Intelligent shelf control method and system based on RFID

Abstract

The invention relates to an RFID-based intelligent shelf control method and system, and the method comprises the steps: firstly building a library according to material characteristics, obtaining a pure RFID signal through multi-band signal collection and filtering processing, and judging the binding of materials and labels based on a deep learning multi-label recognition model; a triangulation positioning algorithm is combined with a signal attenuation model for accurate positioning, and a dynamic shelf mapping algorithm is used for updating the adjusted material position. And an edge calculation signal processing module is further designed to realize rapid signal processing, a software and hardware collaborative optimization framework is constructed, and system performance is improved by a self-adaptive optimization model based on reinforcement learning. The goods and materials can be intelligently managed and tracked in real time, and the goods shelf control efficiency and precision are improved.

Description

Technical Field

[0001] The invention belongs to the field of information technology, and in particular relates to an intelligent shelf control method and system based on RFID. Background Art

[0002] In the intelligent layer shelf system, it is necessary to achieve accurate binding and inductive identification of materials and RFID tags. Due to the complex storage environment, the materials, shapes, and placement of different materials vary greatly, and a variety of different materials may be stored on the same shelf, resulting in severe challenges to the accuracy and stability of RFID inductive identification. In traditional shelf management, the identification, positioning, and management of materials mostly rely on manual operations, which are inefficient, low in precision, and difficult to achieve real-time tracking. At the same time, it becomes more difficult to accurately identify material types, bind tags, and accurately locate materials when multiple materials are mixed, placed in various locations, and there is environmental interference. In addition, when the shelf structure is adjusted, updating the material location information is also challenging, and it is difficult for the system to dynamically optimize performance based on different needs, and it cannot meet the requirements of modern warehousing for intelligence, efficiency, precision, and real-time performance.

[0003] Therefore, how to achieve rapid and reliable identification and positioning tracking of materials in complex and ever-changing actual scenarios, and adapt to changes in shelf adjustments, is a key technical problem that needs to be solved urgently. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide an RFID-based intelligent shelf control method and system to realize automatic identification and precise positioning of materials, overcome the influence of environmental interference on RFID signals, accurately judge the binding relationship between materials and tags, and effectively deal with the position update problem caused by shelf adjustment. Through the coordinated optimization of software and hardware and reinforcement learning and other technologies, the performance and real-time processing capabilities of the system are improved to realize efficient, precise and real-time control and management of intelligent shelf.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: An intelligent shelf control method based on RFID, the steps are as follows: S1. Based on the differences in material and shape of materials, a material feature library is pre-established to obtain the material reflectance coefficient and shape profile data of each material, and the material type is determined by matching the feature library, including: S1.1. Based on the differences in material and shape of materials, a feature database containing multiple material types is pre-established, and each material type corresponds to a set of material reflection coefficient and shape profile data; S1.2. Use an optical sensor to obtain the reflectance spectrum data of the surface of the material to be identified, and obtain the material reflectance coefficient of the material by analyzing the peaks and troughs of the spectrum data; S1.3. Use machine vision technology to image the material to be identified, extract the contour area of ​​the material through the image segmentation algorithm, and obtain the contour data representing the shape of the material; S1.4. Input the acquired material reflectance coefficient and shape profile data into the material feature database, and use the feature similarity matching algorithm to calculate the similarity between the material to be identified and each material type in the database; Feature similarity matching algorithm formula: ; in: : Indicates the similarity between the feature A of the material to be identified and the feature B of the material in the database.

[0006] In the present invention, A can be a feature vector of the material to be identified, which includes the feature representation of the material reflectance and shape profile data of the material, such as ,in It can be a characteristic dimension of the material reflectance. It can be a dimension of shape contour data, etc.; B is the feature vector of a material type in the material feature library, for example .

[0007] : represents the value of the i-th dimension of the feature vector A of the material to be identified. For example, if A contains multiple feature dimensions of material reflectance and shape contour data, It can be a quantitative indicator of the material reflection coefficient. It can be a quantitative indicator of shape contour data (such as length, width, etc.).

[0008] : represents the value of the i-th dimension of a material feature vector B in the database. Similarly, for the material feature vector in the material feature library, , Etc. are quantitative indicators of their corresponding dimensions.

[0009] n: represents the dimension of the feature vector. According to the present invention, it is the total number of feature dimensions of the material reflectance coefficient and shape contour data of the material under consideration, such as multiple indicators of the material reflectance coefficient and multiple indicators of the shape contour data, totaling n dimensions.

[0010] S1.5. Set a similarity threshold. If the similarity between the material to be identified and a certain material type is greater than the threshold, the material is judged to belong to that type, otherwise it is judged to be an unknown type; S1.6. If the material to be identified is determined to be of unknown type, its material reflectance coefficient and shape profile data are added to the material feature database as a new material type, completing the automatic update of the database; S1.7. According to the identification results of material types, the materials to be identified are classified and stored or processed subsequently to realize intelligent management of materials; S2. In view of the placement and environmental interference, multi-band RFID signal acquisition technology is used to obtain signal strength data of different frequency bands, and the environmental noise is removed through the signal filtering algorithm to obtain a pure RFID signal, including: S2.1. Determine the frequency range of the RFID signal to be collected according to the placement of the RFID tag, and set up the corresponding signal receiving antenna and receiver for different frequency bands; S2.2. Use multi-band RFID signal acquisition technology to simultaneously receive RFID signals from multiple frequency bands and obtain RFID tag signal strength data in different frequency bands; S2.3. Preprocess the acquired multi-band RFID signal strength data to remove obvious outliers and invalid data, and obtain a preliminary filtered signal strength data set; S2.4. According to the characteristics of environmental interference, an environmental noise model is established, and an adaptive signal filtering algorithm is used to automatically adjust the filter parameters to dynamically remove the impact of environmental noise on RFID signals; Adaptive signal filtering algorithm formula: ; Where: y[n]: represents the value of the pure RFID signal after adaptive filtering at time n.

