RFID smart label positioning tracking printing method and system

Abstract

The present application relates to the technical field of RFID smart labels, and discloses a method and system for printing RFID smart labels for positioning and tracking, the method comprising: inputting the RFID tag radio frequency signal into a conditional variational autoencoder for convolution processing and attention weighting to obtain a global feature vector of the RFID tag; establishing a hierarchical sensor node network to perform multi-dimensional acquisition of environmental parameters to obtain an environmental spatiotemporal dataset; inputting the RFID tag global feature vector and the environmental spatiotemporal dataset into a deep fusion network for feature fusion to obtain multimodal fusion features; performing sequence modeling based on the multimodal fusion features to obtain an item trajectory prediction model; applying the item trajectory prediction model to a target area grid coordinate system to perform density distribution calculation to obtain a real-time item positioning heat map and a trajectory database, thereby establishing an efficient item positioning heat map generation and trajectory data management mechanism, which facilitates the system to perform real-time location visualization and historical trajectory query.

Description

Technical Field

[0001] The present application relates to the technical field of RFID smart labels, and in particular to a method and system for positioning, tracking and printing RFID smart labels. Background Art

[0002] With the rapid development of the Internet of Things (IoT), RFID technology has been widely used in the field of object positioning and tracking. However, traditional RFID positioning systems face problems such as signal attenuation, multipath effects, and environmental interference in complex environments, resulting in positioning accuracy and stability that cannot meet actual requirements.

[0003] Current RFID positioning systems typically rely on a single RF signal signature for location estimation, lacking comprehensive consideration of environmental factors and unable to effectively cope with dynamic, real-world scenarios. Especially in densely populated environments, signal interference and occlusion can significantly degrade positioning performance. Summary of the Invention

[0004] This application provides an RFID smart label positioning and tracking printing method and system, thereby establishing an efficient object positioning heat map generation and trajectory data management mechanism, which facilitates the system to perform real-time location visualization and historical trajectory query.

[0005] In a first aspect, the present application provides an RFID smart label positioning and tracking printing method, the RFID smart label positioning and tracking printing method comprising:

[0006] The RFID tag radio frequency signal is input into the conditional variational autoencoder for convolution processing to obtain the RFID tag structured features;

[0007] Performing attention weighting on the structured features of the RFID tag to obtain a global feature vector of the RFID tag;

[0008] Establish a hierarchical sensor node network to collect multi-dimensional environmental parameters and obtain an environmental spatiotemporal data set;

[0009] Inputting the RFID tag global feature vector and the environmental spatiotemporal dataset into a deep fusion network for feature fusion to obtain a multimodal fusion feature;

[0010] Perform sequence modeling based on the multimodal fusion features to obtain an item trajectory prediction model;

[0011] The object trajectory prediction model is applied to the target area grid coordinate system to perform density distribution calculation to obtain a real-time object location heat map and trajectory database.

[0012] A second aspect of the present application provides an RFID smart label positioning and tracking printing system, the RFID smart label positioning and tracking printing system comprising:

[0013] The convolution processing module is used to input the RFID tag radio frequency signal into the conditional variational autoencoder for convolution processing to obtain the RFID tag structured features;

[0014] an attention weighting module, configured to perform attention weighting on the structured features of the RFID tag to obtain a global feature vector of the RFID tag;

[0015] The multi-dimensional acquisition module is used to establish a hierarchical sensor node network to collect environmental parameters in multiple dimensions and obtain an environmental spatiotemporal data set;

[0016] A feature fusion module is used to input the RFID tag global feature vector and the environmental spatiotemporal dataset into a deep fusion network for feature fusion to obtain a multimodal fusion feature;

[0017] A sequence modeling module, configured to perform sequence modeling based on the multimodal fusion features to obtain an item trajectory prediction model;

[0018] The distribution calculation module is used to apply the item trajectory prediction model to the target area grid coordinate system to perform density distribution calculation to obtain a real-time item positioning heat map and trajectory database.

[0019] Compared with the existing technology, the present application has the following beneficial effects: by introducing conditional variational autoencoders and attention mechanisms, deep feature extraction and weighted processing of RFID radio frequency signals are performed, the robustness and discriminability of feature expression are improved, and the signal interference problem in complex environments is effectively overcome; by adopting a layered sensor network architecture and a multi-dimensional environmental parameter acquisition strategy, comprehensive perception and real-time monitoring of the environmental status of the target area are achieved, providing rich environmental context information for object positioning; a deep fusion network is designed for multimodal feature fusion, and through dual-branch feature extraction and cross-modal interaction, RFID signal features and environmental parameter information are effectively integrated, enhancing the environmental adaptability of the system; a trajectory prediction method based on sequence modeling and constrained optimization, combined with a temporal attention mechanism and an autoregressive decoding structure, achieves accurate prediction and smoothing of the object's motion trajectory; through gridded density distribution calculation and spatiotemporal index construction, an efficient object positioning heat map generation and trajectory data management mechanism is established, which facilitates the system to perform real-time location visualization and historical trajectory query. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] The structures, proportions, sizes, etc. depicted in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not intended to limit the conditions under which the present invention can be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportional relationships, or adjustments in size should still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and objectives that can be achieved by the present invention.

[0022] Figure 1 This is a flow chart of a method for positioning, tracking, and printing RFID smart labels provided by an embodiment of the present invention;

[0023] Figure 2 This is a schematic block diagram of the structure of the RFID smart label positioning tracking printing system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0025] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined, or partially merged, so the actual execution order may vary depending on the actual situation.

[0026] It should also be understood that the terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit the present application. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0027] It should be further understood that the term "and / or" used in this specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 In one embodiment of the present application, an embodiment of the RFID smart label positioning tracking printing method includes:

[0028] Step 100: Input the RFID tag radio frequency signal into a conditional variational autoencoder for convolution processing to obtain the RFID tag structured features;

[0029] It is understandable that the execution subject of this application can be an RFID smart label positioning tracking printing system, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0030] Specifically, the RFID tag radio frequency signal is fed into the feature extraction layer of a conditional variational autoencoder (CVA). The feature extraction layer performs convolution operations on the input signal to extract multi-scale feature maps. The feature extraction layer contains five parallel convolution kernels, each with a stride of 2. This design captures the multi-scale characteristics of the input signal through different receptive fields. Furthermore, to ensure the stability of the convolution operation and accelerate model convergence, a batch normalization layer and a Reluctant Unit (ReLU) activation function are added after each convolution layer. The batch normalization layer reduces the distribution variation of the input features, while the ReLU activation function effectively introduces nonlinear characteristics, improving the model's expressive power. Max pooling is then performed on the multi-scale feature maps to achieve dimensionality reduction. Max pooling uses a 2×2 pooling window with a stride of 2, compressing the spatial dimensions of the feature maps while retaining the most significant features. Max pooling reduces computational complexity and improves the feature map's ability to express local information, generating a reduced-dimensional feature map. The reduced-dimensional feature map is then fed into the fully connected layer of the CVA for dimensionality conversion, generating a fixed-dimensional feature vector. The fully connected layer consists of three hidden layers, with 512, 256, and 128 neurons in each hidden layer, respectively. This gradually decreasing design helps compress the feature dimensionality while preserving important global information. In the fully connected layer, through weight sharing and nonlinear mapping, the model captures the underlying characteristics of the RFID signal at a deeper level and compresses it into a fixed-length vector representation. To enhance the representational power of the feature vector, adaptive weights are applied to the fixed-dimensional feature vector. These adaptive weights are dynamically calculated based on feature importance scores. The model assesses the contribution of each feature and assigns a weight to each feature. This effectively highlights key features while suppressing redundant information, generating a weighted feature vector. The weighted feature vector is then input into the encoder of a conditional variational autoencoder for compression mapping. The encoder is designed to consist of a mean encoder and a variance encoder, which output the mean and variance vectors of the latent space, respectively. The mean and variance vectors together define the distribution of the latent space features. By modeling the mean and variance, the latent variable distribution of the input signal is captured. Reparameterized sampling techniques are used to sample the latent space features. Reparameterized sampling generates sampling points that follow a standard normal distribution through the mean vector and variance vector, achieving a randomized representation of the latent space features. Through reparameterized sampling, the structural features of the RFID tag are ultimately generated.

