A real-time monitoring system for laser-induced thermal cracking cutting of brittle materials

Abstract

The present invention relates to the field of laser processing technology, and specifically to a real-time monitoring system for laser-induced thermal cracking cutting of brittle materials, comprising a numerical control processing module, a visual acquisition module, an acoustic emission acquisition module, a multimodal information fusion module, and a state evaluation module; wherein: the numerical control processing module is used to generate a cutting laser beam, and control the laser focus to move along a set trajectory according to a preset program; the visual acquisition module is used to capture a high-resolution image of the cutting area in real time; the acoustic emission acquisition module is used to detect the acoustic wave signal generated during the processing; the multimodal information fusion module is used to extract features from the visual signal and the acoustic emission signal. The present invention significantly improves the monitoring accuracy and real-time performance of the laser-induced thermal cracking cutting process by combining multimodal information fusion and deep learning algorithms, thereby optimizing the processing quality, reducing material waste, and enhancing operational safety.

Description

Technical Field

[0001] The invention relates to the technical field of laser processing, and in particular to a real-time monitoring system for a laser-induced thermal cracking cutting process of a brittle material. Background Art

[0002] In modern manufacturing, laser-induced thermal cracking (LITP) cutting technology is widely used for precise cutting of brittle materials such as glass and ceramics. This technology relies on the high temperature generated by the laser to cause local thermal expansion and contraction of the material, thereby inducing thermal stress on the surface of the material, generating and expanding cracks. Although this technology has higher precision and less physical contact damage than traditional mechanical cutting, its monitoring process faces many challenges. Especially in high-speed cutting processes, how to accurately monitor the generation and expansion of cracks in real time to ensure cutting quality and accuracy has always been a technical problem. Existing monitoring methods, such as vision-based monitoring systems, can usually only capture surface crack information, but are often unable to monitor crack depth and microcracks.

[0003] In view of the above problems, the main difficulty of the existing technology is that a single monitoring mode (such as only visual or acoustic emission monitoring) is difficult to fully capture the complex physical phenomena in the cutting process, which limits the response speed and accuracy of the monitoring system. In addition, when processing multiple sensor data, the existing system often lacks an effective data fusion strategy and cannot fully utilize the complementary information provided by different sensors, resulting in unstable monitoring results or large errors.

[0004] Therefore, it is particularly important to develop a system that can integrate multiple monitoring signals and provide accurate monitoring information in real time. Summary of the invention

[0005] Based on the above objectives, the present invention provides a real-time monitoring system for the laser-induced thermal cracking cutting process of brittle materials.

[0006] A real-time monitoring system for laser-induced thermal cracking cutting of brittle materials, comprising a numerical control processing module, a visual acquisition module, an acoustic emission acquisition module, a multimodal information fusion module and a state evaluation module; wherein:

[0007] CNC machining module: including a laser generator and a CNC mobile platform. The laser generator is used to generate a cutting laser beam, and the CNC mobile platform is used to control the laser focus to move along a set trajectory according to a preset program and adjust the laser action position and path;

[0008] Visual acquisition module: including CCD industrial camera and image acquisition card. The CCD industrial camera is used to capture high-resolution images of the cutting area in real time. The image acquisition card is used to convert the captured image signals into digital signals and transmit them to the multimodal information fusion module.

[0009] Acoustic emission acquisition module: including an acoustic emission sensor, a preamplifier and an acoustic emission signal collector, wherein the acoustic emission sensor is used to detect the acoustic wave signal generated during the processing, the preamplifier is used to enhance the acoustic wave signal, and the acoustic emission signal collector is used to convert the enhanced acoustic wave signal into a digital format and transmit it to the multimodal information fusion module;

[0010] Multimodal information fusion module: receives signal inputs from the visual acquisition module and the acoustic emission acquisition module, extracts features from the visual signal and the acoustic emission signal using the window attention and multi-head self-attention mechanisms, and achieves deep fusion of the two modalities through the cross-modal interactive encoder, providing pre-processed feature data for the state assessment module;

[0011] State evaluation module: Receives the output of the multimodal information fusion module, applies the preset deep neural network algorithm, analyzes the fused feature data, adjusts the weights of the multimodal interaction fusion features through the feedforward neural network, and finally calculates the classification probability of the corresponding processing state through the SoftMax layer, and outputs the evaluation results of the processing status.

[0012] Furthermore, the laser generator includes a wavelength modulation unit, a power control unit and a beam focusing unit; wherein:

[0013] Wavelength modulation unit: including variable wavelength laser diode, which can adjust the wavelength of laser according to the optical absorption characteristics of the material, and optimize the interaction between laser and brittle material by precise control of wavelength, thereby reducing the heat affected zone;

[0014] Power control unit: A current controller is used to adjust the power supply of the laser diode, thereby accurately controlling the output power of the laser. The specific power adjustment is optimized according to the thickness of the material and the required cutting speed to ensure the reasonable distribution of energy;

[0015] Beam focusing unit: Utilizes a high-performance optical lens system, including a combination of convex lens and cylindrical lens, to focus the laser beam to the focal point. By adjusting the lens spacing, the spot diameter can be precisely controlled to ensure that the laser energy is concentrated in the smallest area, thus enhancing the cutting accuracy and depth.

[0016] The numerical control mobile platform includes a positioning control unit and a driving execution unit; wherein:

[0017] Positioning control unit: Based on high-precision CNC device, it uses path planning algorithms including interpolation algorithm and trajectory optimization algorithm to calculate the optimal movement trajectory of the laser head, and dynamically adjusts the path to maintain cutting accuracy according to the actual situation in the processing process including material displacement or deformation through real-time feedback mechanism;

[0018] Drive execution unit: It is composed of a stepper motor and a synchronous belt drive device, which is used to achieve precise position control of the laser head. The stepper motor moves the laser head at a preset speed and accuracy by receiving instructions from the positioning control unit. At the same time, the synchronous belt is used to ensure the smoothness and accuracy of the movement to avoid affecting the cutting quality due to mechanical vibration.