[0011] In the present invention, n can be a moment in the time series of collecting RFID signals, and y[n] is the signal strength after filtering at that moment. For example, in the process of collecting RFID signals, for the nth time point, the pure signal strength is obtained after filtering.

[0012] x[n]: represents the value of the input noisy RFID signal at time n. This is the original signal obtained from the multi-band RFID signal acquisition technology, and the signal strength value of different frequency bands at time n.

[0013] : represents the kth filter coefficient at time n. In the present invention, according to the characteristics of environmental interference, It will automatically adjust according to the adaptive signal filtering algorithm to remove environmental noise at different times in different frequency bands. k represents the filter order index, and M represents the filter order. For example, it can be determined based on experience or adaptive algorithms as , are the values ​​of filter coefficients of different orders at time n.

[0014] S2.5. After being processed by the adaptive signal filtering algorithm, the pure RFID signal strength data after removing the environmental noise is obtained, which is used as the input data for subsequent tag positioning and identification; S2.6. Use a multi-band RFID tag positioning algorithm based on signal strength, comprehensively utilize the signal strength data of different frequency bands, and calculate the spatial position coordinates of the RFID tag through methods such as triangulation or fingerprint matching; S2.7. Match the RFID tag location coordinates obtained by positioning with the pre-established environmental map to determine the specific location of the tag, such as the shelf number and area name, to complete the RFID tag positioning and identification process; S3. Based on the scenario of multiple materials, a multi-label recognition model based on deep learning is designed. The pure RFID signal and material feature library data are input to determine the binding relationship between materials and labels, and the binding results are output, including: S3.1. Obtain RFID signal data and material feature library data in a multi-material mixed scenario as input to the deep learning multi-label recognition model; S3.2. Preprocess the acquired RFID signal data to remove noise and interference and extract pure RFID signal features; S3.3. Match the preprocessed RFID signal features with the material feature database data, and find the material features that best match each RFID tag through similarity calculation; S3.4. Based on the matching results, the support vector machine algorithm is used to perform multi-label classification on the materials and RFID tags to obtain a preliminary binding relationship between the materials and the tags; Support vector machine algorithm formula: ; in: : is a classification decision function used to determine the category of input x. In the present invention, x is a pre-processed RFID signal feature, which can be a feature vector extracted from a pure RFID signal, such as ,in These are the various features extracted from the signal.

[0015] : The Lagrange multiplier is a parameter obtained by solving the quadratic programming problem. It is a parameter obtained when training a support vector machine classifier based on a multi-material mixed scene. Its value depends on the selection of training data and kernel function.

[0016] : represents training samples The category label is usually or 1. In the present invention, It can represent a binding relationship between a material and a label, and 1 represents another binding relationship, such as marking different materials and labels according to the material feature library data.

[0017] : Kernel function, used to map samples from low-dimensional space to high-dimensional space. Common kernel functions include linear kernel , polynomial kernel , Radial Basis Kernel In the present invention, for pure RFID signal features and material feature library data, a suitable kernel function can be selected according to the distribution and features of the data. For example, when using the radial basis kernel, is a sample feature vector of a material feature and label binding in the training sample, x is the sample feature vector to be classified, is the parameter of the kernel function and can be adjusted according to the data.

[0018] : The bias term is a parameter obtained by solving a quadratic programming problem and is an important parameter of a support vector machine classifier and is determined during the training process of the multi-label recognition model of the present invention.

[0019] N: the number of training samples. In the present invention, it is the number of samples in a multi-material mixed scenario used to train the support vector machine. These samples contain the characteristics and label information of the binding relationship between different materials and labels.

[0020] S3.5. Using the convolutional neural network algorithm, with material characteristics and RFID signal characteristics as input, further judge and optimize the binding relationship between materials and tags; Convolutional neural network algorithm formula: ; in: : represents the output feature vector of the lth layer. In the multi-label recognition model based on deep learning of the present invention, for the input pure RFID signal and material feature library data, is the output of the network at layer l, for example, in the first layer When , the input is the preprocessed RFID signal features and material features, which are obtained after convolution and activation function. .

[0021] : Activation function, the most common one is Sigmoid function , ReLU function Etc. According to the performance requirements of multi-label recognition in the present invention, a ReLU function can be selected to avoid the gradient vanishing problem.

[0022] : represents the connection weight from the ith input neuron to the jth output neuron in the lth layer. In the present invention, these weights are learned when training a convolutional neural network for multi-label recognition. The weights of different layers are continuously updated according to the network structure and training data. For example, for a simple network structure, the weight from the input layer to the first hidden layer is It will be adjusted based on the training data and optimization algorithm.

[0023] : represents the bias term of the jth neuron in the lth layer, which is adjusted together with the weights when training the convolutional neural network to improve the recognition accuracy of the binding relationship between materials and labels.

