An intelligent detection device and detection method for an anti-corrosion aluminum alloy part

Abstract

The present application relates to the field of intelligent detection technology, and discloses an intelligent detection device and a detection method for corrosion-resistant aluminum alloy parts. The method includes: extracting features from an image sequence of the corrosion-resistant aluminum alloy cutting process to obtain a multi-dimensional feature map; performing three-dimensional space mapping and morphological feature analysis on the multi-dimensional feature map to obtain spatio-temporal feature data of the cutting trajectory; inputting the spatio-temporal feature data of the cutting trajectory into a hierarchical cutting parameter prediction model for hierarchical feature parameter prediction to obtain a cutting parameter prediction result; performing Gaussian mixture modeling and kernel density analysis on the cutting parameter prediction result to obtain probability density distribution data of the cutting trajectory, and optimizing the variable observation domain of the Gaussian process model to obtain an optimized sequence of cutting parameters; inputting the optimized sequence of cutting parameters into the Kalman filter-Hungarian algorithm for state tracking to obtain a control instruction for the cutting parameters, thereby realizing real-time and precise control of the cutting parameters and significantly improving the cutting processing quality.

Description

Technical Field

[0001] This application relates to the field of intelligent detection technology, and particularly to an intelligent detection device and method for corrosion-resistant aluminum alloy parts. Background Art

[0002] During the cutting process of corrosion-resistant aluminum alloy, due to the complexity of material properties and processing environment, traditional cutting parameter control methods are difficult to meet the requirements of high-precision processing.

[0003] Currently, the quality inspection of the corrosion-resistant aluminum alloy cutting process mainly relies on manual experience and fixed parameter control. This method is not only inefficient but also difficult to cope with the dynamic changes during the cutting process. Especially in complex cutting conditions, it is difficult to achieve real-time tracking of the cutting trajectory and parameter optimization, which easily leads to problems such as unstable cutting quality and fluctuating precision. Summary of the Invention

[0004] This application provides an intelligent detection device and method for corrosion-resistant aluminum alloy parts, thereby realizing real-time and precise control of cutting parameters and significantly improving the quality of cutting processing.

[0005] In the first aspect of this application, an intelligent detection method for corrosion-resistant aluminum alloy parts is provided. The intelligent detection method for corrosion-resistant aluminum alloy parts includes:

[0006] Extracting features from the image sequence of the corrosion-resistant aluminum alloy cutting process to obtain a multi-dimensional feature map;

[0007] Performing three-dimensional space mapping and morphological feature analysis on the multi-dimensional feature map to obtain spatio-temporal feature data of the cutting trajectory;

[0008] Inputting the spatio-temporal feature data of the cutting trajectory into a hierarchical cutting parameter prediction model for hierarchical feature parameter prediction to obtain a cutting parameter prediction result;

[0009] Performing Gaussian mixture modeling and kernel density analysis on the cutting parameter prediction result to obtain probability density distribution data of the cutting trajectory, and optimizing the Gaussian process model for variable observation domains to obtain an optimized sequence of cutting parameters;

[0010] Inputting the optimized sequence of cutting parameters into the Kalman filter-Hungarian algorithm for state tracking to obtain a control instruction for cutting parameters.

[0011] In the second aspect of this application, an intelligent detection device for corrosion-resistant aluminum alloy parts is provided. The intelligent detection device for corrosion-resistant aluminum alloy parts includes:

[0012] A feature extraction module for extracting features from the image sequence of the corrosion-resistant aluminum alloy cutting process to obtain a multi-dimensional feature map;

[0013] A spatial mapping module for performing three-dimensional spatial mapping and morphological feature analysis on the multi-dimensional feature map to obtain spatio-temporal feature data of the cutting trajectory;

[0014] A parameter prediction module for inputting the spatio-temporal feature data of the cutting trajectory into a hierarchical cutting parameter prediction model to perform hierarchical feature parameter prediction and obtain a cutting parameter prediction result;

[0015] An optimization module for performing Gaussian mixture modeling and kernel density analysis on the cutting parameter prediction result to obtain probability density distribution data of the cutting trajectory, and performing variable observation domain optimization on the Gaussian process model to obtain an optimized sequence of cutting parameters;

[0016] A state tracking module for inputting the optimized sequence of cutting parameters into the Kalman filter-Hungarian algorithm for state tracking to obtain a control instruction for the cutting parameters.

[0017] Compared with the prior art, the present application has the following beneficial effects: Through the design of the multi-scale cascaded GhostNet network and the spatial positioning and feature generalization enhancement module, the feature extraction ability for complex scenes during the cutting process is significantly enhanced, and the detection accuracy of the cutting trajectory is greatly improved; The hierarchical cutting parameter prediction model is used for hierarchical feature fusion, combined with the design of a multi-level prediction head, to achieve accurate prediction of the cutting parameters and effectively reduce the error of parameter prediction; The Gaussian mixture modeling and kernel density analysis methods are introduced, and through the variable observation domain adaptive optimization strategy, the performance degradation problem of traditional methods in low-speed cutting scenarios is solved, ensuring the stability of cutting parameter optimization; A state tracking framework based on the Kalman filter-Hungarian algorithm is designed, combined with a parameter optimization controller and a trajectory feedback compensation mechanism, to achieve real-time and accurate control of the cutting parameters, significantly improving the cutting quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] The structures, ratios, sizes, etc. shown in the drawings of this specification are only used to cooperate with the content disclosed in the specification for those who are familiar with this technology to understand and read, and are not used to limit the limited conditions under which the present invention can be implemented. Therefore, they do not have a substantial technical meaning. Any modification of the structure, change in the proportional relationship, or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope that can be covered by the technical content disclosed in the present invention.

[0020] Figure 1 It is a schematic flowchart of the intelligent detection method for corrosion-resistant aluminum alloy parts provided by an embodiment of the present invention;

[0021] Figure 2 It is a schematic block diagram of the structure of the intelligent detection device for corrosion-resistant aluminum alloy parts provided by an embodiment of the present invention. Specific Embodiments

[0022] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] The flowchart shown in the accompanying drawings is only an example for illustration, and does not necessarily include all contents and operations / steps, nor does it necessarily need to be executed in the described order. For example, some operations / steps can also be decomposed, combined or partially merged, so the actual execution order may be changed according to the actual situation.

[0024] It should also be understood that the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. As used in the specification of this application and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0025] It should be further understood that the term "and / or" used in the specification of this application and the appended claims refers to any combination and all possible combinations of one or more of the related listed items, and includes these combinations. Please refer to Figure 1 , an embodiment of the intelligent detection method for corrosion-resistant aluminum alloy parts in the embodiments of this application includes:

[0026] Step 100: Extract features from the image sequence of the corrosion-resistant aluminum alloy cutting process to obtain a multi-dimensional feature map;

[0027] It can be understood that the execution subject of this application can be the intelligent detection device for corrosion-resistant aluminum alloy parts, or a terminal or a server, and specifically is not limited here. In the embodiments of this application, the server is taken as the execution subject for illustration.

