High-precision linear guide positioning system and method based on artificial intelligence
Through the multi-sensor integrated system and deep learning technology based on artificial intelligence, the problem of unstable accuracy in harsh environments of traditional linear guide rail positioning methods is solved, and high-precision and high-reliability linear guide rail positioning is achieved to adapt to different working conditions and environments condition.
Patent Information
- Application Number
- CN202411246406.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-09-06
AI Technical Summary
The traditional linear guide rail positioning method is prone to interference in harsh environments, and has high cost, making it difficult to meet the needs of high precision and high reliability.
Adopt a high-precision linear guide rail positioning system based on artificial intelligence, collect multimodal data through a multi-sensor integrated system, perform standardized processing and data enhancement, build a dual-condition generation adversarial network, optimize it with joint loss function, and use deep Q network for intelligence Decision-making and closed-loop optimization to achieve real-time positioning and dynamic error compensation.
It improves the accuracy and reliability of linear guide rail positioning, adapts to different working conditions and environmental conditions, realizes the adaptability and continuous improvement of the positioning system, and ensures performance stability during long-term use.
Smart Images

Figure CN118759862B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of linear guide positioning technology, and in particular to a high-precision linear guide positioning system and method based on artificial intelligence. Background Art
[0002] As a key component in precision mechanical equipment, the positioning accuracy of linear guides directly affects the performance and processing quality of the entire system. With the continuous improvement of precision requirements in modern manufacturing, traditional linear guide positioning methods have been unable to meet the needs of high precision and high reliability. The commonly used positioning methods such as grating scales and magnetic scales are easily disturbed in harsh environments, and the cost is high, making them difficult to be widely used. Summary of the invention
[0003] The present application provides a high-precision linear guide positioning system and method based on artificial intelligence, which is used to achieve high-precision positioning of the linear guide.
[0004] In a first aspect, the present application provides a high-precision linear guide positioning method based on artificial intelligence, and the high-precision linear guide positioning method based on artificial intelligence includes:
[0005] Perform multi-modal data acquisition and preprocessing on the linear guide to obtain a standardized training data set;
[0006] A dual-conditional generative adversarial network is constructed using the standardized training data set to obtain a multimodal deep learning model;
[0007] The multimodal deep learning model is optimized using a joint loss function to obtain optimized model parameters, and the optimized model parameters are iteratively trained and verified to obtain a first positioning prediction model;
[0008] Performing real-time positioning of the linear guide rail by using the first positioning prediction model to obtain real-time positioning data, and performing dynamic error compensation and reliability evaluation on the real-time positioning data to obtain corrected positioning data;
[0009] Intelligent decision-making and closed-loop optimization are performed on the corrected positioning data to obtain a high-precision positioning result, and the first positioning prediction model is updated according to the high-precision positioning result to obtain a second positioning prediction model.
[0010] In a second aspect, the present application provides a high-precision linear guide positioning system based on artificial intelligence, and the high-precision linear guide positioning system based on artificial intelligence includes:
[0011] An acquisition module is used to collect and preprocess multimodal data of the linear guide rail to obtain a standardized training data set;
[0012] A construction module, used to construct a dual-conditional generative adversarial network through the standardized training data set to obtain a multimodal deep learning model;
[0013] A training module, used to optimize the multimodal deep learning model using a joint loss function to obtain optimized model parameters, and iteratively train and verify the optimized model parameters to obtain a first positioning prediction model;
[0014] A positioning module, used for performing real-time positioning of the linear guide rail by using the first positioning prediction model to obtain real-time positioning data, and performing dynamic error compensation and reliability evaluation on the real-time positioning data to obtain corrected positioning data;
[0015] An updating module is used to perform intelligent decision-making and closed-loop optimization on the corrected positioning data to obtain a high-precision positioning result, and to update the first positioning prediction model according to the high-precision positioning result to obtain a second positioning prediction model.
[0016] In the technical solution provided by the present application, the multi-modal data of the linear guide is collected by a multi-sensor integrated system, and standardized processing and data enhancement are performed to improve the quality and diversity of the data. The position estimation and attitude estimation two conditional generative adversarial networks are combined in series to achieve high-precision estimation of the position and attitude of the linear guide, and the generation and discrimination capabilities of the model are improved through adversarial training. Combining adversarial loss, reconstruction loss, geometric consistency loss and multi-task loss, a comprehensive joint loss function is constructed, which effectively improves the training effect and generalization ability of the model. The positioning uncertainty is estimated by the random perturbation method, and filtering is performed to achieve dynamic error compensation of real-time positioning data, and improve positioning accuracy and reliability. The deep Q network is used for intelligent decision-making, and the model is continuously optimized through experience playback and temporal difference learning, so as to achieve the adaptive ability and continuous improvement of the positioning system. The dynamic update of the positioning prediction model is realized, which ensures the performance stability and adaptability of the model during long-term use. The method integrates a variety of advanced artificial intelligence technologies, has strong versatility and robustness, and can adapt to the positioning requirements of linear guides under different working conditions and environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0018] Figure 1 This is a schematic diagram of an embodiment of a high-precision linear guide positioning method based on artificial intelligence in an embodiment of the present application;
[0019] Figure 2 This is a schematic diagram of an embodiment of a high-precision linear guide positioning system based on artificial intelligence in an embodiment of the present application. DETAILED DESCRIPTION
[0020] The embodiments of the present application provide a high-precision linear guide positioning system and method based on artificial intelligence. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0021] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the high-precision linear guide positioning method based on artificial intelligence in the embodiment of the present application includes:
[0022] Step S101, performing multi-modal data collection and preprocessing on the linear guide rail to obtain a standardized training data set;
[0023] It is understandable that the execution subject of the present application can be a high-precision linear guide positioning system based on artificial intelligence, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.
[0024] Specifically, a multi-sensor integrated system is used to collect data from the linear guide under different working conditions. Various sensors such as displacement sensors, acceleration sensors, temperature sensors, etc. installed at various key parts of the guide are used to collect various data of the guide under different environments and operating conditions to obtain the original multimodal data set. The original multimodal data set is cleaned, and a preliminary cleaned data set is obtained by filtering, denoising, and missing value filling. According to the preliminary cleaned data set, different types of data are standardized to ensure the comparability and uniformity of the data. By converting data of different units and magnitudes into the same range, for example, by normalization or standardization methods, the values of the data are unified into a preset range. Data enhancement is performed on the standardized data set to enrich the diversity and robustness of the data set. Data enhancement can be achieved by adding noise, data perturbation, or generating new data samples to obtain an enhanced data set. The time series data in the enhanced data set is converted into the frequency domain, and the signal in the time domain is converted into the frequency domain by using methods such as Fourier transform or wavelet transform to extract frequency domain feature data. For the non-stationary signals contained in the enhanced data set, multi-scale analysis is performed. The signal is decomposed into different scales using multi-scale geometric analysis or wavelet packet transform and other techniques to extract its time-frequency feature data. The frequency domain feature data and time-frequency feature data are subjected to dimensionality reduction processing to reduce the data dimension and retain its main information. Dimensionality reduction methods such as principal component analysis and linear discriminant analysis are used to extract the dimensionality reduction feature data that best represents the original data. Nonlinear dimensionality reduction and visualization are performed based on the dimensionality reduction feature data. By using nonlinear dimensionality reduction techniques such as t-SNE and UMAP, the high-dimensional data is mapped into a two-dimensional space to obtain a two-dimensional feature distribution map. Cluster analysis is performed on the two-dimensional feature distribution map to identify the intrinsic structure and pattern of the data. Commonly used clustering algorithms include k-means and hierarchical clustering. Through these algorithms, data points with similar features are classified into one category to obtain an optimized feature data set. The optimized feature data set is divided into a training set, a validation set, and a test set according to a preset ratio to ensure the effectiveness of model training, validation, and testing, and obtain a standardized training data set.
[0025] Step S102: constructing a dual-condition generative adversarial network using a standardized training data set to obtain a multimodal deep learning model;
[0026] Specifically, the standardized training data set is divided into a position information data set and a posture information data set. According to the position information data set, a position estimation conditional generative adversarial network is constructed, which includes a first generator and a first discriminator. Among them, the first generator adopts a U-Net structure, including an encoder and a decoder. The encoder consists of 5 convolutional layers, each of which uses the LeakyReLU activation function and batch normalization to ensure the effectiveness and stability of feature extraction. The decoder consists of 5 deconvolutional layers, each of which uses the ReLU activation function and batch normalization, and is responsible for gradually restoring the spatial resolution and details of the image. At the same time, the first discriminator adopts a PatchGAN structure, which consists of 4 convolutional layers, each of which uses a normalization layer and a LeakyReLU activation function. This design can effectively discriminate the authenticity of the local area. The position information data set is input into the position estimation conditional generative adversarial network. Through adversarial training, the first generator generates a position estimation result, and the first discriminator discriminates and feedbacks the generated result, continuously optimizes the performance of the generator, and obtains the first position estimation result. According to the pose information dataset and the first position estimation result, a pose estimation conditional generative adversarial network is constructed, which includes a second generator and a second discriminator. The second generator adopts a ResNet structure enhanced by an attention mechanism, which contains 9 residual blocks. Each residual block consists of two convolutional layers and a short-circuit connection, and a self-attention layer is added between the residual blocks to enhance the ability of feature expression and capture global dependencies. The second discriminator adopts a multi-scale discriminator structure, which consists of 3 sub-discriminators of different scales. Each sub-discriminator contains 4 convolutional layers and a fully connected layer to discriminate information of different scales and improve the accuracy of pose estimation. The first position estimation result and the pose information dataset are input into the pose estimation conditional generative adversarial network. Through adversarial training, the second generator generates the pose estimation result. The second discriminator discriminates and feeds back the generated result to optimize the performance of the generator, and finally obtains the first pose estimation result. The position estimation conditional generative adversarial network and the pose estimation conditional generative adversarial network are combined in series, and a domain adaptation layer is added between the two to coordinate the data distribution of the two different information to form a complete multimodal deep learning model. The multimodal deep learning model can comprehensively process position information and posture information to achieve high-precision linear guide positioning.
