Multi-product synchronous ultrasonic bubble removal and scanning analysis method

Abstract

The invention discloses a multi-product synchronous ultrasonic bubble removal and scanning analysis method, and relates to the field of ultrasonic bubble removal. Products are positioned to an ultrasonic treatment groove matrix type station through a six-axis mechanical arm and a hydraulic lifting table; forming three-dimensional standing wave field directional migration broken bubbles by adopting a phase conjugate adaptive focusing technology; the detection precision is improved by integrating ultrasonic wave and phased array scanning and combining a frequency domain synthetic aperture focusing algorithm; extracting bubble boundaries through a local entropy adaptive threshold and a region growing algorithm; tracking a bubble movement track by adopting a Kalman filtering and particle filtering fusion algorithm; constructing a double-attention mechanism deep residual network intelligent classification bubble; and optimizing processing parameters in real time based on a model predictive control algorithm to form a closed-loop system. According to the method, the processing adaptability is improved by dynamically matching ultrasonic parameters, the micro bubble recognition capability is enhanced by a multi-modal scanning technology, the optimal processing effect is ensured by closed-loop feedback control, and an efficient and intelligent quality control solution is provided for industrial production.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic degassing technology, and more particularly to a method for simultaneous ultrasonic degassing and scanning analysis of multiple products. Background Technology

[0002] In modern industrial production, internal air bubbles severely impact product quality and performance. Taking optical lenses, precision castings, and polymer composites as examples, residual air bubbles can lead to decreased material strength, deteriorated optical performance, and even functional failure. Traditional ultrasonic degassing technology typically uses a single frequency and fixed power of ultrasound waves, making it difficult to adapt to the varying characteristics of different materials and structures. For instance, for high-density metal castings and low-density engineering plastics, uniform ultrasonic parameters not only fail to effectively remove air bubbles but may also cause product damage due to concentrated energy. Furthermore, the lack of real-time monitoring mechanisms in a single degassing process makes it impossible to determine whether air bubbles have been completely removed, easily leading to under- or over-treatment.

[0003] Existing technologies for bubble detection and analysis have significant limitations. Conventional ultrasonic scanning techniques have low resolution, making it difficult to identify tiny bubbles (less than 0.1 mm in diameter), and they have blind spots in detecting internal bubbles in complex products. For example, when inspecting multilayer composite materials, ultrasonic echo signals are easily interfered with by interlayer interfaces, leading to misidentification or missed detection of bubbles. Furthermore, traditional detection methods often employ single-modal imaging, relying solely on ultrasonic grayscale images for analysis, lacking the ability to comprehensively assess the physical properties of bubbles (such as elastic modulus and internal pressure), and thus failing to accurately evaluate the potential impact of bubbles on product performance.

[0004] As industrial production moves towards automation and intelligence, the need for parallel processing of multiple products is becoming increasingly urgent. However, most existing degassing and inspection equipment is designed for single-station operation, making it difficult to meet the high-efficiency processing requirements of batch products. Even in the case of multi-station equipment, there is a lack of collaborative control mechanisms between the stations, making it impossible to dynamically adjust processing parameters according to product characteristics. Simultaneously, data processing and analysis are relatively lagging, with inspection results largely relying on manual interpretation. The lack of automated classification and decision-making systems hinders the rapid generation of effective quality assessment reports, severely restricting production efficiency and quality control levels. Summary of the Invention

[0005] The present invention proposes a multi-product synchronous ultrasonic degassing and scanning analysis method to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-product synchronous ultrasonic degassing and scanning analysis method, comprising: Multi-station positioning steps: The product is positioned in the ultrasonic treatment tank matrix station using a six-axis robotic arm. Each station is equipped with an independent hydraulic lifting platform and a vacuum adsorption fixture. The position is calibrated twice using a grating ruler. The ultrasonic bubble removal process involves dynamically matching the ultrasonic frequency based on the product's material density, constructing a three-dimensional standing wave field using phase conjugate focusing technology, adjusting the acoustic field parameters according to the bubble size, and directionally migrating and breaking up the bubbles. Multimodal scanning steps: Integrating ultrasonic C-scan and phased array B-scan, using time-reversal mirror algorithm for multifocal synchronous imaging, and using ultrasonic backscatter integration, frequency domain synthetic aperture focusing technology and adaptive matched filtering algorithm to detect sensitivity; Dynamic thresholding steps: The image is initially segmented using an adaptive threshold adjustment algorithm based on local entropy, noise is removed by combining morphological opening operation, bubble boundaries are extracted by a region growing algorithm, and pseudo bubble regions are removed by topological analysis. Bubble trajectory tracking steps: The bubble removal process is monitored in real time by high-frequency pulse sequences, the bubble motion vector is calculated by optical flow method, a Kalman filter prediction model is introduced, the trajectory is corrected by particle filter, and a multi-scale pyramid optical flow algorithm is used to track high-speed moving bubbles. Intelligent classification decision-making steps: Construct a deep residual network, and fuse time-domain, frequency-domain, and spatial-domain information of ultrasound images in the input layer; Employ a transfer learning strategy, fine-tune the model pre-trained on ImageNet, and balance the samples through a focus loss function; Closed-loop feedback control steps: Based on the bubble dynamics model, the ultrasonic frequency, power and action time are adjusted in real time using the model predictive control algorithm to optimize defoaming efficiency and product damage risk.