[0028] Specifically, a multi-scale pyramid decomposition is performed on the image sequence of the anti-corrosion aluminum alloy cutting process to obtain a high-resolution feature layer, a medium-resolution feature layer, and a low-resolution feature layer. The multi-scale pyramid decomposition decomposes the input image into image levels of different resolutions through downsampling and convolution operations. Each level contains feature representations of different scales from the input image. Among them, the high-resolution feature layer retains more detailed information, the medium-resolution feature layer focuses on the significant patterns in local regions, and the low-resolution feature layer extracts global low-frequency semantic information. The high-resolution feature layer is input into the first-level GhostNet module for feature extraction. The design of the GhostNet module is based on the idea of efficient feature representation, which combines the backbone processing of information with redundant feature decomposition using Ghost bottleneck units. The first-level GhostNet module contains 8 Ghost bottleneck units. Each bottleneck unit consists of a main branch convolutional layer and a parallel branch transformation layer. The main branch convolutional layer extracts basic features, and the parallel branch transformation layer generates redundant features through simple linear operations, thus reducing the computational amount while retaining the detailed information of the image. Through this step, the first-level feature map is extracted from the high-resolution feature layer, capturing the microscopic texture and edge features in the cutting process. The medium-resolution feature layer is input into the second-level GhostNet module for feature extraction. The second-level GhostNet module contains 6 Ghost bottleneck units. To improve the accuracy of feature extraction and the expressive ability of the model, a channel attention mechanism and a spatial attention mechanism are embedded in each bottleneck unit. The channel attention mechanism assigns weights to different channels, enabling the model to pay more attention to the key features related to the cutting process while ignoring redundant information. The spatial attention mechanism enhances the model's ability to capture significant regions in the image by weighting the spatial positions. After the processing of this module, the generated second-level feature map can effectively express the significant patterns in local regions during the cutting process. The low-resolution feature layer is input into the third-level GhostNet module for feature extraction. The third-level GhostNet module consists of 4 Ghost bottleneck units. To improve the ability to model global semantic information, each bottleneck unit is configured with a cross-layer residual connection and an adaptive feature aggregation unit. The cross-layer residual connection can effectively alleviate the problem of gradient disappearance during information transmission and ensure the fusion between low-level features and high-level features. The adaptive feature aggregation unit enhances the global modeling ability of low-resolution features by dynamically adjusting the weights of features at different scales. Through the processing of the third-level GhostNet module, the generated third-level feature map mainly expresses the macroscopic information and global trends in the cutting process. The first-level feature map, the second-level feature map, and the third-level feature map are subjected to feature fusion. A weighted fusion strategy is adopted during the fusion process, enabling the information of features at different resolutions to be complementary and generating a fusion feature map containing multi-scale information. The fusion feature map is input into the feature enhancement module for processing. This module contains 3 depthwise separable convolutional layers and 1 global context encoder.The depthwise separable convolution layer decomposes the standard convolution into depthwise convolution and pointwise convolution, significantly reducing the computational complexity while preserving the image detail information. After each depthwise separable convolution layer, a batch normalization layer is connected to standardize the feature distribution and introduce non-linearity through the PReLU activation function, enhancing the feature representation ability. The global context encoder enhances the model's perception of global information by modeling the global context relationship of the entire image. After being processed by the feature enhancement module, a multi-dimensional feature map is obtained.

[0029] Step 200: Perform three-dimensional spatial mapping and morphological feature analysis on the multi-dimensional feature map to obtain the spatio-temporal feature data of the cutting trajectory;

[0030] Specifically, perform a spatial coordinate transformation on the multi-dimensional feature map. Convert the two-dimensional image features into position information in a three-dimensional space coordinate system. Through a spatial transformation method based on projective geometry, combined with the actual calibration parameters of the cutting device and the geometric relationship of multi-view images, generate a spatial mapping matrix of the cutting trajectory. This matrix takes the pixel position of the cutting trajectory as input and maps its actual position in the three-dimensional coordinate system, thereby accurately characterizing the dynamic changes of the cutting trajectory in three-dimensional space. Construct a geometric feature extraction operator for the cutting trajectory based on the spatial mapping matrix and analyze the geometric features of the cutting trajectory. The geometric feature extraction operator is the core tool for feature analysis, including the directional gradient operator, curvature operator, and continuity operator, etc. Among them, the directional gradient operator is used to calculate the change rate of the cutting trajectory in different directions, so as to extract the edge characteristics of the trajectory; the curvature operator reflects the bending degree of the cutting trajectory based on the local curvature change of the trajectory points; the continuity operator ensures the integrity description of the cutting trajectory by evaluating the connectivity between trajectory points. Through the comprehensive analysis of these geometric operators, generate a geometric feature vector of the cutting trajectory, characterizing the overall shape of the cutting trajectory and its local change characteristics. Perform morphological operations on the geometric feature vector to extract the morphological feature data of the cutting trajectory. Morphological operation is an image processing method based on set theory, and the operations include dilation, erosion, opening operation, and closing operation, etc. In this process, by performing structured operations on the geometric feature vector, enhance the key features of the trajectory, remove redundant information, and make the morphological boundary of the cutting trajectory clearer. Perform temporal feature analysis on the morphological feature data. Based on the time series distribution of the trajectory points, by analyzing the dynamic change pattern of the trajectory points on the time axis, generate a temporal feature sequence of the cutting trajectory. This sequence quantifies the law of the geometric shape of the trajectory changing with time based on time. On this basis, establish a spatio-temporal feature association model based on the temporal feature sequence of the cutting trajectory. This model combines the spatial position of the trajectory with the time evolution process to complete the spatio-temporal association modeling of the cutting trajectory. Adopt a method based on graph neural network or recurrent neural network to jointly model the spatial features and time series features of the cutting trajectory, and extract the association features of the trajectory in both time and space dimensions. Through spatio-temporal association modeling, capture the dynamic behavior pattern and spatial distribution law of the cutting trajectory, and generate the spatio-temporal association features of the cutting trajectory. Combine the spatial mapping matrix, morphological feature data, and spatio-temporal association features of the cutting trajectory to generate the spatio-temporal feature data of the cutting trajectory.

[0031] Step 300: Input the spatio-temporal feature data of the cutting trajectory into a hierarchical cutting parameter prediction model for hierarchical feature parameter prediction to obtain a cutting parameter prediction result;

[0032] It should be noted that a hierarchical cutting parameter prediction model is constructed. The input spatio-temporal feature data is decomposed layer by layer through a feature hierarchical network to extract feature information at different levels. The spatio-temporal feature data of the cutting trajectory is input into the feature hierarchical network. Through a series of convolutional operations and pooling processes, the input feature data is decomposed into a local feature layer, a regional feature layer, and a global feature layer. The local feature layer is used to capture the detailed information of the cutting trajectory. The regional feature layer focuses on the characteristic analysis of local regions, while the global feature layer covers the overall spatial and temporal dynamics of the cutting trajectory. After completing the feature hierarchy, the local feature layer is input into the first prediction head in the hierarchical cutting parameter prediction model for processing. The first prediction head contains 4 convolutional blocks and 1 fully connected layer, and each convolutional block is equipped with a spatial pyramid pooling module. The spatial pyramid pooling module effectively captures the multi-scale information of local features by performing pooling operations on features at multiple scales, enhancing the representational ability of the details of the cutting trajectory. The stacking of convolutional blocks further extracts features in depth, and the fully connected layer maps the extracted features into local cutting parameter prediction data, which are used to describe the microscopic changes of the tool trajectory during the cutting process and the local parameter adjustment requirements. At the same time, the regional feature layer is input into the second prediction head for processing. The second prediction head contains 3 convolutional blocks and 1 fully connected layer, and each convolutional block is configured with a channel concatenation module. The channel concatenation module enhances the model's ability to capture the dependence relationship between channels by establishing a concatenation relationship between different channels, enabling the regional feature layer to better express the significant patterns in local regions of the cutting trajectory. The gradual stacking of convolutional blocks further refines the features to ensure the accurate expression of regional information, and the fully connected layer maps the extracted regional features into regional cutting parameter prediction data. These prediction data are used to guide the parameter optimization in the medium-scale range during the cutting process, such as cutting speed and energy distribution. At the same time, the global feature layer is input into the third prediction head for processing. The third prediction head consists of 2 convolutional blocks and 1 fully connected layer, and each convolutional block is set with a feature recalibration module. The feature recalibration module dynamically adjusts the global features by learning the importance weights of the features, making the model pay more attention to the key information related to the overall state of the cutting trajectory. The introduction of this module can significantly improve the model's sensitivity to global information and ensure the accurate modeling of the overall shape and spatio-temporal dynamics of the cutting trajectory. The fully connected layer maps the global features into global cutting parameter prediction data, which are used to describe the adjustment of the overall strategy during the cutting process, such as the global optimization of the cutting path and the balanced distribution of parameters. The local cutting parameter prediction data, regional cutting parameter prediction data, and global cutting parameter prediction data are input into a parameter fusion network for fusion processing. The parameter fusion network integrates the detail accuracy of local features, the regional information of regional features, and the integrity of global features by combining data at different levels.During the fusion process, strategies such as weighted average, attention mechanism or multi-layer perceptron are adopted to dynamically allocate weights according to the importance of different hierarchical features, and generate the prediction results of cutting parameters.