[0024] :Indicates the The number of neurons in each layer is determined according to the number of designed network layers and the number of neurons in each layer. For example, the input layer has The number of neurons depends on the feature dimension and representation of the input pure RFID signal and material feature library data.

[0025] S3.6. By integrating the judgment results of the support vector machine and the convolutional neural network, the matching degree between the material characteristics and the RFID signal is comprehensively considered to determine the final binding relationship between the material and the tag; S3.7. Output the binding results of materials and labels to complete the label recognition and binding tasks in the mixed scene of multiple materials; S4. To meet the needs of positioning and tracking, a triangulation positioning algorithm based on signal strength is used to obtain the signal strength data of multiple RFID readers and writers, and the precise location of the materials is determined by calculating the signal attenuation model, including: S4.1. Obtain signal strength data of at least three readers according to a pre-established RFID reader layout; S4.2. Based on the acquired signal strength data, a triangulation positioning algorithm is used to calculate the approximate location coordinates of the target material; Triangulation positioning algorithm formula: ; in: : represents the two-dimensional plane position coordinates of the target material. In the present invention, this is the position of the material to be finally determined on the shelf plane, such as the x-axis and y-axis coordinates of the shelf.

[0026] : Respectively represent the distance values ​​after the signal strength conversion of the target material received by three different RFID readers (calculated according to the signal attenuation model). In the present invention, is the distance from the material to the first RFID reader calculated based on the received signal strength and signal attenuation model. Assuming that the first reader is located at position , the received signal strength is , through the signal attenuation model (See the signal attenuation model section for details) Calculate the distance , similarly, and .

[0027] :represent the known distances between the three RFID readers. These distances can be determined based on the preset RFID reader layout. For example, the coordinates of the three readers are , and ,but , similarly we can calculate and .

[0028] S4.3. Calculate the distance between the target material and each reader based on the approximate location coordinates obtained by the triangulation positioning algorithm and the preset signal attenuation model; Signal attenuation model formula: ; Where: r: represents the signal strength received at a distance d. In the present invention, it is the signal strength received from multiple RFID readers, for example, in the positioning and tracking requirements, the signal strength received by a reader at a distance d from the material.

[0029] : Indicates the reference distance (usually the distance from the emission source In the present invention, the signal strength at the reference distance from the RFID tag can be measured in advance. The signal strength measured at a distance of 1 meter is used as .

[0030] : represents path loss. Common path loss models are: , in is the reference distance The path loss at the location can be determined based on experience or field testing in the present invention. The value of; n is the path loss index, which is determined according to the environment where the material is located (such as whether there is metal shielding, signal transmission medium, etc.). It is a Gaussian random variable with a mean of zero, representing the random factors in the environment. In the present invention, its variance can be estimated according to the characteristics of the environmental noise.

[0031] S4.4. If the calculated distance value is less than the preset threshold, the location is determined to be the precise location of the target material; S4.5. If it is greater than the threshold, return to step 1 and continue to obtain signal strength data of more readers; S4.6. Train historical positioning data through machine learning algorithms to obtain optimized signal attenuation model parameters to improve the accuracy of subsequent positioning; S4.7. Bind the precise location coordinates to the material ID, update the material location information in real time, and store it in the database; S4.8. When the location of a specific material needs to be queried, the latest location coordinates of the material are obtained from the database and visualized on the map to achieve real-time positioning and tracking; S5. According to the shelf adjustment changes, a dynamic shelf mapping algorithm is designed to obtain the spatial coordinate data after the shelf structure is adjusted, and the material location information is updated through coordinate transformation, including: S5.1. According to the shelf adjustment changes, obtain the adjusted shelf structure data and spatial coordinate data; S5.2. Use a dynamic mapping algorithm to establish a mapping relationship between shelf structures and coordinates before and after adjustment; S5.3. Calculate the new coordinate position of the material in the adjusted shelf structure through coordinate transformation; S5.4. Update the material location information database according to the new coordinates of the material; S5.5. If the location of the material changes, the automatic update mechanism is triggered to synchronize the updated location information to the relevant systems; S5.6. Obtain updated material location information and generate shelf adjustment change reports to facilitate inventory management; S5.7. Through visualization technology, the material distribution before and after shelf adjustment is displayed in real time, supporting dynamic query and positioning of materials; S6. Based on the real-time requirements, design a signal processing module based on edge computing to obtain real-time signal data from RFID readers and writers, and realize fast signal processing through a distributed computing framework, including: S6.1. Obtain real-time signal data collected by the RFID reader and transmit the data to the edge computing node for preprocessing; S6.2. Perform parallel computation on the preprocessed signal data through a distributed computing framework, and use the fast Fourier transform algorithm to perform frequency domain analysis on the signal to obtain the frequency characteristics of the signal; Fast Fourier transform algorithm formula: ; Where: X[k]: represents the amplitude and phase of the kth frequency component in the frequency domain. In the present invention, for the real-time signal data collected from the RFID reader, X[k] represents the information of different frequency components obtained after fast Fourier transform, which can be used for subsequent analysis of whether the signal contains valid RFID tag information.

[0032] x[n]: represents the value of the discrete signal in the time domain at the nth sampling point, which is the signal strength value of the real-time signal collected from the RFID reader at discrete time point n.

[0033] N: represents the number of sampling points, which is determined according to the duration of the signal acquisition and the sampling frequency in the present invention. For example, if a signal of 1 second is acquired at a sampling frequency of 1000 Hz, then .

[0034] j : imaginary unit, .