[0007] Furthermore, the ultrasonic bubble removal step employs a quantum genetic algorithm to optimize the excitation parameters of the phased array transducer. By constructing a qubit encoding population, the uniformity of the sound field energy density is optimized, and the bubbles are manipulated non-contactly using the acoustic tweezers effect.

[0008] Furthermore, the multimodal scanning step develops an ultrasonic elastography auxiliary module, which measures the tissue displacement field distribution by applying low-frequency modulated stress waves and reconstructs the elastic modulus image by combining it with an inversion algorithm; by fusing elastic information with ultrasonic grayscale image features, bubbles hidden in high-density areas are detected, and areas of stress concentration inside the material caused by bubbles are identified.

[0009] Furthermore, the dynamic thresholding step introduces superpixel segmentation technology to preprocess the image, combines graph cut algorithm to optimize the segmentation boundary, and constructs an energy function by calculating the gray-level difference and texture similarity between adjacent superpixels.

[0010] Furthermore, the bubble motion trajectory tracking step employs an instance segmentation network to separate the target in overlapping bubble scenarios. Overlapping bubbles are segmented by training a synthetic bubble image dataset, and the trajectory association algorithm is used to match the motion trajectory of the separated bubbles.

[0011] Furthermore, the intelligent classification decision-making step design integrates a hybrid model of transfer learning and ensemble learning, integrating a pre-trained EfficientNet-B7 model and a randomly initialized ResNet-34 model, and fusing the model prediction results through a stacked generalization strategy.

[0012] Furthermore, a digital twin real-time simulation system is established for the closed-loop feedback control steps. A multi-physics coupling model is constructed using finite element analysis, and the model parameters are calibrated using ultrasonic echo data. The digital twin is used to predict the processing effects of different parameter combinations, thereby optimizing parameters and energy consumption.

[0013] Furthermore, the multi-station positioning step employs multi-sensor fusion positioning technology, combining LiDAR, visual recognition, and inertial navigation data. By estimating the state through extended Kalman filtering, the positioning time for high-speed moving products is optimized, and positioning errors are controlled.

[0014] Furthermore, a quality traceability system is constructed, using a consortium blockchain architecture to hash and upload product material information, processing parameters, and test results onto the blockchain, and zero-knowledge proof technology is used to protect data privacy.

[0015] Furthermore, it also includes: Human-machine collaborative optimization module: It uses reinforcement learning algorithms to generate parameter optimization schemes, proposes optimization requirements through natural language interaction, and automatically converts human experience into model parameters.

[0016] Compared with existing technologies, the beneficial effects of this invention are: In the debubbling process, the ultrasonic frequency is dynamically matched to the product's material density, and a three-dimensional standing wave field is formed using phase conjugate adaptive focusing technology. This enables precise processing of products with different materials and structures. Whether it's high-density metals or lightweight polymers, the system effectively removes bubbles while avoiding product damage caused by uneven energy distribution, ensuring the safety and reliability of the processing.

[0017] In bubble detection and analysis, multimodal scanning technology combines 1-5MHz ultrasonic C-scanning with 0.5-2MHz phased array B-scanning, along with algorithms such as frequency domain synthetic aperture focusing, significantly improving detection resolution and accuracy. Even tiny bubbles with a diameter of 0.02mm can be effectively identified. Furthermore, through an ultrasonic elastography auxiliary module, the impact of bubbles on the internal stress distribution of materials can be further analyzed, providing more comprehensive data support for product quality assessment.