[0033] Step 400: Perform Gaussian mixture modeling and kernel density analysis on the prediction results of cutting parameters to obtain the probability density distribution data of the cutting trajectory, and optimize the Gaussian process model for variable observation domains to obtain the optimized sequence of cutting parameters;

[0034] Specifically, a Gaussian mixture model is constructed based on the cutting parameter prediction results. By analyzing the spatial characteristics of the cutting trajectory, the trajectory is divided into several local regions, and a Gaussian component is generated for each local region. Each Gaussian component includes a mean vector and a covariance matrix, and these parameters jointly define the position and distribution shape of the component in the cutting trajectory. By constructing a parameter group containing N Gaussian components, the multimodal distribution characteristics of the cutting trajectory are initially described. The expectation-maximization algorithm is executed according to the Gaussian component parameter group to optimize the model parameters. The expectation-maximization algorithm includes two main steps. In the "expectation" step, for each cutting trajectory data point, the posterior probability of its belonging to each Gaussian component is calculated, and these probability values indicate the attribution weights of the data point on different regional components. In the "maximization" step, the parameters of the Gaussian component, including the mean vector and the covariance matrix, are re-estimated using the calculated posterior probabilities, and these parameters are optimized through continuous iteration. This optimization process continuously improves the value of the logarithmic likelihood function of the model in each iteration until convergence. The optimized Gaussian mixture parameters are obtained to accurately describe the probability distribution of the cutting trajectory. The Gaussian mixture parameters are input into the radial basis kernel function for density calculation. The radial basis kernel function can smooth the probability distribution and improve the accuracy of density estimation. The bandwidth parameter matrix of the kernel function is set, and this matrix determines the smoothing degree and local sensitivity of the kernel function. To select the optimal bandwidth parameter, a cross-validation optimization method is adopted. By performing a grid search on the likelihood function values under different bandwidth parameters, the bandwidth parameter that maximizes the likelihood function value is selected to obtain the target kernel density function. After obtaining the target kernel density function, probability compensation is performed for the special requirements of the low-speed cutting region. The low-speed cutting regions are identified through the target kernel density function, and these regions appear as low-value regions in the density distribution. The compensation weights are calculated based on these regions, and the calculation of the weights is based on the local physical characteristics of the trajectory and the cutting stability requirements. The original probability density values are updated using these compensation weights to generate the compensated density function. To ensure the consistency and physical meaning of the compensated density function as a whole, probability integral normalization is performed on it, that is, the density value is divided by the overall density integral to make the density distribution maintain the normalization characteristic globally. The probability density distribution data of the cutting trajectory is obtained to describe the probability characteristics of the cutting trajectory in space. Based on the generated probability density distribution data, the Gaussian process model is optimized in the variable observation domain. In this process, the input space of the Gaussian process model is dynamically adjusted to adapt to the distribution characteristics of different regions in the cutting trajectory. In the low-speed cutting region, the compensated density function is introduced to adjust the weight distribution of the observation domain, thereby enhancing the prediction ability of the model in these regions. At the same time, by optimizing the covariance function of the model, the adaptability and robustness of the Gaussian process model to uncertain regions are improved. The optimized Gaussian process model can more accurately predict the dynamic changes of the cutting parameters. An optimized sequence of cutting parameters is generated through the above optimized Gaussian process model.

[0035] The probability density distribution data of the cutting trajectory is input into the observation domain partitioning module for adaptive partitioning processing. By analyzing the spatial probability distribution characteristics of the cutting trajectory, the cutting trajectory is dynamically partitioned according to the density changes and characteristic requirements in different regions to generate variable observation domain configuration data. This process uses a density threshold and a regional clustering algorithm to divide the cutting trajectory into several sub-regions, each sub-region having independent observation domain characteristics, ensuring that the subsequent optimization process can be precisely adjusted according to the characteristics of the local region. Based on the variable observation domain configuration data, the kernel function of the Gaussian process model is constructed. The design of the kernel function adopts a combined kernel function based on the Markov property. This kernel function captures the characteristics of periodic change patterns and random change patterns in the cutting trajectory by combining the weighted combination of the periodic kernel function and the Gaussian kernel function. The periodic kernel function models the periodic behavior of the trajectory to ensure accurate prediction of repeated cutting patterns, while the Gaussian kernel function uses its smoothing characteristics and local correlation to capture random perturbations and regional differences. During the construction of the kernel function, the combined kernel function can adaptively generate kernel function parameters according to the characteristics of different regions of the cutting trajectory through the dynamic adjustment of the weight parameters. The hyperparameters of the kernel function of the Gaussian process model are optimized. The core of hyperparameter optimization is to ensure the modeling accuracy of the Gaussian process model for the cutting trajectory characteristics by searching for the optimal parameter combination. The optimization process combines the physical constraints of the cutting process for parameter search. The physical constraints include cutting speed, smoothness of the tool path, and limiting conditions of material properties. By methods such as grid search or Bayesian optimization, the key hyperparameters of the kernel function, such as the amplitude factor, correlation scale, and weight coefficient, are gradually adjusted and evaluated according to the log marginal likelihood value of the model to obtain the optimized Gaussian process model. Based on the optimized Gaussian process model, an uncertainty-based active learning strategy is used to select sampling points. The active learning strategy identifies key points that need further sampling by analyzing the prediction uncertainty of the model in different cutting parameter spaces. At these key points, sparse sampling is performed to maximize the information gain, generating a target sampling point sequence. Based on the target sampling point sequence, the cutting quality is evaluated and the sampling points are scored to generate a cutting parameter scoring matrix. The cutting parameter scoring matrix combines the parameter accuracy predicted by the model, the local quality characteristics of the sampling points, and the actual effect of the cutting process, and quantitatively evaluates different sampling points with a multi-dimensional scoring mechanism. The scoring matrix is input into the sequence optimizer for temporal optimization of the cutting parameters. The sequence optimizer selects the optimal parameter path based on the scoring matrix through dynamic programming algorithms or reinforcement learning methods to achieve global optimization of the cutting parameters. During the selection process, the optimizer fully considers the continuity and globality of the cutting parameters to ensure that the final optimized sequence has a smooth change trend at different time points and meets the operation constraints in the actual cutting process. Through the above steps, an optimized sequence of cutting parameters is obtained.

[0036] Step 500: Input the optimized sequence of cutting parameters into the Kalman filter-Hungarian algorithm for state tracking to obtain the control instructions for the cutting parameters.

[0037] Specifically, input the optimized sequence of cutting parameters into the state prediction module to construct a state space model for cutting parameter tracking. The core of the state prediction module lies in establishing the system state equation and the measurement equation. These two equations respectively describe the state transition process and the observation process of the cutting parameters. By constructing the state transition matrix and the observation matrix of the cutting parameters, the dynamic evolution of the cutting parameters in the time series dimension and the relationship between the measured value and the actual state are characterized, forming the state space model of the cutting parameters. Perform Kalman filter calculation on the state space model of the cutting parameters. The first step of the Kalman filter is the prediction stage. In this stage, the predicted state vector of the cutting parameters is calculated through the state transition matrix and the control input, and at the same time, the predicted covariance matrix is estimated to measure the confidence of the system in the predicted state. After obtaining the predicted state vector, it is combined with the actual observation data to correct the system state. Input the optimized sequence and the observation data into the Hungarian algorithm module. This module uses the optimal matching strategy of the bipartite graph to associate multi-objective states. The Hungarian algorithm generates the target association matrix by optimizing the calculation of the cost matrix between the predicted state vector and the observation data, clarifying the best matching relationship between each observation value and the predicted state. Perform measurement update on the association result. The process of measurement update includes calculating the Kalman gain matrix, the posterior state estimate, and the posterior covariance matrix. The Kalman gain matrix is an important parameter that weighs the predicted state and the actual observation value. Its calculation is based on the ratio of the predicted covariance and the observation noise covariance, and is used to dynamically adjust the weight of state correction. After calculating the Kalman gain matrix, apply it to the target association matrix to correct the predicted state, obtain the posterior state estimate, that is, the optimal state sequence of the cutting parameters, and at the same time update the posterior covariance matrix to reflect the uncertainty of the system. Input the optimal state sequence of the cutting parameters into the parameter optimization controller to achieve the joint optimization of the key parameters in the cutting process. The parameter optimization controller analyzes the cutting speed, feed rate, and cutting depth in the optimal state sequence, and uses a multi-objective optimization algorithm to comprehensively adjust these parameters to generate preliminary parameter adjustment instructions. These instructions are designed to ensure a balance between efficiency and quality in the cutting process, while meeting the material characteristics and process requirements. Perform trajectory feedback compensation on the parameter adjustment instructions. The process of trajectory feedback compensation is based on the error calculation between the real-time cutting state and the target trajectory. By comparing the current state with the target state, the compensation value is calculated, which can correct the deviation caused by system errors or external disturbances. After the compensation value is calculated, it is superimposed on the parameter adjustment instructions to generate the final cutting parameter control instructions.

