High-precision self-adaptive needle point dynamic correction module system and online transfer learning method

Abstract

The invention relates to the field of precision manufacturing automation control, and discloses a high-precision self-adaptive needle point dynamic correction module system and an online transfer learning method, and the system comprises a multi-mode sensing module which is configured to synchronously obtain the three-dimensional geometric information of a target needle point and the surface feature information reflecting the physical characteristics of a workpiece; and the data processing and decision-making module is in communication connection with the multi-modal sensing module and is configured to process the three-dimensional geometric information and the surface feature information acquired by the multi-modal sensing module based on a neural network model. According to the invention, the multi-mode sensing technology and the FPGA hardware acceleration neural network are combined, so that high-precision recognition and high-speed real-time correction of the position deviation of the needle tip are realized; meanwhile, a unique online transfer learning mechanism endows the system with self-optimization capability, so that the system can automatically adapt to workpieces with different physical characteristics, and the application flexibility of the system and the stability of machining quality are remarkably enhanced.

Description

Technical Field

[0001] This invention relates to the field of precision manufacturing automation control, and in particular to a high-precision adaptive needle tip dynamic correction module system and an online transfer learning method. Background Technology

[0002] In the fields of precision manufacturing and automation control, ensuring the precise positioning of machining tools, such as probes, needle tips, or micro-tools, during operation is a core element in guaranteeing the quality and consistency of the final product. To achieve this goal, the industry is continuously committed to developing advanced technologies that improve the positioning accuracy and dynamic response speed of machining equipment. These technological advancements have significantly boosted production efficiency and machining precision.

[0003] Specifically, in the field of needle tip dynamic correction, which is directly related to this technical solution, existing mainstream technologies rely on detection systems using a single sensor. For example, some systems use laser displacement sensors to measure the geometric distance deviation between the needle tip and the target reference, while others use machine vision imaging to identify the position of the needle tip in the image coordinate system and adjust its position accordingly. These single-modal sensing and control methods form the basis for current automated correction.

[0004] However, as the manufacturing industry's requirements for processing precision and yield become increasingly stringent, these traditional technologies based on single sensors are gradually revealing their inherent limitations. A significant challenge is that these systems struggle to simultaneously meet the mutually constraining performance indicators of sub-micron level ultra-high detection accuracy and millisecond-level rapid dynamic response. More importantly, existing web correction systems suffer from significant shortcomings in adaptability. Due to the singularity of their sensing dimension, these systems are designed and calibrated for workpieces with specific physical properties. Therefore, when workpieces of different materials, batches, or surface treatments appear on the production line, traditional web correction systems often fail to adjust their internal correction models or control parameters in a timely and accurate manner to adapt to these changes in workpiece characteristics. Their sensing data becomes inaccurate due to differences in reflectivity, and their control strategies cannot cope with differences in material elastic deformation, ultimately leading to fluctuations in processing quality and making it difficult to guarantee high consistency. Summary of the Invention

[0005] The purpose of this invention is to provide a high-precision adaptive needle tip dynamic correction module system and an online transfer learning method, which solves the problem that the existing needle tip detection system is not accurate in identifying submicron level deviations and cannot meet the requirements of high-precision manufacturing.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a high-precision adaptive needle tip dynamic correction module system, comprising: The multimodal sensing module is configured to simultaneously acquire the three-dimensional geometric information of the target needle tip and the surface feature information reflecting the physical properties of the workpiece; The data processing and decision-making module, which is communicatively connected to the multimodal sensing module, is configured to process the three-dimensional geometric information and surface feature information acquired by the multimodal sensing module based on a neural network model, so as to generate a correction command for compensating for the tip position deviation. An online model update module, which is communicatively connected to the data processing and decision module, is configured to update the parameters of the neural network model used by the data processing and decision module online based on deviation data observed during actual processing, so that the system can adapt to workpieces with different physical properties.

[0007] Preferably, the multimodal sensing module includes: A laser triangulation unit is used to acquire the three-dimensional geometric information of the target needle tip; a multispectral imaging unit is used to acquire the surface feature information reflecting the physical properties of the workpiece.