[0018] Furthermore, the system's multi-station parallel processing and closed-loop feedback control mechanism achieve the dual goals of efficient production and precise quality control. The 8x8 matrix workstation layout, coupled with independent hydraulic lifting platforms, supports simultaneous processing of multiple products. The model-predictive control-based closed-loop system dynamically adjusts ultrasonic parameters based on real-time detection results, ensuring optimal bubble removal for each product. Simultaneously, the intelligent classification decision module automatically classifies bubbles using an integrated learning model, generating inspection reports that include location, size, and risk level, significantly reducing manual interpretation time and substantially improving production efficiency and the level of intelligent quality control. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of a multi-product synchronous ultrasonic degassing and scanning analysis method proposed in this invention; Figure 2 A diagram comparing the bubble removal rate of products made of different materials; Figure 3 This is a schematic diagram illustrating the relationship between bubble size and detection accuracy. Figure 4 This is a diagram comparing system energy consumption and processing efficiency. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 4 A method for simultaneous ultrasonic degassing and scanning analysis of multiple products, comprising: Multi-station positioning process: An ABB IRB 6700 six-axis robotic arm is used as the positioning execution unit, with a load capacity of 150kg and a repeatability of ±0.05mm. The robotic arm's end effector is equipped with a customized vacuum adsorption fixture, using an SMC ZSE30A vacuum generator to provide a stable adsorption force of -90kPa, and a pressure sensor (accuracy ±0.1kPa) monitors the adsorption status in real time. The ultrasonic treatment tank is designed with an 8×8 matrix layout, with each station equipped with an independent electro-hydraulic servo lifting platform (MOONS'DSH series), a stroke range of 0-50mm, and a resolution of 0.01mm. A grating ruler (HEIDENHAIN LIC 4117) is used for position feedback to ensure liquid level control accuracy of ±0.05mm. During positioning, the Keyence CV-X600 vision system identifies the QR code on the product surface (containing information such as product material and dimensions), and a binocular structured light camera (Intel RealSense D455) acquires the product's 3D point cloud data. After preprocessing with the PCL library, the point cloud data is registered with a preset model using the Iterative Closest Point (ICP) algorithm, achieving a registration accuracy of ±0.02mm. The robotic arm adjusts its posture based on the registration results, precisely placing the product at the target workstation. The entire positioning process takes ≤30 seconds.

[0024] The ultrasonic bubble removal process employs a 256-element phased array transducer (Piezosystem Jena GmbH), with each element measuring 1mm × 1mm, operating at a frequency range of 20-40kHz, and a maximum output power of 500W. The system uses an X-ray fluorescence spectrometer (Thermo Scientific Niton XL3t) to measure the product's material density in real time. Based on a pre-trained neural network model (input density values, output optimal frequency parameters), the ultrasonic frequency is dynamically matched. For example, for aluminum alloy (density approximately 2.7g / cm³), the system automatically selects 32kHz as the operating frequency; for engineering plastics (density approximately 1.2g / cm³), 25kHz is selected. For sound field control, a MATLAB and COMSOL co-simulation platform is used to establish a three-dimensional sound field model based on the acoustic boundary element method (BEM). The element excitation parameters are optimized using a quantum genetic algorithm, with a population size of 100 and a quantum rotating gate update step size of 0.01π. After 50 iterations, the sound field energy density uniformity is achieved to be ≥95%. In practical applications, the system acquires an ultrasonic echo signal every 5 seconds and dynamically adjusts the excitation parameters by calculating the time-domain energy distribution of the signal to ensure that the acoustic intensity gradient is always controlled within ±0.1W / cm².

[0025] Multimodal scanning steps: A 5MHz linear array probe (Philips L12-5) and a 2MHz phased array probe (Siemens P4-2) are integrated, and a self-developed ultrasound imaging platform (based on FPGA+DSP architecture) is used to achieve synchronous acquisition of multimodal data. During scanning, the probe performs a spiral scan at a 0.05mm interval, covering the entire product volume. The phased array probe employs dynamic focusing technology with up to 16 focal points, enabling simultaneous imaging of areas at different depths. In signal processing, frequency-domain synthetic aperture focusing (F-SAF) technology is introduced, performing Fast Fourier Transform (FFT) on the echo signal, phase compensation in the frequency domain, and reconstructing the focused image through Inverse Fourier Transform (IFFT). Simultaneously, ultrasound backscatter integration (UBI) technology is used to extract tissue characteristic information, calculate the signal power spectral density in the 3-5MHz frequency band, and identify bubble regions by comparing with preset thresholds. For noisy environments, an adaptive matched filtering algorithm is adopted. Based on the minimum mean square error (LMS) criterion, with a step size factor μ=0.05, it can effectively suppress noise and improve the signal-to-noise ratio by 15dB.