[0038] In the embodiments of the present application, through the design of the multi-scale cascaded GhostNet network and the spatial positioning and feature generalization enhancement module, the feature extraction ability for complex scenarios during the cutting process is significantly enhanced, greatly improving the cutting trajectory detection accuracy; the hierarchical cutting parameter prediction model is used for hierarchical feature fusion, combined with the design of a multi-level prediction head, to achieve accurate prediction of cutting parameters and effectively reduce the error of parameter prediction; the Gaussian mixture modeling and kernel density analysis methods are introduced, and through the variable observation domain adaptive optimization strategy, the performance degradation problem of traditional methods in low-speed cutting scenarios is solved, ensuring the stability of cutting parameter optimization; a state tracking framework based on the Kalman filter-Hungarian algorithm is designed, combined with a parameter optimization controller and a trajectory feedback compensation mechanism, to achieve real-time and accurate control of cutting parameters, significantly improving the cutting quality.

[0039] In a specific embodiment, the process of executing step 100 may specifically include the following steps:

[0040] Perform multi-scale pyramid decomposition on the image sequence of the anti-corrosion aluminum alloy cutting process to obtain a high-resolution feature layer, a medium-resolution feature layer, and a low-resolution feature layer;

[0041] Input the high-resolution feature layer into the first-level GhostNet module for feature extraction. The first-level GhostNet module contains 8 Ghost bottleneck units, and each Ghost bottleneck unit is composed of a main branch convolutional layer and a parallel branch transformation layer to obtain a first-level feature map;

[0042] Input the medium-resolution feature layer into the second-level GhostNet module for feature extraction. The second-level GhostNet module contains 6 Ghost bottleneck units, and each Ghost bottleneck unit is provided with a channel attention mechanism and a spatial attention mechanism to obtain a second-level feature map;

[0043] Input the low-resolution feature layer into the third-level GhostNet module for feature extraction. The third-level GhostNet module contains 4 Ghost bottleneck units, and each Ghost bottleneck unit is configured with a cross-layer residual connection and an adaptive feature aggregation unit to obtain a third-level feature map;

[0044] Perform feature fusion on the first-level feature map, the second-level feature map, and the third-level feature map to obtain a fused feature map, and input the fused feature map into the feature enhancement module for processing. The feature enhancement module contains 3 depthwise separable convolutional layers and 1 global context encoder. After each depthwise separable convolutional layer, a batch normalization layer and a PReLU activation function are connected to obtain a multi-dimensional feature map.

[0045] Specifically, the image sequence of the cutting process is used as input data for multi-scale pyramid decomposition. The original image data is decomposed into a high-resolution feature layer, a medium-resolution feature layer, and a low-resolution feature layer through downsampling and Gaussian filtering, corresponding to the detailed information, local patterns, and global structure of the image respectively. Assume the input image sequence is , where and represent the spatial coordinates of the image, and represents the time frame. The multi-scale decomposition is expressed by the formula:

[0046] ;

[0047] ;

[0048] ;

[0049] where is the high-resolution feature layer, is the medium-resolution feature layer, is the low-resolution feature layer, DownSample() is the downsampling operation, implemented using convolution with a stride of 2 or average pooling. The high-resolution feature layer is input into the first-level GhostNet module for feature extraction. The GhostNet module reduces the computational cost through Ghost bottleneck units while maintaining the feature representation ability. Each Ghost bottleneck unit consists of a main branch convolutional layer and a parallel branch transformation layer. The main branch convolutional layer is used to generate the basic features, denoted as , and the parallel branch generates redundant features through a simple linear transformation, denoted as . The feature extraction process is expressed as:

[0050] ;

[0051] ;

[0052] ;

[0053] where and are the convolutional kernel and transformation weight matrix, is the bias term, and are the non-linear activation function and linear transformation function respectively, * represents the convolution operation. After being processed by 8 Ghost bottleneck units, the high-resolution feature map is obtained. Similarly, the medium-resolution feature layer Input the second - level GhostNet module. The second - level GhostNet module contains 6 Ghost bottleneck units, and channel attention mechanism and spatial attention mechanism are introduced in each unit. The channel attention mechanism realizes the enhancement of key channels by learning the importance weights of each feature channel , achieving the enhancement of key channels:

[0054] ;

[0055] where GAP() is the global average pooling operation, and are the weights and biases of the fully - connected layer. The spatial attention mechanism enhances the features of the significant regions by generating spatial weights :

[0056] ;

[0057] The final feature fusion is:

[0058] ;

[0059] Obtain the medium - resolution feature map . For the low - resolution feature layer , input it into the third - level GhostNet module, which contains 4 Ghost bottleneck units. Each unit is configured with cross - layer residual connections and adaptive feature aggregation units. The cross - layer residual connection realizes the direct flow of information by short - circuiting the input and output, defined as:

[0060] ;

[0061] And the adaptive feature aggregation unit realizes the fusion by dynamically adjusting the weights of features with different resolutions:

[0062] ;

[0063] where and are the adaptive weight parameters, and Pooling() is the downsampling operation. After processing, obtain the low - resolution feature map . After completing the feature extraction of the three resolutions, perform feature fusion on the first - level feature map , the second - level feature map and the third - level feature map . The feature fusion is achieved through weighted accumulation:

[0064] ;

[0065] where is the fusion weight, used to adjust the contribution ratio of features with different resolutions. The fused feature map The input feature enhancement module processes. The feature enhancement module consists of 3 depthwise separable convolutional layers and 1 global context encoder. The depthwise separable convolutional layer reduces the computational complexity by decomposing the convolution operation, and its operation is defined as:

[0066] ;

[0067] ;

[0068] where and are the weights of the depth convolution and the pointwise convolution respectively. The global context encoder enhances the features by modeling the global feature relationships, and finally generates a multi-dimensional feature map .

[0069] In a specific embodiment, the process of executing step 200 may specifically include the following steps:

[0070] Perform a spatial coordinate transformation on the multi-dimensional feature map to convert the two-dimensional image features into position information in a three-dimensional space coordinate system, and obtain the spatial mapping matrix of the cutting trajectory;

[0071] Construct a geometric feature extraction operator for the cutting trajectory according to the spatial mapping matrix. The geometric feature extraction operator includes a directional gradient operator, a curvature operator, and a continuity operator, and perform geometric feature analysis on the cutting trajectory to obtain the geometric feature vector of the cutting trajectory;

[0072] Perform morphological operations on the geometric feature vector to obtain the morphological feature data of the cutting trajectory, and perform temporal feature analysis on the morphological feature data to obtain the temporal feature sequence of the cutting trajectory;

[0073] Establish a spatio-temporal feature correlation model based on the temporal feature sequence of the cutting trajectory, and model the spatial position and temporal evolution of the cutting trajectory to obtain the spatio-temporal correlation features of the cutting trajectory;

[0074] Combine the spatial mapping matrix of the cutting trajectory, the morphological feature data of the cutting trajectory, and the spatio-temporal correlation features of the cutting trajectory to obtain the spatio-temporal feature data of the cutting trajectory.