[0035] S6.3. According to the preset frequency characteristic threshold, determine whether the signal contains valid RFID tag information, if it contains, extract the tag ID, otherwise discard the signal data; S6.4. For the extracted RFID tag ID, obtain the corresponding item information by querying the pre-established mapping table of tag ID and item information; S6.5. Associating the acquired item information with the corresponding timestamp and reader ID to generate structured item tracking data; S6.6. Use incremental learning algorithms to analyze item tracking data in real time and determine the current location and status of items based on their movement trajectory and status changes; Incremental learning algorithm formula: ; in: : represents the updated model parameters. In the present invention, the incremental learning algorithm is used to perform real-time analysis on the item tracking data. It can be a model parameter used to describe the movement trajectory and state change of an object, such as the updated value of parameters such as the object's movement speed and position offset.

[0036] : Represents the current model parameters, the parameters of the object tracking model at different time points t, such as the initial speed estimation, position estimation, etc.

[0037] : learning rate, controlling the step size of the update. In the present invention, the learning rate can be adjusted according to the movement characteristics of the object and the data update frequency. For example, for fast-moving items, a larger learning rate can be used, and for slow-moving items, a smaller learning rate can be used.

[0038] : represents the gradient calculated at time t. In the present invention, the gradient of the object position and state, such as the rate of change of the object position over time, can be calculated according to the characteristics and analysis objectives of the object tracking data.

[0039] S6.7. Push the analyzed item location and status information to upper-level applications in real time for business systems to make decisions and optimize, thus achieving real-time tracking and management of items; S7. Based on the algorithm design and hardware design, build a software and hardware collaborative optimization framework, obtain algorithm processing results and hardware performance data, and optimize system operation efficiency through resource scheduling strategies, including: S7.1. Based on the design principles of the hardware-software collaborative optimization framework, a comprehensive optimization system is constructed by combining algorithm design and hardware design. S7.2. Obtain algorithm processing result data and hardware operation performance data through the system and use them as input for resource scheduling strategy optimization; S7.3. Based on the obtained algorithm results and hardware performance data, use a machine learning algorithm to optimize the resource scheduling strategy to obtain an optimized resource scheduling strategy; S7.4. Apply the optimized resource scheduling strategy to the hardware-software co-optimization framework to improve system operation efficiency by dynamically adjusting the allocation of system resources; S7.5. During system operation, continuously monitor changes in algorithm processing results and hardware performance data. If system operation efficiency decreases, re-optimization of resource scheduling strategies is triggered; S7.6. Dynamically adjust the allocation of software and hardware resources according to the optimized resource scheduling strategy to ensure that the system operation efficiency is maintained at a high level; S7.7. Continue to iterate and optimize the above process, and achieve dynamic optimization of system operation efficiency through adaptive adjustment of the software and hardware collaborative optimization framework to ensure the stability and improvement of system performance; S8. Design an adaptive optimization model based on reinforcement learning to meet the accuracy requirements of sensor recognition and positioning tracking, obtain the error data of recognition and positioning, adjust the algorithm parameters through feedback mechanism, and improve the system performance, including: S8.1. Obtain the error data generated during the sensing and positioning tracking process, use it as the input of the reinforcement learning model, and evaluate the performance of the model through the reward function; S8.2. Dynamically adjust key parameters in the sensing recognition and positioning tracking algorithms, such as the threshold of the feature extraction algorithm and the similarity threshold of the matching algorithm, based on the output results of the reinforcement learning model, to meet the accuracy requirements in different scenarios; S8.3. In the process of adjusting algorithm parameters, an adaptive optimization strategy is used to automatically select appropriate optimization algorithms, such as gradient descent and Newton method, according to the distribution characteristics and change trends of error data to accelerate parameter convergence; S8.4. By setting a reasonable reward function, the reinforcement learning model is guided to optimize in the direction of reducing recognition and positioning errors, while avoiding overfitting and underfitting problems and improving the generalization ability of the model; S8.5. During the training process of the reinforcement learning model, an experience replay mechanism is used to store historical error data in the replay buffer and randomly select samples for training to break the correlation between data and improve the stability of the model. S8.6. Regularly evaluate the performance of the reinforcement learning model, determine whether the model has converged to the optimal state by calculating indicators such as the mean and variance of the recognition and localization errors, and adjust the model structure and hyperparameters if necessary; S8.7. Apply the optimized algorithm parameters to the actual sensing recognition and positioning tracking system, continuously monitor the system performance, collect new error data, form a closed-loop feedback mechanism, and achieve adaptive optimization and continuous improvement of the algorithm; An intelligent shelf control system based on RFID, the system comprising: Material feature library module, used to perform S1 related operations; Multi-band RFID signal acquisition module, used to perform related operations of S2; Multi-tag identification module, used to perform S3 related operations; Triangulation positioning module, used to perform S4 related operations; Dynamic shelf mapping module, used to perform S5 related operations; Edge computing signal processing module, used to perform S6 related operations; Software and hardware collaborative optimization module, used to perform S7 related operations; The adaptive optimization module is used to perform S8 related operations.

[0040] The present invention can achieve the following beneficial effects: 1. The present invention realizes automatic identification and precise positioning of materials, overcomes the influence of environmental interference on RFID signals, accurately determines the binding relationship between materials and tags, effectively copes with the position update problem caused by shelf adjustment, and improves the system's performance and real-time processing capabilities through software and hardware collaborative optimization and reinforcement learning and other technologies, so as to realize efficient, precise and real-time control and management of intelligent shelf shelves.

