adiabatic index, etc. to calculate the vibration velocity of the molten pool, thereby improving the accuracy of molten pool state monitoring; considering the error bias of the optical sensor can reduce the impact of measurement errors on the calculation results, thereby improving the accuracy of the data.

[0067] Step S3 includes:

[0068] S31, constructing an initial deep belief network model based on the vibration velocity and spectral irradiation of the molten pool, obtaining historical molten pool image data, and dividing the historical molten pool image data into a training set and a validation set;

[0069] S32, iteratively training the initial deep belief network model through the training set, configuring the loss function and hyperparameters, until the iteration stop condition is reached, and obtaining the trained deep belief network model;

[0070] S33, using the validation set to test the trained deep belief network model, and performing model evaluation on the model. When the model evaluation does not meet the requirements, return to step S32 to adjust the hyperparameters and re-train the model iteratively.

[0071] Specifically, this embodiment uses historical melt pool image data for training and verification, and uses a deep belief network model to more accurately extract and identify the characteristics of the melt pool, thereby improving the prediction accuracy and anti-interference ability of the model in the actual SLM process.

[0072] Based on iterative training and hyperparameter adjustment, the model can adaptively optimize its structure and parameters to ensure good performance under different process parameters and material conditions, thereby improving the adaptability of the model.

[0073] By dividing the data into training and validation sets and performing model evaluation, we ensure that the model not only performs well on the training data, but also maintains high performance under unseen SLM process conditions, thereby enhancing the model's ability to be applied in actual production.

[0074] Based on validation set testing and model evaluation, an effective feedback mechanism is established so that the model can be adjusted and improved in a timely manner according to the evaluation results, ensuring that the model meets the expected performance standards in practical applications.

[0075] By utilizing the multi-layer structure of the deep belief network, complex nonlinear features can be extracted from the vibration velocity and spectral irradiation, providing richer and more accurate information for subsequent molten pool state judgment, thereby improving the accuracy of SLM process monitoring.

[0076] Step S31 includes:

[0077] According to the vibration speed and spectral irradiation of the molten pool, the input layer structure of the initial deep belief network model is determined, the network weights and biases of the initial deep belief network model are initialized, the historical molten pool image data are collected and preprocessed, and the preprocessed historical molten pool image data are divided into a training set and a validation set with a ratio of 7:3. The hidden layer structure of the initial deep belief network model is determined, including the number of hidden layers and the number of neurons in each hidden layer.

[0078] Specifically, this embodiment optimizes the network structure according to the vibration speed and spectral irradiation of the molten pool, so that the model can more accurately capture the characteristics of the molten pool in the SLM process, thereby significantly improving the accuracy of molten pool status monitoring.

[0079] Reasonable initialization of network weights and biases significantly shortens the model training time, enables the model to be applied to actual production more quickly, improves production efficiency, preprocesses historical melt pool image data, effectively removes interference factors, provides more reliable training data, and thus improves the stability of the model in practical applications.

[0080] A training set and validation set ratio of 7:3 is adopted, so that the model can maintain good performance under different SLM process parameters and material conditions, and enhance the adaptability of the model in actual production. By modifying the hidden layer structure, the model can better extract the complex features of the SLM melt pool and improve the ability to recognize abnormal conditions, which helps to timely discover and prevent potential quality problems.

[0081] Step S32 includes:

[0082] The initial hyperparameters are set, including the number of pre-training iterations and the number of layers of the restricted Boltzmann machine. The initial deep belief network model uses the cross entropy loss function and the Adam optimization algorithm. For each training batch, the network output is calculated by forward propagation to check whether the iteration stopping condition is reached. The iteration stopping condition includes reaching the maximum number of iterations and the loss function convergence.

[0083] Specifically, this embodiment provides a good starting point for the model by setting initial hyperparameters, such as the number of pre-training iterations and the number of layers of the restricted Boltzmann machine, which helps to quickly achieve a high performance level in the SLM melt pool monitoring task.

[0084] The Adam optimization algorithm can adaptively adjust the learning rate, speed up the convergence of the model, reduce the time required for model training, and enable the model to be applied to actual production more quickly.