[0027] Step S103: optimizing the multimodal deep learning model using a joint loss function to obtain optimized model parameters, and iteratively training and verifying the optimized model parameters to obtain a first positioning prediction model;
[0028] Specifically, the training output results of the multimodal deep learning model are obtained, and the adversarial loss of the training output results and the true label is calculated. Through adversarial training between the generator and the discriminator in the generative adversarial network, the generator can generate prediction results that are consistent with the real data distribution. According to the training output results and the true label, the weighted sum of L1 loss and perceptual loss is calculated to obtain the reconstruction loss. L1 loss is used to measure the absolute error between the prediction result and the true label, while perceptual loss extracts high-level features through a pre-trained neural network to evaluate the difference between the prediction result and the true label in the feature space. The weighted sum of the two can more comprehensively reflect the reconstruction quality. The spatial transformation network is used to perform geometric transformation on the first position estimation result and the first attitude estimation result of the multimodal deep learning model, and the consistency error before and after the transformation is calculated to obtain the geometric consistency loss. The spatial transformation network is used to perform transformation operations such as rotation and translation on the estimation results to ensure the consistency and accuracy of the position and attitude estimation results. The multi-task loss of the multimodal deep learning model in the position estimation task, attitude estimation task and reliability prediction task is calculated. The multi-task loss combines the loss values of different tasks and comprehensively considers the training effects of each task to ensure the balanced performance of the model on each task. The adversarial loss, reconstruction loss, geometric consistency loss and multi-task loss are weighted and combined, and the L2 regularization term and gradient penalty term are introduced to form a joint loss function. The L2 regularization term prevents the model from overfitting by penalizing the sum of squares of the model parameters, while the gradient penalty term is used to stabilize the training process of the generative adversarial network. The joint loss function is optimized by gradient descent, and the parameters of the multimodal deep learning model are updated to obtain the optimized model parameters. The gradient descent optimization gradually adjusts the model parameters through the back propagation algorithm to minimize the joint loss function, thereby improving the prediction accuracy of the model. According to the optimized model parameters, the multimodal deep learning model is fine-tuned and parameter optimized. Through further refined training, the good performance of the model on different data distributions and tasks is ensured. The adjusted model is iteratively trained and verified. Through multiple rounds of training and verification, the generalization ability and accuracy of the model are continuously improved to obtain the first positioning prediction model.
[0029] Step S104: positioning the linear guide rail in real time by using the first positioning prediction model to obtain real-time positioning data, and performing dynamic error compensation and reliability evaluation on the real-time positioning data to obtain corrected positioning data;
[0030] Specifically, real-time multimodal data of the linear guide rail is collected. These data are collected under different working conditions through multiple sensors such as displacement sensors, acceleration sensors and temperature sensors to ensure real-time and diversity. The collected real-time multimodal data is preprocessed, including data cleaning, normalization and denoising steps, to generate real-time input data. The preprocessing step can eliminate noise and outliers in the data and ensure the quality of the input data. The real-time input data is input into the first generator of the position estimation condition generation adversarial network of the first positioning prediction model. The generator adopts a U-Net structure, which is composed of 5 convolutional layers and 5 deconvolutional layers. Through the layer-by-layer processing and feature extraction of these layers, the second position estimation result is generated. The second position estimation result is input into the first discriminator, which adopts a PatchGAN structure with 4 convolutional layers. Through the processing of these convolutional layers, the authenticity of the position estimation result is evaluated to generate the discriminant result of the position estimation. The evaluation process helps the generator optimize the output result through the feedback of the discriminator to make it closer to the real data. The second position estimation result and the real-time input data are input into the second generator of the attitude estimation conditional generative adversarial network of the first positioning prediction model. The generator adopts the ResNet structure enhanced by the attention mechanism, which consists of 9 residual blocks and self-attention layers. Through the layer-by-layer processing of these layers, the second attitude estimation result is generated. This attitude estimation result is input into the second discriminator, which contains 3 sub-discriminators of different scales. The sub-discriminators perform multi-scale authenticity evaluation on the attitude estimation result to generate the discrimination result of the attitude estimation. Multi-scale evaluation can capture the features at different scales and improve the accuracy of attitude estimation. According to the discrimination results of the position estimation and the discrimination results of the attitude estimation, the second position estimation result and the second attitude estimation result are combined, and the real-time positioning data is generated through the processing of the domain adaptation layer. The domain adaptation layer ensures the consistency and coordination of the position and attitude data by adjusting the differences between different data distributions. The random perturbation method is used to perform multiple small random perturbations on the real-time positioning data, and the positioning uncertainty is estimated by calculating the changes in the output results. According to the positioning uncertainty estimation, the real-time positioning data is filtered to eliminate the errors caused by noise and disturbance and obtain the corrected positioning data.
[0031] The second position estimation result is preprocessed and converted into a format suitable for the first discriminator input to obtain the discriminator input data, ensuring that the input data meets the requirements of the discriminator in terms of size and format, so that the subsequent convolution operation can be carried out smoothly. The discriminator input data is input into the first convolution layer of the first discriminator, and convolution operation is performed through 64 3x3 convolution kernels, and the LeakyReLU activation function is applied to obtain the first feature map. The LeakyReLU activation function ensures the capture of nonlinear features while avoiding the "death" of neurons, thereby improving the expression ability of the model. The first feature map is instance normalized to obtain a normalized first feature map. By standardizing the data in the feature map, the differences between batches are eliminated, ensuring the stability and convergence of the training process. The normalized first feature map is input into the second convolution layer of the first discriminator, and convolution operation is performed through 128 3x3 convolution kernels, and the LeakyReLU activation function is applied to obtain the second feature map. The second feature map is instance normalized to obtain a normalized second feature map, which is then fed into the third convolutional layer of the first discriminator. The third feature map is obtained by convolution operation with 256 3x3 convolution kernels and LeakyReLU activation function is applied. The generalization ability of the model is enhanced by standardizing the feature map. The normalized third feature map is fed into the fourth convolutional layer of the first discriminator, convolution operation is performed with 512 3x3 convolution kernels and LeakyReLU activation function is applied to obtain the fourth feature map. The layer-by-layer convolution operation enables the model to gradually extract complex features in the input data while ensuring the effective capture of features at different levels. The fourth feature map is subjected to global average pooling operation to compress it into a one-dimensional vector to obtain a discriminant feature vector. Global average pooling generates a fixed-length vector by averaging all values in the feature map, thereby reducing the number of parameters, preventing overfitting, and retaining key information. The discriminant feature vector is fed into the fully connected layer, and the output is mapped to between 0 and 1 through the Sigmoid activation function to obtain the discriminant result of the position estimation. The Sigmoid activation function can compress the output value into a standard probability range, making the output result easy to interpret and analyze.
[0032] The second position estimation result and the real-time input data are feature fused, and the channels are aligned through a 1x1 convolution layer to obtain a fused feature map, ensuring the consistency of data from different sources in the channel dimension. The fused feature map is input into the first convolution layer of the second generator, downsampled through 64 7x7 convolution kernels, and instance normalization and ReLU activation functions are applied to obtain the initial feature map. Instance normalization is used to standardize the data in the feature map to ensure that the distribution of each batch of data is consistent, thereby stabilizing the training process. The initial feature map is downsampled twice, each time using 128 3x3 convolution kernels with a step size of 2, and instance normalization and ReLU activation functions are applied to obtain the downsampled feature map. Downsampling convolution improves the efficiency and hierarchy of feature extraction by reducing the spatial size of the feature map. The downsampled feature map is input into the first residual block, processed through two 3x3 convolution layers and short-circuit connections, and instance normalization and ReLU activation functions are applied to obtain the output of the first residual block. The design of residual blocks can effectively solve the gradient vanishing problem in deep networks. The short-circuit connection (i.e., jump connection) makes the gradient transfer of the previous and next layers smoother. The output of the first residual block is continuously input into the second residual block to the eighth residual block. The structure of each residual block is the same as that of the first residual block, and the output of the eighth residual block is obtained. The design of stacking multiple residual blocks can ensure the stability and effectiveness of feature extraction while maintaining the depth of the network. The output of the eighth residual block is input into the self-attention layer to calculate the global correlation of the feature map to obtain the attention weighted feature map. The self-attention layer generates a weighted feature map by calculating the relationship between each position in the feature map and other positions, thereby enhancing the network's ability to capture global dependencies. The attention weighted feature map is input into the ninth residual block for the last residual processing to obtain the final residual output. The final residual output is subjected to two upsampling deconvolutions, each using 64 3x3 deconvolution kernels with a step size of 2, and instance normalization and ReLU activation functions are applied to obtain the upsampled feature map. Upsampling deconvolution gradually restores the resolution of the original input by increasing the spatial size of the feature map. The upsampled feature map is input into the last convolution layer of the second generator, processed by three 7x7 convolution kernels to obtain the output feature map of the original size, and the tanh activation function is operated on the output feature map to obtain the second pose estimation result. The tanh activation function limits the value range of the output feature map to between -1 and 1, which helps to stabilize the output result and adapt to subsequent processing steps.