[0019] Furthermore, the CCD industrial camera is equipped with a CCD sensor for capturing high-resolution images of the cutting area in real time. The CCD sensor can capture clear images under various lighting conditions. The CCD industrial camera is also equipped with an optical filtering device that can effectively filter laser scattered or reflected light to ensure that the image meets predetermined standards. The camera's exposure time and frame rate can be adjusted according to the speed and accuracy requirements of laser cutting to capture each stage of the cutting process; the image acquisition card has a built-in analog-to-digital converter for converting analog signals into digital signals. The converter has a high sampling rate and high data width to ensure that the details of the image are not lost during the conversion process, thereby maintaining the high resolution and depth of the image. After the conversion is completed, the acquisition card transmits the digital image data to the multimodal information fusion module via a high-speed data interface.

[0020] Furthermore, the acoustic emission sensor adopts a highly sensitive piezoelectric sensor, which is used to generate a sound wave of a predetermined frequency when a tiny crack expands in the material during the laser-induced thermal cracking cutting process. The piezoelectric sensor can convert the mechanical vibration of the sound wave of the predetermined frequency into an electrical signal; after the preamplifier receives the weak electrical signal output by the acoustic emission sensor, it amplifies it through a low-noise amplifier circuit. The amplifier has the characteristics of high gain and wide bandwidth, which can ensure that the signal maintains its original frequency characteristics and amplitude information during the amplification process, thereby amplifying the acoustic emission signal without distortion. In addition, the preamplifier is also equipped with a filtering function, which can filter out noise in the non-target frequency range and enhance signal clarity; the acoustic emission signal collector includes a high-precision analog-to-digital converter, which is used to convert the analog acoustic emission signal output by the preamplifier into a digital format. The converter has high resolution and fast sampling rate, and can accurately capture and convert high-speed changing sound wave signals to ensure that the details and dynamic range of the signal are retained during the digitization process. The converted digital signal is transmitted to the multimodal information fusion module through a high-speed data interface. The interface supports real-time transmission of data to ensure that the acoustic emission data can be processed synchronously with the visual data.

[0021] Furthermore, the multimodal information fusion module includes a visual feature extraction unit, an acoustic emission feature extraction unit and a cross-modal interactive fusion unit; wherein:

[0022] Visual feature extraction unit: It applies the window attention mechanism to extract features. Specifically, it defines multiple small windows on the image and focuses on analyzing the local image features within each window to enhance the understanding of the starting point and expansion path of the crack in the cutting area. The size and step length of the window are dynamically adjusted according to the resolution of the actual image and the importance of the cutting features to optimize the efficiency and accuracy of feature capture.

[0023] Acoustic emission feature extraction unit: A multi-head self-attention mechanism is used to process the acoustic emission signal. Specifically, the attention is distributed to multiple predetermined parts of the acoustic emission signal, and the acoustic wave activities at multiple frequencies and time points are focused on at the same time, so as to comprehensively monitor the complex material stress response. Each attention head independently extracts a set of features from the signal, and then forms a comprehensive, multi-dimensional acoustic emission signal feature representation by fusing all the extracted signal features;

[0024] Cross-modal interactive fusion unit: It is used to deeply fuse the features output by the visual feature extraction unit and the acoustic emission feature extraction unit. It uses a cross-modal interactive encoder to process and integrate data from different sensor sources. The encoder uses a series of fusion layers, including sharing layers and interactive layers, to deeply analyze and integrate the data of the two modes. This layer uses nonlinear transformation and adaptive weight adjustment to accurately reconcile the information in the visual and acoustic emission signals, and finally produces a fused feature set.

[0025] Furthermore, the visual feature extraction unit includes:

[0026] Window definition and parameter setting: First, dynamically set the window size W and step size S according to the high-resolution image data provided by the visual acquisition module. The window size W determines the area range of each analysis, and the step size S determines the distance the window moves. These two parameters jointly affect the granularity and processing speed of feature extraction;

[0027] Application of window attention mechanism: For each window area, an improved window attention mechanism is applied. This mechanism is based on the Softmax function to assign weights to each pixel in the window. The attention score calculation formula is:

[0028] Among them, A ij is the attention weight of the pixel at position (i, j), W ij It is the original weight value calculated by the feedforward neural network. exp represents the exponential function, which is used to enhance the nonlinear expression ability of the feature.

[0029] Feature weighting and aggregation: The feature vector of each pixel will be weighted according to its corresponding attention weight and then aggregated within the window to form a representative feature vector for the window, specifically through the following formula: F 窗口 =∑ i,j∈窗口A ij ·X ij , where F 窗口 is the aggregate feature vector of the window, X ij is the pixel feature at position (i, j), ∑ represents the summation operation;

[0030] Feature mapping and optimization: The aggregated window feature vectors are fed into a preset deep neural network for nonlinear mapping and optimization to extract deep visual information. The network includes multi-layer perceptrons and activation functions to enhance the expressiveness and adaptability of the model.