[0041] 2. Through the automated material identification, positioning and tracking process, manual operations are reduced and the efficiency of shelf management is significantly improved. Technologies such as multi-band signal acquisition, deep learning and signal attenuation model are used to overcome environmental interference, accurately determine material types and binding relationships, achieve accurate positioning of materials, and meet the needs of mixed multi-material scenarios.

[0042] 3. Dynamic shelf mapping algorithm ensures that the material location information can be quickly updated when the shelf is adjusted, and the system can adapt to changes in shelf structure. Based on edge computing signal processing and software and hardware collaborative optimization framework, fast signal processing and optimal configuration of system resources are achieved to improve system operation efficiency.

[0043] 4. Through the adaptive optimization model of reinforcement learning, the algorithm parameters can be adjusted according to the error data to achieve continuous improvement of system performance and ensure high accuracy of sensing recognition and positioning tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 is a flow chart of the method of the present invention; Figure 2 It is the data processing logic diagram of the present invention; Figure 3 This is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0045] The preferred solution is Figures 1 to 3 As shown, a method and system for controlling intelligent shelf racks based on RFID, S1. According to the differences in material and shape of materials, a material feature library is established in advance to obtain the material reflection coefficient and shape profile data of each material, and the material type is determined by matching the feature library.

[0046] According to the differences in the materials and shapes of materials, a feature database containing multiple material types is pre-established, and each material type corresponds to a set of material reflectance coefficients and shape contour data. Optical sensors are used to obtain the reflectance spectrum data of the surface of the material to be identified, and the material reflectance coefficient of the material is obtained by analyzing the peaks, troughs and other features of the spectrum data. Machine vision technology is used to image the material to be identified, and the contour area of ​​the material is extracted through the image segmentation algorithm to obtain the contour data representing the shape of the material. The acquired material reflectance coefficient and shape contour data are input into the material feature database, and the feature similarity matching algorithm is used to calculate the similarity between the material to be identified and each material type in the database. A similarity threshold is set. If the similarity between the material to be identified and a certain material type is greater than the threshold, the material is judged to belong to this type, otherwise it is judged to be an unknown type. If the material to be identified is judged to be an unknown type, its material reflectance coefficient and shape contour data are added to the material feature database as a new material type to complete the automatic update of the database. According to the identification results of the material type, the material to be identified is classified and stored or subjected to subsequent processing operations to realize the intelligent management of materials.

[0047] Specifically, a feature database containing 100 common material types is pre-established, and each material type corresponds to a set of material reflection coefficient vectors consisting of 10 band reflectances and a shape contour vector containing 1000 pixel coordinates. A spectrometer is used to obtain the reflectance spectrum data within the 380-780nm band of the surface of the material to be identified, and the 10 extreme points of the spectrum curve are analyzed by Gaussian function fitting to obtain the material reflection coefficient vector of the material. A CCD camera is used to image the material to be identified, and the edge contour of the material is extracted by the Canny algorithm, and the shape contour vector containing 1000 pixel coordinates is obtained using the Freeman chain code. The obtained material reflection coefficient vector and shape contour vector are input into the material feature database, and the KNN algorithm is used to calculate the Euclidean distance between the material to be identified and each material type in the database. K=5 is taken, and weighted voting is performed on the 5 material types with the closest distance. The similarity threshold is set to 8. If the similarity between the material to be identified and a certain material type is greater than the threshold, the material is judged to belong to this type, otherwise it is judged to be an unknown type. If the material to be identified is judged as an unknown type, its material reflection coefficient vector and shape contour vector are added as new material types to the material feature database, and the K-means clustering algorithm is used to reclassify the database to complete the automatic update of the database. According to the identification results of the material type, the materials to be identified are classified and stored on the corresponding RFID smart shelves, and the materials are automatically transported and stacked by AGV carts to realize intelligent management of materials.

[0048] S102. In view of the placement and environmental interference, multi-band RFID signal acquisition technology is used to obtain signal strength data of different frequency bands, and environmental noise is removed through a signal filtering algorithm to obtain a pure RFID signal.

[0049] According to the placement of the RFID tag, determine the RFID signal frequency band range that needs to be collected, and set the corresponding signal receiving antenna and receiver for different frequency bands. Adopt multi-band RFID signal acquisition technology to receive RFID signals of multiple frequency bands at the same time and obtain RFID tag signal strength data under different frequency bands. Preprocess the acquired multi-band RFID signal strength data to remove obvious outliers and invalid data, and obtain a preliminary filtered signal strength data set. According to the characteristics of environmental interference, establish an environmental noise model, adopt an adaptive signal filtering algorithm, automatically adjust the filter parameters, and dynamically remove the influence of environmental noise on RFID signals. After processing by the adaptive signal filtering algorithm, obtain the pure RFID signal strength data after removing the environmental noise, which is used as the input data for subsequent tag positioning and identification. Adopt a multi-band RFID tag positioning algorithm based on signal strength, comprehensively utilize the signal strength data of different frequency bands, and calculate the spatial position coordinates of the RFID tag through triangulation positioning or fingerprint matching. Match the RFID tag position coordinates obtained by positioning with the pre-established environmental map to determine the specific location of the tag, such as shelf number, area name, etc., and complete the positioning and identification process of the RFID tag.