[0033] The second pose estimation result is downsampled three times to generate input data of different scales, including original size input data, 1 / 2 size input data, and 1 / 4 size input data. The original size input data is input into the first convolutional layer of the first sub-discriminator, which is convolved by 64 4x4 convolution kernels with a step size of 2 and LeakyReLU activation function is applied to generate the first feature map of the first sub-discriminator. The LeakyReLU activation function helps to introduce nonlinear features while preventing the problem of "dead neurons". The first feature map is continuously input into the second to fourth convolutional layers, and each convolutional layer uses 128, 256 and 512 4x4 convolution kernels with a step size of 2, and instance normalization and LeakyReLU activation function are applied to gradually extract deep features to obtain the final feature map of the first sub-discriminator. Instance normalization ensures the consistency and stability of feature maps in different batches. At the same time, the 1 / 2 size input data is input to the second sub-discriminator for processing. The structure of the second sub-discriminator is the same as that of the first sub-discriminator. Through similar convolution operations and activation functions, the final feature map of the second sub-discriminator is generated. Similarly, the 1 / 4 size input data is input to the third sub-discriminator for processing. The structure of the third sub-discriminator is also the same as that of the first sub-discriminator. Through the same number of convolution kernels and step size settings, the final feature map of the third sub-discriminator is obtained. The multi-scale processing method can capture the details and global information of the features at different resolutions, and improve the discriminative ability and robustness of the discriminator. The final feature maps of the first sub-discriminator, the second sub-discriminator, and the third sub-discriminator are respectively subjected to global average pooling to obtain three discriminative feature vectors. Global average pooling compresses the feature map into a one-dimensional vector by taking the average of all values in the feature map, thereby reducing the number of parameters and computational complexity while retaining global features. The three discriminative feature vectors are respectively input to the corresponding fully connected layers, and the output is mapped to between 0 and 1 through the Sigmoid activation function to obtain the discrimination results of three scales. The Sigmoid activation function limits the output results to a standard probability range, making them easy to interpret and analyze. The discrimination results of these three scales are weighted averaged to obtain the final posture estimation discrimination result. The weighted average combines the discrimination results of different scales and comprehensively considers the weight and importance of features at each scale to provide a more accurate and reliable evaluation result. Through the multi-scale authenticity evaluation method, it is possible to effectively capture and analyze features at different scales, improve the reliability and accuracy of posture estimation results, and ensure the high-precision positioning requirements for linear guides in practical applications.
[0034] Step S105: perform intelligent decision-making and closed-loop optimization on the corrected positioning data to obtain a high-precision positioning result, and update the first positioning prediction model according to the high-precision positioning result to obtain a second positioning prediction model.
[0035] Specifically, feature extraction is performed on the corrected positioning data, and a comprehensive feature vector in the data is extracted using a deep neural network. The comprehensive feature vector can represent the key features and patterns of the positioning data, and is input into a deep Q network to calculate the Q values of various possible actions to obtain the action value evaluation results. The deep Q network approximates the action-state value through a neural network to evaluate the value of each possible action. The optimal action is selected according to the action value evaluation result to obtain an intelligent decision result, and the intelligent decision result is executed to obtain a decision execution result. The execution result can feedback whether the action taken by the system in the current state is effective. The reward value is calculated according to the decision execution result, and the reward value indicates the degree of superiority or inferiority of the selected action in the current state. The reward value, current state and next state are stored in the experience replay pool to obtain updated experience data. The experience replay pool avoids training bias caused by data correlation by storing and reusing past experience data. Experience data is randomly sampled from the experience replay pool, and the Q network parameters are updated using the temporal difference learning method to obtain an optimized Q network. Temporal difference learning gradually adjusts the network parameters by comparing the difference between the current Q value and the target Q value, so that the Q network can more accurately evaluate the long-term value of the action. The optimized Q network can process the corrected positioning data more efficiently and obtain more accurate high-precision positioning results. The high-precision positioning results are compared and analyzed with historical data to determine whether to trigger a model update. A model update trigger signal is generated based on the analysis results. If the model needs to be updated, the parameters of the first positioning prediction model are updated through the incremental learning method. Incremental learning ensures that the model adapts to new data distribution and features without forgetting existing knowledge by gradually accumulating gradients and dynamically adjusting the learning rate. The updated model parameters can better reflect the latest data characteristics and system status. The updated model parameters are applied to the first positioning prediction model, and the second positioning prediction model is obtained through model structure fine-tuning and knowledge distillation. Model structure fine-tuning improves the expression ability and adaptability of the model by adjusting the number of network layers, the number of nodes, and the connection method. Knowledge distillation ensures that the new model inherits the advantages of the old model while improving its performance and accuracy by migrating the knowledge of the old model to the new model.
[0036] In the embodiment of the present application, the multimodal data of the linear guide is collected by a multi-sensor integrated system, and standardized processing and data enhancement are performed to improve the quality and diversity of the data. The position estimation and attitude estimation two conditional generative adversarial networks are combined in series to achieve high-precision estimation of the position and attitude of the linear guide, and the generation and discrimination capabilities of the model are improved through adversarial training. Combining adversarial loss, reconstruction loss, geometric consistency loss and multi-task loss, a comprehensive joint loss function is constructed, which effectively improves the training effect and generalization ability of the model. The positioning uncertainty is estimated by the random perturbation method, and filtering is performed to achieve dynamic error compensation of real-time positioning data, and improve positioning accuracy and reliability. The deep Q network is used for intelligent decision-making, and the model is continuously optimized through experience playback and temporal difference learning, so as to achieve the adaptive ability and continuous improvement of the positioning system. The dynamic update of the positioning prediction model is realized, which ensures the performance stability and adaptability of the model during long-term use. The method integrates a variety of advanced artificial intelligence technologies, has strong versatility and robustness, and can adapt to the linear guide positioning requirements under different working conditions and environmental conditions.
[0037] In a specific embodiment, the process of executing step S101 may specifically include the following steps:
[0038] A multi-sensor integrated system is used to collect data of the linear guide under different working conditions to obtain an original multi-modal data set, and the original multi-modal data set is cleaned to obtain a preliminary cleaned data set;
[0039] According to the initially cleaned data set, different types of data are standardized and unified into the same range to obtain a standardized data set, and the standardized data set is enhanced to obtain an enhanced data set;
[0040] Perform frequency domain conversion on the time series data in the enhanced data set to obtain frequency domain feature data, and perform multi-scale analysis on the non-stationary signal in the enhanced data set to obtain time-frequency feature data;
[0041] Performing dimensionality reduction processing on the frequency domain feature data and the time-frequency feature data to obtain dimensionality reduction feature data, and performing nonlinear dimensionality reduction and visualization based on the dimensionality reduction feature data to obtain a two-dimensional feature distribution map;
[0042] Cluster analysis is performed on the two-dimensional feature distribution map to obtain an optimized feature data set, and the optimized feature data set is divided into a training set, a validation set, and a test set according to a preset ratio to obtain a standardized training data set.
[0043] Specifically, multiple sensors installed at different positions on the linear guide rail, such as acceleration sensors, displacement sensors, temperature sensors, and force sensors, can collect data under various working conditions. The data collected by these sensors under different operating conditions constitute the original multimodal data set. For example, in a typical industrial environment, the acceleration data of the guide rail , displacement data , temperature data Heli Data are recorded simultaneously. These data may contain many outliers and noisy data due to sensor noise, environmental interference, etc. The original multimodal data set is cleaned, and methods such as denoising filtering and outlier detection are used to remove and correct bad data to obtain a preliminary cleaned data set. According to the preliminary cleaned data set, different types of data are standardized to unify the data into the same range, eliminate the impact of different data types and dimensions, improve the training effect of the model, and obtain a standardized data set. Data enhancement is performed on the standardized data set. More training samples are generated by adding noise, data transformation (such as rotation, scaling), etc. to obtain an enhanced data set. The time series data in the enhanced data set is converted to the frequency domain to obtain frequency domain feature data. Frequency domain conversion is usually achieved through Fourier transform, and the formula is:
[0044] ;
[0045] in is the frequency domain signal, is the time domain signal, is the frequency, is an imaginary unit. Fourier transform can convert time domain signals to frequency domain and extract the frequency characteristics of the signal. For the non-stationary signals in the enhanced data set, multi-scale analysis is performed, usually using wavelet transform, the formula is:
[0046] ;
[0047] in are the wavelet coefficients, is the time domain signal, is the mother wavelet function, and They are scale and translation parameters respectively. Wavelet transform can analyze signals at different scales and time positions and extract time-frequency feature data. Dimensionality reduction processing is performed on frequency domain feature data and time-frequency feature data to obtain dimension-reduced feature data. Commonly used dimensionality reduction methods include principal component analysis and linear discriminant analysis. For example, principal component analysis projects the original data onto the principal component space through eigenvalue decomposition, and the formula is: ,in is the data after dimensionality reduction, is the original data matrix, is the principal component eigenvector matrix. Dimensionality reduction can reduce the dimension of data, retain the main information, and improve computational efficiency. Nonlinear dimensionality reduction and visualization based on dimensionality reduction feature data are usually performed using t-SNE or UMAP methods. t-SNE maps data to two-dimensional or three-dimensional space by minimizing the difference in probability distribution between high-dimensional data and low-dimensional embedded data to obtain a two-dimensional feature distribution map. Cluster analysis of two-dimensional feature distribution maps is performed, and commonly used methods include K-means clustering and hierarchical clustering. For example, Kmeans clustering distributes data points to In clusters, the similarity of data points in the cluster is maximized. The formula is:
[0048] ;
[0049] in It is Clusters, It is The centroid of each cluster. Through cluster analysis, data points with similar features are grouped into one category to obtain an optimized feature data set. The optimized feature data set is divided into training set, validation set, and test set according to the preset ratio to obtain a standardized training data set to ensure the effectiveness of model training, validation, and testing. The training set is used for model training, the validation set is used for model hyperparameter tuning and performance evaluation, and the test set is used for the final evaluation and verification of the model. Through these steps, a high-quality standardized training data set is constructed.