[0031] Step 200: Perform attention weighting on the RFID tag structured features to obtain the RFID tag global feature vector;

[0032] Specifically, the structured features of the RFID tag are input into a multi-head attention mechanism for feature decomposition. The multi-head attention mechanism processes different feature subspaces through parallel attention heads to more comprehensively capture the contextual information of the RFID signal. The multi-head attention mechanism consists of eight parallel attention heads, each of which independently processes the input features and outputs a 64-dimensional feature subspace. This design not only ensures sufficient model expressiveness but also effectively decomposes and captures the importance of the input features across different dimensions, generating multiple sets of attention feature subspaces with diverse characteristics. These sets of attention feature subspaces are then input into a multi-layer perceptron for nonlinear transformation to generate a feature map matrix. The multi-layer perceptron consists of two hidden layers, each with dimensions of 512 and 256, respectively. Each layer uses LeakyReLU as the activation function. Compared to traditional ReLU, LeakyReLU has a certain negative slope, which can retain some gradients when the input values are negative, alleviating the vanishing gradient problem. Through the nonlinear transformation of the multi-layer perceptron, the model can learn more complex feature interactions and generate a feature map matrix with deep information. The feature map matrix is processed using a dot product operation and a softmax normalization function to calculate the importance score of each feature, known as the attention score. The dot product operation quantifies the relationship between the feature map matrix and the structured features into a weighted signal, while the softmax function normalizes these weighted signals to ensure that the sum of the importance scores of all features is 1. The attention score reflects the contribution of different features to the global context. The structured features of the RFID tag are weightedly summed with the attention score to generate a global context vector. This weighted summation is performed using matrix multiplication, making it computationally efficient. The global context vector is then input into a residual connection module for feature fusion, resulting in a fused feature vector. The residual connection module includes layer normalization layers and skip connections. The layer normalization layers normalize the feature vectors to eliminate internal covariate shift, while the skip connections preserve the original information of the input features and superimpose them with the nonlinearly transformed features. The fused feature vector undergoes a dimensionality reduction transformation. A 1×1 convolutional layer is used to reduce the feature dimensionality, mapping the high-dimensional fused feature vector into a 128-dimensional global feature space. The 1×1 convolutional layer reduces feature dimensionality and uses the weight matrix to learn the optimal combination of different feature dimensions, thereby improving the compactness and effectiveness of feature representation. The above steps ultimately generate a 128-dimensional global feature vector.

[0033] Step 300: Establish a hierarchical sensor node network to collect environmental parameters in multiple dimensions and obtain an environmental spatiotemporal data set;

[0034] It should be noted that multiple sensor nodes are deployed in a grid structure to construct a hierarchical sensor network topology. This grid deployment approach aims to cover all parts of the target area. By rationally planning the location and spacing of nodes, the sensor network can fully perceive environmental changes while avoiding node redundancy. Each sensor node is equipped with a temperature sensor, a humidity sensor, a light intensity sensor, and a motion sensor. The combination of these sensors can collect multi-dimensional environmental parameters, including temperature distribution, humidity fluctuations, light intensity, and the motion state of objects. The hierarchical sensor network topology is divided into master and slave nodes to establish a hierarchical data collection system. The master node is designated as the central node responsible for data aggregation, while the slave nodes focus on collecting local environmental parameters. The master and slave node division is optimized based on the node's geographic location, hardware performance, and network coverage to maximize network transmission efficiency and stability. The master node centrally processes data collected by the slave nodes, achieving efficient data management across the entire network while reducing transmission latency and bandwidth usage. Environmental parameters are sampled synchronously based on a hierarchical data collection system, ensuring temporal consistency across sensor nodes. This prevents bias or errors introduced by asynchronous data collection. During this process, environmental parameters such as temperature, humidity, light intensity, and object motion trajectory are collected to form a raw environmental data stream. This raw data stream is arranged in chronological order, reflecting the dynamic changes and spatial distribution of environmental parameters within the target area. The raw environmental data stream is then input into the data compression module for compression encoding. An efficient data compression algorithm is used to remove redundant information, reducing data storage and transmission overhead. Compression encoding effectively reduces data volume while maintaining the integrity and recoverability of critical information. After data compression, the compressed data packets undergo error correction encoding to enhance data robustness during transmission. Error correction encoding effectively corrects errors caused by network fluctuations or interference during data transmission, ensuring data reliability and integrity. The compressed and error-corrected data stream is decoded and reconstructed to generate a dataset that can be further utilized. During the decoding process, each data entry is timestamped and identified by its location. The introduction of timestamps marks the time of collection for each piece of data, enabling time-series analysis of environmental parameters. Location identifiers, used to identify the geographic location or node number corresponding to the data, allow environmental data to be accurately mapped to the grid coordinates of the target area. This step transforms the environmental data from a disordered and scattered state into a well-organized environmental spatiotemporal dataset with clear spatiotemporal information.

[0035] Step 400: Input the RFID tag global feature vector and the environmental spatiotemporal dataset into a deep fusion network for feature fusion to obtain multimodal fusion features;

[0036] Specifically, the RFID tag's global feature vector is time-aligned to ensure temporal consistency between the feature data and the environmental spatiotemporal data. Relying on the timestamp information included in the environmental spatiotemporal dataset, the sampling times of the RFID tag features and environmental parameters are compared and matched, adjusting the temporal distribution of the feature sequence to synchronize it with the environmental parameters. This time alignment avoids feature misalignment caused by temporal data asynchrony. The aligned feature sequence is input into the first feature extraction branch of a deep fusion network, while the environmental spatiotemporal dataset is input into the second feature extraction branch. Both branches perform feature extraction in parallel. The first and second feature extraction branches utilize the same architecture, each consisting of three convolutional layers and two fully connected layers. The convolutional layers extract spatial features from the input data, mining local patterns and relationships through layer-by-layer operations, while the fully connected layers transform the extracted features into high-level semantic representations. The two branches operate independently, generating deep feature representations for the RFID tag features and environmental parameters, respectively. Cross-modal interaction is then performed on the bimodal feature representations to capture potential correlations between the RFID tag features and environmental information. Cross-modal interaction computation maps the features of both modalities into the same feature space through matrix operations, generating an interaction feature matrix that reveals the complex relationship between RFID tag signals and environmental factors, such as the interference of ambient temperature and humidity on RF signal propagation. The interaction feature matrix is then processed in a feature enhancement module to generate an enhanced feature vector. The feature enhancement module uses residual blocks as its basic building blocks. Each residual block consists of two 3×3 convolutional layers and a skip connection. The residual block design introduces nonlinear transformations into the feature stream while preserving the original information of the input features through skip connections, thereby improving feature expressiveness while avoiding the vanishing gradient problem. Stacking multiple residual blocks enables deep feature reconstruction, making the enhanced feature vector more expressive and stable. Multi-scale feature extraction is performed on the enhanced feature vector to construct a multi-scale feature pyramid. Multi-scale feature extraction uses convolution operations at different scales to capture both local details and global information, ensuring that the features are adaptable to task requirements at different granularities. The feature pyramid structure effectively integrates information from different scales, providing more comprehensive contextual support for localization and tracking. After the feature pyramid is constructed, the features of each layer are fused through splicing operations to generate multimodal fusion features.