[0050] Specifically, according to the placement of the RFID tag, the frequency band range of the RFID signal to be collected is determined, which usually includes low frequency (125kHz-134kHz), high frequency (156MHz), ultra-high frequency (860MHz-960MHz), etc. For different frequency bands, corresponding signal receiving antennas and receivers are set, such as using logarithmic periodic antennas to receive RFID signals in the 860MHz-960MHz frequency band. Multi-band RFID signal acquisition technology is used to simultaneously receive RFID signals in multiple frequency bands to obtain RFID tag signal strength data in different frequency bands. The acquired multi-band RFID signal strength data is preprocessed, and the median filtering algorithm is used to remove obvious outliers. Data that exceeds the normal range by more than 20% is regarded as invalid data and eliminated, and a preliminary filtered signal strength data set is obtained. According to the characteristics of environmental interference, an environmental noise model is established, and an adaptive signal filtering algorithm, such as the Kalman filtering algorithm, is used to adaptively remove the influence of environmental noise on RFID signals by dynamically adjusting the covariance matrix parameters of the filter. After being processed by the adaptive signal filtering algorithm, the pure RFID signal strength data after removing the environmental noise is obtained, and the signal-to-noise ratio is improved by more than 10dB, providing high-quality input data for subsequent tag positioning and identification. A multi-band RFID tag positioning algorithm based on signal strength is adopted, and the signal strength data of different frequency bands are comprehensively utilized to calculate the spatial position coordinates of the RFID tag through the triangulation positioning method. Specifically, at least 3 receivers are selected, and a nonlinear equation group is constructed according to the distance between the receivers and the received signal strength. The equation group is solved by the least squares method to obtain the three-dimensional coordinates of the RFID tag, and the positioning accuracy can reach 5m. The RFID tag position coordinates obtained by positioning are matched with the pre-established environmental map, and the minimum distance between the tag coordinates and each position in the map is found through Euclidean distance calculation, and the specific location of the tag is determined, such as shelf A-16, area B, etc., to complete the positioning and identification process of the RFID tag, and the recognition accuracy rate reaches more than 95%.

[0051] S3. According to the scenario of mixed materials, a multi-label recognition model based on deep learning is designed, which inputs pure RFID signals and material feature library data, determines the binding relationship between materials and labels, and outputs the binding results.

[0052] The RFID signal data and material feature library data in a mixed multi-material scenario are obtained as inputs to the deep learning multi-label recognition model. The obtained RFID signal data is preprocessed to remove noise and interference, and the pure RFID signal features are extracted. The preprocessed RFID signal features are matched with the material feature library data, and the material features that best match each RFID tag are found through similarity calculation. According to the matching results, the support vector machine algorithm is used to perform multi-label classification on the materials and RFID tags to obtain a preliminary binding relationship between materials and tags. The convolutional neural network algorithm is used to further judge and optimize the binding relationship between materials and tags with material features and RFID signal features as inputs. By integrating the judgment results of the support vector machine and the convolutional neural network, the matching degree of the material features and the RFID signal is comprehensively considered to determine the final binding relationship between materials and tags. The binding relationship results between materials and tags are output to complete the tag recognition and binding tasks in a mixed multi-material scenario.

[0053] Specifically, in a mixed scenario of multiple materials, a large amount of RFID tag signal data is collected through the RFID reader, and the characteristic information of each material, including the size, weight, material, etc. of the material, is extracted from the material feature library as the input of the deep learning multi-tag recognition model. The collected RFID signal data is subjected to wavelet transform and Fourier transform to remove high-frequency noise and low-frequency interference, and extract pure RFID signal features. The cosine similarity algorithm is used to match the preprocessed RFID signal features with the data in the material feature library, calculate the similarity between each RFID tag and the material feature, and select the material with the highest similarity as the material corresponding to the tag. Based on the matching results, the support vector machine algorithm is used, with material features and RFID signal features as input, and multiple label classifications are set, such as material type, material ID, etc., to train the SVM model and obtain the preliminary binding relationship between materials and labels. The convolutional neural network algorithm is further used to construct a CNN model containing convolutional layers, pooling layers and fully connected layers. With material features and RFID signal features as input, the binding relationship between materials and labels is judged and optimized to improve the accuracy of binding. Based on the judgment results of the SVM and CNN models, the threshold is set to 8. When the material-label binding relationships given by the two models are consistent and the confidence levels are both greater than 8, the final binding relationship is determined. Otherwise, the binding relationship between the material and the label is considered uncertain. The determined material-label binding relationship results are output, and the label information of the material in the material feature library is updated to complete the label recognition and binding tasks in the multi-material mixed scenario.

[0054] S4. In response to positioning and tracking needs, a triangulation positioning algorithm based on signal strength is used to obtain signal strength data from multiple RFID readers and writers, and the precise location of the materials is determined by calculating the signal attenuation model.

[0055] According to the pre-established RFID reader layout, the signal strength data of at least three readers are obtained; for the multiple signal strength data obtained, the triangulation positioning algorithm is used to calculate the approximate location coordinates of the target material; based on the approximate location coordinates obtained by the triangulation positioning algorithm, combined with the preset signal attenuation model, the distance between the target material and each reader is calculated; if the calculated distance value is less than the preset threshold, the position is judged to be the precise location of the target material; if it is greater than the threshold, return to step 1 and continue to obtain signal strength data of more readers; the historical positioning data is trained by the machine learning algorithm to obtain the optimized signal attenuation model parameters to improve the accuracy of subsequent positioning; the precise positioning location coordinates are bound to the material ID, the material location information is updated in real time, and stored in the database; when the location of a specific material needs to be queried, the latest location coordinates of the material are obtained from the database, and visualized on the map to achieve real-time positioning and tracking.