[0008] Preferably, the data processing and decision-making module is further configured to: execute a spatiotemporal registration algorithm before processing the three-dimensional geometric information and the surface feature information, so as to fuse the data collected by the laser triangulation unit and the multispectral imaging unit in a unified spatiotemporal coordinate system.

[0009] Preferably, the neural network model in the data processing and decision-making module is a convolutional neural network model.

[0010] Preferably, the data processing and decision-making module includes a field-programmable gate array (FPGA) hardware platform, and the convolutional neural network model is deployed on the FPGA hardware platform for hardware acceleration processing.

[0011] Preferably, the convolutional neural network model is a lightweight model, and the lightweight processing method includes at least one of sparse connection or low bit quantization techniques.

[0012] Preferably, the online model update module is configured to use an online transfer learning strategy to update the parameters of the neural network model, that is, to fine-tune the model using the bias data based on a pre-trained model.

[0013] Preferably, the online model update module is specifically configured as follows: The surface feature information, three-dimensional geometric information, and corresponding actual processing deviations obtained by the multimodal perception module are continuously collected and together constitute incremental training samples. Using the incremental training samples, the weight parameters of the neural network model are updated online via mini-batch gradient descent.

[0014] Preferably, the online model update module uses the surface feature information collected by the multimodal perception module as the main basis for distinguishing the physical properties of different workpieces, so as to drive the adaptive update of the neural network model parameters.

[0015] A high-precision adaptive tip dynamic correction online transfer learning method includes the following steps: S1. Through the multimodal sensing module, the three-dimensional geometric information of the target needle tip and the surface feature information reflecting the physical properties of the workpiece are acquired simultaneously. S2. Based on a preset neural network model, the three-dimensional geometric information and the surface feature information are fused to generate a correction command for compensating for the tip position deviation. S3. During the processing, the parameters of the neural network model are updated online and incrementally based on the observed actual deviations, so that the correction strategy of the method can adapt to workpieces with different physical properties.

[0016] In summary, the present invention has at least one of the following beneficial technical effects: 1. This invention significantly improves the overall accuracy of needle tip position deviation identification by employing a composite sensing method combining multispectral imaging and laser triangulation. Specifically, this scheme not only acquires high-precision three-dimensional geometric information of the target needle tip through the laser triangulation unit, but also captures surface feature information reflecting the physical properties of the workpiece through the multispectral imaging unit. By fusing these two different dimensions of information, the system's judgment of needle tip position deviation becomes more comprehensive and sufficient, thereby effectively improving the overall accuracy and reliability of the correction control.

[0017] 2. This invention significantly improves the real-time response speed of the system by deploying a lightweight convolutional neural network model on a Field-Programmable Gate Array (FPGA) hardware platform, meeting the stringent timeliness requirements of dynamic error correction. On one hand, the lightweight network architecture itself reduces the computational complexity of the decision model; on the other hand, the dedicated hardware acceleration capabilities of the FPGA, through its inherent parallel and pipelined processing mechanisms, fundamentally ensure ultra-low end-to-end latency from data input to error correction command output. This deep co-design of the algorithm and hardware ensures that the error correction loop can be completed in an extremely short time, thereby effectively improving the dynamic response performance of the system.

[0018] 3. This invention endows the system with advanced adaptive capabilities and broad applicability by designing a dynamic model update module based on an online transfer learning strategy. This approach enables the system to continuously and incrementally fine-tune and optimize its internal neural network model without interrupting the machining process, using newly acquired data related to the current workpiece characteristics (e.g., different elastic moduli or surface reflectivity). This dynamic learning and evolutionary capability frees the system from static dependence on specific workpiece parameters, allowing it to automatically adapt to different materials and batches of processed objects, thereby greatly enhancing the flexibility of correction control and its robustness to complex and changing working conditions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1 The present invention will be further described in detail below.