[0050] In a specific embodiment, the process of executing step S102 may specifically include the following steps:
[0051] The standardized training data set is divided into a position information data set and a posture information data set;
[0052] Constructing a position estimation conditional generative adversarial network according to the position information data set, wherein the position estimation conditional generative adversarial network includes a first generator and a first discriminator, wherein the first generator adopts a U-Net structure, and the first discriminator adopts a PatchGAN structure;
[0053] The first generator includes an encoder and a decoder. The encoder contains 5 convolutional layers, each of which uses the LeakyReLU activation function and batch normalization. The decoder contains 5 deconvolutional layers, each of which uses the ReLU activation function and batch normalization. The first discriminator includes 4 convolutional layers, each of which uses a normalization layer and a LeakyReLU activation function.
[0054] Inputting the position information data set into the position estimation conditional generative adversarial network, and obtaining a first position estimation result through adversarial training;
[0055] Constructing a posture estimation conditional generative adversarial network according to the posture information data set and the first position estimation result, the posture estimation conditional generative adversarial network includes a second generator and a second discriminator, the second generator adopts a ResNet structure enhanced by an attention mechanism, and the second discriminator adopts a multi-scale discriminator structure;
[0056] The second generator consists of 9 residual blocks, each of which contains two convolutional layers and a short-circuit connection, and a self-attention layer is added between the residual blocks; the second discriminator consists of 3 sub-discriminators of different scales, each of which contains 4 convolutional layers and a fully connected layer;
[0057] Inputting the first position estimation result and the posture information data set into a posture estimation conditional generative adversarial network, and obtaining a first posture estimation result through adversarial training;
[0058] The position estimation conditional generative adversarial network and the pose estimation conditional generative adversarial network are combined in series, and a domain adaptation layer is added between the position estimation conditional generative adversarial network and the pose estimation conditional generative adversarial network to obtain a multimodal deep learning model.
[0059] Specifically, information related to position, such as position signals and motion trajectories recorded by sensors, is extracted from the standardized training data set to form a position information data set. Similarly, information related to posture, such as angle changes and rotation matrices, is extracted to form a posture information data set. The separation of these two data sets enables us to focus on the characteristics of position and posture respectively when building the model, providing targeted processing and optimization. According to the position information data set, a position estimation conditional generative adversarial network is constructed, which includes a first generator and a first discriminator. The first generator adopts a U-Net structure, and captures multi-scale features in position information through the symmetrical design of the encoder and decoder. The encoder of the first generator contains 5 convolutional layers, each of which uses the LeakyReLU activation function and batch normalization to ensure the stability of the training process and accelerate convergence. The formula is expressed as:
[0060] ;
[0061] in is the output of the layer, is the input feature map, and are the convolution kernel and the bias term respectively. The decoder also contains 5 deconvolution layers, each of which uses the ReLU activation function and batch normalization to gradually restore the spatial resolution of the feature map and generate an output with the same size as the original input. The formula is expressed as:
[0062] ;
[0063] in The decoder is The output of the layer, and They are the deconvolution kernel and the bias term respectively. The first discriminator adopts the PatchGAN structure and improves the detail fidelity of the generated results through the local discrimination method. The discriminator contains 4 convolutional layers, each of which uses a normalization layer and a LeakyReLU activation function. The formula is expressed as:
[0064] ;
[0065] in It is the discriminator The output of the layer, and are the convolution kernel and the bias term respectively. The position information dataset is input into the position estimation conditional generative adversarial network, and the first position estimation result is obtained through adversarial training of the generator and the discriminator. The generator attempts to generate position information similar to the real data, while the discriminator evaluates the difference between the generated result and the real data. Through this adversarial process, the generation ability of the generator and the discrimination ability of the discriminator are continuously improved. According to the posture information dataset and the first position estimation result, a posture estimation conditional generative adversarial network is constructed. The network includes a second generator and a second discriminator. The second generator adopts a ResNet structure enhanced by the attention mechanism, and captures the global correlation in the posture information through residual blocks and self-attention layers. The second generator contains 9 residual blocks, each of which contains two convolutional layers and a short-circuit connection, and the formula is expressed as:
[0066] ;
[0067] in is the output of the residual block, is the input feature map, represents the nonlinear transformation through two convolutional layers, is the convolution kernel. A self-attention layer is added between the residual blocks to generate a weighted feature map by calculating the relationship between each position in the feature map and other positions. The formula is expressed as:
[0068] ;
[0069] in They are the query, key, and value matrices of the feature graph, respectively. The global dependency in the feature graph is enhanced through the self-attention mechanism. The second discriminator adopts a multi-scale discriminator structure, which contains 3 sub-discriminators of different scales. Each sub-discriminator contains 4 convolutional layers and a fully connected layer. The multi-scale discriminator can capture the details and global information of features at different scales and improve the discriminative ability of the discriminator. The formula is expressed as:
[0070] ;
[0071] in It is the sub-discriminator The output of the layer, and are the convolution kernel and the bias term respectively. The first position estimation result and the pose information data set are input into the pose estimation conditional generative adversarial network, and the first pose estimation result is obtained through adversarial training of the generator and the discriminator. The generator tries to generate an estimation result similar to the true pose, while the discriminator evaluates the difference between the generated result and the true data. Through this adversarial process, the generation ability of the generator and the discrimination ability of the discriminator are continuously improved. The position estimation conditional generative adversarial network and the pose estimation conditional generative adversarial network are combined in series, and a domain adaptation layer is added between the two to obtain a multimodal deep learning model. The domain adaptation layer adjusts the differences between different data distributions so that the position and pose estimation results can be coordinated and the overall performance and stability of the model are improved.
[0072] In a specific embodiment, the process of executing step S103 may specifically include the following steps:
[0073] Obtain the training output results of the multimodal deep learning model, and calculate the adversarial loss of the training output results and the true labels. Based on the training output results and the true labels, calculate the weighted sum of the L1 loss and the perceptual loss to obtain the reconstruction loss.
[0074] A spatial transformation network is used to geometrically transform the first position estimation result and the first posture estimation result of the multimodal deep learning model, and the consistency error before and after the transformation is calculated to obtain the geometric consistency loss.
[0075] Calculate the multi-task loss corresponding to the position estimation task, attitude estimation task, and reliability prediction task of the multimodal deep learning model;
[0076] The adversarial loss, reconstruction loss, geometric consistency loss and multi-task loss are weighted together, and the L2 regularization term and gradient penalty term are added to obtain the joint loss function;
[0077] Perform gradient descent optimization on the joint loss function, update the parameters of the multimodal deep learning model, and obtain the optimized model parameters;
[0078] The multimodal deep learning model is fine-tuned and parameter optimized according to the optimized model parameters to obtain an adjusted model, and the adjusted model is iteratively trained and verified to obtain a first positioning prediction model.
[0079] Specifically, we define the calculation method of adversarial loss. Adversarial loss is used to evaluate the difference between the results generated by the generator and the real data. The result generated is , discriminator Distinguish the generated results and real data. The adversarial loss can be expressed as:
[0080] ;
[0081] in, It is real data. is the noise vector, and Represent the probability density of data distribution and noise distribution respectively. After obtaining the adversarial loss, the weighted sum of L1 loss and perceptual loss is calculated based on the training output and the true label to obtain the reconstruction loss. L1 loss is used to measure the absolute error between the generated result and the true label, and the formula is:
[0082] ;
[0083] in, is the true label, is the generated result. The perceptual loss extracts high-level features through a pre-trained neural network and evaluates the difference between the generated result and the true label in the feature space. The formula is:
[0084] ;
[0085] in, represents the feature extraction network. The reconstruction loss can be expressed as the weighted sum of the two:
[0086] ;
[0087] in, and is a weight parameter. The spatial transformer network is used to geometrically transform the first position estimation result and the first posture estimation result of the multimodal deep learning model. The function of the spatial transformer network is to perform spatial transformation on the feature map, such as rotation, translation and scaling, so as to evaluate the consistency of the features before and after the transformation. The formula of the spatial transformer network is:
[0088] ;
[0089] in, represents the spatial transformation operation, is the transformation parameter, is the input feature map. Calculate the consistency error before and after the transformation to get the geometric consistency loss, the formula is:
[0090] ;
[0091] Calculate the multi-task loss corresponding to the position estimation task, attitude estimation task, and reliability prediction task of the multimodal deep learning model. The multi-task loss combines the loss values of different tasks and comprehensively considers the training effect of each task. The formula is:
[0092] ;
[0093] in, is the position estimation loss, is the pose estimation loss, is the reliability prediction loss, , and is a weight parameter. The adversarial loss, reconstruction loss, geometric consistency loss, and multi-task loss are weighted together, and the L2 regularization term and gradient penalty term are added to obtain the joint loss function. The L2 regularization term prevents the model from overfitting by penalizing the sum of the squares of the model parameters. The formula is:
[0094] ;
[0095] in, is a model parameter. The gradient penalty term is used to stabilize the training process of the generative adversarial network, and the formula is:
[0096] ;
[0097] in, is the interpolation of the data distribution and the generated distribution. The joint loss function can be expressed as:
[0098] ;
[0099] in, , , , and is the weight parameter. The joint loss function is optimized by gradient descent to update the parameters of the multimodal deep learning model. The formula is: ,in, are model parameters, is the learning rate, Represents the gradient of the model parameters. Through the back-propagation algorithm, the model parameters are gradually adjusted to minimize the joint loss function, thereby improving the prediction accuracy of the model. According to the optimized model parameters, the multimodal deep learning model is fine-tuned and parameter optimization is performed. Through refined training, the good performance of the model on different data distributions and tasks is ensured. The model is fine-tuned through the incremental learning method. Incremental learning ensures that the model adapts to new data distributions and features without forgetting existing knowledge by gradually accumulating gradients and dynamically adjusting the learning rate. The adjusted model is trained and verified multiple times to continuously improve the generalization ability and accuracy of the model, and obtain the first positioning prediction model.