[0085] Using the cross entropy loss function can avoid the gradient vanishing problem, making the model more stable during training and improving the reliability of the final model in actual SLM process monitoring.

[0086] By setting the iteration stop conditions, including the maximum number of iterations and the loss function convergence criteria, precise control of model training is ensured, overfitting or underfitting is avoided, and the generalization ability of the model in practical applications is improved.

[0087] Through batch training and dynamic adjustment, the model can better adapt to various complex situations in the SLM process and improve the ability to recognize the molten pool state under different process parameters and material conditions.

[0088] Step S33 includes:

[0089] The mAP value of the trained deep belief network model is calculated on the validation set, and the current mAP value is compared with the preset mAP value. When the current mAP value is not less than the preset mAP value, the model evaluation is completed and the deep belief network model is obtained.

[0090] Specifically, this embodiment evaluates the model through the mAP value, and the mAP value is calculated as follows:

[0091]

[0092] Where d is the number of data categories of the deep belief network model, R(t) is the recall rate of the current timestamp, R(t+1) is the recall rate of the next timestamp, and precision(t,t+1) is the accuracy of the current timestamp and the next timestamp;

[0093] And specify the following parameters:

[0094] TP: The number of samples correctly classified as positive samples; they are actually positive samples and are also classified as positive samples by the model;

[0095] FP: The number of samples that are misclassified as positive samples; they are actually negative samples, but are classified as positive samples by the model;

[0096] TN: The number of samples correctly classified as negative samples; actually negative samples, but also classified as negative samples by the model;

[0097] FN: The number of samples that are misclassified as negative samples; they are actually positive samples, but are classified as negative samples by the model;

[0098] Precision = TP / (TP+FP), recall = TP / (TP+FN);

[0099] Taking the Precision and Recall of each type as the horizontal and vertical axes respectively, finding the AP value is actually equivalent to finding the area enclosed by Precision and Recall. Since the curve area is very jittery, interpolation is used here to smooth the curve area. If the current recall value is R(t), the interpolation value is the maximum Precision value between the next position recall R(t+1). Where t represents the current position. The mAP value is the average of all types of AP values, where d is the number of data categories in the deep belief network model.

[0100] This embodiment calculates the mAP value to comprehensively evaluate the detection and classification capabilities of the model in different categories, ensuring that the model can maintain a high level of performance in various melt pool states. Using the mAP value as an evaluation criterion can effectively identify the performance differences of the model in different categories, ensuring that the final selected model has high accuracy and reliability in practical applications.

[0101] By comparing the current mAP value with the preset mAP value, it is possible to quickly determine whether the model meets the expected performance standards, help select the optimal model configuration, and improve the application effect of the model in SLM melt pool monitoring. If the current mAP value does not meet the preset standard, it can provide a clear direction and basis for further optimization of the model, help adjust the model parameters and structure to improve performance. By interpolating and smoothing the Precision and Recall curves, the impact of curve jitter on the evaluation results is reduced, and the accuracy of mAP value calculation is improved.

[0102] Step S4 includes:

[0103] The calculation formula of the molten pool state discriminant function is:

[0104]

[0105] Among them, f(x) is the molten pool state discriminant function, sign(·) is the signal function, and x i is the i-th sample of the best hyperplane, D is the number of samples of the best hyperplane, η i is the Lagrange multiplier of the i-th sample of the optimal hyperplane, y i For sample x i The observed value, G(x i ,x) is the kernel function, b1 and b2 are the biases of the first molten pool image feature and the second molten pool image feature, respectively.

[0106] Specifically, this embodiment achieves high-precision classification of the molten pool state in the SLM process by constructing a molten pool state discrimination function using a support vector machine algorithm, thereby effectively distinguishing between normal and abnormal molten pool states and improving the quality control capability in the SLM manufacturing process.

[0107] By introducing the kernel function G(x i ,x), which enables the model to perform nonlinear classification in high-dimensional feature space, enhances the model's ability to handle complex molten pool state characteristics, and enables it to adapt to various complex process parameters and material conditions in the SLM process.

[0108] Through the concept of the optimal hyperplane, the model can find the optimal decision boundary in the feature space, improve the accuracy of melt pool state classification, enhance the robustness of the model to boundary conditions, and reduce the possibility of misjudgment.