[0026] Dynamic thresholding segmentation steps: A hybrid algorithm architecture is adopted, using the SLIC superpixel algorithm to over-segment the original ultrasound image. The superpixel size is set to 10×10, the compactness coefficient is 0.1, and approximately 2000 superpixels are generated. An adaptive thresholding algorithm based on local entropy calculates the optimal segmentation threshold for each superpixel, with a local window size of 11×11 pixels. To improve segmentation accuracy, a graph cut algorithm is combined to optimize the segmentation results. An energy function is constructed by calculating the gray-level difference (weight coefficient α=0.7) and texture similarity (weight coefficient β=0.3) between adjacent superpixels. In the morphological processing stage, a 5×5 disk structuring element is used for opening operations to remove small noise points, and an 8-connected region growing algorithm is used for refined extraction of bubble boundaries. The seed point selection strategy selects points in the thresholding results whose gray-level values ​​are within the mean ± standard deviation range and have large gradient magnitudes as seed points. For bubble segmentation in complex backgrounds, this algorithm improves accuracy by 20% and processing speed by 3 times compared to the traditional Otsu algorithm, with a boundary localization error ≤0.03mm.

[0027] Bubble trajectory tracking steps: A multi-sensor fusion scheme is adopted, with four high-speed cameras (Basler ace acA2040-90um, frame rate 100fps, resolution 2048×1088) deployed around the ultrasonic processing tank, synchronized with the ultrasonic system in time. First, the foreground of the bubble motion is extracted using a background subtraction algorithm (ViBe algorithm, 20 samples, matching threshold 20). Then, the bubble motion vector is calculated using optical flow (Farneback algorithm, pyramid level 3, window size 15×15). To improve tracking accuracy, a Kalman filter prediction model is introduced. The state vector includes the bubble position (x, y, z), velocity (vx, vy, vz), and acceleration (ax, ay, az). The state transition matrix is ​​constructed using the bubble dynamics equations. For high-speed moving bubbles (velocity > 1 mm / s), a multi-scale pyramid optical flow algorithm is used to calculate the optical flow field at three scales (original, 1 / 2, and 1 / 4 resolution), and then weighted fusion is performed to obtain the final result. For overlapping bubble scenes, Mask R-CNN was used for instance segmentation with an accuracy of 96.5%, and the Hungarian algorithm was used for trajectory association with an association threshold of 0.7.

[0028] Intelligent classification decision-making steps: An ensemble learning architecture is adopted, consisting of one pre-trained EfficientNet-B7 model and five randomly initialized ResNet-34 models. The input data simultaneously includes temporal features (echo intensity sequence), frequency domain features (power spectral density), and spatial domain features (morphological parameters) of ultrasound images. During the training phase, 100,000 synthetic bubble images (generated through COMSOL simulation) are used for pre-training, followed by fine-tuning on a real dataset. The optimizer uses AdamW with an initial learning rate of 0.001, a weight decay coefficient of 1e-4, a batch size of 32, and a training duration of 100 epochs. To balance the sample imbalance problem, a focal loss function is used, where α=0.25 and γ=2. Finally, the ensemble model fuses the prediction results of each sub-model through a stacking generalization strategy, achieving a classification accuracy of 99.2% for five types of bubbles (normal bubbles, adhered bubbles, dense bubbles, microbubbles, and pseudo-bubbles) on the test set, with a false positive rate of <0.3%. When deploying the model, it is optimized using TensorRT, with inference time <50ms, meeting the requirements for real-time processing.

[0029] Closed-loop feedback control steps: A Model Predictive Control (MPC) architecture is adopted, and a predictive model is constructed based on a bubble dynamics model (combining Rayleigh-Plesset equations and Navier-Stokes equations). The control objective is to achieve a bubble removal rate >95% in the shortest time (≤60 seconds) while ensuring a product damage risk <0.1%. System state variables include bubble number, average size, and distribution density, while control variables are ultrasonic frequency, power, and duration. The optimization solution uses a Sequential Quadratic Programming (SQP) algorithm, with N=10 steps in the prediction time domain and M=5 steps in the control time domain. The current state is collected every 50ms, and the optimal control sequence is calculated using the MPC algorithm. Only the first control action is executed, then the state is updated and the calculation is repeated. To improve real-time performance, a graphics processor (NVIDIA RTX 3080) is used to accelerate matrix operations, making the single optimization time <20ms. In practical applications, the system response latency is <50ms, the single closed-loop control cycle is ≤200ms, and the processing parameters can be adjusted in real time according to the dynamic changes of the bubbles.