[0075] Specifically, perform a spatial coordinate transformation on the multi-dimensional feature map. Assume that the input two-dimensional image feature is , where and are the spatial coordinates of the image. Map the image coordinates to a three-dimensional space coordinate system through the intrinsic matrix and extrinsic matrix of the camera. The intrinsic matrix describes the focal length and principal point offset of the camera, and the extrinsic matrix includes the rotation matrix and the translation vector The mapping relationship is expressed as:

[0076] ;

[0077] where is a point in three-dimensional space, is the homogeneous coordinate of a point in the image, is the internal parameter matrix, is the external parameter matrix. Through the above transformation, each two-dimensional image feature point is mapped to the position information in three-dimensional space, generating the spatial mapping matrix of the cutting trajectory, and its form is:

[0078] ;

[0079] where is the total number of feature points. After obtaining the spatial mapping matrix, a geometric feature extraction operator for the cutting trajectory is constructed to perform geometric feature analysis on the trajectory. The geometric feature extraction operator includes a direction gradient operator, a curvature operator, and a continuity operator. Among them, the direction gradient operator is used to calculate the local direction change of each point on the trajectory and is defined as:

[0080] ;

[0081] where and represent the spatial change between adjacent points, is the point 's gradient direction. The curvature operator is used to calculate the local bending degree of the points on the trajectory, and its formula is:

[0082] ;

[0083] where is the curvature of the point , indicating the bending degree of its trajectory. The continuity operator is used to evaluate the coherence of the trajectory points and is defined as:

[0084] ;

[0085] where is the Euclidean distance from the point to the next point, reflecting the smoothness of the trajectory. By applying the above operators to all points on the trajectory, the geometric feature vector , describe the direction change, curvature, and coherence characteristics of the trajectory. After completing the geometric feature analysis, morphological operations are performed on the geometric feature vector. Morphological operations include operations such as dilation, erosion, and skeletonization to enhance the morphological features of the trajectory. For example, the local edges of the trajectory are extended through dilation operations, noise points are removed through erosion operations, and the skeleton path of the trajectory is extracted through skeletonization. Let the structuring element of the morphological operation be , and the trajectory feature be , the dilation operation is expressed as:

[0086] ;

[0087] while the erosion operation is:

[0088] ;

[0089] After the morphological operation, the morphological feature data of the cutting trajectory is obtained . Temporal feature analysis is performed on the morphological feature data to extract the time dynamic change pattern of the trajectory. Temporal feature analysis is completed by constructing a temporal difference vector:

[0090] ;

[0091] where and are the morphological features at times and respectively, and is the corresponding temporal feature vector. Based on the temporal feature vector of the cutting trajectory, a spatio-temporal feature correlation model is established. By modeling the spatial position and time evolution of the trajectory, the spatio-temporal correlation features of the trajectory are extracted. The spatio-temporal correlation model is represented by a spatio-temporal graph, with the nodes being the trajectory points and the edges being the spatio-temporal correlation weights between the points. The spatio-temporal correlation weight is defined as:

[0092] ;

[0093] where and are the spatial positions of the trajectory points and respectively, and is a parameter that controls the correlation range. Combine the spatial mapping matrix of the cutting trajectory, the morphological feature data and the spatio-temporal correlation features to generate the spatio-temporal feature data For example, during the cutting process, if irregular high-frequency vibrations occur in a certain area, the abnormal area is detected through curvature and timing features, and the abnormal points are associated with their neighborhoods through a spatio-temporal correlation model, and finally a global feature matrix reflecting the dynamic changes of the trajectory is generated.

[0094] In a specific embodiment, the process of executing step 300 may specifically include the following steps:

[0095] Input the spatio-temporal feature data of the cutting trajectory into the feature grading network in the hierarchical cutting parameter prediction model for feature grading to obtain a local feature layer, a regional feature layer, and a global feature layer;

[0096] Input the local feature layer into the first prediction head in the hierarchical cutting parameter prediction model for feature processing to obtain local cutting parameter prediction data. The first prediction head includes 4 convolutional blocks and 1 fully connected layer, and each convolutional block is provided with a spatial pyramid pooling module;

[0097] Input the regional feature layer into the second prediction head in the hierarchical cutting parameter prediction model for feature processing to obtain regional cutting parameter prediction data. The second prediction head includes 3 convolutional blocks and 1 fully connected layer, and each convolutional block is configured with a channel concatenation module;

[0098] Input the global feature layer into the third prediction head in the hierarchical cutting parameter prediction model for feature processing to obtain global cutting parameter prediction data. The third prediction head includes 2 convolutional blocks and 1 fully connected layer, and each convolutional block is provided with a feature recalibration module;

[0099] Input the local cutting parameter prediction data, the regional cutting parameter prediction data, and the global cutting parameter prediction data into the parameter fusion network for parameter fusion processing to obtain the cutting parameter prediction result.

[0100] Specifically, the input spatio-temporal feature data of the cutting trajectory is passed into the feature grading network for hierarchical processing. The core of the feature grading network is to decompose the spatio-temporal feature data into three levels: local, regional, and global through a series of convolutional operations and pooling operations, to obtain the local feature layer , the regional feature layer and the global feature layer . The feature grading process is represented by the following formula:

[0101] ;

[0102] ;

[0103] ;

[0104] Where It is a convolution operation used to extract local features; and are pooling operations for regional and global features respectively, using max pooling or average pooling to reduce the resolution and expand the receptive field. After completing the feature hierarchy, the local feature layer is input into the first prediction head in the hierarchical cutting parameter prediction model for processing. The first prediction head contains 4 convolutional blocks and 1 fully connected layer, and each convolutional block integrates a spatial pyramid pooling (SPP) module for extracting multi-scale information. For the th convolutional block, its output is expressed as:

[0105] ;

[0106] ;

[0107] where represents the pooling operation at different scales, and concat[] represents concatenating the features at each scale. After passing through the fully connected layer , the local cutting parameter prediction data is output:

[0108] ;

[0109] Meanwhile, the regional feature layer is input into the second prediction head, which contains 3 convolutional blocks and 1 fully connected layer, and each convolutional block is configured with a channel concatenation module. The channel concatenation module enhances the feature expression ability by constructing the dependency relationship between channels, and its formula is:

[0110] ;

[0111] ;

[0112] where is the activation function (such as Sigmoid), is the global average pooling operation, and are the weight and bias respectively. The regional features pass through the fully connected layer and output the regional cutting parameter prediction data:

[0113] ;

[0114] The global feature layer is passed to the third prediction head, which contains 2 convolutional blocks and 1 fully connected layer, and each convolutional block is configured with a feature recalibration module. The feature recalibration module performs dynamic weighting by adjusting the feature importance, and its process is:

[0115] ;

[0116] Among them, and are weight parameters obtained through learning, is a pooling operation. The global feature passes through a fully connected layer to output global cutting parameter prediction data:

[0117] ;

[0118] After completing the above steps, the local, regional, and global cutting parameter prediction data , and are input into a parameter fusion network for fusion processing. The parameter fusion network adopts a multi-layer perceptron structure and realizes the final prediction by weighted accumulation of features of each layer. The fusion process is as follows:

[0119] ;

[0120] Among them, is the fusion weight, which is obtained through optimization training. For example, assuming that the speed and depth of the cutting tool need to be optimized during the cutting process, the local feature reflects local overheating in a certain area, the regional feature captures the heat distribution relationship between this area and adjacent areas, and the global feature provides the trend information of overall heat diffusion.

[0121] In a specific embodiment, the process of executing step 400 may specifically include the following steps:

[0122] Construct N Gaussian components based on the cutting parameter prediction results. The N Gaussian components include a mean vector and a covariance matrix, where each Gaussian component corresponds to a local area of the cutting trajectory, and a Gaussian component parameter group is obtained;

[0123] Perform expectation-maximization iterative operation according to the Gaussian component parameter group. The expectation-maximization iterative operation includes calculating the posterior probability of each data point belonging to each Gaussian component and updating the parameters of the Gaussian component, and optimized Gaussian mixture parameters are obtained;

[0124] Input the optimized Gaussian mixture parameters into a radial basis kernel function for density calculation, and set the kernel function bandwidth parameter matrix to obtain an initial kernel density function. Perform cross-validation optimization on the initial kernel density function, perform grid search on the likelihood function values under different bandwidth parameters, and select the bandwidth parameter corresponding to the maximum likelihood function value to obtain the target kernel density function;

[0125] Perform probability compensation on the low-speed cutting area based on the target kernel density function. The probability compensation process includes identifying the low-speed cutting area, calculating the compensation weight, and updating the probability density value to obtain the compensated density function, and performing probability integral normalization on the compensated density function. Divide the probability density value by the overall density integral to obtain the probability density distribution data of the cutting trajectory;

[0126] Optimize the Gaussian process model according to the probability density distribution data of the cutting trajectory to obtain the optimized sequence of cutting parameters.