[0109] By introducing two bias terms, b1 and b2, the model can flexibly process feature information from different sources (such as the first molten pool image features and the second molten pool image features), thus improving the model's ability to utilize multi-source data and making state discrimination more comprehensive and accurate.

[0110] Based on this discriminant function, the system can make real-time judgments on the state of the molten pool during the SLM process. This real-time monitoring capability is crucial for timely detection and correction of abnormal conditions in the manufacturing process, effectively improving production efficiency and product quality.

[0111] By using the Lagrange multiplier η i and sample observations y i ,The model can better handle the sample imbalance problem and effectively reduce the false alarm rate.

[0112] Step S5 includes:

[0113] Based on the molten pool state discriminant function, the molten pool image data and the molten pool acoustic signal data are identified to obtain molten pool state identification parameters, wherein the molten pool state identification parameters include the predicted molten pool depth, the predicted molten pool width, the predicted molten pool temperature, the predicted molten pool vibration frequency and the molten pool radiation intensity. A multidimensional vector is constructed based on the predicted molten pool depth, the predicted molten pool width, the predicted molten pool temperature, the predicted molten pool vibration frequency and the molten pool radiation intensity. The molten pool image data and the molten pool acoustic signal data are updated in real time to update the multidimensional vector in real time to obtain the real-time state of the molten pool, and the quality of the SLM formed part is predicted based on the real-time state of the molten pool.

[0114] Specifically, this embodiment uses a molten pool state discrimination function, combined with molten pool image data and acoustic signal data, to comprehensively identify various state parameters of the molten pool, including depth, width, temperature, vibration frequency, and radiation intensity, thereby providing all-round monitoring of the molten pool state.

[0115] By updating the constructed multi-dimensional vector in real time, the changes in the molten pool state can be captured dynamically, improving the prediction accuracy of the molten pool state, thereby more accurately reflecting the actual situation in the SLM process. By updating the molten pool image and acoustic signal data in real time, the system can monitor the molten pool state in real time, thereby promptly discovering and correcting abnormal conditions in the manufacturing process. Based on the real-time molten pool state, the quality of SLM formed parts can be more accurately predicted, which helps to identify potential quality problems in advance and reduce scrap rate and rework costs.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for monitoring a molten pool during a SLM melting process, characterized in that: The following steps are involved: S1, real-time acquisition of molten pool image data and molten pool acoustic signal data during the SLM process; S2, calculating the vibration velocity of the molten pool based on the molten pool acoustic signal data, and calculating the spectral irradiance of the molten pool based on the molten pool image data; S3, constructing a deep belief network model based on the vibration velocity and spectral irradiation of the molten pool, and extracting features of the molten pool image data according to the deep belief network model to obtain a first molten pool image feature and a second molten pool image feature; S4, combining the first molten pool image features and the second molten pool image features through a support vector machine algorithm, searching for the best hyperplane to obtain a molten pool state discriminant function; S5, identifying the molten pool image data and the molten pool acoustic signal data based on the molten pool state discrimination function to obtain the real-time state of the molten pool, and predicting the quality of the SLM formed part based on the real-time state of the molten pool.

2. A method for monitoring a molten pool in a SLM melting process according to claim 1, characterized in that: Step S1 includes: By arranging an optical sensor and an acoustic sensor around the molten pool, the molten pool image data in the SLM process is collected based on the optical sensor, and the molten pool acoustic signal data in the SLM process is collected based on the acoustic sensor.

3. A method for monitoring a molten pool in a SLM melting process according to claim 2, characterized in that: Step S1 also includes: The molten pool image data includes video images, molten pool optical signal wavelength, and roughly measured temperature of the molten pool. The molten pool acoustic signal data includes acoustic impedance and molten pool acoustic wave velocity. The optical signal wavelength and the roughly measured temperature of the molten pool are used to calculate the spectral irradiation of the molten pool, and the acoustic impedance and the molten pool acoustic wave velocity are used to calculate the vibration velocity of the molten pool.