[0100] In a specific embodiment, the process of executing step S104 may specifically include the following steps:
[0101] Collecting real-time multimodal data of the linear guide rail, and preprocessing the real-time multimodal data to obtain real-time input data;
[0102] Input the real-time input data into the first generator of the position estimation conditional generative adversarial network of the first positioning prediction model, and process it through a U-Net structure with 5 convolutional layers and 5 deconvolutional layers to obtain a second position estimation result;
[0103] The second position estimation result is input into the first discriminator, and the authenticity is evaluated through the PatchGAN structure of 4 convolutional layers to obtain the discrimination result of the position estimation;
[0104] According to the second position estimation result and the real-time input data, the second generator of the pose estimation conditional generative adversarial network of the first positioning prediction model is input, and the second pose estimation result is obtained through processing by 9 residual blocks and self-attention layers;
[0105] The second posture estimation result is input into the second discriminator, and multi-scale authenticity evaluation is performed through three sub-discriminators of different scales to obtain the discrimination result of the posture estimation;
[0106] Combining the second position estimation result and the second attitude estimation result according to the position estimation result and the attitude estimation result, and obtaining real-time positioning data through processing by a domain adaptation layer;
[0107] The random perturbation method is used to perform multiple tiny random perturbations on the real-time positioning data, and the changes in the output results are calculated to obtain the positioning uncertainty estimate. Based on the positioning uncertainty estimate, the real-time positioning data is filtered to obtain the corrected positioning data.
[0108] Specifically, multiple sensors installed at different positions of the linear guide rail, such as acceleration sensors, displacement sensors, temperature sensors, and force sensors, collect various types of data in the real-time environment of the guide rail operation. The data collected by these sensors under different working conditions include acceleration data , displacement data , temperature data Heli Data , these data form the original real-time multimodal data set. The real-time multimodal data is preprocessed, including data cleaning, normalization and denoising. Through filtering and outlier detection, noise and abnormal data are eliminated to ensure the quality of the input data. Normalization processing unifies different types of data into the same range to eliminate the impact of dimensional differences on model training. The preprocessed data forms real-time input data. The real-time input data is input into the first generator of the position estimation conditional generative adversarial network of the first positioning prediction model. The first generator adopts the U-Net structure, and captures the multi-scale features in the data through the symmetrical design of 5 convolutional layers and 5 deconvolutional layers. The encoder part of U-Net contains 5 convolutional layers, each of which uses the LeakyReLU activation function and batch normalization to ensure the stability of the training process and accelerate convergence. The formula is:
[0109] ;
[0110] in, It is The output of the layer, is the input feature map, and are the convolution kernel and the bias term respectively. The decoder part contains 5 deconvolution layers, each of which uses the ReLU activation function and batch normalization to gradually restore the spatial resolution of the feature map and generate an output with the same size as the original input. The formula is:
[0111] ;
[0112] in, The decoder is The output of the generator after processing is the second position estimation result. The second position estimation result is input into the first discriminator for authenticity evaluation. The first discriminator adopts the PatchGAN structure and performs local discrimination on the generated results through 4 convolutional layers to improve the detail fidelity of the generated results. Each convolutional layer uses a normalization layer and a LeakyReLU activation function, and the formula is:
[0113] ;
[0114] in, It is the discriminator The output of the discriminator is the discriminant result of the position estimation. According to the second position estimation result and the real-time input data, the second generator of the pose estimation conditional generative adversarial network of the first positioning prediction model is input. The second generator adopts the ResNet structure enhanced by the attention mechanism, and captures the global correlation and detail features in the data through 9 residual blocks and self-attention layers. Each residual block contains two convolutional layers and a short-circuit connection, and the formula is:
[0115] ;
[0116] in, is the output of the residual block, is the input feature map, represents the nonlinear transformation through two convolutional layers, is the convolution kernel. A self-attention layer is added between the residual blocks to generate a weighted feature map by calculating the relationship between each position in the feature map and other positions. The formula is:
[0117] ;
[0118] in, They are the query, key, and value matrices of the feature graph, respectively. The output of the second generator is the second pose estimation result. The second pose estimation result is input into the second discriminator for authenticity assessment. The second discriminator adopts a multi-scale discriminator structure, and performs multi-scale evaluation on the generated results through three sub-discriminators of different scales. Each sub-discriminator contains four convolutional layers and one fully connected layer, which improves the discriminant ability of the discriminator by capturing multi-scale features. The formula is:
[0119] ;
[0120] in, It is the sub-discriminator The output of the layer. According to the discrimination results of the position estimation and the discrimination results of the attitude estimation, the second position estimation result and the second attitude estimation result are combined, and the real-time positioning data is obtained through processing by the domain adaptation layer. The domain adaptation layer ensures the consistency and coordination of the position and attitude estimation results by adjusting the differences between different data distributions. The random perturbation method is used to perform multiple small random perturbations on the real-time positioning data, and the changes in the output results are calculated to obtain the positioning uncertainty estimate. By adding a small amount of random noise to the real-time positioning data and observing the changes in the output results, the formula is:
[0121] ;
[0122] in, is the uncertainty estimate, is the number of disturbances, It is The output result after perturbation is: is the mean of the output results. According to the positioning uncertainty estimation, the real-time positioning data is filtered to eliminate the errors caused by noise and disturbance, and the corrected positioning data is obtained.
[0123] In a specific embodiment, the execution step inputs the second position estimation result into the first discriminator, and the authenticity is evaluated through the PatchGAN structure of 4 convolutional layers. The process of obtaining the discrimination result of the position estimation can specifically include the following steps:
[0124] Preprocessing the second position estimation result and converting it into a format suitable for input of the first discriminator to obtain discriminator input data;
[0125] The discriminator input data is input into the first convolution layer of the first discriminator, convolved through 64 3x3 convolution kernels, and the LeakyReLU activation function is applied to obtain the first feature map;
[0126] Perform instance normalization on the first feature map to obtain a normalized first feature map, and input the normalized first feature map into the second convolution layer of the first discriminator, perform convolution operation through 128 3x3 convolution kernels, and apply the LeakyReLU activation function to obtain a second feature map;
[0127] Perform instance normalization on the second feature map to obtain a normalized second feature map, and input the normalized second feature map into the third convolution layer of the first discriminator, perform convolution operation through 256 3x3 convolution kernels, and apply the LeakyReLU activation function to obtain the third feature map;
[0128] Perform instance normalization on the third feature map to obtain a normalized third feature map, and input the normalized third feature map into the fourth convolution layer of the first discriminator, perform convolution operation through 512 3x3 convolution kernels, and apply the LeakyReLU activation function to obtain a fourth feature map;
[0129] A global average pooling operation is performed on the fourth feature map, and the fourth feature map is compressed into a one-dimensional vector to obtain a discriminant feature vector. The discriminant feature vector is input into the fully connected layer, and the output is mapped to between 0 and 1 through the Sigmoid activation function to obtain the discriminant result of the position estimation.
[0130] Specifically, the second position estimation result is preprocessed, including adjusting the image size, standardizing pixel values, and converting data types. The resolution of the image is adjusted by interpolation method to ensure that the input image size matches the input requirements of the discriminator. The pixel values of the image are normalized to between 0 and 1 to eliminate the influence of different data dimensions and ensure the numerical stability of the data. After completing the preprocessing operation, the discriminator input data is obtained. The preprocessed discriminator input data is input into the first convolution layer of the first discriminator for preliminary feature extraction. The first convolution layer uses 64 3x3 convolution kernels to extract local features of the input data through convolution operation. The convolution operation formula is:
[0131] ;
[0132] in, is the output feature map of the first convolutional layer, is the input data, is the convolution kernel, is the bias term. After the convolution operation, the LeakyReLU activation function is applied to introduce nonlinear characteristics. The formula is:
[0133] ;
[0134] in, is the activated feature map. The first feature map is instance normalized to eliminate differences between batches and ensure the stability of training. Instance normalization normalizes each feature map so that its mean is 0 and its standard deviation is 1. The normalized first feature map is input into the second convolution layer of the first discriminator to continue feature extraction. The second convolution layer uses 128 3x3 convolution kernels to further extract the deep features of the data through convolution operations. The formula is:
[0135] ;
[0136] in, is the output feature map of the second convolutional layer, is the convolution kernel, is the bias term. After the convolution operation, the LeakyReLU activation function is applied to obtain the second feature map: . Perform instance normalization on the second feature map, the formula is: The normalized second feature map is input into the third convolution layer, and 256 3x3 convolution kernels are used for convolution operation. The formula is:
[0137] ;
[0138] in, is the output feature map of the third convolutional layer, is the convolution kernel, is the bias term. After the convolution operation, the LeakyReLU activation function is applied to obtain the third feature map: . Perform instance normalization on the third feature map, the formula is: The normalized third feature map is input into the fourth convolution layer, and 512 3x3 convolution kernels are used for convolution operation. The formula is: ,in, is the output feature map of the fourth convolutional layer, is the convolution kernel, is the bias term. After the convolution operation, the LeakyReLU activation function is applied to obtain the fourth feature map: The fourth feature map is subjected to global average pooling to compress it into a one-dimensional vector. Global average pooling reduces the dimension of the feature map by taking the average value of each channel of the feature map while retaining the global information. The formula is:
[0139] ;
[0140] in, is a one-dimensional vector, and is the height and width of the feature map. The discriminant feature vector is input into the fully connected layer, and the output is mapped to between 0 and 1 through the Sigmoid activation function. The formula is:
[0141] ;
[0142] in, is the final judgment result. is the weight of the fully connected layer, is a bias term. The role of the Sigmoid activation function is to limit the output result between 0 and 1 to represent the authenticity probability of the generated result.