[0021] This invention provides a high-precision adaptive needle tip dynamic correction module system, comprising: The multimodal sensing module is configured to simultaneously acquire the three-dimensional geometric information of the target needle tip and the surface feature information reflecting the physical properties of the workpiece; In this embodiment, the multimodal sensing module serves as the sensing front-end of the correction system of this invention. Its core responsibility is to provide the entire closed-loop control system with rich, accurate, and complementary raw physical world information. This module is configured to capture, in parallel and in real-time, two types of information that are fundamentally different in nature but crucial for achieving high-precision adaptive correction: one is the precise geometric shape of the target needle tip in three-dimensional space, and the other is surface feature information that indirectly reflects the physical properties of the workpiece it contacts or is adjacent to. The rational design of this module and the completeness of its information output are the foundation and prerequisite for the successful implementation of subsequent data processing, intelligent decision-making, and online adaptive learning.

[0022] In one specific embodiment, the multimodal sensing module may consist of the following two functional units: a laser triangulation unit and a multispectral imaging unit.

[0023] Laser triangulation unit The laser triangulation ranging unit in this embodiment aims to acquire the three-dimensional coordinate data of the target needle tip surface in a non-contact and high-precision manner. Specifically, this unit may include a laser emitter, a high-speed scanning galvanometer system, and a high-resolution position-sensitive detector (preferably a CMOS image sensor).

[0024] The unit operates using the classic laser triangulation method. Specifically, a laser emitter emits a collimated laser beam with a stable wavelength. This laser beam is precisely deflected by a high-speed scanning galvanometer system, guided, and projected onto the surface of the target needle tip, forming a tiny, high-brightness laser spot. Simultaneously, a CMOS image sensor is positioned at a fixed point at a predetermined angle to the laser emission axis, and observes the needle tip surface from this specific angle to capture the imaging position of the laser spot on the sensor's photosensitive target surface.

[0025] Since the laser emitter's emission point, the laser spot on the needle tip surface, and the optical center of the CMOS sensor form a triangle in space, and the relative position and attitude relationship (i.e., baseline length and angle) between the laser emitter and the sensor are fixed parameters known after offline system calibration, by accurately detecting the pixel coordinates of the spot on the CMOS target surface, the three-dimensional spatial coordinates of the center point of the spot in the sensor's own coordinate system can be accurately deduced through geometric calculations.

[0026] To acquire the complete three-dimensional topography of the entire needle tip or its key areas, a high-speed scanning galvanometer system drives a laser beam to perform rapid, pre-defined scanning (e.g., raster scanning) across the target surface. During this process, a CMOS sensor synchronously and continuously acquires images of the laser spot at a high frame rate and calculates the three-dimensional coordinates of the corresponding points. In this way, the unit can efficiently generate a series of high-density three-dimensional point cloud data representing the geometry of the needle tip surface. Each point in this point cloud data, for example, P... l All of these can be determined by their coordinates in the laser sensor coordinate system (C). l The three-dimensional coordinate vector [x] under ) l ,y l ,z l ] T These factors, together, form the basis for quantifying the macroscopic geometric deviation of the needle tip.

[0027] Multispectral imaging unit The multispectral imaging unit in this embodiment undertakes tasks far beyond conventional visual imaging; its core objective is to acquire surface spectral feature information that reveals the intrinsic physical properties of the workpiece. This information is indispensable for realizing the workpiece adaptive function of the online model update module. This unit preferably includes a high-speed industrial camera and a programmable composite light source system that works in conjunction with it.

[0028] In one specific embodiment, the composite light source system integrates multiple light sources in different wavelength bands. For example, it may include an array of red (R), green (G), and blue (B) LED light sources in the visible light band, as well as an array of LED light sources in the infrared (IR) band, which is invisible to the human eye.

[0029] The working principle of this unit is based on the following physical fact: the surfaces of workpieces with different physical properties (e.g., different material compositions, elastic moduli, surface roughness, heat treatment states, or the presence of micron-level coatings) will exhibit significant and quantifiable differences in their absorption, reflection, and scattering characteristics for different wavelengths of light. This difference constitutes the unique "spectral signature" of the workpiece.

[0030] During system operation, the programmable composite light source system, according to a preset control sequence, excites one or more wavelengths of light to illuminate the target area, which typically includes the needle tip and the surface of the workpiece it is about to contact or is currently contacting. Simultaneously, a high-speed industrial camera captures the reflected images of the target area under each specific lighting condition. Thus, the system acquires not a single, limited grayscale or RGB color image, but a multidimensional image dataset composed of multiple channels (each channel corresponding to a spectral band).