[0031] Furthermore, the acoustic emission feature extraction unit includes:

[0032] Signal preprocessing: Receive the raw analog signal from the acoustic emission acquisition module, convert it into a digital signal through an analog-to-digital converter, and then perform signal preprocessing, including denoising and normalization, to optimize signal quality and prepare for advanced feature extraction;

[0033] Application of multi-head self-attention mechanism: The preprocessed signal is input into the multi-head self-attention mechanism. Specifically, the original acoustic emission signal is passed through a linear transformation layer to generate query Q, key K and value V matrices. The specific formulas are: Q = W Q ·X;K=W K X and V = W V ·X; where X is the input feature matrix, W Q ,W K ,W V is a learnable weight matrix; the attention score is then calculated using the dot product of the query and the key and scaled by Softmax. The specific formula is:

[0034] Among them, d k is the dimension of the key vector;

[0035] Feature aggregation and fusion: In the multi-head self-attention mechanism, each head processes the signal independently and captures the features of the acoustic emission signal from different angles. The feature vectors output by each head are then aggregated. The aggregation formula is: MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O , where each head i =Attention(Q i ,k i ,V i ),W O is the weight matrix of the output layer; this aggregate integrates the information extracted by different heads to form a comprehensive representation of the acoustic emission signal features;

[0036] Output feature optimization: The aggregated multi-dimensional acoustic emission signal features are processed through preset deep neural network layers, including activation functions and batch normalization, to optimize the robustness and expressiveness of the signal representation;

[0037] Feature output: After processing and optimization, the comprehensive acoustic emission signal features formed represent the final output.

[0038] Furthermore, the cross-modal interaction fusion unit includes:

[0039] Feature reception and preliminary integration: The cross-modal interactive fusion unit first receives the visual features F from the visual feature extraction unit. 视觉 and the acoustic emission feature F of the acoustic emission feature extraction unit 声发射 , specifically in the form of vectors, containing the information of each mode;

[0040] Feature conversion and dimension upgrade: In order to effectively integrate the features of the two modes in the same space, the features are first converted and dimension upgraded. The specific formula is:

[0041] F 视觉 ′=W 视觉 ·F 视觉 ;

[0042] F 声发射 ′=W 声发射 ·F 声发射 , where W 视觉 and W 声发射 It is a learnable weight matrix used to adjust the original features to a form suitable for fusion;

[0043] Feature interaction fusion: Next, the interactive fusion layer is used to perform deep interaction and fusion of visual and acoustic emission features. The interactive fusion layer uses nonlinear transformation and adaptive weight adjustment to accurately reconcile the features of the two modes. The specific formula is: F 融合 =σ(W 交互 ·(F 视觉 ′+F 声发射 ′)+b 交互 ), where w 交互 is the weight matrix of the interactive fusion layer, b 交互 is a bias term, and σ is a nonlinear activation function used to increase the expressiveness and nonlinear characteristics of the model;

[0044] Feature optimization and output: Feature F after processing by the interactive fusion layer 融合 Optimization is performed through a multi-layer neural network, including batch normalization and further nonlinear transformation to improve the generalization and robustness of the features; finally, the optimized features are output as the fused feature set F 最终, the feature set F 最终 It will be used in the subsequent status assessment module to provide support for monitoring and decision-making of the cutting process.

[0045] Furthermore, the status assessment module includes:

[0046] Feature reception and preliminary processing: The state assessment module first receives the fused feature set F output by the multimodal information fusion module. 最终 , providing a data basis for status assessment;

[0047] Application of feedforward neural network: The feature data is first input into the preset feedforward neural network for analysis and processing. The feedforward neural network includes multiple hidden layers, and each layer uses the following calculation formula for feature conversion and activation: F 层 =σ(W·F 上一层 +b), where W and b are the weight matrix and bias vector respectively, and σ is a nonlinear activation function used to enhance the network's ability to handle complex nonlinear relationships;

[0048] Feature weight adjustment: In a feedforward neural network, the output of each layer is adjusted according to the learned weights to optimize its contribution to subsequent layers;

[0049] SoftMax layer classification probability calculation: The feature data processed by the multi-layer feedforward network is sent to the SoftMax layer to convert the real number vector output by the previous layer into a probability distribution. The value of each element represents the probability that the data belongs to the corresponding category. The specific calculation formula is: Among them, P(y i ) is the probability of category i, F i is the score of the feature value output by the previous layer corresponding to category i;

[0050] Output the evaluation results of the processing status: The classification probability output by the SoftMax layer is used to determine the final processing status, specifically whether there is unqualified crack extension.

[0051] Beneficial effects of the present invention:

[0052] The present invention significantly improves the monitoring accuracy and real-time performance during laser-induced thermal cracking cutting by introducing multimodal information fusion technology and deep learning algorithms. By comprehensively utilizing visual and acoustic emission monitoring data, the system can more comprehensively capture the generation and expansion of cracks, thereby ensuring precise control of the cutting process. This high-precision monitoring capability is particularly suitable for demanding precision machining fields, such as the optoelectronic industry and precision device manufacturing, and effectively improves machining quality and product consistency.

[0053] The present invention helps to reduce material waste and secondary processing caused by monitoring errors and reduce production costs. The use of a multimodal fusion analysis framework optimizes the data processing process, improves the system's adaptability and response speed to complex situations, and enables the operator to adjust parameters in a timely manner during the cutting process to prevent unnecessary material loss.

[0054] The present invention, through real-time and accurate monitoring and evaluation, the system can instantly detect potential cutting failures or non-conformance to specifications, and make timely interventions and adjustments, which not only ensures the safety of operators, but also provides strong support for enterprises to maintain high-quality production standards. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the present invention or 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 in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0056] Figure 1 A schematic diagram of a high-precision real-time monitoring system according to an embodiment of the present invention;

[0057] Figure 2 Schematic diagram of a multimodal information fusion module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific implementation cases.

[0059] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0060] like Figure 1-Figure 2As shown, a real-time monitoring system for laser-induced thermal cracking cutting of brittle materials includes a numerical control processing module, a visual acquisition module, an acoustic emission acquisition module, a multimodal information fusion module and a state evaluation module; wherein:

[0061] CNC machining module: including laser generator and CNC mobile platform. The laser generator is used to generate cutting laser beam. The CNC mobile platform is used to control the laser focus to move along the set trajectory according to the preset program and adjust the laser action position and path to ensure accurate cutting of brittle materials.