[0127] Specifically, perform local partitioning on the trajectory based on the prediction result of the cutting trajectory to construct Gaussian components, and each Gaussian component is used to describe the probability distribution characteristics of a local area in the cutting trajectory. Assume that the cutting trajectory data is , where represents the -dimensional feature of the th data point in the trajectory. The probability density function of the Gaussian component is expressed as:

[0128] ;

[0129] where, is the mean vector of the th Gaussian component, is the covariance matrix, represents its determinant, is the parameter group. Through local partitioning of the cutting trajectory, Gaussian components are combined to form a Gaussian mixture model (GMM). The overall probability density of the GMM is:

[0130] ;

[0131] where is the weight of the th Gaussian component, satisfying , is the parameter set of all Gaussian components. After constructing the Gaussian components, it is necessary to optimize their parameters. Introduce the expectation maximization (EM) algorithm, and maximize the log-likelihood function by iteratively updating :

[0132] ;

[0133] The EM algorithm is divided into two stages: the expectation (E) step and the maximization (M) step. In the E step, calculate the posterior probability (responsibility) that each data point belongs to the th Gaussian component:

[0134] ;

[0135] Among them represents the probability weight of the data point on the th Gaussian component. In the M step, according to update the model parameters:

[0136] ;

[0137] ;

[0138] By repeatedly iterating the above steps until the model converges, the optimized Gaussian mixture parameters are obtained. Input the optimized Gaussian mixture parameters into the radial basis kernel function (RBF) for density calculation. The form of RBF is:

[0139] ;

[0140] Among them is the bandwidth parameter of the kernel function, which is used to control the smoothness of density estimation. The initial kernel density function is obtained through the weighted combination of the kernel function and the Gaussian components:

[0141] ;

[0142] To optimize the bandwidth parameter , cross-validate the initial kernel density functions with different bandwidth values, and select the bandwidth parameter that maximizes the value of the log-likelihood function through grid search. The optimized density function is:

[0143] ;

[0144] Based on the target kernel density function, perform probability compensation on the low-speed cutting area. The low-speed cutting area is identified by the density threshold . When satisfies , this point is determined to be a low-speed area. In these areas, calculate the compensation weight:

[0145] ;

[0146] and update the probability density value:

[0147] ;

[0148] Subsequently, perform integral normalization on the compensated density function to ensure that the total density is 1:

[0149] ;

[0150] The probability density distribution data of the cutting trajectory is used for the optimization of the variable observation domain of the Gaussian process model. By adjusting the observation domain range of the Gaussian process, the kernel function and covariance function of the model are optimized according to the density values in the local area, and an optimized sequence of cutting parameters is generated.

[0151] In a specific embodiment, the process of performing the step of optimizing the variable observation domain of the Gaussian process model according to the probability density distribution data of the cutting trajectory to obtain an optimized sequence of cutting parameters may specifically include the following steps:

[0152] Input the probability density distribution data of the cutting trajectory into the observation domain partitioning module for adaptive partitioning of the cutting trajectory to obtain variable observation domain configuration data;

[0153] Construct a kernel function for the Gaussian process model according to the variable observation domain configuration data. The kernel function is a combined kernel function based on the Markov property, including a weighted combination of a periodic kernel function and a Gaussian kernel function, to obtain the kernel function parameters of the Gaussian process model;

[0154] Optimize the hyperparameters of the kernel function parameters of the Gaussian process model, and perform parameter search based on the physical constraints of the cutting process to obtain an optimized Gaussian process model;

[0155] Based on the optimized Gaussian process model, adopt an uncertainty-based active learning strategy to select sampling points, perform sparse sampling on the cutting parameter space, and obtain a target sampling point sequence;

[0156] Evaluate the cutting quality and score the sampling points based on the target sampling point sequence to obtain a cutting parameter scoring matrix, and input the cutting parameter scoring matrix into the sequence optimizer for temporal optimization of the cutting parameters, select the optimal parameter path, and obtain an optimized sequence of cutting parameters.

[0157] Specifically, input the probability density distribution data of the cutting trajectory into the observation domain partitioning module for adaptive partitioning of the cutting trajectory. The probability density distribution data of the cutting trajectory is expressed as , where is a spatial point in the trajectory. By setting a density threshold , the cutting trajectory is divided into a high-density area and a low-density area. Assume that the high-density area and the low-density area satisfy:

[0158] ;

[0159] To refine the observation domain partitioning, each area is sub-regionally partitioned based on a clustering algorithm (such as K-Means or DBSCAN), and variable observation domain configuration data is generated in combination with the temporal characteristics of the trajectory. , which is used to describe the range and characteristics of each sub-region. After the observation domain division is completed, according to the variable observation domain configuration data Construct the kernel function of the Gaussian process model. The kernel function is the core of the Gaussian process and defines the similarity between data points. To adapt to the dynamic changes of the cutting trajectory, the kernel function adopts a combined form based on the Markov property, including the periodic kernel function and the Gaussian kernel function weighted combination:

[0160] ;

[0161] where and are weight parameters that satisfy . The periodic kernel function is used to capture the periodic pattern of the cutting trajectory, and its form is:

[0162] ;

[0163] where is the amplitude parameter, is the length scale, is the period. The Gaussian kernel function is used to model the smooth characteristics of the local area in the trajectory:

[0164] ;

[0165] According to the variable observation domain configuration data, optimize the parameters of the kernel function for each sub-region separately to generate a set of region-adapted kernel function parameters. After obtaining the kernel function of the Gaussian process model, perform hyperparameter optimization on the parameters of the kernel function. The optimization goal is to maximize the log marginal likelihood function of the model:

[0166] ;

[0167] where is the observed value, is the input data, is the covariance matrix generated by the kernel function, is the hyperparameter of the kernel function. Search for the optimal parameter combination through gradient descent or Bayesian optimization, and at the same time combine the physical constraints of the cutting process (such as the limitations of cutting speed and depth) to ensure the physical rationality of the model and obtain an optimized Gaussian process model. Based on the optimized Gaussian process model, an active learning strategy based on uncertainty is adopted to select sampling points. The prediction uncertainty of the Gaussian process model is represented by the prediction mean and the prediction variance , where:

[0168] ;

[0169] ;

[0170] By sampling by selecting points with high prediction variance, the information gain of the model can be maximized to generate a target sampling point sequence . Evaluate the cutting quality based on the target sampling point sequence and assign a score to each sampling point. The scoring basis includes cutting accuracy, surface quality, and the size of the heat-affected area, and the comprehensive scoring matrix is expressed as:

[0171] ;

[0172] where are the scoring functions for accuracy, surface quality, and heat influence respectively, is the weight parameter. Input the scoring matrix into the sequence optimizer, and the sequence optimizer finds the optimal parameter path in the time series through dynamic programming or reinforcement learning algorithms to generate the final optimized sequence of cutting parameters. For example, during the cutting process, if the probability density of a certain area is low and the corresponding heat influence is large, the active learning strategy will select this area for more sampling point analysis. Through the Gaussian process model to optimize and evaluate these points, the sequence optimizer finally generates a parameter adjustment strategy, such as reducing the tool speed and depth, to ensure the improvement of cutting quality and the effective protection of materials.

[0173] In a specific embodiment, the process of executing step 500 may specifically include the following steps:

[0174] Input the optimized sequence of cutting parameters into the state prediction module. The state prediction module includes a system state equation and a measurement equation, constructs the state transition matrix and the observation matrix of the cutting parameters, and obtains the state space model of the cutting parameters;

[0175] Perform Kalman filtering calculation on the state space model of the cutting parameters to obtain the predicted state vector of the cutting parameters, and input the predicted state vector of the cutting parameters and the actual observation data into the Hungarian algorithm. The Hungarian algorithm module uses the optimal matching strategy of the bipartite graph to associate multi-objective states to obtain the target association matrix;

[0176] Perform measurement update on the target association matrix. The measurement update includes calculating the Kalman gain matrix, posterior state estimation, and posterior covariance matrix to obtain the optimal state sequence of the cutting parameters;

[0177] Input the optimal state sequence of cutting parameters into the parameter optimization controller to jointly optimize the cutting speed, feed rate, and cutting depth, obtain the parameter adjustment instruction, perform trajectory feedback compensation on the parameter adjustment instruction to obtain the compensation value, and superimpose the compensation value on the parameter adjustment instruction to obtain the control instruction for cutting parameters.