4. A method for monitoring a molten pool in a SLM melting process according to claim 3, characterized in that: Step S2 includes: The vibration velocity of the molten pool is calculated based on the molten pool acoustic signal data, and the spectral irradiance of the molten pool is calculated based on the molten pool image data. The calculation formulas for the vibration velocity and spectral irradiance of the molten pool are respectively: Where v is the vibration speed of the molten pool, α is the laser absorption coefficient, β is the maximum energy distribution coefficient, γ is the gas adiabatic index, and v 声 is the acoustic wave velocity of the molten pool, Z is the acoustic impedance, f is the spectral irradiance of the molten pool, h and c are Planck constants, k is the Boltzmann constant, λ is the wavelength of the optical signal of the molten pool, T is the rough temperature of the molten pool, and σ is the error bias of the optical sensor.

5. A method for monitoring a molten pool in a SLM melting process according to claim 4, characterized in that: Step S3 includes: S31, constructing an initial deep belief network model based on the vibration velocity and spectral irradiation of the molten pool, obtaining historical molten pool image data, and dividing the historical molten pool image data into a training set and a validation set; S32, iteratively training the initial deep belief network model through the training set, configuring the loss function and hyperparameters, until the iteration stop condition is reached, and obtaining the trained deep belief network model; S33, using the validation set to test the trained deep belief network model, and performing model evaluation on the model. When the model evaluation does not meet the requirements, return to step S32 to adjust the hyperparameters and re-train the model iteratively.

6. A method for monitoring a molten pool in a SLM melting process according to claim 5, characterized in that: Step S31 includes: According to the vibration speed and spectral irradiation of the molten pool, the input layer structure of the initial deep belief network model is determined, the network weights and biases of the initial deep belief network model are initialized, the historical molten pool image data are collected and preprocessed, and the preprocessed historical molten pool image data are divided into a training set and a validation set with a ratio of 7:

3. The hidden layer structure of the initial deep belief network model is determined, including the number of hidden layers and the number of neurons in each hidden layer.

7. A method for monitoring a molten pool in a SLM melting process according to claim 6, characterized in that: Step S32 includes: The initial hyperparameters are set, including the number of pre-training iterations and the number of layers of the restricted Boltzmann machine. The initial deep belief network model uses the cross entropy loss function and the Adam optimization algorithm. For each training batch, the network output is calculated by forward propagation to check whether the iteration stopping condition is reached. The iteration stopping condition includes reaching the maximum number of iterations and the loss function convergence.

8. A method for monitoring a molten pool in a SLM melting process according to claim 7, characterized in that: Step S33 includes: The mAP value of the trained deep belief network model is calculated on the validation set, and the current mAP value is compared with the preset mAP value. When the current mAP value is not less than the preset mAP value, the model evaluation is completed and the deep belief network model is obtained.

9. A method for monitoring a molten pool in a SLM melting process according to claim 8, characterized in that: Step S4 includes: The calculation formula of the molten pool state discriminant function is: Among them, f(x) is the molten pool state discriminant function, sign(·) is the signal function, and x i is the i-th sample of the best hyperplane, D is the number of samples of the best hyperplane, η i is the Lagrange multiplier of the i-th sample of the optimal hyperplane, y i For sample x i The observed value, G(x i ,x) is the kernel function, b1 and b2 are the biases of the first molten pool image feature and the second molten pool image feature, respectively.

10. A method for monitoring a molten pool in a SLM melting process according to claim 9, characterized in that: Step S5 includes: Based on the molten pool state discriminant function, the molten pool image data and the molten pool acoustic signal data are identified to obtain molten pool state identification parameters, wherein the molten pool state identification parameters include the predicted molten pool depth, the predicted molten pool width, the predicted molten pool temperature, the predicted molten pool vibration frequency and the molten pool radiation intensity. A multidimensional vector is constructed based on the predicted molten pool depth, the predicted molten pool width, the predicted molten pool temperature, the predicted molten pool vibration frequency and the molten pool radiation intensity. The molten pool image data and the molten pool acoustic signal data are updated in real time to update the multidimensional vector in real time to obtain the real-time state of the molten pool, and the quality of the SLM formed part is predicted based on the real-time state of the molten pool.

Images