[0030] In this invention, to further improve the accuracy and efficiency of ultrasonic bubble removal, a quantum genetic algorithm is used to optimize the excitation parameters (including delay, amplitude, and phase) of the phased array transducer. Specifically, the quantum genetic algorithm uses a qubit-encoded population (set to a population size of 100, with the update step size set to 0.01π through precise control of the quantum rotation gate) to simulate the superposition and entanglement characteristics of quantum states, continuously optimizing the excitation parameters over 50 iterations. During each iteration, the state of the qubits is adjusted based on feedback from the uniformity of the acoustic field energy density. After multiple generations of evolution, an acoustic field energy density uniformity of ≥95% can be achieved, allowing the ultrasonic energy to be more evenly distributed within the product and acting more stably on the bubbles. Simultaneously, utilizing the acoustic tweezers effect (based on the unique diffraction-free, self-healing propagation characteristics of Bessel beams), tiny acoustic radiation force traps are formed to non-contactly manipulate microbubbles with a diameter <0.05mm, precisely capturing and driving the bubbles to move and break up. This overcomes the limitations of traditional methods for handling microbubbles and improves the overall bubble removal effect.

[0031] In this invention, the multimodal scanning step overcomes the limitations of traditional ultrasonic scanning in detecting bubbles in complex scenarios by developing an ultrasonic elastography auxiliary module. A low-frequency modulated stress wave is applied using a high-precision vibration generator (frequency strictly controlled at 50Hz, amplitude precisely adjusted to 0.1mm), and a high-sensitivity displacement sensor array is used to measure the displacement field distribution of the tissue under stress. Based on linear elasticity theory, combined with a regularized inversion algorithm (introducing a Tikhonov regularization term to balance stability and accuracy), the internal elastic modulus distribution of the material is calculated, reconstructing a high-resolution elastic modulus image. Subsequently, feature fusion is performed using DS evidence theory. Texture and edge features from the ultrasonic grayscale image are extracted separately from the stiffness and uniformity features of the elastic modulus image. By constructing a basic probability allocation function, evidence from different feature channels is synthesized, effectively linking the intrinsic relationship between bubbles and material elasticity changes. Extensive experimental verification has shown that this module improves the accuracy of bubble detection in high-density areas (such as the dense structure inside metal alloys and the multi-layer interface of composite materials) to 97.8%. At the same time, based on the abnormal distribution of elastic modulus, it can identify stress concentration areas inside materials caused by the presence of bubbles, providing key basis for evaluating the mechanical properties and service reliability of products.

[0032] In this invention, a superpixel segmentation technique (SLIC algorithm) is introduced in the dynamic threshold segmentation step. The superpixel size is set to 10×10 pixel units, and the compactness coefficient is adjusted to 0.1. This coefficient balances the spatial compactness of the superpixel and color similarity, allowing the algorithm to construct a similarity metric based on LAB color space distance and pixel coordinate distance in the image space. This pre-segmentation divides the image into several locally consistent superpixel blocks, reducing the amount of data required for subsequent processing. After preprocessing, the segmentation boundary is optimized using a graph cut algorithm. By calculating the grayscale difference between adjacent superpixels (using absolute grayscale difference as a metric) and texture similarity (using LBP texture operator to extract features and calculate similarity), an energy function containing data and smoothing terms is constructed. The data term reflects the matching degree between the superpixel and the bubble / background categories, while the smoothing term constrains the segmentation consistency of adjacent superpixels. The minimum value of the energy function is solved using the maximum flow-minimum cut algorithm, achieving segmentation boundary optimization. Practical verification shows that this process improves bubble segmentation accuracy by 20% in complex backgrounds (such as areas with texture interference and uneven grayscale). Due to the reduction of computation nodes caused by superpixel pre-segmentation, the processing speed is 3 times faster than traditional methods.

[0033] In this invention, the bubble trajectory tracking step targets overlapping bubble scenarios (bubbles overlap due to aggregation, similar size, etc.). A deep learning-based instance segmentation network (MaskR-CNN, with ResNeXt-101 as the backbone) is used for target separation. ResNeXt-101, leveraging grouped convolutions and cardinality to enhance feature representation, first extracts features from the bubble image, generating candidate regions containing bubble positions and outlines. The Mask branch then predicts a binary mask for each candidate region, accurately segmenting individual bubbles. When constructing the dataset, physical simulation and image synthesis techniques are used to simulate the distribution of bubbles of different sizes (0.01-1mm) and densities (single bubbles, multiple overlapping bubbles). Combined with real ultrasonic scanning noise, 100,000 labeled synthetic bubble images covering various overlapping patterns are generated. After training, the network achieves a 96.5% accuracy rate in segmenting overlapping bubbles. After segmentation, a trajectory association algorithm (Hungarian algorithm) is used to construct a cost matrix with bubble center coordinates, size changes, and other association features. The optimal matching is then solved to ensure continuous tracking of the bubble's motion trajectory after separation, and to handle complex motion states such as bubble merging and splitting.