[0037] Step 500: Perform sequence modeling based on multimodal fusion features to obtain an item trajectory prediction model;

[0038] Specifically, the multimodal fusion features are segmented in time series to generate fixed-length feature sequence segments. Continuous feature data is segmented into segments with fixed time windows to better capture the dynamic changes of items over different time periods. The length of the feature sequence segments is adjusted based on the actual task requirements to balance the integrity of the time series information with computational complexity. After time series segmentation, these segments are input into a sequence encoder for feature encoding, generating a hidden state sequence. The sequence encoder employs a bidirectional long short-term memory (Bi-LSTM) architecture, which simultaneously captures the dynamic characteristics of the forward and backward dependencies in the time series through forward and backward information flow. The Bi-LSTM effectively avoids information loss and fully preserves the global dependencies of item trajectories in the temporal dimension. A temporal attention calculation is performed on the hidden state sequence to generate a weighted context vector. The temporal attention mechanism assigns weights to each time step in the hidden state sequence, highlighting the feature information of key time points while suppressing interference from non-critical time points. The attention weights are obtained by calculating the dot product between the hidden state and the context query vector and then performing softmax normalization. The weighted context vector integrates important information from the time series features. The weighted context vector is input into the trajectory generator for trajectory decoding, generating a trajectory sequence representation. The trajectory generator employs an autoregressive decoding architecture consisting of four Transformer decoding layers. The Transformer decoding layer efficiently decodes input features using a multi-head attention mechanism combined with a feedforward neural network. The multi-head attention mechanism captures global dependencies between features, while the feedforward neural network further models nonlinearities during the decoding process. The autoregressive decoding architecture ensures high trajectory sequence generation accuracy by progressively predicting trajectory points at each time step. The generated trajectory sequence representation undergoes constrained optimization to obtain a smooth trajectory sequence. This constraint optimization process includes position continuity constraints and motion smoothness constraints. Position continuity constraints ensure that the spatial position changes between adjacent trajectory points conform to physical laws, avoiding jumps or unreasonable trajectories. Motion smoothness constraints constrain the velocity and acceleration of the trajectory, making the generated trajectory smoother and more physically accurate. The combination of these two constraints effectively improves trajectory stability and realism, making it more suitable for practical applications. The smoothed trajectory sequence is input into the prediction head network for parameter regression to generate the object trajectory prediction model. The prediction head network consists of three fully connected layers. The first and second layers are used for feature mapping and dimensionality reduction, respectively, while the final layer regresses trajectory parameters. The output layer's dimensionality matches that of the trajectory parameters, ensuring that the model can directly generate key trajectory parameters such as position coordinates, velocity vectors, and angular changes. Through parameter regression, the trajectory prediction model converts smooth trajectory sequences into trajectory information that can be directly used by downstream tasks.

[0039] Step 600: Apply the item trajectory prediction model to the target area grid coordinate system to perform density distribution calculation to obtain a real-time item location heat map and trajectory database.

[0040] Specifically, coordinate mapping is performed on the target area to generate a grid coordinate matrix for the target area. By dividing the target area into uniformly sized grid cells, the actual physical space is converted into a discretized grid representation. The size of the grid cells is adjusted based on the resolution requirements of the application scenario, achieving a balance between spatial accuracy and computational complexity. The result of coordinate mapping is a multidimensional grid coordinate matrix, in which each grid cell corresponds to a specific spatial location. The trajectory parameter sequence output by the object trajectory prediction model is input into the coordinate transformation module for grid projection processing. The trajectory parameter sequence includes the object's three-dimensional spatial coordinates (e.g., x, y, z) and motion state parameters (e.g., velocity, acceleration, and azimuth). The coordinate transformation module maps these continuous spatial trajectory points to the corresponding grid cells according to the target area's grid division rules, generating a grid spatial trajectory sequence. Density accumulation calculation is performed on the grid spatial trajectory sequence to generate an initial density distribution map. The density accumulation calculation estimates the grid usage based on the dwell time and appearance frequency of each object in each grid cell. Whenever an item passes through a grid cell, a weight is added to the density value of that cell based on its dwell time. Furthermore, the frequency of occurrence is used as a key factor in density accumulation to reflect the activity of items in the area. Based on this, an initial density distribution map is generated, a two-dimensional matrix reflecting the distribution of items within the target area's grid cells. This initial density distribution map is then fed into the density smoothing module for Gaussian filtering. Gaussian filtering is a smoothing technique that effectively suppresses noise and smoothes local density fluctuations by taking a weighted average of density values within a neighborhood, making the distribution map more coherent and natural. After Gaussian filtering, a dynamic density matrix is generated, reflecting the dynamic distribution trends of items within the area. The dynamic density matrix is normalized to a uniform range (e.g., [0, 1]). Threshold segmentation is then performed to divide grid cells into high-density and low-density areas based on a preset density threshold, generating a real-time item location heatmap. To efficiently store and retrieve trajectory data and the dynamic density matrix, a spatiotemporal index is constructed. The grid-space trajectory sequence and the dynamic density matrix are organized using a quadtree structure to generate a trajectory database. A quadtree is a hierarchical index structure suitable for spatial data. By recursively dividing space into four quadrants, it significantly improves data storage and retrieval efficiency. The leaf nodes of a quadtree store trajectory fragments and density information, while the internal nodes record spatial extent and density statistics, supporting fast range queries and spatiotemporal analysis.

[0041] In the embodiment of the present application, by introducing conditional variational autoencoders and attention mechanisms, deep feature extraction and weighted processing of RFID radio frequency signals are performed, which improves the robustness and discriminability of feature expression and effectively overcomes the signal interference problem in complex environments; by adopting a layered sensor network architecture and a multi-dimensional environmental parameter acquisition strategy, comprehensive perception and real-time monitoring of the environmental status of the target area are achieved, providing rich environmental context information for object positioning; a deep fusion network is designed for multimodal feature fusion, and through dual-branch feature extraction and cross-modal interaction, RFID signal features and environmental parameter information are effectively integrated, enhancing the environmental adaptability of the system; a trajectory prediction method based on sequence modeling and constrained optimization, combined with a temporal attention mechanism and an autoregressive decoding structure, achieves accurate prediction and smoothing of the object's motion trajectory; through gridded density distribution calculation and spatiotemporal index construction, an efficient object positioning heat map generation and trajectory data management mechanism is established, which facilitates the system to perform real-time location visualization and historical trajectory query.