[0031] The rich spectral information contained in this multidimensional image data provides a crucial basis for subsequent neural network models, transcending purely geometric dimensions. This allows the system not only to "see" the positional deviation of the needle tip but also to "sense" changes in the workpiece material, thereby driving the online learning module to adjust model parameters accordingly, ultimately achieving smooth and automatic adaptation from one workpiece to another.

[0032] In summary, this multimodal perception module, through the collaborative work of the two units mentioned above, is able to capture the macroscopic geometric information of the needle tip (via point cloud data P). l The module outputs both the physical characteristics (represented by the surface-level data) and the microscopic, material-related physical features (represented by multispectral image data) in parallel. To ensure the effectiveness of subsequent data fusion, the module is also configured to append a high-precision timestamp to each frame of output data (i.e., each point cloud image and each set of multispectral images), providing the necessary time reference for subsequent spatiotemporal registration processing.

[0033] The data processing and decision-making module, which communicates with the multimodal sensing module, is configured to process the three-dimensional geometric information and surface feature information acquired by the multimodal sensing module based on a neural network model, in order to generate correction instructions for compensating for needle tip position deviation. In this embodiment, the data processing and decision-making module constitutes the computational and decision-making core of the correction system of the present invention, playing a crucial role in the system architecture. This module communicates with the aforementioned multimodal sensing module, and its core responsibility is to receive heterogeneous, multi-source raw data streams output by the sensing module. Through a series of efficient data processing, fusion, and intelligent analysis processes, it ultimately generates correction commands that can directly drive the actuators to perform precise compensation. The design of this module comprehensively considers the advancement of the algorithm and the efficiency of the hardware implementation, and is key to ensuring the real-time response and intelligent decision-making capabilities of the entire system.

[0034] In one specific embodiment, the function of the data processing and decision-making module can be decomposed into several tightly coupled sub-processes, including spatiotemporal registration and feature fusion, bias decision based on convolutional neural networks, and the underlying high-performance hardware acceleration platform.

[0035] Spatiotemporal registration and feature fusion Since the data output by the multimodal sensing module originates from two independent units with different physical locations and operating times (i.e., the laser triangulation unit and the multispectral imaging unit), these raw data must be strictly spatiotemporally aligned before any meaningful analysis can be performed.

[0036] This module first performs temporal synchronization. Specifically, it utilizes the high-precision hardware timestamps attached to each frame of data (i.e., each 3D point cloud and each set of multispectral images) by the multimodal sensing module, and associates the most temporally close data pairs through a lookup and matching algorithm. This ensures that the geometric and surface feature information used for subsequent analysis accurately corresponds to the system state at the same physical moment, avoiding analysis errors caused by data latency inconsistencies.

[0037] Subsequently, this module performs spatial registration. This process aims to align the coordinate system (Ci) of the laser ranging unit with the coordinate system of the laser ranging unit. i The 3D point cloud data acquired under the current conditions are uniformly transformed to the camera coordinate system (C) of the multispectral imaging unit. c Below this, the transformation relationship can be described by a fixed rigid transformation, defined by a 3×3 rotation matrix R and a 3×1 translation vector T. This transformation occurs at a three-dimensional point P in the laser coordinate system. l The point P in the camera coordinate system can be mapped using the following formula. c :P c =R·P l +T; The rotation matrix R and translation vector T can be pre-calculated and stored in the system through a one-time, high-precision offline calibration process. During system operation, this module applies this transformation to all point cloud data output by the laser ranging unit in real time, thereby achieving precise alignment of the two modal data in spatial dimensions.

[0038] After completing spatiotemporal registration, this module deeply fuses the aligned 3D geometric information (point cloud data) with the corresponding multispectral feature information (multichannel image data) to construct a unified, high-dimensional multimodal feature vector, denoted as x. This vector x serves as an information carrier, encoding both the precise geometric position and orientation of the target needle tip at the current moment, and also containing the unique physical material properties of the surrounding workpiece surface. This feature vector will serve as the unique and complete input for subsequent neural network model analysis and decision-making.