[0143] In a specific embodiment, the execution step inputs the second generator of the pose estimation conditional generative adversarial network of the first positioning prediction model according to the second position estimation result and the real-time input data, and the process of obtaining the second pose estimation result through the processing of 9 residual blocks and the self-attention layer can specifically include the following steps:
[0144] Perform feature fusion on the second position estimation result and the real-time input data, perform channel alignment through a 1x1 convolution layer, and obtain a fused feature map;
[0145] The fused feature map is input into the first convolution layer of the second generator, downsampled by 64 7x7 convolution kernels, and instance normalization and ReLU activation function are applied to obtain the initial feature map;
[0146] The initial feature map is downsampled and convolved twice, each time using 128 3x3 convolution kernels with a step size of 2, and instance normalization and ReLU activation function are applied to obtain the downsampled feature map;
[0147] The downsampled feature map is input into the first residual block, processed through two 3x3 convolutional layers and short-circuit connections, and instance normalization and ReLU activation function are applied to obtain the output of the first residual block;
[0148] The output of the first residual block is continuously input into the second residual block to the eighth residual block, and the structure of each residual block is the same as that of the first residual block, and finally the output of the eighth residual block is obtained;
[0149] The output of the eighth residual block is input into the self-attention layer to calculate the global correlation of the feature map to obtain the attention-weighted feature map, and the attention-weighted feature map is input into the ninth residual block for the last residual processing to obtain the final residual output;
[0150] The final residual output is upsampled and deconvolved twice, each time using 64 3x3 deconvolution kernels with a step size of 2, and instance normalization and ReLU activation function are applied to obtain the upsampled feature map;
[0151] The upsampled feature map is input into the last convolution layer of the second generator, processed by three 7x7 convolution kernels to obtain the output feature map of the original size, and the tanh activation function is performed on the output feature map to obtain the second posture estimation result.
[0152] Specifically, the second position estimation result and the real-time input data are stacked together to form a multi-channel input tensor. In order to unify the number of channels of the feature map and perform feature fusion, a 1x1 convolution layer is used for channel alignment. The role of 1x1 convolution is to integrate the information of different channels through linear transformation. The formula is: ,in, is the output feature map, is a 1x1 convolution kernel, is the input feature map, is the bias term. After processing the 1x1 convolution layer, a fused feature map is obtained. The fused feature map is input into the first convolution layer of the second generator, which uses 64 7x7 convolution kernels for downsampling, while retaining important features, reducing the spatial resolution of the feature map. The formula for downsampling convolution is:
[0153] ;
[0154] in, is the output feature map, is the fusion feature map, is a 7x7 convolution kernel, is the bias term with a step size of 2. After the convolution operation, instance normalization and ReLU activation function are applied, and the formula is: ,in, is the normalized and activated feature map. The initial feature map is downsampled twice, each time using 128 3x3 convolution kernels with a step size of 2. The purpose of downsampling convolution is to reduce the spatial resolution of the feature map while extracting higher-level features. The formula is: ,in, is the output feature map of the second convolution, is a 3x3 convolution kernel, is the bias term. After the convolution operation, instance normalization and ReLU activation function are also applied, and the formula is: . After two downsampling convolutions, the downsampled feature map is obtained. The downsampled feature map is input into the first residual block for processing. The residual block is implemented by two 3x3 convolutional layers and a short-circuit connection. The residual block can effectively alleviate the gradient vanishing problem in deep networks while maintaining the richness and accuracy of features. The output of the first residual block is continuously input into the second residual block to the eighth residual block. The structure of each residual block is the same as that of the first residual block. Through the layer-by-layer processing of multiple residual blocks, the deep features are gradually extracted and integrated to obtain the output of the eighth residual block. The output of the eighth residual block is input into the self-attention layer to calculate the global correlation of the feature map. The self-attention layer generates a weighted feature map by calculating the relationship between each position in the feature map and other positions. The formula is: ,in, (Query), (key) and (values) are the linear transformations of the feature maps, is the feature map after attention weighting. The self-attention mechanism can capture the global dependencies in the feature map and enhance the expressiveness of the features. The attention-weighted feature map is input into the ninth residual block for the last residual processing to obtain the final residual output. The structure of the residual block is the same as the previous residual block, which is processed by two 3x3 convolutional layers and a short-circuit connection. The final residual output is upsampled and deconvolved twice, each time using 64 3x3 deconvolution kernels with a step size of 2. The purpose of upsampling deconvolution is to gradually restore the spatial resolution of the feature map. The formula is:
[0155]
[0156] in, is the output feature map of the first upsampling, is the deconvolution kernel, is the bias term. Repeat the above steps to get the upsampled feature map. The upsampled feature map is input into the last convolution layer of the second generator and processed by three 7x7 convolution kernels. The formula is: ,in, is the output feature map, is the convolution kernel, is the bias term. The tanh activation function is operated on the output feature map to limit the output value range to between -1 and 1, and the second posture estimation result is obtained.
[0157] In a specific embodiment, the execution step inputs the second posture estimation result into the second discriminator, and the multi-scale authenticity evaluation is performed through three sub-discriminators of different scales. The process of obtaining the discrimination result of the posture estimation can specifically include the following steps:
[0158] The second posture estimation result is downsampled three times to obtain original size input data, 1 / 2 size input data and 1 / 4 size input data;
[0159] The original size input data is input into the first convolutional layer of the first sub-discriminator, convolved with 64 4x4 convolution kernels with a step size of 2, and the LeakyReLU activation function is applied to obtain the first feature map of the first sub-discriminator;
[0160] The first feature map of the first sub-discriminator is continuously input into the second to fourth convolutional layers. Each convolutional layer uses 128, 256, and 512 4x4 convolution kernels, respectively, with a step size of 2, and applies instance normalization and LeakyReLU activation function to obtain the final feature map of the first sub-discriminator.
[0161] Input the 1 / 2 size input data into the second sub-discriminator for processing to obtain the final feature map of the second sub-discriminator, which has the same structure as the first sub-discriminator;
[0162] Input the 1 / 4 size input data into the third sub-discriminator for processing to obtain the final feature map of the third sub-discriminator, which has the same structure as the first sub-discriminator;
[0163] Perform global average pooling on the final feature map of the first sub-discriminator, the final feature map of the second sub-discriminator, and the final feature map of the third sub-discriminator to obtain three discriminant feature vectors;
[0164] The three discriminant feature vectors are respectively input into the corresponding fully connected layer, and the output is mapped to between 0 and 1 through the Sigmoid activation function to obtain the discrimination results of the three scales. The discrimination results of the three scales are weighted averaged to obtain the discrimination result of the posture estimation.
[0165] Specifically, the second pose estimation result is downsampled three times, and the original image is reduced to 1 / 2 and 1 / 4 size respectively, so as to facilitate multi-scale discrimination. Downsampling can be achieved by bilinear interpolation or other interpolation methods to obtain image data of different resolutions. The original size input data is input into the first convolution layer of the first sub-discriminator, and convolution operation is performed through 64 4x4 convolution kernels with a step size of 2. The formula of the convolution operation is: ,in, is the output feature map, is the input image, is the convolution kernel, is the bias term. After the convolution operation, the LeakyReLU activation function is applied to introduce nonlinear characteristics. The formula is: The first feature map of the first sub-discriminator is continuously input into the second convolution layer to the fourth convolution layer. The second convolution layer uses 128 4x4 convolution kernels with a step size of 2. The formula is: ,in, is the output feature map of the second convolutional layer, is the convolution kernel, is the bias term. After the convolution operation, instance normalization and LeakyReLU activation function are applied, and the formula is: . Similarly, the third convolution layer uses 256 4x4 convolution kernels with a stride of 2, and the fourth convolution layer uses 512 4x4 convolution kernels with a stride of 2. Instance normalization and LeakyReLU activation function are applied after each convolution operation to obtain the feature maps of the third and fourth sub-discriminators, respectively, and the final feature map of the first sub-discriminator is obtained. The 1 / 2 size input data is input to the second sub-discriminator for processing. The second sub-discriminator has the same structure as the first sub-discriminator, containing four convolution layers, using 64, 128, 256 and 512 4x4 convolution kernels, respectively, with a stride of 2. Instance normalization and LeakyReLU activation function are applied after each convolution operation to obtain the final feature map of the second sub-discriminator. Similarly, the 1 / 4 size input data is input to the third sub-discriminator for processing. The third sub-discriminator has the same structure as the first sub-discriminator, and contains four convolutional layers, using 64, 128, 256 and 512 4x4 convolution kernels respectively, with a step size of 2. Instance normalization and LeakyReLU activation function are applied after each convolution operation. The formula is similar to the processing steps of the first sub-discriminator mentioned above, and the final feature map of the third sub-discriminator is obtained. Global average pooling is performed on the final feature map of the first sub-discriminator, the final feature map of the second sub-discriminator, and the final feature map of the third sub-discriminator respectively. Global average pooling reduces the dimension of the feature map by taking the average of each channel of the feature map while retaining the global information. The formula is:
[0166] ;
[0167] in, is a one-dimensional vector, and is the height and width of the feature map. The three discriminant feature vectors are input into the corresponding fully connected layer respectively, and the output is mapped to between 0 and 1 through the Sigmoid activation function. The formula is:
[0168] ;
[0169] in, is the final judgment result. is the weight of the fully connected layer, is the bias term. The function of the Sigmoid activation function is to limit the output result between 0 and 1 to indicate the authenticity probability of the generated result. The discrimination results of the three scales are weighted averaged to obtain the discrimination result of the posture estimation. The weighted average formula is:
[0170] ;
[0171] in, is the final attitude estimation result. They are the discrimination results of three scales. is the corresponding weight parameter.
[0172] In a specific embodiment, the process of executing step S105 may specifically include the following steps:
[0173] Extract features from the corrected positioning data to obtain a comprehensive feature vector, and input the comprehensive feature vector into the deep Q network to calculate the Q value of various possible actions to obtain the action value evaluation result;
[0174] Select the optimal action according to the action value evaluation result to obtain the intelligent decision result, and execute the intelligent decision result to obtain the decision execution result;
[0175] The reward value is calculated according to the decision execution result, and the reward value, current state and next state are stored in the experience replay pool to obtain updated experience data. Experience data is randomly sampled from the experience replay pool, and the Q network parameters are updated through temporal difference learning to obtain the optimized Q network.