[0042] In a specific embodiment, the process of executing step 100 may specifically include the following steps:

[0043] The RFID tag radio frequency signal is input into the feature extraction layer of the conditional variational autoencoder for convolution operation to obtain a multi-scale feature map. The feature extraction layer includes 5 parallel convolution kernels, each with a stride of 2, and each convolution layer is connected to a batch normalization layer and a ReLU activation function.

[0044] Perform a maximum pooling operation on the multi-scale feature map to obtain a dimensionality-reduced feature map, where the maximum pooling operation uses a 2×2 pooling window with a step size of 2;

[0045] The reduced-dimensional feature map is input into the fully connected layer of the conditional variational autoencoder for dimension conversion to obtain a fixed-dimensional feature vector. The fully connected layer includes three hidden layers, and the number of neurons in each hidden layer is 512, 256, and 128, respectively.

[0046] Apply adaptive weights to the fixed-dimensional feature vector to obtain a weighted feature vector. The adaptive weights are dynamically calculated based on the feature importance score.

[0047] The weighted feature vector is input into the encoder of the conditional variational autoencoder for compression mapping to obtain the latent space feature. The encoder includes a mean encoder and a variance encoder, which output the mean vector and variance vector of the latent space respectively.

[0048] The latent space features are reparameterized and sampled to obtain the RFID tag structured features, wherein the reparameterized sampling generates sampling points that obey the standard normal distribution based on the mean vector and the variance vector.

[0049] Specifically, the RFID tag's radio frequency signal is represented as a time series data matrix ,in represents the number of time steps, Represents the feature dimension of each time step. This signal is passed as input to the feature extraction layer of the conditional variational autoencoder. The feature extraction layer contains 5 parallel convolution kernels, each with a step size of 2. The sizes of the convolution kernels are , representing different receptive field sizes. Multi-scale convolution design can capture the characteristic patterns of RF signals in different time ranges. The mathematical expression of convolution is:

[0050] ;

[0051] in Indicates the The feature map extracted by the convolution kernel, For the The weight matrix of the convolution kernel, is the bias term, * represents the convolution operation, BN( ) is the batch normalization operation, ReLU( ) is the ReLU activation function. Through this step, multiple multi-scale feature maps are generated, and the local and global features of the signal are captured by integrating the outputs of different convolution kernels. The multi-scale feature maps are subjected to the maximum pooling operation to achieve dimensionality reduction. The maximum pooling operation is performed by The pooling window and step size of 2 compress the spatial dimension of the feature map, reduce the computational complexity, and retain the significant features. The mathematical expression is:

[0052] ;

[0053] in It is The window represents the range of feature values within the pooling window. The reduced feature map is input to the fully connected layer for dimensionality conversion, generating a fixed-length feature vector. The fully connected layer consists of three hidden layers, with 512, 256, and 128 neurons, respectively. Each layer processes the input features through a linear transformation and an activation function, which can be expressed mathematically as follows:

[0054] ;

[0055] in Indicates the The input features of the layer, and are weight and bias parameters respectively, and It is Layer and The feature dimension of the layer. The generated fixed-dimensional feature vector Adaptive weights are applied to generate weighted feature vectors. The adaptive weights are calculated based on the feature importance scores , calculated through the normalization mechanism, the mathematical expression is:

[0056] ;

[0057] in Indicates the The importance score of a feature, Represents the corresponding normalized weight. The final weighted feature vector is:

[0058] ;

[0059] in is the weighted vector. The weighted feature vector is input into the encoder part of the conditional variational autoencoder for compression mapping to obtain the latent space feature. The encoder contains a mean encoder and a variance encoder, which output the mean vector of the latent space respectively. and variance vector This process is expressed as

[0060] ;

[0061] in and are the weights and biases of the linear transformation. Reparameterized sampling is performed on the latent space features to generate sample points that conform to the standard normal distribution. Reparameterized sampling is completed by the following formula:

[0062] ;

[0063] in is the sampled latent vector, is random noise sampled from a standard normal distribution, is the square root of the variance.

[0064] In this embodiment, the following steps are also included between the RFID tag structured features and the attention weighting: the RFID tag structured features are input into the feature decomposition module for subspace decomposition to obtain multi-scale feature components, the feature decomposition module uses wavelet transform to decompose the feature space into 8 different frequency subbands; the multi-scale feature components are adaptively thresholded to obtain a noise threshold parameter matrix, the adaptive threshold learning constructs a loss function based on the feature statistical distribution, and the threshold parameters are optimized by back propagation; the multi-scale feature components and the noise threshold parameter matrix are input into the soft threshold processing unit for feature filtering to obtain a denoised feature map, and the soft threshold processing uses a differentiable sigmoid function to replace the traditional hard threshold; the denoised feature map is feature reconstructed to obtain preliminary purified features, and the feature reconstruction is performed. The inverse wavelet transform is used to reorganize multi-scale features into a unified feature space; the preliminary purified features are input into the feature enhancement network for non-negative constraint reconstruction to obtain enhanced feature representation. The feature enhancement network contains 3 fully connected layers, each followed by a PReLU activation function; an adaptive gain matrix is constructed based on the enhanced feature representation to obtain the feature gain coefficient. The adaptive gain matrix calculates the importance score of each feature component through the attention mechanism; the enhanced feature representation and the feature gain coefficient are adaptively enhanced to obtain the final purified features. The adaptive enhancement uses element-by-element multiplication to amplify the features; the final purified features are input into the feature calibration module for distribution alignment to obtain the calibrated RFID tag structured features. The distribution alignment uses batch renormalization technology to ensure the stability of the feature distribution.

[0065] In a specific embodiment, the process of executing step 200 may specifically include the following steps:

[0066] Perform multi-head attention decomposition on the structured features of RFID tags to obtain multiple sets of attention feature subspaces. The multi-head attention decomposition includes 8 parallel attention heads, and each attention head outputs a 64-dimensional feature subspace.

[0067] Multiple sets of attention feature subspaces are input into a multilayer perceptron for nonlinear transformation to obtain a feature mapping matrix. The multilayer perceptron includes two hidden layers with dimensions of 512 and 256 respectively. Each layer uses the LeakyReLU activation function.

[0068] Calculate the attention score based on the feature mapping matrix to obtain the feature weight coefficient. The attention score is calculated by dot product operation and softmax normalization function;

[0069] Perform weighted summation on the RFID tag structured features and feature weight coefficients to obtain the global context vector. The weighted summation is implemented using matrix multiplication.

[0070] The global context vector is input into the residual connection module for feature fusion to obtain a fused feature vector. The residual connection module includes a layer normalization layer and a skip connection.

[0071] The fusion feature vector is subjected to dimension reduction transformation to obtain the global feature vector of the RFID tag. The dimension reduction transformation uses a 1×1 convolutional layer to reduce the feature dimension to 128 dimensions.