[0178] Specifically, establish a state prediction module. The core of the state prediction module is to construct the system state equation and measurement equation, which are used to describe the dynamic changes of cutting parameters and the relationship between observations and the true state. Assume that the cutting parameter optimization sequence is , where respectively represent the cutting speed, feed rate, and cutting depth at time . The system state vector is expressed as , where each state component represents the state value of the th target state at moment. The state transition equation describes the dynamic changes of the state over time and has the form:

[0179] ;

[0180] where is the state transition matrix, representing the time dependence of the system state, is the control input matrix, representing the influence of the optimization parameters on the state, is the process noise, is the process noise covariance matrix. The measurement equation defines the relationship between the observed value and the true state and has the form:

[0181] ;

[0182] where is the observation vector, is the observation matrix, representing the mapping relationship between the observed value and the state, is the observation noise, is the observation noise covariance matrix. By constructing the above state transition matrix and observation matrix , a state space model of cutting parameters is formed to describe the dynamic behavior of the system. Perform Kalman filtering calculation on the state space model to estimate the predicted state vector of cutting parameters. Kalman filtering includes two stages: prediction and update. In the prediction stage, predict the current state and covariance according to the state transition equation:

[0183] ;

[0184] ;

[0185] where is the predicted state vector, is the predicted covariance matrix. In the update stage, combining the observation value to perform state correction and calculate the Kalman gain matrix:

[0186] ;

[0187] Then update the state estimate and covariance:

[0188] ;

[0189] ;

[0190] where is the Kalman gain matrix, representing the correction weight of the observation value to the state estimate. Input the predicted state vector and the observed data into the Hungarian algorithm module, and use the optimal matching strategy of the bipartite graph to associate the multi-target states to generate the target association matrix. Assume the matching cost matrix is , and its value represents the cost between the predicted state and the observation value . The Hungarian algorithm minimizes the total cost:

[0191] ;

[0192] where is an element of the matching matrix (taking values of 0 or 1), indicating whether and are matched. After generating the target association matrix, perform measurement update to obtain the optimal state sequence . Input the optimal state sequence into the parameter optimization controller to jointly optimize the cutting speed, feed rate, and cutting depth. The optimization controller uses a multi-objective optimization algorithm to find the optimal solution of the control parameters, and the objective function is:

[0193] ;

[0194] where is the error of the cutting speed, feed rate, and depth, is the weight. Perform trajectory feedback compensation on the parameter adjustment instruction, and calculate the compensation value by comparing the current trajectory and the target trajectory:

[0195] ;

[0196] where is the feedback gain. The compensation value is superimposed on the parameter adjustment instruction to generate the final control instruction:

[0197] ;

[0198] For example, when the cutting trajectory shows abnormal speed, the change trend is quickly predicted through Kalman filtering. The Hungarian algorithm correlates and corrects the actual observed value with the predicted value, adjusts the speed and depth parameters in the optimization controller, and combines the trajectory feedback compensation to correct the instruction to ensure the cutting quality and stability. The generated control instruction can dynamically adapt to complex cutting conditions and achieve intelligent real-time control.

[0199] The intelligent detection method for the corrosion-resistant aluminum alloy parts in the embodiments of the present application is described above. Next, the intelligent detection device 10 for the corrosion-resistant aluminum alloy parts in the embodiments of the present application will be described. Please refer to Figure 2 , an embodiment of the intelligent detection device 10 for the corrosion-resistant aluminum alloy parts in the embodiments of the present application includes:

[0200] A feature extraction module 11, configured to extract features from the image sequence of the corrosion-resistant aluminum alloy cutting process to obtain a multi-dimensional feature map;

[0201] A spatial mapping module 12, configured to perform three-dimensional spatial mapping and morphological feature analysis on the multi-dimensional feature map to obtain the spatio-temporal feature data of the cutting trajectory;

[0202] A parameter prediction module 13, configured to input the spatio-temporal feature data of the cutting trajectory into a hierarchical cutting parameter prediction model for hierarchical feature parameter prediction to obtain a cutting parameter prediction result;

[0203] An optimization module 14, configured to perform Gaussian mixture modeling and kernel density analysis on the cutting parameter prediction result to obtain the probability density distribution data of the cutting trajectory, and perform variable observation domain optimization on the Gaussian process model to obtain an optimized sequence of cutting parameters;

[0204] A state tracking module 15, configured to input the optimized sequence of cutting parameters into the Kalman filtering-Hungarian algorithm for state tracking to obtain a control instruction for the cutting parameters.

[0205] Through the collaborative cooperation of the above-mentioned various components, and through the design of the multi-scale cascaded GhostNet network and the spatial positioning and feature generalization enhancement module, the feature extraction ability for complex scenarios during the cutting process is significantly enhanced, greatly improving the detection accuracy of the cutting trajectory; the hierarchical cutting parameter prediction model is adopted for hierarchical feature fusion, combined with the design of a multi-level prediction head, to achieve accurate prediction of the cutting parameters and effectively reduce the error of parameter prediction; the Gaussian mixture modeling and kernel density analysis methods are introduced, and through the variable observation domain adaptive optimization strategy, the performance degradation problem of traditional methods in low-speed cutting scenarios is solved, ensuring the stability of cutting parameter optimization; a state tracking framework based on the Kalman filter-Hungarian algorithm is designed, combined with a parameter optimization controller and a trajectory feedback compensation mechanism, to achieve real-time and accurate control of the cutting parameters, significantly improving the cutting quality.

[0206] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated here.

[0207] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing an electronic device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0208] As mentioned above, the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of various embodiments of this application.

Claims

1. An intelligent detection method for corrosion-resistant aluminum alloy parts, characterized in that: The method comprises: Feature extraction is performed on the image sequence of the anti-corrosion aluminum alloy cutting process to obtain a multi-dimensional feature map; Performing three-dimensional spatial mapping and morphological feature analysis on the multi-dimensional feature graph to obtain spatiotemporal feature data of the cutting trajectory; Inputting the spatiotemporal characteristic data of the cutting trajectory into a hierarchical cutting parameter prediction model to perform hierarchical characteristic parameter prediction to obtain a cutting parameter prediction result; Gaussian mixture modeling and kernel density analysis are performed on the cutting parameter prediction results to obtain probability density distribution data of the cutting trajectory, and the Gaussian process model is optimized in a variable observation domain to obtain an optimized sequence of cutting parameters; the optimized sequence of cutting parameters obtained includes: inputting the probability density distribution data of the cutting trajectory into an observation domain partitioning module to perform adaptive partitioning of the cutting trajectory to obtain variable observation domain configuration data; constructing a kernel function for the Gaussian process model according to the variable observation domain configuration data, wherein the kernel function is a combined kernel function based on Markov properties, including a weighted combination of a periodic kernel function and a Gaussian kernel function, to obtain the Gaussian process model. Kernel function parameters; hyperparameter optimization is performed on the kernel function parameters of the Gaussian process model, and parameter search is performed based on the physical constraints of the cutting process to obtain an optimized Gaussian process model; based on the optimized Gaussian process model, sampling points are selected using an uncertainty-based active learning strategy, and sparse sampling is performed on the cutting parameter space to obtain a target sampling point sequence; based on the target sampling point sequence, the cutting quality is evaluated and the sampling points are scored to obtain a cutting parameter scoring matrix, and the cutting parameter scoring matrix is ​​input into a sequence optimizer to perform cutting parameter timing optimization, select the optimal parameter path, and obtain an optimized sequence of cutting parameters; The optimized sequence of cutting parameters is input into the Kalman filter-Hungarian algorithm for state tracking to obtain the control instructions of the cutting parameters.

2. The intelligent detection method for corrosion-resistant aluminum alloy parts according to claim 1 is characterized in that: The image sequence of the corrosion-resistant aluminum alloy cutting process is subjected to feature extraction to obtain a multi-dimensional feature map, including: Perform multi-scale pyramid decomposition on the image sequence of the anti-corrosion aluminum alloy cutting process to obtain high-resolution feature layer, medium-resolution feature layer and low-resolution feature layer; Inputting the high-resolution feature layer into the first-level GhostNet module for feature extraction, the first-level GhostNet module comprises 8 Ghost bottleneck units, each Ghost bottleneck unit is composed of a main branch convolution layer and a parallel branch transformation layer, and obtaining a first-level feature map; Inputting the medium-resolution feature layer into the second-level GhostNet module for feature extraction, the second-level GhostNet module includes 6 Ghost bottleneck units, each Ghost bottleneck unit is set with a channel attention mechanism and a spatial attention mechanism, and a second-level feature map is obtained; Inputting the low-resolution feature layer into a third-level GhostNet module for feature extraction, wherein the third-level GhostNet module comprises four Ghost bottleneck units, each of which is configured with a cross-layer residual connection and an adaptive feature aggregation unit to obtain a third-level feature map; The first-level feature map, the second-level feature map and the third-level feature map are subjected to feature fusion to obtain a fused feature map, and the fused feature map is input into a feature enhancement module for processing. The feature enhancement module comprises three depth-wise separable convolutional layers and one global context encoder. Each depth-wise separable convolutional layer is followed by a batch normalization layer and a PReLU activation function to obtain a multidimensional feature map.