[0039] Bias Decision Based on Convolutional Neural Networks The core decision-making mechanism of this module is an optimized, lightweight convolutional neural network (CNN) model, denoted as f. The model's role is to construct a three-dimensional prediction and correction compensation vector from the input multimodal feature vector x. The complex nonlinear mapping relationship between them. This mapping relationship can be expressed as: Here, θ represents the set of all learnable parameters in the CNN model, such as the weights and biases of the convolutional kernels in each layer of the network. The output vector... This directly represents the positional compensation amount that the system predicts needs to apply to the needle tip in three spatial dimensions.

[0040] To enable the neural network model to be deployed on resource-constrained industrial hardware with extremely high latency requirements, this module preferably employs a series of lightweight processing techniques in the design of the CNN model. In one embodiment, the lightweight processing may include at least one of the following methods: Sparse connections and pruning: By analyzing the weight parameters of the network model after training, connections with values ​​close to zero or contributing little to the model output are permanently removed. This significantly reduces the number of multiply-accumulate operations required during inference without substantially affecting model accuracy, thereby directly improving computational speed.

[0041] Low-bit quantization: The native, typically 32-bit single-precision floating-point weights and activation values ​​in the model are converted into 16-bit half-precision, 8-bit integers, or even lower-bit-width fixed-point numbers for storage and computation. This technique not only significantly compresses the model size and reduces memory bandwidth usage, but more importantly, it allows computations to be performed by simpler, faster, and more energy-efficient fixed-point hardware.

[0042] FPGA hardware acceleration platform To fundamentally guarantee the ultra-low latency performance of this data processing and decision-making module, all its core algorithms are deployed on a hardware acceleration platform based on a Field-Programmable Gate Array (FPGA). This deployment is not a simple software code porting, but a deep algorithm-hardware co-design.

[0043] Specifically, this module is implemented by analyzing and reconstructing the aforementioned spatiotemporal registration algorithm, feature fusion process, and inference process of the lightweight CNN model at the algorithmic level, and mapping the computationally intensive operations to highly customized, parallel digital logic circuits within the FPGA. For example, matrix multiplication operations used for spatial registration can be implemented as a dedicated parallel multiplier array; large-scale convolution operations in the CNN model can be unfolded into a deep pipeline structure; and dedicated hardware processing units can also be designed for irregular memory access patterns with sparse connections and fixed-point arithmetic operations after low-bit quantization.

[0044] In this way, once the data stream enters this module, it is processed within the FPGA in a highly parallel and pipelined manner, minimizing the time overhead caused by data handling, instruction decoding, and serial computation. This hardware acceleration platform is the physical foundation and core guarantee for this module and even the entire system to achieve end-to-end millisecond-level response from physical perception to decision output.

[0045] The online model update module, which communicates with the data processing and decision module, is configured to update the parameters of the neural network model used by the data processing and decision module online based on the deviation data observed during the actual processing, so that the system can adapt to workpieces with different physical properties. In this embodiment, the online model update module is the core functional unit that endows the correction system of this invention with the ability to "learn" and "evolve," and is the technological cornerstone for realizing the advanced adaptability of the system. This module has a close two-way communication connection with the data processing and decision-making module, constructing a dynamic closed-loop learning loop with the processing effect as the final feedback. Its core responsibility is to continuously monitor the decision-making performance of the system and, when necessary, automatically and incrementally optimize the neural network model used by the data processing and decision-making module online. This allows the entire system to proactively adapt to uncertainties caused by changes in workpiece material, batch, or processing environment without manual intervention or downtime for recalibration.

[0046] In a specific embodiment, the function and principle of the online model update module can be explained in detail through its online transfer learning strategy, dynamic update mechanism, and unique adaptive driving basis.

[0047] Application of online transfer learning strategies The online model update module in this embodiment does not use the traditional from-scratch training method, but preferably adopts an efficient online transfer learning strategy.