[0056] Specifically, first, according to the pre-established RFID reader layout, the signal strength data of at least three readers are obtained through the RS232 interface, such as the signal strength of reader A is -60dBm, the signal strength of reader B is -65dBm, and the signal strength of reader C is -70dBm. Then, for the multiple signal strength data obtained, the Chan algorithm is used for triangulation positioning, and the approximate position coordinates of the target material are calculated as (x, y). Then, according to the approximate position coordinates obtained by the triangulation positioning algorithm, combined with the preset logarithmic distance path loss model, the distance between the target material and each reader is calculated, such as the distance to reader A is 5m, the distance to reader B is 7m, and the distance to reader C is 9m. If the calculated distance values ​​are all less than the preset threshold of 10m, the position is judged to be the precise position of the target material; if there is a distance value greater than the threshold, the signal strength data of more readers is continued to be obtained through the RS485 bus, and the above positioning process is repeated. At the same time, the historical positioning data is trained through the support vector machine algorithm to obtain the optimized signal attenuation model parameters, such as the path loss index is optimized from 5 to 8, which is used to improve the accuracy of subsequent positioning. Finally, the precise positioning location coordinates (x, y) are bound to the material ID through the MQTT protocol, and the material location information is updated in real time and stored in the MongoDB database. When the location of a specific material needs to be queried, the latest location coordinates of the material are obtained from the database through the RESTful API, and visualized on the Web GIS platform to achieve real-time positioning and tracking.

[0057] S5. According to the shelf adjustment changes, a dynamic shelf mapping algorithm is designed to obtain the spatial coordinate data after the shelf structure is adjusted, and the material location information is updated through coordinate transformation.

[0058] According to the shelf adjustment changes, the adjusted shelf structure data and spatial coordinate data are obtained; the dynamic mapping algorithm is used to establish the mapping relationship between the shelf structure and coordinates before and after the adjustment; the new coordinate position of the material in the adjusted shelf structure is calculated through coordinate transformation; the material location information database is updated according to the new coordinate position of the material; if the material location changes, the automatic update mechanism is triggered to synchronize the updated location information to the relevant system; the updated material location information is obtained and a shelf adjustment change report is generated to facilitate inventory management; through visualization technology, the material distribution before and after the shelf adjustment is displayed in real time, and dynamic query and positioning of materials are supported.

[0059] S6. Based on the real-time requirement, a signal processing module based on edge computing is designed to obtain the real-time signal data of the RFID reader and realize fast signal processing through a distributed computing framework.

[0060] Acquire the real-time signal data collected by the RFID reader and transfer the data to the edge computing node for preprocessing. Perform parallel calculations on the preprocessed signal data through the distributed computing framework, and use the fast Fourier transform algorithm to perform frequency domain analysis on the signal to obtain the frequency characteristics of the signal. According to the preset frequency characteristic threshold, determine whether the signal contains valid RFID tag information. If it does, extract the tag ID, otherwise discard the signal data. For the extracted RFID tag ID, query the pre-established mapping table of tag ID and item information to obtain the corresponding item information. Associate the acquired item information with the corresponding timestamp and reader ID to generate structured item tracking data. Use the incremental learning algorithm to analyze the item tracking data in real time, and determine the current location and status of the item according to the movement trajectory and state changes of the item. Push the analyzed item location and status information to the upper-level application in real time for the business system to make decisions and optimization, and realize real-time tracking and management of items.

[0061] Specifically, the RFID reader transmits electromagnetic waves through the antenna, activates the tag and receives the electromagnetic wave signal reflected by the tag. The receiving circuit inside the reader demodulates and amplifies the received high-frequency signal and converts the analog signal into a digital signal. The digital signal is transmitted to the edge computing node in real time through the network transmission protocol. The edge node uses the Storm distributed streaming computing framework to divide the data stream into multiple subtasks for parallel processing. Each data packet is decoded to extract characteristic parameters such as carrier frequency and received signal strength. The FFT algorithm is used to transform the time domain signal to obtain the frequency domain feature. The energy detection algorithm is used to determine whether the frequency domain feature meets the energy threshold of the valid RFID tag. If the energy exceeds -60dBm, it is considered to be a valid tag signal and the 96bit ID of the tag is extracted. Otherwise, it is considered to be an invalid signal and discarded. The extracted tag ID is matched with the pre-imported tag and item mapping table. The mapping table is stored in a distributed NoSQL database and supports millisecond-level query response. After matching, the SKU, name, specification and other information of the item are obtained, which are associated with the signal reception timestamp, reader ID, and antenna ID to form structured tracking data. The Flink streaming computing platform is used to analyze the tracking data in real time. Through the nearest neighbor dynamic time warping algorithm, the location change trajectory and state change of the item are comprehensively analyzed with a time window of 200ms, the tracking data is clustered, and the current location of the item is inferred. If the location of the item changes, the status is updated. The location and status information of the item is pushed to the upper-level application in real time through WebSocket. The business system makes intelligent decisions based on preset rules, generates sorting task orders, guides the AGV car or robotic arm to pick up the goods at the designated cargo location, and optimizes the path to realize the automated and intelligent tracking management of items.

[0062] S7. Based on algorithm design and hardware design, build a software-hardware collaborative optimization framework, obtain algorithm processing results and hardware performance data, and optimize system operation efficiency through resource scheduling strategies.