[0072] Specifically, the structural features of the RFID tag are represented as a matrix ,in represents the number of features, Indicates the dimension of each feature. In order to improve the model’s learning ability for different feature dimensions, a multi-head attention mechanism is used to Decompose. Multi-head attention includes parallel attention heads, each of which performs an independent linear transformation on the input features and generates query ,key Sum matrix:

[0073] ;

[0074] in It is The linear transformation weight matrix of the attention heads, Represents the dimension of the feature subspace output by each attention head. Through linear transformation, the input features are projected into multiple sets of subspaces to capture feature relationships of different dimensions. The dot product operation is performed on the query and key matrices of each attention head and normalized by the softmax function to calculate the attention score. :

[0075] ;

[0076] in Indicates the The correlation weight matrix between features in the attention head, is a scaling factor used to stabilize gradient updates. The attention scores are then combined with the value matrix Multiply them together to get the weighted feature representation:

[0077] ;

[0078] in It is The output feature subspace of the attention heads. The output of all attention heads is concatenated and linearly transformed to obtain the output of the multi-head attention:

[0079] ;

[0080] in Represents the features after splicing, Is the output weight matrix. The features output by multi-head attention The input is sent to the multilayer perceptron for nonlinear transformation to generate a feature map matrix. The multilayer perceptron includes two hidden layers with dimensions of 512 and 256 respectively. Each layer uses the LeakyReLU activation function, which is mathematically expressed as:

[0081] ;

[0082] in It is The input features of the layer, and are the weight matrix and bias vector respectively, and It is Layer and The feature dimension of the layer. After processing by the multi-layer perceptron, the feature mapping matrix can capture the nonlinear relationship of the input features. Based on the feature mapping matrix , calculate the attention score through dot product operation and softmax function :

[0083] ;

[0084] in Is the feature weight coefficient matrix. and weight coefficient Perform weighted summation to generate the global context vector:

[0085] ;

[0086] in is the fused global feature representation. In order to optimize the feature representation, the global context vector The input is sent to the residual connection module for feature fusion. The residual connection module contains layer normalization and skip connection, which is implemented by the following formula:

[0087] ;

[0088] in is the fused feature vector, LayerNorm It is a layer normalization operation used to eliminate distribution offset. Perform dimensionality reduction transformation to reduce the feature dimension to 128. Dimension reduction is achieved by The convolution operation is expressed as:

[0089] ;

[0090] in and is the convolution weight and bias. After the dimension reduction transformation, the final feature vector The dimension is Global feature representation of RFID tags.

[0091] In a specific embodiment, the process of executing step 300 may specifically include the following steps:

[0092] Deploy multiple sensor nodes in a grid structure to obtain a hierarchical sensor network topology. The sensor nodes include temperature sensors, humidity sensors, light intensity sensors, and motion sensors.

[0093] The hierarchical sensor network topology is divided into master and slave nodes to obtain a data collection hierarchy system, where the master node is responsible for data aggregation and the slave node is responsible for environmental parameter collection;

[0094] Based on the data acquisition hierarchical system, environmental parameters are synchronously sampled to obtain the original environmental data stream, where the environmental parameters include temperature distribution, humidity distribution, light intensity distribution and object movement trajectory information;

[0095] Input the original environment data stream into the data compression module for compression coding to obtain a compressed data packet, and perform error correction coding processing on the compressed data packet to obtain a transmission data stream;

[0096] The transmission data stream is decoded and reconstructed and timestamps and location identifiers are added to obtain the environmental spatiotemporal dataset.

[0097] Specifically, a reasonable spatial partitioning scheme is designed to divide the target area into regular grid cells, each corresponding to one or more sensor nodes. Each sensor node is equipped with a temperature sensor, a humidity sensor, a light intensity sensor, and a motion sensor to collect multi-dimensional environmental parameters. These nodes are physically installed or virtualized logically to form a hierarchical topology. The mathematical model of the grid partitioning is expressed as:

[0098] ;

[0099] in, Represents the network topology, is a collection of sensor nodes, is the communication edge between nodes. The coordinates of the nodes Represents its location within the grid cell. The hierarchical sensor network topology is divided into master and slave nodes. The master node is selected as the central node for data aggregation and transmission, while the slave nodes are responsible for collecting environmental parameters. The master-slave node division is based on factors such as the node's geographic location, computing power, and communication distance, and is achieved through an optimal sensor placement algorithm. The mathematical expression for the master-slave relationship is:

[0100] ;

[0101] in Represents the master node set, Indicates that the master node The set of connected slave nodes, is the number of slave nodes corresponding to each master node. Through the hierarchical mechanism, the master node collects data from the slave nodes to reduce the communication overhead of the entire network. After the master and slave nodes are divided, the environmental parameters are synchronously sampled based on the data collection hierarchy system. The goal of synchronous sampling is to ensure the consistency of data from different nodes in the time dimension and avoid data distortion caused by time delay or frequency inconsistency. Assume that each node is at time Collect data , which is expressed as:

[0102] ;

[0103] in 、 、 and Respectively indicate time The collected temperature, humidity, light intensity and motion parameters are transmitted from the node to the corresponding master node after data collection. The master node performs time alignment and preliminary aggregation on the data to form a multi-dimensional raw environmental data stream. :

[0104] ;

[0105] in The original environment data stream is input into the data compression module for compression coding, which reduces the data transmission bandwidth while maintaining data integrity. Data compression uses lossless compression algorithms (such as Huffman coding) or lossy compression algorithms (such as predictive coding). For a data packet , and its compression process is expressed as:

[0106] ;

[0107] in is the compression function, It is a compressed data packet. To enhance transmission reliability, error correction coding is performed on the compressed data packet to generate a transmission data stream that is resistant to noise interference. Error correction coding uses convolutional coding or low-density parity check coding, and its mathematical description is:

[0108] ;

[0109] in It is a transmission data packet that has been error-corrected. The transmission data stream is sent through the communication network to the data center, where it is decoded and reconstructed. The decoding process includes inverse error correction and decompression operations:

[0110] ;

[0111] in Indicates the reconstructed data packet. To achieve spatiotemporal annotation of data, timestamps and location identifiers are added to the reconstructed data. Timestamps Record sampling time and location identification Corresponding to the node location of the collected data. Get the environmental spatiotemporal dataset :

[0112] ;

[0113] Each record Contains environmental parameters, collection time and node location.

[0114] In a specific embodiment, the process of executing step 400 may specifically include the following steps:

[0115] Perform time alignment on the global feature vectors of RFID tags to obtain an aligned feature sequence. The time alignment is synchronized based on the timestamp information in the environmental spatiotemporal dataset.

[0116] The aligned feature sequence is input into the first feature extraction branch of the deep fusion network, and the environmental spatiotemporal dataset is input into the second feature extraction branch of the deep fusion network for parallel feature extraction to obtain a bimodal feature representation. The first and second feature extraction branches both include 3 convolutional layers and 2 fully connected layers.

[0117] Perform cross-modal interaction calculations on the bimodal feature representation to obtain an interaction feature matrix, which is then input into the feature enhancement module for feature reconstruction to obtain an enhanced feature vector. The feature enhancement module contains multiple residual blocks, each of which consists of two 3×3 convolutional layers and a skip connection.

[0118] Multi-scale feature extraction is performed on the enhanced feature vector to obtain a multi-scale feature pyramid, and the multi-scale feature pyramid is subjected to feature splicing and fusion to obtain multimodal fusion features.