[0048] Specifically, the initial weight parameters θ0 of the neural network model in the data processing and decision-making module are not randomly set during the initial deployment of the system, but are obtained through pre-training on a large-scale, diverse dataset. This pre-trained model has learned and solidified general and basic feature extraction and deviation analysis capabilities for the needle tip localization task.

[0049] The core task of this module is to perform online, condition-specific fine-tuning of the pre-trained model. When the system encounters new workpieces in the actual production environment that were not fully learned during the pre-training phase, this module does not discard the existing model. Instead, it uses a small amount of newly collected, high-value sample data under the current conditions to make minor but precise adjustments to the parameters of the pre-trained model. This transfer learning strategy greatly improves the efficiency and convergence speed of the learning process, enabling the system to adapt to new tasks in a very short time, demonstrating high flexibility and practicality.

[0050] Dynamic Model Update Mechanism and Process: This module establishes a continuously running dynamic learning loop, and its implementation process is as follows: First, this module continuously monitors the system's correction performance. It receives the predicted correction compensation vector output by the data processing and decision-making module. Simultaneously, the corresponding true deviation vector d is obtained from an externally set high-precision reference measurement system (e.g., a higher-precision vision or displacement sensor).

[0051] Secondly, this module compares the predicted values ​​with the actual values ​​in real time to calculate the correction error. When the magnitude of this error (e.g., Euclidean distance)... When a preset tolerance threshold is exceeded, the system determines that the current model performance can no longer meet the accuracy requirements and then triggers an online update process.

[0052] After the update process is triggered, this module will update the input data that caused this large error, namely the multimodal feature vector x used by the data processing and decision-making module. (i) , and the corresponding true deviation vector d (i) Together, they constitute a completely new training sample (x) with ground truth labels. (i) ,d (i)The system will collect a small batch of new samples, numbered m.

[0053] Then, this module employs mini-batch gradient descent to incrementally update the parameters θ of the neural network model based on this small batch of new samples. The core of this process lies in minimizing a loss function J(θ) that measures the difference between the model's predicted and actual values. In a preferred embodiment, mean squared error (MSE) can be used as the loss function. Where, f(x) (i) ;θ) is the current model's response to input x (i) The goal of learning is to find new parameters that minimize the value of J(θ). This is achieved by calculating the loss function on each parameter θ of the current model. t The partial derivative (i.e., gradient) This is achieved by updating the parameters in the opposite direction of the gradient. The parameter update iteration rule is as follows: Where, θ t+1 These are the updated parameters. η is a hyperparameter called the learning rate, which controls the step size of each parameter update to ensure stable convergence of the learning process. The updated parameters θ t+1 It will be immediately sent back to the data processing and decision-making module to replace its old model parameters, so that the system can demonstrate the effect of learning and evolution in the next decision.

[0054] Physically-based adaptive driving Unlike traditional methods that rely solely on passive correction of final geometric errors, a significant innovation of the online model update module of this invention is that it uses the surface feature information (i.e., multispectral image data) collected by the multimodal perception module as a key and forward-looking physical basis for driving adaptive model updates.

[0055] Specifically, this module is configured to analyze the time-series changes of the sub-vectors representing surface spectral features in the input multimodal feature vector x. When a significant and sustained drift in this spectral feature is detected, even if the geometric processing deviation has not yet accumulated to the point of triggering an update threshold, this module can infer with high probability that the physical properties of the workpiece (such as material, surface treatment, etc.) have changed. In this case, the system can proactively and more sensitively trigger the online learning process, or dynamically adjust hyperparameters such as the learning rate η, thereby achieving a more predictive proactive adaptation rather than a lagging passive compensation.

[0056] To ensure that the online learning process itself does not become a bottleneck for system response, the computationally intensive tasks involved in this module, especially gradient comprehension... The computation (i.e., the backpropagation process of the neural network) is also deployed on an FPGA hardware acceleration platform, with a dedicated hardware computing unit designed for it. This ensures that the model's "learning" and "evolution" processes can be completed at extremely high speeds, thereby maintaining the end-to-end real-time performance of the entire closed-loop control system.