[0062] Visual acquisition module: including CCD industrial camera and image acquisition card. The CCD industrial camera is used to capture high-resolution images of the cutting area in real time. The image acquisition card is used to convert the captured image signals into digital signals and transmit them to the multimodal information fusion module to provide data support for subsequent image processing and analysis.

[0063] Acoustic emission acquisition module: including an acoustic emission sensor, a preamplifier and an acoustic emission signal collector, wherein the acoustic emission sensor is used to detect the acoustic wave signal generated during the processing, the preamplifier is used to enhance the acoustic wave signal, and the acoustic emission signal collector is used to convert the enhanced acoustic wave signal into a digital format and transmit it to the multimodal information fusion module;

[0064] Multimodal information fusion module: built in Linux environment, using deep learning model based on Pytorch framework, receiving signal input from visual acquisition module and acoustic emission acquisition module, using window attention and multi-head self-attention mechanism to extract features of visual signal and acoustic emission signal, and realizing deep fusion of two modalities through cross-modal interactive encoder, providing pre-processed feature data for state assessment module;

[0065] State evaluation module: It is also based on the Pytorch framework. It receives the output of the multimodal information fusion module, applies the preset deep neural network algorithm, analyzes the fused feature data, adjusts the weights of the multimodal interaction fusion features through the feedforward neural network, and finally calculates the classification probability of the corresponding processing state through the SoftMax layer, and outputs the evaluation results of the processing status.

[0066] The laser generator includes a wavelength modulation unit, a power control unit and a beam focusing unit; wherein:

[0067] Wavelength modulation unit: including variable wavelength laser diode, which can adjust the wavelength of laser according to the optical absorption characteristics of the material, and optimize the interaction between laser and brittle materials (such as glass or ceramics) through precise control of wavelength, thereby reducing the heat affected zone and improving cutting quality;

[0068] Power control unit: A current controller is used to adjust the power supply of the laser diode, thereby accurately controlling the output power of the laser. The specific power adjustment is optimized according to the thickness of the material and the required cutting speed to ensure the reasonable distribution of energy and achieve efficient and clear cutting effects;

[0069] Beam focusing unit: uses a high-performance optical lens system, including a combination of convex lenses and cylindrical lenses, to focus the laser beam to a focal point as small as possible. By adjusting the lens spacing, the spot diameter can be precisely controlled to ensure that the laser energy is concentrated in the smallest area, thereby enhancing the cutting accuracy and depth.

[0070] The CNC mobile platform includes a positioning control unit and a drive execution unit; wherein:

[0071] Positioning control unit: Based on high-precision CNC device, it uses path planning algorithms including interpolation algorithm and trajectory optimization algorithm to calculate the optimal movement trajectory of the laser head, and dynamically adjusts the path to maintain cutting accuracy according to the actual situation in the processing process including material displacement or deformation through real-time feedback mechanism;

[0072] Drive execution unit: It is composed of a stepper motor and a synchronous belt drive device, which is used to achieve precise position control of the laser head. The stepper motor moves the laser head at a preset speed and accuracy by receiving instructions from the positioning control unit. At the same time, the synchronous belt is used to ensure the smoothness and accuracy of the movement to avoid the impact of mechanical vibration on the cutting quality. The design and coordinated work of the above units and components ensure that while the laser generator generates a laser beam suitable for cutting brittle materials, the CNC mobile platform can accurately control the laser head to move along an optimized trajectory, thereby maximizing the quality and safety of material processing while ensuring cutting efficiency and accuracy.

[0073] CCD industrial cameras are equipped with CCD sensors to capture high-resolution images of the cutting area in real time. With its high sensitivity and low noise characteristics, CCD sensors can capture clear images under various lighting conditions, especially in the complex lighting environment during laser cutting. In addition, CCD industrial cameras are also equipped with optical filtering devices that can effectively filter laser scattered or reflected light to ensure that the image meets the predetermined standards. The camera's exposure time and frame rate can be adjusted according to the speed and accuracy requirements of laser cutting to capture each stage of the cutting process; the image acquisition card has a built-in analog-to-digital converter (ADC) to convert analog signals into digital signals. The converter has a high sampling rate and high data width to ensure that the details of the image are not lost during the conversion process, thereby maintaining the high resolution and depth of the image. After the conversion is completed, the acquisition card transmits the digital image data to the multimodal information fusion module through a high-speed data interface (such as a PCIe interface). This interface supports large-bandwidth data transmission, ensuring that image data can be transmitted in real time even at high resolution, meeting the system's real-time requirements. The above details illustrate the specific technical implementation of the visual acquisition module in capturing cutting area images and image signal conversion and transmission, and demonstrates how the module effectively provides high-quality and real-time visual data for the multimodal information fusion module, supporting the system to perform high-precision cutting process monitoring.

[0074] The acoustic emission sensor uses a highly sensitive piezoelectric sensor, which is used in the process of laser-induced thermal cracking cutting. When a tiny crack in the material expands, a sound wave of a predetermined frequency is generated. The piezoelectric sensor can convert the mechanical vibration of the sound wave of the predetermined frequency into an electrical signal; after the preamplifier receives the weak electrical signal output by the acoustic emission sensor, it amplifies it through a low-noise amplifier circuit. The amplifier has the characteristics of high gain and wide bandwidth, which can ensure that the signal maintains its original frequency characteristics and amplitude information during the amplification process, thereby amplifying the acoustic emission signal without distortion. In addition, the preamplifier is also equipped with a filtering function, which can filter out noise in the non-target frequency range and enhance signal clarity; the acoustic emission signal collector includes a high-precision analog-to-digital converter (ADC), Used to convert the analog acoustic emission signal output by the preamplifier into a digital format, the converter has high resolution and fast sampling rate, can accurately capture and convert high-speed changing acoustic wave signals, ensure that the details and dynamic range of the signal are retained during the digitization process, and the converted digital signal is transmitted to the multimodal information fusion module through a high-speed data interface. The interface supports real-time data transmission to ensure that the acoustic emission data can be processed synchronously with the visual data; the design and implementation of the above parts ensure that the acoustic emission acquisition module can efficiently and accurately capture, amplify and convert the acoustic wave signals generated during the laser-induced thermal cracking cutting process, and provide key acoustic monitoring data for the multimodal information fusion module, thereby supporting the system to conduct more comprehensive cutting process analysis and status evaluation.