[0034] In this invention, the intelligent classification decision-making step employs a hybrid model integrating transfer learning and ensemble learning. Utilizing transfer learning, an EfficientNet-B7 model pre-trained on a large-scale image dataset is selected. Through a composite scaling strategy, the network is optimized across different resolutions, depths, and widths while retaining general feature extraction capabilities. Simultaneously, five ResNet-34 models are initialized with randomly assigned initial weights to ensure model diversity. In the ensemble stage, a stacking generalization strategy is employed. Ultrasonic images are input into each model; EfficientNet-B7 outputs high-dimensional abstract features, while the ResNet-34 series outputs multi-scale features. The prediction results of these models (such as the probability distribution of bubble categories) are used as new features to construct a secondary training set. A multilayer perceptron (MLP) is designed as a meta-classifier, learning on the secondary training set and integrating the advantages of each model. In testing with new material products, due to the integration of complementary information from multiple models, the model's generalization ability is improved by 30% compared to a single model. When facing unknown types of bubbles (such as special-shaped bubbles introduced by new materials), the recognition accuracy remains consistently above 95%, effectively addressing the classification challenges posed by material diversity.

[0035] In this invention, a closed-loop feedback control step establishes a real-time simulation system for a digital twin, overcoming the limitations of traditional control in understanding the dynamic characteristics of complex physical fields. Using COMSOL Multiphysics finite element analysis software, a multi-physics coupling model of the product and ultrasonic field is constructed, comprehensively considering the propagation, reflection, and refraction of ultrasound in the product material, as well as the acoustic-fluid-structure interaction effect when interacting with bubbles. The model precisely defines the mechanical parameters of the product material, such as elastic modulus, density, and damping; the excitation parameters of the ultrasonic transducer, such as frequency, power, and array arrangement; and the boundary conditions, such as the initial distribution, size, and motion characteristics of the bubbles. In the model parameter calibration stage, combined with real-time acquired ultrasonic echo data, a particle swarm optimization algorithm is used for iterative optimization. The particle swarm uses the error between the echo signal and the simulation prediction signal (such as root mean square error) as the fitness function, and continuously adjusts the model parameters (such as material acoustic parameter correction and transducer equivalent impedance optimization) through particle position updates (the velocity iteration formula includes inertia weight, cognitive term, and social term). The convergence error is strictly controlled to <0.01, ensuring a high degree of consistency between the digital twin and the actual physical system. With the aid of a calibrated digital twin, the bubble movement, breakup process, and product response can be simulated in advance under different combinations of ultrasonic frequencies (dynamically adjustable from 20-40kHz), power (finely controlled according to product tolerance), and action times. By comparing the defoaming efficiency and product damage risk of multiple simulation results, the optimal parameters can be quickly selected, improving parameter optimization efficiency by 40% compared to traditional trial-and-error methods. At the same time, by avoiding invalid parameter experiments and reducing redundant energy output, energy consumption is reduced by 15%, ensuring that closed-loop feedback control operates in a highly efficient and low-consumption state, adapting to complex scenarios involving simultaneous processing of multiple products.

[0036] In this invention, the multi-station positioning step, adapted to the high-speed, high-precision product positioning requirements, employs multi-sensor fusion positioning technology. The lidar, with a measurement range of 0-5m and an accuracy of ±1mm, quickly acquires distance information between the product and the workstation. By emitting laser pulses and receiving echoes, and utilizing the Time-of-Flight (ToF) principle, it can perform thousands of distance measurements per second, constructing a rough product position outline. The visual recognition module, relying on a 4096×3072 high-resolution camera, acquires product appearance feature images with a pixel accuracy of 0.01mm. After image preprocessing (including noise reduction, enhancement, and distortion correction), feature extraction (such as SIFT and ORB algorithms to identify product surface markers), and matching, it accurately locates the product's coordinates in a two-dimensional plane. Inertial navigation, using a high-precision gyroscope (drift rate <0.01° / h), senses the angular velocity and angular acceleration of the product in real time during its movement. Combined with initial position information, it calculates the product's trajectory and attitude changes through integral calculations. Data from three types of sensors is input into an Extended Kalman Filter (EKF) for state estimation. First, a system state equation is established, incorporating position, velocity, and attitude. Considering the ranging noise of the lidar (following a Gaussian distribution, with variance calibrated based on actual measurements), the feature matching error of visual recognition (related to image resolution and lighting conditions), and the integral drift error of inertial navigation (accumulating over time), corresponding observation equations are constructed. The EKF employs a two-step iterative prediction-update process, fusing multi-source heterogeneous data in real time: in the prediction phase, the current state is inferred based on the state equation and the state estimate from the previous moment; in the update phase, the predicted value is corrected using the observation equation and real-time sensor data, outputting the optimal state estimation result. Practical verification shows that this technology can reduce the positioning time to 100ms for high-speed moving products (speed > 50mm / s, such as continuously transported workpieces on automated production lines), with positioning errors strictly controlled to <0.03mm, meeting the high-precision workstation docking requirements for simultaneous ultrasonic processing of multiple products.