[0057] A high-precision adaptive tip dynamic correction online transfer learning method includes the following steps: S1. Through the multimodal sensing module, the three-dimensional geometric information of the target needle tip and the surface feature information reflecting the physical properties of the workpiece are acquired simultaneously. In this embodiment, the multimodal perception module includes a laser triangulation ranging unit and a multispectral vision acquisition unit. Both are used to provide information sources of different dimensions to jointly construct a multimodal feature space describing the state of the target needle tip.

[0058] The laser triangulation ranging unit is preferably positioned on one side of the machining equipment spindle, enabling it to output real-time three-dimensional geometric coordinates representing the spatial position of the workpiece or reference surface when the needle tip approaches it. This ranging process is based on the principle of triangulation, recording the projection position of the reflected laser beam on the imaging surface and performing spatial inversion by combining it with a preset measurement geometry.

[0059] Meanwhile, the images acquired by the multispectral vision acquisition unit are not limited to the visible light band, but include image channels across multiple wavelength bands to capture spectral response characteristics such as reflectivity, roughness, and dispersion of the needle tip surface and surrounding area. This information can effectively characterize the surface treatment state and material composition of the workpiece.

[0060] Finally, the multimodal perception module concatenates the laser ranging data (denoted as g) with the multispectral image feature vector (denoted as s) into a set of composite input features x = [g; s], and sends it to the data processing module for subsequent decision-making and reasoning.

[0061] S2. Based on a preset neural network model, the three-dimensional geometric information and surface feature information are fused to generate a correction command to compensate for the deviation of the needle tip position. In this embodiment, the composite input feature x is input into a preset neural network model, which is preferably a lightweight convolutional neural network architecture with good embedding ability and real-time response capability.

[0062] The model fuses features from different modalities through a series of feature transformation operations (such as convolution, pooling, activation, and fully connected mapping) to extract the latent space representation most strongly correlated with the needle tip deviation. The model ultimately outputs a displacement vector for deviation correction. The magnitude and direction represent the estimated deviation of the needle tip in three spatial dimensions, respectively.

[0063] In a preferred embodiment, the displacement vector The error is sent directly to the CNC system to correct the position of the end effector, thereby achieving real-time machining path compensation. The generation speed of this correction command is supported by an FPGA hardware acceleration platform, enabling completion within extremely short latency and ensuring rapid response to dynamic deviations.

[0064] To improve its applicability to different types of workpieces, the aforementioned neural network model is not trained on a specific type of workpiece with a fixed material or shape. Instead, it is pre-trained on diverse training datasets to give it a certain degree of cross-working-condition generalization ability. Its initial weight parameters θ0 are obtained through offline training, exhibiting stable initialization performance.

[0065] S3. During the processing, the parameters of the neural network model are updated online and incrementally based on the observed actual deviations, so that the method's correction strategy can adapt to workpieces with different physical properties. In this embodiment, to enable the above model to have long-term adaptability under complex working conditions, the system introduces an online transfer learning mechanism to achieve continuous and autonomous optimization of model parameters.

[0066] During the processing, the system continuously monitors the prediction deviation vector. The error between the actual deviation vector d and the actual deviation vector d. The actual deviation d can be obtained through an external high-precision calibration system, such as a high-resolution interferometer or a spindle feedback displacement sensor. The error is measured in Euclidean distance as follows: When E exceeds the allowable threshold ε, the system triggers the model update mechanism.

[0067] Specifically, the composite input feature x corresponding to the current time step (i) and the true deviation d (i) The samples are cached as labeled samples, and the system uses them to build a new set of samples. Subsequently, the neural network parameters θ are incrementally updated using mini-batch gradient descent, with the update objective being to minimize the loss function. The parameter update steps are as follows: Where η is the adaptive learning rate, the updated model immediately replaces the old model and is put into operation, achieving continuous optimization without downtime.

[0068] It should be noted that in model updates, the collected surface spectral features s also serve as an important basis for determining whether to trigger learning in advance. When a stable drift trend of this feature is detected on the time axis, even if the deviation has not yet reached the threshold, the system can predict changes in material properties and initiate parameter fine-tuning in advance, demonstrating the system's early warning and adaptive behavior.