[0119] Specifically, the global feature vector of the RFID tag Perform time alignment processing, where represents the sequence length of the feature vector, Represents feature dimensions. Based on timestamp information in the environmental spatiotemporal dataset , ensuring that RFID features and environmental data are strictly synchronized in the time dimension. Assume that the environmental spatiotemporal data is represented as , each record Include timestamp and corresponding environmental characteristics , time alignment is performed using the following formula:

[0120] ;

[0121] in Indicates the RFID features after alignment, is the target time step, It is the timestamp in the environment data, ensuring that each time point Find the closest environmental data time point Through this process, the aligned RFID feature sequence and the corresponding environmental feature sequence are generated. Input the first feature extraction branch of the deep fusion network and the aligned environment data Enter the second feature extraction branch. Both branches consist of 3 convolutional layers and 2 fully connected layers to extract high-level modal features. For the first branch, feature extraction is performed using the following convolution formula:

[0122] ;

[0123] in Indicates the The input features of the convolution layer, is the convolution kernel weight, is the bias, BN( is the activation function. After three convolutional layers, the generated high-dimensional features are input into the fully connected layer for further processing:

[0124] ;

[0125] in and is the parameter of the fully connected layer. The second branch processes the environmental data similarly to the first branch and also generates high-dimensional feature representations. The features output by the two branches are and These two sets of features are input into the cross-modal interaction module to calculate the interaction feature matrix. The interaction calculation is implemented by the following dot product formula:

[0126] ;

[0127] in Represents the correlation matrix between modalities, which is used to capture the potential interaction between RFID features and environmental features. Input feature enhancement module for reconstruction. The feature enhancement module consists of multiple residual blocks, each of which contains two The update formula of the convolutional layer and a skip connection is:

[0128] ;

[0129] in are input features, and is the convolution weight and bias, and the jump connection ensures the stability of gradient transmission in the deep network. Input the multi-scale feature extraction module, which captures the local and global patterns of features through receptive fields of different sizes and generates a multi-scale feature pyramid. The feature representation of each scale is:

[0130] ;

[0131] in It is The characteristics of the scale, Indicates the pooling window size. Features of each scale are concatenated and fused:

[0132] ;

[0133] Where [;] represents the concatenation operation in the dimension to generate the final multimodal fusion features.

[0134] In a specific embodiment, the process of executing step 500 may specifically include the following steps:

[0135] The multimodal fusion features are segmented into time series to obtain feature sequence segments of fixed length, and the feature sequence segments are input into the sequence encoder for feature encoding to obtain the hidden state sequence. The sequence encoder adopts a bidirectional long short-term memory network structure;

[0136] Perform temporal attention calculation on the hidden state sequence to obtain a weighted context vector, which is then input into the trajectory generator for trajectory decoding to obtain a trajectory sequence representation. The trajectory generator adopts an autoregressive decoding structure and contains 4 layers of Transformer decoding layers.

[0137] The trajectory sequence representation is constrained optimized to obtain a smooth trajectory sequence, where the constraint optimization process includes position continuity constraints and motion smoothness constraints;

[0138] The smoothed trajectory sequence is input into the prediction head network for parameter regression to obtain the item trajectory prediction model. The prediction head network consists of three fully connected layers, and the output layer dimension matches the trajectory parameter dimension.

[0139] Specifically, the multimodal fusion features Input timing segmentation module, where represents the total length of the time step, Represents the dimension of the feature. Split the long time series into fixed-length feature segments so that the sequence encoder can process them more efficiently. Assume that the length of each segment is , segmented by the sliding window method, each sliding step is The segmentation formula is:

[0140] ;

[0141] in Indicates the Characteristic sequence fragments, is the fragment index, the total number of fragments is Time series segmentation ensures that each segment covers a fixed time range. Input the bidirectional long short-term memory network (Bi-LSTM) for sequence encoding to generate a hidden state sequence ,in is the dimension of the hidden state. Through bidirectional processing, the hidden state of each time step contains information from the previous and next time steps, forming a hidden state sequence Perform temporal attention calculation on the hidden state sequence to generate a weighted context vector Temporal attention highlights key features by calculating the importance weight of each time step. The calculation formula is:

[0142] ;

[0143] in It is a hidden state and query vector The correlation score of is a learnable weight matrix is the query vector. The weighted context vector is:

[0144] ;

[0145] in Represents the global context representation of the current sequence segment. The context vector The input trajectory generator is decoded to generate a trajectory sequence representation. The trajectory generator adopts an autoregressive structure and contains 4 layers of Transformer decoding layers. The core of each layer of Transformer decoding is a multi-head attention mechanism and a feedforward network. Its update formula is:

[0146]

[0147] in are query, key, and value matrices respectively, is the attention dimension, is the parameter of the feedforward network. The autoregressive structure gradually generates a trajectory sequence to ensure the dynamic relationship between time steps. The generated trajectory sequence is represented as Enter the constrained optimization module to smooth the trajectory. Position continuity constraints are applied by limiting the displacement between adjacent points:

[0148] ;

[0149] in is the maximum displacement allowed. The motion smoothness constraint is implemented by limiting the change in acceleration:

[0150] ;

[0151] in is the threshold for acceleration change. These constraints optimize the smoothness and physical plausibility of the trajectory sequence. The smoothed trajectory sequence is input into the prediction head network for parameter regression to obtain the object trajectory prediction model. The prediction head network consists of three fully connected layers, and the dimensions of its output layer match the dimensions of the trajectory parameters, such as position coordinates and velocity vectors. The prediction formula is:

[0152] ;

[0153] in are the weights and biases of the fully connected layer, is the activation function.

[0154] In a specific embodiment, the process of executing step 600 may specifically include the following steps:

[0155] Perform coordinate mapping on the target area to obtain the target area grid coordinate matrix;

[0156] Based on the target area grid coordinate matrix, the trajectory parameter sequence output by the object trajectory prediction model is input into the coordinate transformation module for grid projection processing to obtain a grid space trajectory sequence. The trajectory parameter sequence includes the three-dimensional space coordinates and motion state parameters of the object;

[0157] Perform density accumulation calculation on the grid space trajectory sequence to obtain the initial density distribution map, where the density accumulation calculation is based on the residence time and appearance frequency of each item in each grid cell;

[0158] The initial density distribution map is input into the density smoothing module for Gaussian filtering to obtain a dynamic density matrix;

[0159] The dynamic density matrix is normalized and threshold segmented to obtain a real-time object positioning heat map. The grid space trajectory sequence and the dynamic density matrix are subjected to spatiotemporal indexing to obtain a trajectory database. The spatiotemporal index uses a quadtree structure to organize the trajectory data.