[0075] The multimodal information fusion module includes a visual feature extraction unit, an acoustic emission feature extraction unit, and a cross-modal interactive fusion unit; wherein:

[0076] Visual feature extraction unit: It applies the window attention mechanism to extract features. Specifically, it defines multiple small windows on the image and focuses on analyzing the local image features within each window to enhance the understanding of the starting point and expansion path of the crack in the cutting area. The size and step length of the window are dynamically adjusted according to the resolution of the actual image and the importance of the cutting features to optimize the efficiency and accuracy of feature capture.

[0077] Acoustic emission feature extraction unit: A multi-head self-attention mechanism is used to process the acoustic emission signal. Specifically, the attention is distributed to multiple predetermined parts of the acoustic emission signal, and the acoustic wave activities at multiple frequencies and time points are focused on at the same time, so as to comprehensively monitor the complex material stress response. Each attention head independently extracts a set of features from the signal, and then forms a comprehensive, multi-dimensional acoustic emission signal feature representation by fusing all the extracted signal features, which is crucial for understanding and predicting crack propagation behavior.

[0078] Cross-modal interactive fusion unit: used to deeply fuse the features output by the visual feature extraction unit and the acoustic emission feature extraction unit. By using the cross-modal interactive encoder, a deep learning-based fusion architecture specially designed to process and integrate data from different sensor sources, the encoder deeply analyzes and integrates the two modes of data through a series of fusion layers, including sharing layers and interaction layers. This layer uses nonlinear transformations and adaptive weight adjustments to accurately reconcile the information in the visual and acoustic emission signals, and finally produces a fused feature set. These features have higher predictive ability and robustness; each unit in the multimodal information fusion module is interdependent and works in collaboration, which together improves the real-time monitoring performance of the entire system during laser-induced thermal cracking cutting of brittle materials. This structural design not only enhances the accuracy of feature extraction, but also optimizes the efficiency and practicality of the data processing process.

[0079] The visual feature extraction unit includes:

[0080] Window definition and parameter setting: First, dynamically set the window size W and step size S according to the high-resolution image data provided by the visual acquisition module. The window size W determines the area range of each analysis, and the step size S determines the distance the window moves. These two parameters jointly affect the granularity and processing speed of feature extraction;

[0081] Application of window attention mechanism: For each window area, an improved window attention mechanism is applied. This mechanism is based on the Softmax function to assign weights to each pixel in the window. The attention score calculation formula is:

[0082] Among them, A i is the attention weight of the pixel at position (i, j), W ij It is the original weight value calculated by the feedforward neural network, which represents the importance of the pixel for crack monitoring in the cutting area. exp represents the exponential function, which is used to enhance the nonlinear expression ability of the feature.

[0083] Feature weighting and aggregation: The feature vector of each pixel will be weighted according to its corresponding attention weight and then aggregated within the window to form a representative feature vector for the window, specifically through the following formula: F 窗口 =∑ i,j∈窗口 A ij ·X ij , where F 窗口 is the aggregate feature vector of the window, X ij is the pixel feature at position (i, j). This calculation reflects the weighted sum of the contributions of each pixel in the local area to the overall crack identification. ∑ represents the summation operation, ensuring that all features in the window are considered;

[0084] Feature mapping and optimization: The aggregated window feature vectors are fed into a preset deep neural network for nonlinear mapping and optimization processing to extract deep visual information. The network includes a multi-layer perceptron (MLP) and an activation function to enhance the expressiveness and adaptability of the model. After the above processing, the feature vector generated by each window will be encoded and standardized to meet the needs of the cross-modal interactive fusion unit in the multimodal fusion module. This step ensures that the information extracted from the visual signal can be effectively combined with the characteristics of the acoustic emission signal.

[0085] The acoustic emission feature extraction unit includes:

[0086] Signal preprocessing: Receive the raw analog signal from the acoustic emission acquisition module, convert it into a digital signal through an analog-to-digital converter, and then perform signal preprocessing, including denoising and normalization, to optimize signal quality and prepare for advanced feature extraction;

[0087] Application of multi-head self-attention mechanism: The preprocessed signal is input into the multi-head self-attention mechanism. Specifically, the original acoustic emission signal is passed through a linear transformation layer to generate query Q, key K and value V matrices. The specific formulas are: Q = W Q ·X;K=WW K X and V = W V ·X; where X is the input feature matrix, W Q ,W K ,W Vis a learnable weight matrix; the attention score is then calculated using the dot product of the query and the key and scaled by Softmax. The specific formula is:

[0088] Among them, d k is the dimension of the key vector, this scaling factor helps stabilize the gradient;

[0089] Feature aggregation and fusion: In the multi-head self-attention mechanism, each head processes the signal independently and captures the features of the acoustic emission signal from different angles. The feature vectors output by each head are then aggregated. The aggregation formula is: MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O , where each head i =Attention(Q i ,K i ,V i ),W O is the weight matrix of the output layer; this aggregate integrates the information extracted by different heads to form a comprehensive representation of the acoustic emission signal features;

[0090] Output feature optimization: The aggregated multi-dimensional acoustic emission signal features are processed through preset deep neural network layers, including activation functions and batch normalization, to optimize the robustness and expressiveness of the signal representation;

[0091] Feature output: After processing and optimization, the comprehensive acoustic emission signal feature representation is formed as the final output for further use by the multimodal information fusion module for state assessment and monitoring decision support; the above steps ensure the efficient processing of the acoustic emission signal through the multi-head self-attention mechanism, allowing the system to fully capture the complex dynamics of material stress changes, thereby improving the monitoring system's responsiveness and accuracy to crack propagation.