3. The intelligent detection method for corrosion-resistant aluminum alloy parts according to claim 1 is characterized in that: The performing three-dimensional spatial mapping and morphological feature analysis on the multi-dimensional feature graph to obtain spatiotemporal feature data of the cutting trajectory includes: Performing spatial coordinate transformation on the multi-dimensional feature map, converting the two-dimensional image features into position information in a three-dimensional spatial coordinate system, and obtaining a spatial mapping matrix of the cutting trajectory; According to the spatial mapping matrix, a geometric feature extraction operator of the cutting trajectory is constructed, wherein the geometric feature extraction operator includes a directional gradient operator, a curvature operator and a continuity operator, and a geometric feature analysis is performed on the cutting trajectory to obtain a geometric feature vector of the cutting trajectory; Performing morphological operations on the geometric feature vectors to obtain morphological feature data of the cutting trajectory, and performing time series feature analysis on the morphological feature data to obtain a time series feature sequence of the cutting trajectory; A spatiotemporal feature association model is established based on the temporal feature sequence of the cutting trajectory, and the spatial position and time evolution of the cutting trajectory are modeled to obtain the spatiotemporal association features of the cutting trajectory; The spatial mapping matrix of the cutting trajectory, the morphological feature data of the cutting trajectory and the spatiotemporal correlation features of the cutting trajectory are combined to obtain the spatiotemporal feature data of the cutting trajectory.

4. The intelligent detection method for corrosion-resistant aluminum alloy parts according to claim 1, characterized in that: The step of inputting the spatiotemporal characteristic data of the cutting trajectory into a hierarchical cutting parameter prediction model to perform hierarchical characteristic parameter prediction to obtain a cutting parameter prediction result includes: Inputting the spatiotemporal feature data of the cutting trajectory into a feature classification network in a hierarchical cutting parameter prediction model for feature classification to obtain a local feature layer, a regional feature layer and a global feature layer; Inputting the local feature layer into the first prediction head in the hierarchical cutting parameter prediction model for feature processing to obtain local cutting parameter prediction data, wherein the first prediction head comprises 4 convolution blocks and 1 fully connected layer, and each convolution block is provided with a spatial pyramid pooling module; Inputting the regional feature layer into the second prediction head in the hierarchical cutting parameter prediction model for feature processing to obtain regional cutting parameter prediction data, wherein the second prediction head comprises 3 convolution blocks and 1 fully connected layer, and each convolution block is configured with a channel cascade module; Inputting the global feature layer into the third prediction head in the hierarchical cutting parameter prediction model for feature processing to obtain global cutting parameter prediction data, wherein the third prediction head comprises 2 convolution blocks and 1 fully connected layer, and each convolution block is provided with a feature recalibration module; The local cutting parameter prediction data, the regional cutting parameter prediction data and the global cutting parameter prediction data are input into a parameter fusion network for parameter fusion processing to obtain a cutting parameter prediction result.

5. The intelligent detection method for corrosion-resistant aluminum alloy parts according to claim 1, characterized in that: The cutting parameter prediction results are subjected to Gaussian mixture modeling and kernel density analysis to obtain probability density distribution data of the cutting trajectory, and the Gaussian process model is subjected to variable observation domain optimization to obtain an optimization sequence of cutting parameters, including: Based on the cutting parameter prediction result, N Gaussian components are constructed, wherein the N Gaussian components include a mean vector and a covariance matrix, wherein each Gaussian component corresponds to a local area of ​​the cutting trajectory, and a Gaussian component parameter group is obtained; Performing an expectation maximization iterative operation according to the Gaussian component parameter group, wherein the expectation maximization iterative operation includes calculating the posterior probability that each data point belongs to each Gaussian component and updating the parameters of the Gaussian components to obtain optimized Gaussian mixture parameters; The optimized Gaussian mixture parameters are input into the radial basis kernel function for density calculation, and the kernel function bandwidth parameter matrix is ​​set to obtain an initial kernel density function, the initial kernel density function is cross-validated and optimized, a grid search is performed on the likelihood function values ​​under different bandwidth parameters, and the bandwidth parameter corresponding to the maximized likelihood function value is selected to obtain a target kernel density function; Based on the target kernel density function, probability compensation is performed on the low-speed cutting area. The probability compensation process includes identifying the low-speed cutting area, calculating the compensation weight and updating the probability density value to obtain the compensation density function, and performing probability integral normalization processing on the compensation density function. The probability density value is divided by the overall density integral to obtain the probability density distribution data of the cutting trajectory; The Gaussian process model is optimized in a variable observation domain according to the probability density distribution data of the cutting trajectory to obtain an optimized sequence of cutting parameters.

6. The intelligent detection method for corrosion-resistant aluminum alloy parts according to claim 1, characterized in that: The optimized sequence of cutting parameters is input into the Kalman filter-Hungarian algorithm for state tracking to obtain the control instructions of the cutting parameters, including: Inputting the optimized sequence of cutting parameters into a state prediction module, wherein the state prediction module includes a system state equation and a measurement equation, constructing a state transfer matrix and an observation matrix of the cutting parameters, and obtaining a state space model of the cutting parameters; Performing Kalman filter calculation on the state space model of the cutting parameters to obtain a predicted state vector of the cutting parameters, and inputting the predicted state vector of the cutting parameters and actual observation data into a Hungarian algorithm. The Hungarian algorithm module uses a bipartite graph optimal matching strategy to associate multiple target states to obtain a target association matrix; Performing measurement update on the target association matrix, wherein the measurement update includes calculating a Kalman gain matrix, a posterior state estimate, and a posterior covariance matrix to obtain an optimal state sequence of cutting parameters; The optimal state sequence of the cutting parameters is input into the parameter optimization controller, the cutting speed, feed rate and cutting depth are jointly optimized to obtain parameter adjustment instructions, and trajectory feedback compensation is performed on the parameter adjustment instructions to obtain compensation values, and the compensation values ​​are superimposed on the parameter adjustment instructions to obtain control instructions for the cutting parameters.

7. An intelligent detection device for corrosion-resistant aluminum alloy parts, characterized in that: The intelligent detection method for the corrosion-resistant aluminum alloy part according to claim 1 is used to perform the intelligent detection device for the corrosion-resistant aluminum alloy part, comprising: A feature extraction module is used to extract features from the image sequence of the corrosion-resistant aluminum alloy cutting process to obtain a multi-dimensional feature map; A spatial mapping module, used for performing three-dimensional spatial mapping and morphological feature analysis on the multi-dimensional feature map to obtain spatiotemporal feature data of the cutting trajectory; A parameter prediction module, used for inputting the spatiotemporal characteristic data of the cutting trajectory into a hierarchical cutting parameter prediction model to perform hierarchical characteristic parameter prediction to obtain a cutting parameter prediction result; The optimization module is used to perform Gaussian mixture modeling and kernel density analysis on the cutting parameter prediction results to obtain the probability density distribution data of the cutting trajectory, and perform variable observation domain optimization on the Gaussian process model to obtain the optimization sequence of cutting parameters; the optimization sequence of cutting parameters obtained includes: inputting the probability density distribution data of the cutting trajectory into the observation domain partitioning module to perform adaptive partitioning of the cutting trajectory to obtain variable observation domain configuration data; constructing a kernel function for the Gaussian process model according to the variable observation domain configuration data, wherein the kernel function is a combined kernel function based on Markov properties, including a weighted combination of a periodic kernel function and a Gaussian kernel function to obtain a Gaussian process The kernel function parameters of the model are obtained by performing hyperparameter optimization on the kernel function parameters of the Gaussian process model, and performing parameter search based on the physical constraints of the cutting process to obtain an optimized Gaussian process model; based on the optimized Gaussian process model, sampling points are selected using an active learning strategy based on uncertainty, and sparse sampling is performed on the cutting parameter space to obtain a target sampling point sequence; based on the target sampling point sequence, the cutting quality is evaluated and the sampling points are scored to obtain a cutting parameter scoring matrix, and the cutting parameter scoring matrix is ​​input into a sequence optimizer to perform cutting parameter timing optimization, select an optimal parameter path, and obtain an optimized sequence of cutting parameters; The state tracking module is used to input the optimized sequence of cutting parameters into the Kalman filter-Hungarian algorithm for state tracking to obtain the control instructions of the cutting parameters.

Images