[0160] Specifically, the target area is divided into spatial grids to generate a grid coordinate matrix. Assume that the target area is a rectangular space, and its boundary is defined as The area is divided into The size of each grid cell is and The grid coordinate matrix is expressed as:

[0161] ;

[0162] in It is -row, first - The center coordinates of the grid cells. The trajectory parameter sequence output by the object trajectory prediction model is input into the coordinate transformation module for grid projection processing. Contains items at time The three-dimensional position of ,speed and acceleration The goal of grid projection is to map the three-dimensional space coordinates of the object to the two-dimensional grid coordinates. For each trajectory point , find its corresponding grid cell index:

[0163] ;

[0164] in Indicates a round-down operation. is the index position of the trajectory point in the grid. Through the above operation, the three-dimensional trajectory sequence Converted into grid space trajectory sequence Perform density accumulation calculation on the grid space trajectory sequence to generate the initial density distribution map. The core of density accumulation is to count the residence time and frequency of each grid cell. Assume that the object is in the grid cell The residence time in , the total number of occurrences is The formula for density accumulation is:

[0165] ;

[0166] in is a grid cell The initial density distribution is represented as a matrix , where each element reflects the activity intensity of the corresponding grid item. Since the initial density distribution is noisy or non-smooth, it is input into the density smoothing module for Gaussian filtering. Gaussian filtering achieves smoothing by taking a weighted average of the density values in a local neighborhood. The filtering formula is:

[0167] ;

[0168] in is the smoothed density value, are the Gaussian kernel weights, is the standard deviation, is the filter window size. After Gaussian filtering, the dynamic density matrix More smooth and coherent. For dynamic density matrices Normalization and threshold segmentation are performed to generate a real-time object location heat map. The goal of normalization is to map the density value to the interval [0,1]. The normalization formula is:

[0169] ;

[0170] in and They are matrices The normalized density matrix is segmented by threshold to generate a heat map, where the segmentation formula is:

[0171] ;

[0172] in is the set density threshold, The generated binary heat map shows high-density activity areas in the target area. The grid space trajectory sequence and the dynamic density matrix are indexed in a spatiotemporal manner to form a trajectory database. The spatiotemporal index is organized using a quadtree structure, where each node represents a spatial range and stores the trajectory points and density values within that range. The quadtree is constructed by recursively partitioning the space:

[0173] ;

[0174] in The current region The trajectory database supports efficient range queries and spatiotemporal analysis through the quadtree.

[0175] In this embodiment, obtaining the real-time item location heat map and trajectory database further includes the following steps: performing density analysis on the real-time item location heat map to obtain the RFID reader deployment position matrix, and the density analysis calculates the optimal coverage point set based on the item distribution density map; inputting the RFID reader deployment position matrix into the path planning module to generate an initial path to obtain an initial inspection path set, and the initial path generation uses an improved ant colony algorithm to plan a path for each RFID reader; dynamically updating the initial inspection path set to obtain an optimized path sequence, and the dynamic update process includes two goals: path length optimization and coverage balance; inputting the optimized path sequence into the evolutionary algorithm module for first-stage optimization to obtain a local optimal path sequence. The evolutionary algorithm adopts adaptive crossover and mutation operations with a population size of 100. The local optimal path is optimized in the second stage to obtain the global optimal path. The second stage optimization introduces a taboo search strategy to avoid falling into the local optimum. The workload of the RFID reader is calculated based on the global optimal path to obtain the load distribution matrix. The workload includes path length, coverage area and acquisition frequency. The load distribution matrix is balanced in real time to obtain the task allocation scheme. The balance calculation evenly distributes the workload through the minimum-maximum criterion. The task allocation scheme is converted into motion control instructions for the RFID reader to obtain a multi-agent collaborative scheduling strategy. The motion control instructions include speed, direction and sampling interval.

[0176] The above describes the RFID smart label positioning tracking printing method in the embodiment of the present application. The following describes the RFID smart label positioning tracking printing system 10 in the embodiment of the present application. Figure 2 In one embodiment of the present application, an RFID smart label positioning tracking printing system 10 includes:

[0177] The convolution processing module 11 is used to input the RFID tag radio frequency signal into the conditional variational autoencoder for convolution processing to obtain the RFID tag structured features;

[0178] An attention weighting module 12 is used to perform attention weighting on the structured features of the RFID tag to obtain a global feature vector of the RFID tag;

[0179] The multi-dimensional acquisition module 13 is used to establish a hierarchical sensor node network to perform multi-dimensional acquisition of environmental parameters to obtain an environmental spatiotemporal data set;

[0180] The feature fusion module 14 is used to input the RFID tag global feature vector and the environmental spatiotemporal dataset into the deep fusion network for feature fusion to obtain multimodal fusion features;

[0181] A sequence modeling module 15 is used to perform sequence modeling based on multimodal fusion features to obtain an item trajectory prediction model;

[0182] The distribution calculation module 16 is used to apply the object trajectory prediction model to the target area grid coordinate system to perform density distribution calculation to obtain a real-time object positioning heat map and trajectory database.

[0183] Through the collaborative cooperation of the above components, by introducing conditional variational autoencoders and attention mechanisms, deep feature extraction and weighted processing of RFID radio frequency signals are performed, which improves the robustness and discriminability of feature expression and effectively overcomes the signal interference problem in complex environments; adopting a layered sensor network architecture and a multi-dimensional environmental parameter acquisition strategy, comprehensive perception and real-time monitoring of the environmental status of the target area are achieved, providing rich environmental context information for object positioning; a deep fusion network is designed for multimodal feature fusion, and through dual-branch feature extraction and cross-modal interaction, RFID signal features and environmental parameter information are effectively integrated, enhancing the environmental adaptability of the system; a trajectory prediction method based on sequence modeling and constrained optimization, combined with a temporal attention mechanism and an autoregressive decoding structure, achieves accurate prediction and smoothing of the object's motion trajectory; through gridded density distribution calculation and spatiotemporal index construction, an efficient object positioning heat map generation and trajectory data management mechanism is established, facilitating the system to perform real-time location visualization and historical trajectory query.

[0184] Those skilled in the art will 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.

[0185] 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, 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 and includes several instructions for enabling an electronic 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, and other media that can store program code.

[0186] 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above 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 method for printing RFID smart label positioning and tracking, characterized in that: The method comprises: The RFID tag radio frequency signal is input into the conditional variational autoencoder for convolution processing to obtain the RFID tag structured features; specifically comprising: inputting the RFID tag radio frequency signal into the feature extraction layer of the conditional variational autoencoder for convolution operation to obtain a multi-scale feature map, wherein the feature extraction layer includes 5 parallel convolution kernels, each convolution kernel has a step size of 2, and each convolution layer is connected to a batch normalization layer and a ReLU activation function; performing a maximum pooling operation on the multi-scale feature map to obtain a reduced dimension feature map, wherein the maximum pooling operation adopts a 2×2 pooling window with a step size of 2; inputting the reduced dimension feature map into the fully connected layer of the conditional variational autoencoder for dimensionality conversion to obtain a fixed-dimensional feature vector , the fully connected layer includes three hidden layers, and the number of neurons in each hidden layer is 512, 256, and 128 respectively; applying adaptive weights to the fixed-dimensional feature vector to obtain a weighted feature vector, and the adaptive weights are dynamically calculated based on the feature importance score; inputting the weighted feature vector into the encoder of the conditional variational autoencoder for compression mapping to obtain a latent space feature, and the encoder includes a mean encoder and a variance encoder, which respectively output the mean vector and variance vector of the latent space; reparameterizing and sampling the latent space feature to obtain the RFID tag structured feature, wherein the reparameterizing sampling generates sampling points that obey the standard normal distribution based on the mean vector and the variance vector; Performing attention weighting on the structured features of the RFID tag to obtain a global feature vector of the RFID tag; Establish a hierarchical sensor node network to collect multi-dimensional environmental parameters and obtain an environmental spatiotemporal data set; Inputting the RFID tag global feature vector and the environmental spatiotemporal dataset into a deep fusion network for feature fusion to obtain a multimodal fusion feature; Perform sequence modeling based on the multimodal fusion features to obtain an item trajectory prediction model; The object trajectory prediction model is applied to the target area grid coordinate system to perform density distribution calculation to obtain a real-time object location heat map and trajectory database.