[0092] The cross-modal interaction fusion unit includes:

[0093] Feature reception and preliminary integration: The cross-modal interactive fusion unit first receives the visual features F from the visual feature extraction unit. 视觉 and the acoustic emission feature F of the acoustic emission feature extraction unit 声发射 , specifically in the form of vectors, containing the information of each mode;

[0094] Feature conversion and dimension upgrade: In order to effectively integrate the features of the two modes in the same space, the features are first converted and dimension upgraded. The specific formula is:

[0095] F 视觉 ′=W 视觉 ·F 视觉 ;

[0096] F 声发射 ′=W 声发射 ·F 声发射 , where W 视觉 and W 声发射 It is a learnable weight matrix used to adjust the original features to a form suitable for fusion;

[0097] Feature interaction fusion: Next, the interactive fusion layer is used to perform deep interaction and fusion of visual and acoustic emission features. The interactive fusion layer uses nonlinear transformation and adaptive weight adjustment to accurately reconcile the features of the two modes. The specific formula is: F 融合 =σ(W 交互 ·(f 视觉 ′+F 声发射 ′)+b 交互 ), where W 交互 is the weight matrix of the interactive fusion layer, b 交互 is a bias term, and σ is a nonlinear activation function (such as ReLU or Sigmoid), which is used to increase the expressiveness and nonlinear characteristics of the model;

[0098] Feature optimization and output: Feature F after processing by the interactive fusion layer 融合 Optimization is performed through a multi-layer neural network, including batch normalization and further nonlinear transformation to improve the generalization and robustness of the features; finally, the optimized features are output as the fused feature set F 最终 , the feature set F 最终 It will be used in the subsequent state assessment module to provide support for the monitoring and decision-making of the cutting process; the above steps elaborate on how to effectively fuse the visual and acoustic emission signals through advanced interactive fusion technology to generate a feature set with high information integration, so as to ensure that the system can respond accurately and quickly in a complex monitoring environment.

[0099] The condition assessment modules include:

[0100] Feature reception and preliminary processing: The state assessment module first receives the fused feature set F output by the multimodal information fusion module. 最终 ,These feature sets contain key information extracted and fused from visual and acoustic emission signals, providing a data basis for state assessment;

[0101] Application of feedforward neural network: The feature data is first input into the preset feedforward neural network for analysis and processing. The feedforward neural network includes multiple hidden layers, and each layer uses the following calculation formula for feature conversion and activation: F 层 =σ(W·F 上一层+b), where W and b are the weight matrix and bias vector respectively, and σ is a nonlinear activation function (usually ReLU or Sigmoid), which is used to enhance the network's ability to handle complex nonlinear relationships;

[0102] Feature weight adjustment: In a feedforward neural network, the output of each layer is adjusted according to the learned weights to optimize the contribution to subsequent layers and enhance the model's ability to recognize various states during the laser cutting process;

[0103] SoftMax layer classification probability calculation: The feature data processed by the multi-layer feedforward network is sent to the SoftMax layer to convert the real number vector output by the previous layer into a probability distribution. The value of each element represents the probability that the data belongs to the corresponding category. The specific calculation formula is: Among them, P(y i ) is the probability of category i, F i is the score of the feature value output by the previous layer corresponding to category i;

[0104] Output the evaluation results of the processing status: The classification probability output by the SoftMax layer is used to determine the final processing status, specifically whether there is unqualified crack extension. The status evaluation module outputs the final processing status based on these probabilities and provides accurate feedback information to the operator or automatic control system. Through the above steps, the status evaluation module can accurately analyze and evaluate various possible states in the laser cutting process, provide real-time and accurate monitoring output, and ensure quality control and safety monitoring of the cutting process.