[0176] The corrected positioning data is processed through the optimized Q network to obtain high-precision positioning results;
[0177] Compare and analyze the high-precision positioning results with historical data to obtain a model update trigger signal, and perform incremental learning on the first positioning prediction model according to the model update trigger signal, and adjust and update the model parameters through gradient accumulation and dynamic learning rate to obtain updated model parameters;
[0178] The updated model parameters are applied to the first positioning prediction model, and the second positioning prediction model is obtained through model structure fine-tuning and knowledge distillation.
[0179] Specifically, feature extraction is performed on the corrected positioning data, and high-dimensional features that can characterize the position and posture are extracted from the data using a convolutional neural network or other deep learning model. The extracted feature vectors are comprehensive feature vectors that can represent various information about the current state. The formula is: ,in, is the comprehensive feature vector, is the feature extraction function, is the corrected positioning data. The comprehensive feature vector is input into the deep Q network, and the Q value of various possible actions is calculated through the deep Q network to obtain the action value evaluation result. The input of the deep Q network is the comprehensive feature vector, and the output is the Q value corresponding to each possible action. The formula is: ,in, represents the Q-value function, is the current state, It's action. is a network parameter. The expected benefits of performing different actions in the current state are evaluated through the Q value function. The optimal action is selected based on the action value evaluation result to obtain the intelligent decision result. The optimal action is the action with the highest Q value, and the formula is: . Execute the intelligent decision result to obtain the decision execution result. For example, in an autonomous driving system, the optimal action may be a specific operation to adjust the direction or speed of the vehicle, and the decision execution result is the actual operating state of the vehicle. Calculate the reward value based on the decision execution result, and store the reward value, current state, and next state in the experience replay pool to obtain updated experience data. The reward value reflects the effect after executing a specific action, and the formula is: ,in, is the reward value, is the reward function. Deposit into the experience replay pool, where is the next state. Randomly sample experience data from the experience replay pool, update the Q network parameters through temporal difference learning, and obtain the optimized Q network. Temporal difference learning learns by comparing the difference between the current Q value and the target O value. The formula is:
[0180] ;
[0181] in, is the timing difference error, is the discount factor, is the target network parameter. By minimizing the timing difference error, the Q network parameters are updated as follows:
[0182] ;
[0183] in, is the learning rate. The corrected positioning data is processed by the optimized Q network to obtain a high-precision positioning result. The optimized Q network can more accurately evaluate the value of each action, thereby selecting the optimal action and improving positioning accuracy. The high-precision positioning result is compared and analyzed with the historical data to obtain a model update trigger signal, and the first positioning prediction model is incrementally learned based on the model update trigger signal. Incremental learning ensures that the model can adapt to the changing environment by gradually accumulating new training data and adjusting model parameters. Through gradient accumulation and dynamic learning rate adjustment, the formula is:
[0184] ;
[0185] in, is the dynamically adjusted learning rate, is the loss function. The updated model parameters are applied to the first positioning prediction model, and the second positioning prediction model is obtained through model structure fine-tuning and knowledge distillation. Model structure fine-tuning improves the model's expressiveness and adaptability by adjusting the number of network layers, nodes, and connection methods. Knowledge distillation ensures that the new model inherits the advantages of the old model while improving its performance and accuracy by transferring the knowledge of the old model to the new model. The formula is:
[0186] ;
[0187] in, is the loss function of knowledge distillation, is the cross entropy loss, is the knowledge distillation loss, and is the weight parameter, is the prediction result, and are the outputs of the teacher model and the student model respectively.
[0188] The above describes the high-precision linear guide positioning method based on artificial intelligence in the embodiment of the present application. The following describes the high-precision linear guide positioning system based on artificial intelligence in the embodiment of the present application. Figure 2 In the embodiments of the present application, an embodiment of a high-precision linear guide positioning system based on artificial intelligence includes:
[0189] The acquisition module 201 is used to perform multi-modal data acquisition and preprocessing on the linear guide rail to obtain a standardized training data set;
[0190] A construction module 202 is used to construct a dual-conditional generative adversarial network using a standardized training data set to obtain a multimodal deep learning model;
[0191] The training module 203 is used to optimize the multimodal deep learning model by using a joint loss function to obtain optimized model parameters, and iteratively train and verify the optimized model parameters to obtain a first positioning prediction model;
[0192] The positioning module 204 is used to perform real-time positioning of the linear guide rail by using the first positioning prediction model to obtain real-time positioning data, and to perform dynamic error compensation and reliability evaluation on the real-time positioning data to obtain corrected positioning data;
[0193] The updating module 205 is used to perform intelligent decision-making and closed-loop optimization on the corrected positioning data to obtain a high-precision positioning result, and to update the first positioning prediction model according to the high-precision positioning result to obtain a second positioning prediction model.
[0194] Through the cooperation of the above components, the multi-modal data of the linear guide is collected through a multi-sensor integrated system, and standardized processing and data enhancement are performed to improve the quality and diversity of the data. The position estimation and attitude estimation conditional generative adversarial networks are combined in series to achieve high-precision estimation of the position and attitude of the linear guide. At the same time, the generation and discrimination capabilities of the model are improved through adversarial training. A comprehensive joint loss function is constructed by combining adversarial loss, reconstruction loss, geometric consistency loss and multi-task loss, which effectively improves the training effect and generalization ability of the model. The positioning uncertainty is estimated by random perturbation method and filtered to achieve dynamic error compensation of real-time positioning data, thereby improving the positioning accuracy and reliability. The deep Q network is used for intelligent decision-making, and the model is continuously optimized through experience replay and temporal difference learning, which realizes the adaptive ability and continuous improvement of the positioning system. The dynamic update of the positioning prediction model is realized to ensure the performance stability and adaptability of the model during long-term use. This method integrates a variety of advanced artificial intelligence technologies, has strong versatility and robustness, and can adapt to the linear guide positioning needs under different working conditions and environmental conditions.
[0195] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0196] 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 the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.
[0197] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A high-precision linear guide positioning method based on artificial intelligence, characterized in that: The method comprises: Perform multi-modal data acquisition and preprocessing on the linear guide to obtain a standardized training data set; A dual-conditional generative adversarial network is constructed through the standardized training data set to obtain a multimodal deep learning model; specifically comprising: dividing the standardized training data set into a position information data set and a posture information data set; constructing a position estimation conditional generative adversarial network according to the position information data set, the position estimation conditional generative adversarial network comprising a first generator and a first discriminator, the first generator adopts a U-Net structure, and the first discriminator adopts a PatchGAN structure; the first generator comprises an encoder and a decoder, the encoder comprises 5 convolutional layers, each convolutional layer uses a LeakyReLU activation function and batch normalization, the decoder comprises 5 deconvolutional layers, each deconvolutional layer uses a ReLU activation function and batch normalization; the first discriminator comprises 4 convolutional layers, each convolutional layer uses a normalization layer and a LeakyReLU activation function; the position information data set is input into the position estimation conditional generative adversarial network, and a first position estimation result is obtained through adversarial training; according to The posture information data set and the first position estimation result construct a posture estimation conditional generative adversarial network, the posture estimation conditional generative adversarial network includes a second generator and a second discriminator, the second generator adopts a ResNet structure enhanced by an attention mechanism, and the second discriminator adopts a multi-scale discriminator structure; the second generator includes 9 residual blocks, each residual block includes two convolutional layers and a short-circuit connection, and a self-attention layer is added between the residual blocks; the second discriminator includes 3 sub-discriminators of different scales, each sub-discriminator includes 4 convolutional layers and a fully connected layer; the first position estimation result and the posture information data set are input into the posture estimation conditional generative adversarial network, and the first posture estimation result is obtained through adversarial training; the position estimation conditional generative adversarial network and the posture estimation conditional generative adversarial network are combined in series, and a domain adaptation layer is added between the position estimation conditional generative adversarial network and the posture estimation conditional generative adversarial network to obtain a multimodal deep learning model; The multimodal deep learning model is optimized using a joint loss function to obtain optimized model parameters, and the optimized model parameters are iteratively trained and verified to obtain a first positioning prediction model; The linear guide is positioned in real time by using the first positioning prediction model to obtain real-time positioning data, and the real-time positioning data is dynamically error compensated and reliability evaluated to obtain corrected positioning data; specifically comprising: collecting real-time multimodal data of the linear guide, and preprocessing the real-time multimodal data to obtain real-time input data; inputting the real-time input data into the first generator of the position estimation condition generative adversarial network of the first positioning prediction model, and processing it through the U-Net structure of 5 convolutional layers and 5 deconvolutional layers to obtain a second position estimation result; inputting the second position estimation result into the first discriminator, and performing authenticity evaluation through the PatchGAN structure of 4 convolutional layers to obtain a discriminant result of the position estimation; based on the second position estimation result and the real-time input data, The posture estimation condition of the first positioning prediction model is input into the second generator of the generative adversarial network, and a second posture estimation result is obtained through processing with 9 residual blocks and a self-attention layer; the second posture estimation result is input into the second discriminator, and a multi-scale authenticity evaluation is performed through 3 sub-discriminators of different scales to obtain a discrimination result of the posture estimation; according to the discrimination result of the position estimation and the discrimination result of the posture estimation, the second position estimation result and the second posture estimation result are combined, and real-time positioning data is obtained through processing with a domain adaptation layer; the real-time positioning data is subjected to multiple small random perturbations by a random perturbation method, and the change of the output result is calculated to obtain a positioning uncertainty estimate, and according to the positioning uncertainty estimate, the real-time positioning data is filtered to obtain corrected positioning data; Perform intelligent decision-making and closed-loop optimization on the corrected positioning data to obtain a high-precision positioning result, and update the first positioning prediction model according to the high-precision positioning result to obtain a second positioning prediction model; specifically including: extracting features from the corrected positioning data to obtain a comprehensive feature vector, and inputting the comprehensive feature vector into a deep Q network to calculate the Q values of various possible actions to obtain an action value evaluation result; selecting the optimal action according to the action value evaluation result to obtain an intelligent decision result, and executing the intelligent decision result to obtain a decision execution result; calculating a reward value according to the decision execution result, and storing the reward value, the current state and the next state in an experience replay pool to obtain a more accurate positioning result. The modified experience data is randomly sampled from the experience replay pool, and the Q network parameters are updated through time difference learning to obtain an optimized Q network; the corrected positioning data is processed through the optimized Q network to obtain a high-precision positioning result; the high-precision positioning result is compared and analyzed with the historical data to obtain a model update trigger signal, and according to the model update trigger signal, the first positioning prediction model is incrementally learned, and the model parameters are updated through gradient accumulation and dynamic learning rate adjustment to obtain updated model parameters; the updated model parameters are applied to the first positioning prediction model, and a second positioning prediction model is obtained through model structure fine-tuning and knowledge distillation.