[0037] In this invention, a blockchain-based quality traceability system is constructed throughout the entire processing flow. A consortium blockchain architecture (Hyperledger Fabric) is adopted, deploying four endorsing nodes. Based on the Raft consensus mechanism, network consensus efficiency and stability are ensured. The nodes encompass participants from the product manufacturing end, processing equipment end, and quality inspection end, forming a trusted consortium ecosystem. For each product, starting from the input stage, a hash algorithm (such as SHA-256) is used to generate a unique hash value for on-chain storage, based on material information (elemental composition, density, etc. measured by equipment such as X-ray fluorescence spectrometer), processing parameters (real-time data of closed-loop control such as ultrasonic frequency, power, and action time), and inspection results (bubble size and distribution from multimodal scanning, intelligent classification conclusions, etc.). Simultaneously, zero-knowledge proof (ZKP) technology is introduced. Taking the zk-SNARKs protocol as an example, the data provider generates a proof, and the verifier can verify the authenticity and integrity of the data without obtaining the original data, achieving data privacy protection and adapting to sensitive product quality data scenarios. During system operation, data verification is automatically executed through smart contracts, calling the hash index stored on the chain to match verification requests in less than 1 second. Due to the immutability of blockchain, combined with the hash chain storage structure, the tamper detection accuracy reaches 100%, providing a reliable basis for the traceability of product quality throughout its entire lifecycle. From raw material entry to degassing and testing before leaving the factory, data at each stage is traceable and verifiable, ensuring the traceability and reliability of product quality and meeting the stringent quality control requirements of high-end manufacturing, medical electronics, and other fields.

[0038] This invention also includes the following steps: Human-Machine Collaborative Optimization Module: At the algorithm-driven level, a Proximal Policy Optimization (PPO) reinforcement learning algorithm is adopted to construct an intelligent decision-making environment that includes a state space, action space, and reward mechanism. The state space incorporates process parameters such as ultrasonic frequency, power, and action time, as well as features such as product material density and initial bubble distribution. The action space is defined as continuous / discrete operations for parameter adjustment. The reward function innovatively integrates multi-dimensional indicators: defoaming efficiency is calculated through real-time scanning of bubble residue rate, processing time is correlated with production line cycle time, and equipment losses are calculated based on transducer working time and power fluctuations. Through weighted summation and dynamic feedback, the intelligent agent is guided to explore the optimal parameter combination and automatically generate optimization solutions adapted to different products. At the human interaction level, a Natural Language Interaction (NLP) system based on the BERT model is built. For operator-inputted optimization requests (such as "improving the microbubble removal rate" and "shortening the processing cycle"), the BERT model first performs word vector encoding, captures semantic associations through a multi-head attention mechanism, and combines it with a pre-trained industry corpus (covering ultrasonic processing terminology and experiential language) to parse the request intent and convert it into parameter adjustment instructions that the algorithm can recognize. The system has a built-in experience mapping engine that breaks down human experience (such as the parameter debugging skills of senior operators and strategies for dealing with abnormal working conditions) into a set of rules and integrates them with the solutions generated by intelligent algorithms to achieve accurate conversion of experience into model parameters.