[0105] The present invention is intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A real-time monitoring system for laser-induced thermal cracking cutting of brittle materials, characterized in that: It includes CNC machining module, visual acquisition module, acoustic emission acquisition module, multimodal information fusion module and state assessment module; among which: CNC machining module: including a laser generator and a CNC mobile platform. The laser generator is used to generate a cutting laser beam, and the CNC mobile platform is used to control the laser focus to move along a set trajectory according to a preset program and adjust the laser action position and path; Visual acquisition module: including CCD industrial camera and image acquisition card. The CCD industrial camera is used to capture high-resolution images of the cutting area in real time. The image acquisition card is used to convert the captured image signals into digital signals and transmit them to the multimodal information fusion module. Acoustic emission acquisition module: including an acoustic emission sensor, a preamplifier and an acoustic emission signal collector, wherein the acoustic emission sensor is used to detect the acoustic wave signal generated during the processing, the preamplifier is used to enhance the acoustic wave signal, and the acoustic emission signal collector is used to convert the enhanced acoustic wave signal into a digital format and transmit it to the multimodal information fusion module; Multimodal information fusion module: receives signal inputs from the visual acquisition module and the acoustic emission acquisition module, extracts features from the visual signal and the acoustic emission signal using the window attention and multi-head self-attention mechanisms, and achieves deep fusion of the two modalities through the cross-modal interactive encoder, providing pre-processed feature data for the state assessment module; The multimodal information fusion module includes a visual feature extraction unit, an acoustic emission feature extraction unit and a cross-modal interactive fusion unit; wherein: Visual feature extraction unit: It applies the window attention mechanism to extract features. Specifically, it defines multiple small windows on the image and focuses on analyzing the local image features within each window to enhance the understanding of the starting point and expansion path of the crack in the cutting area. The size and step length of the window are dynamically adjusted according to the resolution of the actual image and the importance of the cutting features to optimize the efficiency and accuracy of feature capture. Acoustic emission feature extraction unit: A multi-head self-attention mechanism is used to process the acoustic emission signal. Specifically, the attention is distributed to multiple predetermined parts of the acoustic emission signal, and the acoustic wave activities at multiple frequencies and time points are focused on at the same time, so as to comprehensively monitor the complex material stress response. Each attention head independently extracts a set of features from the signal, and then forms a comprehensive, multi-dimensional acoustic emission signal feature representation by fusing all the extracted signal features; Cross-modal interactive fusion unit: used to deeply fuse the features output by the visual feature extraction unit and the acoustic emission feature extraction unit. The cross-modal interactive encoder is used to process and integrate data from different sensor sources. The encoder deeply analyzes and integrates the data of the two modes through a series of fusion layers, including sharing layers and interactive layers. This layer uses nonlinear transformation and adaptive weight adjustment to accurately reconcile the information in the visual and acoustic emission signals, and finally generates a fused feature set. The visual feature extraction unit comprises: Window definition and parameter setting: First, dynamically set the window size W and step size S according to the high-resolution image data provided by the visual acquisition module. The window size W determines the area range of each analysis, and the step size S determines the distance the window moves. These two parameters jointly affect the granularity and processing speed of feature extraction; Application of window attention mechanism: For each window area, an improved window attention mechanism is applied. This mechanism is based on the Softmax function to assign weights to each pixel in the window. The attention score calculation formula is: W ij It is the original weight value calculated by the feedforward neural network. exp represents the exponential function, which is used to enhance the nonlinear expression ability of the feature. Feature weighting and aggregation: The feature vector of each pixel will be weighted according to its corresponding attention weight and then aggregated within the window to form a representative feature vector for the window, specifically through the following formula: F 窗口 =∑ i,j∈窗口 A ij ·X ij , where F 窗口 is the aggregate feature vector of the window, X ij is the pixel feature at position (i, j), ∑ represents the summation operation; Feature mapping and optimization: The aggregated window feature vectors are fed into a preset deep neural network for nonlinear mapping and optimization to extract deep visual information. The network includes a multi-layer perceptron and activation functions to enhance the model’s expressiveness and adaptability. The acoustic emission feature extraction unit comprises: Signal preprocessing: Receive the raw analog signal from the acoustic emission acquisition module, convert it into a digital signal through an analog-to-digital converter, and then perform signal preprocessing, including denoising and normalization, to optimize signal quality and prepare for advanced feature extraction; Application of multi-head self-attention mechanism: The preprocessed signal is input into the multi-head self-attention mechanism. Specifically, the original acoustic emission signal is passed through a linear transformation layer to generate query Q, key K and value W matrices. The specific formulas are: Q = V Q ·X;K=W K X and V = W V ·X; where X is the input feature matrix, W Q ,W K ,W V is a learnable weight matrix; the attention score is then calculated using the dot product of the query and the key and scaled by Softmax. The specific formula is: Among them, d k is the dimension of the key vector; Feature aggregation and fusion: In the multi-head self-attention mechanism, each head processes the signal independently and captures the features of the acoustic emission signal from different angles. The feature vectors output by each head are then aggregated. The aggregation formula is: MultiHead(Q,k,V)=Concat(head1,head2,…,head h )W O , where each head i =Attention(Q i ,K i ,V i ),W O is the weight matrix of the output layer; this aggregate integrates the information extracted by different heads to form a comprehensive representation of the acoustic emission signal features; Output feature optimization: The aggregated multi-dimensional acoustic emission signal features are processed through preset deep neural network layers, including activation functions and batch normalization, to optimize the robustness and expressiveness of the signal representation; Feature output: After processing and optimization, the comprehensive acoustic emission signal features formed represent the final output; State evaluation module: Receives the output of the multimodal information fusion module, applies the preset deep neural network algorithm, analyzes the fused feature data, adjusts the weights of the multimodal interaction fusion features through the feedforward neural network, and finally calculates the classification probability of the corresponding processing state through the SoftMax layer, and outputs the evaluation results of the processing status.

2. A real-time monitoring system for laser-induced thermal cracking cutting of brittle materials according to claim 1, characterized in that: The laser generator comprises a wavelength modulation unit, a power control unit and a beam focusing unit; wherein: Wavelength modulation unit: including variable wavelength laser diode, which can adjust the wavelength of laser according to the optical absorption characteristics of the material, and optimize the interaction between laser and brittle material by precise control of wavelength, thereby reducing the heat affected zone; Power control unit: A current controller is used to adjust the power supply of the laser diode, thereby accurately controlling the output power of the laser. The specific power adjustment is optimized according to the thickness of the material and the required cutting speed to ensure the reasonable distribution of energy; Beam focusing unit: Utilizes a high-performance optical lens system, including a combination of convex lens and cylindrical lens, to focus the laser beam to the focal point. By adjusting the lens spacing, the spot diameter can be precisely controlled to ensure that the laser energy is concentrated in the smallest area, thus enhancing the cutting accuracy and depth. The numerical control mobile platform includes a positioning control unit and a driving execution unit; wherein: Positioning control unit: Based on high-precision CNC device, it uses path planning algorithms including interpolation algorithm and trajectory optimization algorithm to calculate the optimal movement trajectory of the laser head, and dynamically adjusts the path to maintain cutting accuracy according to the actual situation in the processing process including material displacement or deformation through real-time feedback mechanism; Drive execution unit: It is composed of a stepper motor and a synchronous belt drive device, which is used to achieve precise position control of the laser head. The stepper motor moves the laser head at a preset speed and accuracy by receiving instructions from the positioning control unit. At the same time, the synchronous belt is used to ensure the smoothness and accuracy of the movement to avoid affecting the cutting quality due to mechanical vibration.