2. The high-precision linear guide positioning method based on artificial intelligence according to claim 1 is characterized in that: The multi-modal data acquisition and preprocessing of the linear guide rail to obtain a standardized training data set includes: Using a multi-sensor integrated system to collect data on the linear guide under different working conditions to obtain an original multi-modal data set, and performing data cleaning on the original multi-modal data set to obtain a preliminary cleaned data set; According to the initially cleaned data set, different types of data are standardized to unify the data into the same range to obtain a standardized data set, and data enhancement is performed on the standardized data set to obtain an enhanced data set; Performing frequency domain conversion on the time series data in the enhanced data set to obtain frequency domain feature data, and performing multi-scale analysis on the non-stationary signal in the enhanced data set to obtain time-frequency feature data; Performing dimensionality reduction processing on the frequency domain feature data and the time-frequency feature data to obtain dimensionality reduction feature data, and performing nonlinear dimensionality reduction and visualization based on the dimensionality reduction feature data to obtain a two-dimensional feature distribution map; A cluster analysis is performed on the two-dimensional feature distribution map to obtain an optimized feature data set, and the optimized feature data set is divided into a training set, a validation set and a test set according to a preset ratio to obtain a standardized training data set.
3. The high-precision linear guide positioning method based on artificial intelligence according to claim 1 is characterized in that: The method of optimizing the multimodal deep learning model by using a joint loss function to obtain optimized model parameters, and iteratively training and verifying the optimized model parameters to obtain a first positioning prediction model includes: Obtaining the training output result of the multimodal deep learning model, and calculating the adversarial loss of the training output result and the true label, and calculating the weighted sum of the L1 loss and the perceptual loss based on the training output result and the true label to obtain the reconstruction loss; Using a spatial transformation network to perform geometric transformation on the first position estimation result and the first posture estimation result of the multimodal deep learning model, calculating the consistency error before and after the transformation, and obtaining a geometric consistency loss; Calculating multi-task losses corresponding to the position estimation task, the attitude estimation task, and the reliability prediction task of the multimodal deep learning model; The adversarial loss, the reconstruction loss, the geometric consistency loss and the multi-task loss are weightedly combined, and an L2 regularization term and a gradient penalty term are added to obtain a joint loss function; Performing gradient descent optimization on the joint loss function, updating the parameters of the multimodal deep learning model, and obtaining optimized model parameters; The multimodal deep learning model is fine-tuned and parameter optimized according to the optimized model parameters to obtain an adjusted model, and the adjusted model is iteratively trained and verified to obtain a first positioning prediction model.
4. The high-precision linear guide positioning method based on artificial intelligence according to claim 1 is characterized in that: The step of inputting the second position estimation result into the first discriminator and performing authenticity evaluation through the PatchGAN structure of four convolutional layers to obtain a discriminant result of the position estimation includes: Preprocessing the second position estimation result to convert it into a format suitable for input to the first discriminator to obtain discriminator input data; Inputting the discriminator input data into the first convolution layer of the first discriminator, performing convolution operation through 64 3x3 convolution kernels, and applying LeakyReLU activation function to obtain a first feature map; Performing instance normalization processing on the first feature map to obtain a normalized first feature map, and inputting the normalized first feature map into the second convolution layer of the first discriminator, performing a convolution operation through 128 3x3 convolution kernels, and applying a LeakyReLU activation function to obtain a second feature map; Performing instance normalization processing on the second feature map to obtain a normalized second feature map, and inputting the normalized second feature map into the third convolution layer of the first discriminator, performing a convolution operation through 256 3x3 convolution kernels, and applying a LeakyReLU activation function to obtain a third feature map; Performing instance normalization processing on the third feature map to obtain a normalized third feature map, and inputting the normalized third feature map into the fourth convolution layer of the first discriminator, performing a convolution operation through 512 3x3 convolution kernels, and applying a LeakyReLU activation function to obtain a fourth feature map; A global average pooling operation is performed on the fourth feature map, the fourth feature map is compressed into a one-dimensional vector, a discriminant feature vector is obtained, and the discriminant feature vector is input into a fully connected layer, and the output is mapped to between 0 and 1 through a Sigmoid activation function to obtain a discriminant result of position estimation.
5. The high-precision linear guide positioning method based on artificial intelligence according to claim 1 is characterized in that: The method of inputting the second generator of the pose estimation conditional generative adversarial network of the first positioning prediction model according to the second position estimation result and the real-time input data, and obtaining the second pose estimation result through processing by 9 residual blocks and a self-attention layer, includes: Performing feature fusion on the second position estimation result and the real-time input data, performing channel alignment through a 1x1 convolution layer, and obtaining a fused feature map; Input the fused feature map into the first convolution layer of the second generator, downsample it through 64 7x7 convolution kernels, and apply instance normalization and ReLU activation function to obtain an initial feature map; Performing two downsampling convolutions on the initial feature map, each time using 128 3x3 convolution kernels with a step size of 2, and applying instance normalization and ReLU activation function to obtain a downsampled feature map; Input the downsampled feature map into the first residual block, process it through two 3x3 convolutional layers and short-circuit connections, and apply instance normalization and ReLU activation function to obtain the output of the first residual block; Continuously inputting the output of the first residual block into the second residual block to the eighth residual block, wherein the structure of each residual block is the same as that of the first residual block, and finally obtaining the output of the eighth residual block; Inputting the output of the eighth residual block into the self-attention layer to perform global correlation calculation of the feature map to obtain an attention-weighted feature map, and inputting the attention-weighted feature map into the ninth residual block for the last residual processing to obtain a final residual output; Perform two upsampling deconvolutions on the final residual output, each time using 64 3x3 deconvolution kernels with a step size of 2, and apply instance normalization and ReLU activation function to obtain an upsampled feature map; The upsampled feature map is input into the last convolution layer of the second generator, processed by three 7x7 convolution kernels to obtain an output feature map of the original size, and a tanh activation function operation is performed on the output feature map to obtain a second posture estimation result.
6. The high-precision linear guide positioning method based on artificial intelligence according to claim 1 is characterized in that: The step of inputting the second posture estimation result into the second discriminator, performing multi-scale authenticity evaluation through three sub-discriminators of different scales, and obtaining a discrimination result of the posture estimation includes: Downsampling the second posture estimation result three times to obtain original size input data, 1 / 2 size input data and 1 / 4 size input data; Inputting the original size input data into the first convolution layer of the first sub-discriminator, performing a convolution operation through 64 4x4 convolution kernels with a step size of 2, and applying a LeakyReLU activation function to obtain a first feature map of the first sub-discriminator; The first feature map of the first sub-discriminator is continuously input into the second convolution layer to the fourth convolution layer, each convolution layer uses 128, 256 and 512 4x4 convolution kernels respectively, with a step size of 2, and applies instance normalization and LeakyReLU activation function to obtain the final feature map of the first sub-discriminator; Inputting the 1 / 2 size input data into a second sub-discriminator for processing to obtain a final feature map of the second sub-discriminator, wherein the second sub-discriminator has the same structure as the first sub-discriminator; Inputting the 1 / 4 size input data into a third sub-discriminator for processing to obtain a final feature map of the third sub-discriminator, wherein the third sub-discriminator has the same structure as the first sub-discriminator; Performing global average pooling on the final feature map of the first sub-discriminator, the final feature map of the second sub-discriminator, and the final feature map of the third sub-discriminator, respectively, to obtain three discriminant feature vectors; The three discriminant feature vectors are respectively input into the corresponding fully connected layer, and the output is mapped to between 0 and 1 through the Sigmoid activation function to obtain the discrimination results of three scales, and the discrimination results of the three scales are weighted averaged to obtain the discrimination result of posture estimation.
7. A high-precision linear guide positioning system based on artificial intelligence, characterized in that: A system for executing a high-precision linear guide positioning method based on artificial intelligence as claimed in any one of claims 1 to 6, comprising: An acquisition module is used to collect and preprocess multimodal data of the linear guide rail to obtain a standardized training data set; A construction module, used to construct a dual-conditional generative adversarial network through the standardized training data set to obtain a multimodal deep learning model; A training module, used to optimize the multimodal deep learning model using a joint loss function to obtain optimized model parameters, and iteratively train and verify the optimized model parameters to obtain a first positioning prediction model; A positioning module, used for performing real-time positioning of the linear guide rail by using the first positioning prediction model to obtain real-time positioning data, and performing dynamic error compensation and reliability evaluation on the real-time positioning data to obtain corrected positioning data; An updating module is used to perform intelligent decision-making and closed-loop optimization on the corrected positioning data to obtain a high-precision positioning result, and to update the first positioning prediction model according to the high-precision positioning result to obtain a second positioning prediction model.