[0039] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for simultaneous ultrasonic degassing and scanning analysis of multiple products, characterized in that, Includes the following steps: Multi-station positioning steps: The product is positioned in the ultrasonic treatment tank matrix station using a six-axis robotic arm. Each station is equipped with an independent hydraulic lifting platform and a vacuum adsorption fixture. The position is calibrated twice using a grating ruler. The ultrasonic bubble removal process involves dynamically matching the ultrasonic frequency based on the product's material density, constructing a three-dimensional standing wave field using phase conjugate focusing technology, adjusting the acoustic field parameters according to the bubble size, and directionally migrating and breaking up the bubbles. Multimodal scanning steps: Integrating ultrasonic C-scan and phased array B-scan, using time-reversal mirror algorithm for multifocal synchronous imaging, and using ultrasonic backscatter integration, frequency domain synthetic aperture focusing technology and adaptive matched filtering algorithm to detect sensitivity; Dynamic thresholding steps: The image is initially segmented using an adaptive threshold adjustment algorithm based on local entropy, noise is removed by combining morphological opening operation, bubble boundaries are extracted by a region growing algorithm, and pseudo bubble regions are removed by topological analysis. Bubble trajectory tracking steps: The bubble removal process is monitored in real time by high-frequency pulse sequences, the bubble motion vector is calculated by optical flow method, a Kalman filter prediction model is introduced, the trajectory is corrected by particle filter, and a multi-scale pyramid optical flow algorithm is used to track high-speed moving bubbles. Intelligent classification decision-making steps: Construct a deep residual network, and fuse time-domain, frequency-domain, and spatial-domain information of ultrasound images in the input layer; Employ a transfer learning strategy, fine-tune the model pre-trained on ImageNet, and balance the samples through a focus loss function; Closed-loop feedback control steps: Based on the bubble dynamics model, the ultrasonic frequency, power and action time are adjusted in real time using the model predictive control algorithm to optimize defoaming efficiency and product damage risk.

2. The multi-product synchronous ultrasonic degassing and scanning analysis method according to claim 1, characterized in that, The ultrasonic bubble removal process employs a quantum genetic algorithm to optimize the excitation parameters of the phased array transducer. By constructing a quantum bit encoding population, the uniformity of the sound field energy density is optimized, and the bubbles are manipulated non-contactly using the acoustic tweezers effect.

3. The multi-product synchronous ultrasonic degassing and scanning analysis method according to claim 1, characterized in that, The multimodal scanning process developed an ultrasonic elastography auxiliary module, which measures the tissue displacement field distribution by applying low-frequency modulated stress waves and reconstructs the elastic modulus image by combining it with an inversion algorithm. By fusing elastic information with ultrasonic grayscale image features, bubbles hidden in high-density areas are detected, and areas of stress concentration inside the material caused by bubbles are identified.

4. The multi-product synchronous ultrasonic degassing and scanning analysis method according to claim 1, characterized in that, The dynamic thresholding step introduces superpixel segmentation technology to preprocess the image, combines graph cut algorithm to optimize the segmentation boundary, and constructs an energy function by calculating the gray-level difference and texture similarity between adjacent superpixels.

5. The multi-product synchronous ultrasonic degassing and scanning analysis method according to claim 1, characterized in that, The bubble motion trajectory tracking steps employ an instance segmentation network to separate targets in overlapping bubble scenarios. Overlapping bubbles are segmented using a synthetic bubble image dataset trained on the network, and the motion trajectory of the separated bubbles is matched using a trajectory association algorithm.

6. The multi-product synchronous ultrasonic degassing and scanning analysis method according to claim 1, characterized in that, The intelligent classification decision-making process is designed to integrate a hybrid model of transfer learning and ensemble learning, combining a pre-trained EfficientNet-B7 model with a randomly initialized ResNet-34 model, and fusing the model prediction results through a stacked generalization strategy.

7. The multi-product synchronous ultrasonic degassing and scanning analysis method according to claim 1, characterized in that, The closed-loop feedback control process establishes a real-time simulation system using a digital twin. A multi-physics coupled model is constructed using finite element analysis, and the model parameters are calibrated using ultrasonic echo data. The digital twin is used to predict the processing effects of different parameter combinations, thereby optimizing parameters and energy consumption.

8. The method for simultaneous ultrasonic degassing and scanning analysis of multiple products according to claim 1, characterized in that, The multi-station positioning process employs multi-sensor fusion positioning technology, combining LiDAR, visual recognition, and inertial navigation data. It uses extended Kalman filtering to estimate the state, optimizes the positioning time for high-speed moving products, and controls positioning errors.

9. The multi-product synchronous ultrasonic degassing and scanning analysis method according to claim 1, characterized in that, A quality traceability system is built, which adopts a consortium blockchain architecture to hash product material information, processing parameters, and test results onto the blockchain, and uses zero-knowledge proof technology to protect data privacy.

10. The multi-product synchronous ultrasonic degassing and scanning analysis method according to claim 1, characterized in that, Also includes: Human-machine collaborative optimization module: It uses reinforcement learning algorithms to generate parameter optimization schemes, proposes optimization requirements through natural language interaction, and automatically converts human experience into model parameters.

Images