3. A real-time monitoring system for laser-induced thermal cracking cutting of brittle materials according to claim 1, characterized in that: The CCD industrial camera is equipped with a CCD sensor for capturing high-resolution images of the cutting area in real time. The CCD sensor can capture clear images under various lighting conditions. The CCD industrial camera is also equipped with an optical filtering device that can effectively filter laser scattered or reflected light to ensure that the image meets predetermined standards. The camera's exposure time and frame rate can be adjusted according to the speed and accuracy requirements of laser cutting to capture each stage of the cutting process; the image acquisition card has a built-in analog-to-digital converter for converting analog signals into digital signals. The converter has a high sampling rate and a high data width to ensure that the details of the image are not lost during the conversion process, thereby maintaining the high resolution and depth of the image. After the conversion is completed, the acquisition card transmits the digital image data to the multimodal information fusion module via a high-speed data interface.

4. A real-time monitoring system for laser-induced thermal cracking cutting of brittle materials according to claim 1, characterized in that: The acoustic emission sensor adopts a high-sensitivity piezoelectric sensor, which is used in the process of laser-induced thermal cracking cutting. When a small crack expands in the material, a sound wave of a predetermined frequency is generated. The piezoelectric sensor can convert the mechanical vibration of the sound wave of the predetermined frequency into an electrical signal; after the preamplifier receives the weak electrical signal output by the acoustic emission sensor, it amplifies it through a low-noise amplifier circuit. The amplifier has the characteristics of high gain and wide bandwidth, which can ensure that the signal maintains its original frequency characteristics and amplitude information during the amplification process, thereby amplifying the acoustic emission signal without distortion. In addition, the preamplifier is also equipped with a filtering function, which can filter out noise in the non-target frequency range and enhance signal clarity; the acoustic emission signal collector includes a high-precision analog-to-digital converter, which is used to convert the analog acoustic emission signal output by the preamplifier into a digital format. The converter has high resolution and fast sampling rate, and can accurately capture and convert high-speed changing sound wave signals to ensure that the details and dynamic range of the signal are retained during the digitization process. The converted digital signal is transmitted to the multimodal information fusion module through a high-speed data interface. The interface supports real-time transmission of data to ensure that the acoustic emission data can be processed synchronously with the visual data.

5. The real-time monitoring system for laser-induced thermal cracking cutting of brittle materials according to claim 1 is characterized in that: The cross-modal interaction fusion unit includes: Feature reception and preliminary integration: The cross-modal interactive fusion unit first receives the visual features F from the visual feature extraction unit. 视觉 and the acoustic emission feature F of the acoustic emission feature extraction unit 声发射 , specifically in the form of vectors, containing the information of each mode; Feature conversion and dimension upgrade: In order to effectively integrate the features of the two modes in the same space, the features are first converted and dimension upgraded. The specific formula is: F 视觉 ′=W 视觉 ·F 视觉 ; F 声发射 ′=W 声发射 ·F 声发射 , where W 视觉 and W 声发射 It is a learnable weight matrix used to adjust the original features to a form suitable for fusion; Feature interaction fusion: Next, the interactive fusion layer is used to perform deep interaction and fusion of visual and acoustic emission features. The interactive fusion layer uses nonlinear transformation and adaptive weight adjustment to accurately reconcile the features of the two modes. The specific formula is: F 融合 =σ(W 交互 ·(f 视觉 ′+F 声发射 ′)+b 交互 ), where W 交互 is the weight matrix of the interactive fusion layer, b 交互 is a bias term, and σ is a nonlinear activation function used to increase the expressiveness and nonlinear characteristics of the model; Feature optimization and output: Feature F after processing by the interactive fusion layer 融合 Optimization is performed through a multi-layer neural network, including batch normalization and further nonlinear transformation to improve the generalization and robustness of the features; finally, the optimized features are output as the fused feature set F 最终 , the feature set F 最终 It will be used in the subsequent status assessment module to provide support for monitoring and decision-making of the cutting process.

6. The real-time monitoring system for laser-induced thermal cracking cutting of brittle materials according to claim 1 is characterized in that: The status assessment module includes: Feature reception and preliminary processing: The state assessment module first receives the fused feature set F output by the multimodal information fusion module. 最终 , providing a data basis for status assessment; Application of feedforward neural network: The feature data is first input into the preset feedforward neural network for analysis and processing. The feedforward neural network includes multiple hidden layers, and each layer uses the following calculation formula for feature conversion and activation: F 层 =σ(W·F 上一层 +b), where W and b are the weight matrix and bias vector respectively, and σ is a nonlinear activation function used to enhance the network's ability to handle complex nonlinear relationships; Feature weight adjustment: In a feedforward neural network, the output of each layer is adjusted according to the learned weights to optimize its contribution to subsequent layers; SoftMax layer classification probability calculation: The feature data processed by the multi-layer feedforward network is sent to the SoftMax layer to convert the real number vector output by the previous layer into a probability distribution. The value of each element represents the probability that the data belongs to the corresponding category. The specific calculation formula is: Among them, P(y i ) is the probability of category i, F i is the score of the feature value output by the previous layer corresponding to category i; Output the evaluation results of the processing status: The classification probability output by the SoftMax layer is used to determine the final processing status, specifically whether there is unqualified crack extension.

Images