2. The RFID smart label positioning tracking printing method according to claim 1, characterized in that: The step of performing attention weighting on the structured features of the RFID tag to obtain a global feature vector of the RFID tag includes: Performing multi-head attention decomposition on the structured features of the RFID tag to obtain multiple groups of attention feature subspaces, wherein the multi-head attention decomposition includes 8 parallel attention heads, and each attention head outputs a 64-dimensional feature subspace; Inputting the multiple sets of attention feature subspaces into a multilayer perceptron for nonlinear transformation to obtain a feature mapping matrix, wherein the multilayer perceptron includes two hidden layers with hidden layer dimensions of 512 and 256 respectively, and each layer uses a LeakyReLU activation function; Calculating an attention score based on the feature mapping matrix to obtain a feature weight coefficient, wherein the attention score is calculated by a dot product operation and a softmax normalization function; Performing a weighted sum operation on the RFID tag structured features and the feature weight coefficients to obtain a global context vector, wherein the weighted sum operation is implemented using matrix multiplication; Inputting the global context vector into a residual connection module for feature fusion to obtain a fused feature vector, wherein the residual connection module includes a layer normalization layer and a skip connection; A dimension reduction transformation is performed on the fused feature vector to obtain a global feature vector of the RFID tag. The dimension reduction transformation uses a 1×1 convolutional layer to reduce the feature dimension to 128 dimensions.

3. The RFID smart label positioning tracking printing method according to claim 2, characterized in that: The method of establishing a hierarchical sensor node network to collect environmental parameters in multiple dimensions and obtain an environmental spatiotemporal data set includes: Deploy multiple sensor nodes in a grid structure to obtain a hierarchical sensor network topology, wherein the sensor nodes include temperature sensors, humidity sensors, light intensity sensors, and motion sensors; The hierarchical sensor network topology is divided into master and slave nodes to obtain a data collection hierarchical system, where the master node is responsible for data aggregation and the slave node is responsible for environmental parameter collection; Synchronously sampling environmental parameters based on the data acquisition hierarchical system to obtain an original environmental data stream, wherein the environmental parameters include temperature distribution, humidity distribution, light intensity distribution, and object movement trajectory information; Inputting the original environment data stream into a data compression module for compression coding to obtain a compressed data packet, and performing error correction coding processing on the compressed data packet to obtain a transmission data stream; The transmission data stream is decoded and reconstructed and a timestamp and a location identifier are added to obtain an environmental spatiotemporal data set.

4. The RFID smart label positioning tracking printing method according to claim 3, characterized in that: The step of inputting the RFID tag global feature vector and the environmental spatiotemporal dataset into a deep fusion network for feature fusion to obtain a multimodal fusion feature includes: Performing time alignment processing on the RFID tag global feature vector to obtain an alignment feature sequence, wherein the time alignment processing is synchronized based on timestamp information in the environmental spatiotemporal dataset; Inputting the aligned feature sequence into a first feature extraction branch of a deep fusion network, and inputting the environmental spatiotemporal dataset into a second feature extraction branch of the deep fusion network for parallel feature extraction to obtain a bimodal feature representation, wherein both the first feature extraction branch and the second feature extraction branch include three convolutional layers and two fully connected layers; Performing cross-modal interaction calculation on the bimodal feature representation to obtain an interaction feature matrix, and inputting the interaction feature matrix into a feature enhancement module for feature reconstruction to obtain an enhanced feature vector, wherein the feature enhancement module includes multiple residual blocks, each of which is composed of two 3×3 convolutional layers and a skip connection; Multi-scale feature extraction is performed on the enhanced feature vector to obtain a multi-scale feature pyramid, and feature splicing and fusion are performed on the multi-scale feature pyramid to obtain multimodal fusion features.

5. The RFID smart label positioning tracking printing method according to claim 4, characterized in that: The method of performing sequence modeling based on the multimodal fusion features to obtain an item trajectory prediction model includes: The multimodal fusion features are temporally segmented to obtain feature sequence segments of fixed length, and the feature sequence segments are input into a sequence encoder for feature encoding to obtain a hidden state sequence. The sequence encoder adopts a bidirectional long short-term memory network structure; Perform temporal attention calculation on the hidden state sequence to obtain a weighted context vector, and input the weighted context vector into a trajectory generator for trajectory decoding to obtain a trajectory sequence representation. The trajectory generator adopts an autoregressive decoding structure and includes 4 layers of Transformer decoding layers; performing constrained optimization on the trajectory sequence representation to obtain a smooth trajectory sequence, wherein the constrained optimization process includes position continuity constraints and motion smoothness constraints; The smoothed trajectory sequence is input into the prediction head network for parameter regression to obtain an item trajectory prediction model. The prediction head network consists of three fully connected layers, and the output layer dimension matches the trajectory parameter dimension.

6. The RFID smart label positioning tracking printing method according to claim 5, characterized in that: The method of applying the object trajectory prediction model to the target area grid coordinate system to perform density distribution calculation to obtain a real-time object location heat map and trajectory database includes: Perform coordinate mapping on the target area to obtain the target area grid coordinate matrix; Based on the target area grid coordinate matrix, the trajectory parameter sequence output by the item trajectory prediction model is input into a coordinate transformation module for grid projection processing to obtain a grid space trajectory sequence, wherein the trajectory parameter sequence includes the three-dimensional space coordinates and motion state parameters of the item; Performing density accumulation calculation on the grid space trajectory sequence to obtain an initial density distribution map, wherein the density accumulation calculation is based on the residence time and appearance frequency of each item in each grid cell; Inputting the initial density distribution map into a density smoothing module for Gaussian filtering to obtain a dynamic density matrix; The dynamic density matrix is normalized and threshold segmented to obtain a real-time object positioning heat map, and the grid space trajectory sequence and the dynamic density matrix are subjected to spatiotemporal index construction to obtain a trajectory database, wherein the spatiotemporal index adopts a quadtree structure to organize trajectory data.

7. An RFID smart label positioning tracking printing system, characterized in that: Used to execute the RFID smart label positioning tracking printing method according to any one of claims 1 to 6, the RFID smart label positioning tracking printing system comprises: The convolution processing module is used to input the RFID tag radio frequency signal into the conditional variational autoencoder for convolution processing to obtain the RFID tag structured features; an attention weighting module, configured to perform attention weighting on the structured features of the RFID tag to obtain a global feature vector of the RFID tag; The multi-dimensional acquisition module is used to establish a hierarchical sensor node network to collect environmental parameters in multiple dimensions and obtain an environmental spatiotemporal data set; A feature fusion module is used to input the RFID tag global feature vector and the environmental spatiotemporal dataset into a deep fusion network for feature fusion to obtain a multimodal fusion feature; A sequence modeling module, configured to perform sequence modeling based on the multimodal fusion features to obtain an item trajectory prediction model; The distribution calculation module is used to apply the item trajectory prediction model to the target area grid coordinate system to perform density distribution calculation to obtain a real-time item positioning heat map and trajectory database.

Images