[0069] To ensure real-time performance, the forward inference and backward propagation processes of the above parameter update algorithm are implemented in parallel in the FPGA. Its control and caching structure communicates with the main control module via DMA to minimize processing latency.

[0070] As described above, through multimodal information fusion for correction prediction and a dynamic model update mechanism based on online transfer learning, the method provided in this embodiment can effectively improve the accuracy and stability of needle tip pose control while ensuring the continuity of the machining process. Furthermore, it can quickly adapt the model parameters when faced with workpieces of different batches, materials, or surface conditions, thereby ensuring machining consistency and system versatility.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A high-precision adaptive needle tip dynamic correction module system, characterized in that, include: The multimodal sensing module is configured to simultaneously acquire the three-dimensional geometric information of the target needle tip and the surface feature information reflecting the physical properties of the workpiece; The data processing and decision-making module, which is communicatively connected to the multimodal sensing module, is configured to process the three-dimensional geometric information and surface feature information acquired by the multimodal sensing module based on a neural network model, so as to generate a correction command for compensating for the tip position deviation. An online model update module, which is communicatively connected to the data processing and decision module, is configured to update the parameters of the neural network model used by the data processing and decision module online based on deviation data observed during actual processing, so that the system can adapt to workpieces with different physical properties.

2. The high-precision adaptive needle tip dynamic correction module system according to claim 1, characterized in that, The multimodal sensing module includes: A laser triangulation unit is used to acquire the three-dimensional geometric information of the target needle tip; a multispectral imaging unit is used to acquire the surface feature information reflecting the physical properties of the workpiece.

3. The high-precision adaptive needle tip dynamic correction module system according to claim 2, characterized in that, The data processing and decision-making module is further configured to execute a spatiotemporal registration algorithm before processing the three-dimensional geometric information and the surface feature information, so as to fuse the data collected by the laser triangulation unit and the multispectral imaging unit in a unified spatiotemporal coordinate system.

4. The high-precision adaptive needle tip dynamic correction module system according to claim 1, characterized in that, The neural network model described in the data processing and decision-making module is a convolutional neural network model.

5. The high-precision adaptive needle tip dynamic correction module system according to claim 4, characterized in that, The data processing and decision-making module includes a field-programmable gate array (FPGA) hardware platform, and the convolutional neural network model is deployed on the FPGA hardware platform for hardware acceleration processing.

6. The high-precision adaptive needle tip dynamic correction module system according to claim 5, characterized in that, The convolutional neural network model is a lightweight model, and its lightweight processing method includes at least one of sparse connection or low bit quantization techniques.

7. The high-precision adaptive needle tip dynamic correction module system according to claim 1, characterized in that, The online model update module is configured to use an online transfer learning strategy to update the parameters of the neural network model, that is, to fine-tune the model using the bias data based on a pre-trained model.

8. The high-precision adaptive needle tip dynamic correction module system according to claim 1, characterized in that, The online model update module is specifically configured as follows: The surface feature information, three-dimensional geometric information, and corresponding actual processing deviations obtained by the multimodal perception module are continuously collected and together constitute incremental training samples. Using the incremental training samples, the weight parameters of the neural network model are updated online via mini-batch gradient descent.

9. The high-precision adaptive needle tip dynamic correction module system according to claim 1, characterized in that, The online model update module uses the surface feature information collected by the multimodal perception module as the main basis for distinguishing the physical properties of different workpieces, so as to drive the adaptive update of the neural network model parameters.

10. An online transfer learning method for high-precision adaptive needle tip dynamic correction, used in the high-precision adaptive needle tip dynamic correction module system according to any one of claims 1-9, characterized in that, Includes the following steps: S1. Through the multimodal sensing module, the three-dimensional geometric information of the target needle tip and the surface feature information reflecting the physical properties of the workpiece are acquired simultaneously. S2. Based on a preset neural network model, the three-dimensional geometric information and the surface feature information are fused to generate a correction command for compensating for the tip position deviation. S3. During the processing, the parameters of the neural network model are updated online and incrementally based on the observed actual deviations, so that the correction strategy of the method can adapt to workpieces with different physical properties.

Images