[0063] According to the design principles of the software-hardware collaborative optimization framework, a comprehensive optimization system is constructed by combining algorithm design and hardware design. The result data of algorithm processing and the performance data of hardware operation are obtained through this system and used as the input for resource scheduling strategy optimization. According to the obtained algorithm results and hardware performance data, the resource scheduling strategy is optimized by machine learning algorithm to obtain an optimized resource scheduling strategy. The optimized resource scheduling strategy is applied to the software-hardware collaborative optimization framework to improve the system operation efficiency by dynamically adjusting the allocation of system resources. During the operation of the system, the changes in the algorithm processing results and hardware performance data are continuously monitored. If the system operation efficiency is found to decrease, the re-optimization of the resource scheduling strategy is triggered. According to the re-optimized resource scheduling strategy, the allocation of software and hardware resources is dynamically adjusted to ensure that the system operation efficiency is maintained at a high level. The above process is continuously iterated and optimized, and the dynamic optimization of system operation efficiency is achieved through the adaptive adjustment of the software-hardware collaborative optimization framework to ensure the stability and improvement of system performance.

[0064] Specifically, in the design of the software and hardware collaborative optimization framework, the convolutional neural network algorithm can be used to process image data, and the algorithm can be optimized to achieve acceleration on the FPGA hardware platform. The system collects data such as the time for the algorithm to process 1,000 256×256 pixel images and the FPGA resource utilization rate through comprehensive optimization, and uses it as the input for resource scheduling strategy optimization. The Q-Learning reinforcement learning algorithm is used to obtain the optimized resource scheduling strategy through 1,000 iterative trainings, and the resource allocation ratio of DSP, BRAM and other resources on the FPGA is dynamically adjusted to the optimal ratio, which shortens the algorithm processing time by 20% and improves the resource utilization rate by 15%. During the operation of the system, the algorithm processing time and resource utilization data are collected every 10 minutes. When it is found that the processing time exceeds the threshold of 50ms or the resource utilization rate is less than 80%, the re-optimization process is triggered, and the system performance is restored to the optimal state through 5 iterative optimizations. Through software and hardware collaborative optimization, the system can dynamically adjust resource allocation according to the real-time status and maintain efficient operation continuously.

[0065] S8. Aiming at the accuracy requirements of sensing recognition and positioning tracking, an adaptive optimization model based on reinforcement learning is designed to obtain the error data of recognition and positioning, and the algorithm parameters are adjusted through the feedback mechanism to improve the system performance.

[0066] The error data generated during the process of sensing, identifying, and positioning tracking are obtained and used as the input of the reinforcement learning model. The performance of the model is evaluated through the reward function. According to the output results of the reinforcement learning model, the key parameters in the sensing, identifying, and positioning tracking algorithms are dynamically adjusted, such as the threshold of the feature extraction algorithm and the similarity threshold of the matching algorithm, to meet the accuracy requirements in different scenarios. In the process of adjusting the algorithm parameters, an adaptive optimization strategy is adopted to automatically select appropriate optimization algorithms, such as gradient descent and Newton's method, according to the distribution characteristics and change trends of the error data to accelerate the convergence of parameters. By setting a reasonable reward function, the reinforcement learning model is guided to optimize in the direction of reducing the recognition and positioning errors, while avoiding overfitting and underfitting problems and improving the generalization ability of the model. In the training process of the reinforcement learning model, the experience replay mechanism is adopted to store the historical error data in the replay buffer, and randomly extract samples for training to break the correlation between the data and improve the stability of the model. The performance of the reinforcement learning model is regularly evaluated, and the mean and variance of the recognition and positioning errors are calculated to determine whether the model has converged to the optimal state, and the model structure and hyperparameters are adjusted when necessary. Apply the optimized algorithm parameters to actual sensing recognition and positioning tracking systems, continuously monitor system performance, collect new error data, form a closed-loop feedback mechanism, and achieve adaptive optimization and continuous improvement of the algorithm.

[0067] Specifically, in the sensing recognition and positioning tracking system, the error data of the object's motion trajectory is obtained by collecting sensor data, such as accelerometers, gyroscopes, etc. These error data are used as the input of the reinforcement learning model. By designing a reasonable reward function, such as giving a positive reward of 0 when the average error is less than 5 meters and a negative reward of -0 when the error is greater than 1 meter, the model is guided to learn and optimize the algorithm parameters. The output of the model can be the adjustment value of key parameters such as the threshold of the feature extraction algorithm and the similarity threshold of the matching algorithm. For example, the similarity threshold of feature point matching is adjusted from 8 to 85 to improve the matching accuracy. In the parameter adjustment process, according to the distribution characteristics of the error data, such as the error follows a normal distribution, with a mean of 6 meters and a variance of 2 meters, the gradient descent algorithm is automatically selected, and the parameters are iteratively optimized with a learning rate of 0.1 and a momentum factor of 9 until the error mean is reduced to less than 4 meters. At the same time, through the experience replay mechanism, 1,000 samples are randomly selected from the historical error data and trained with an exploration probability of 1 to improve the stability and generalization ability of the model. Every 1000 iterations, the model performance is evaluated, and the mean and variance of the recognition and positioning errors are calculated. If the error index changes less than 0.1 for 5 consecutive evaluations, the model is considered to have converged, and the current algorithm parameters are applied to the actual system. In the actual application process, the system performance is continuously monitored, new error data is collected every 10 minutes, and the playback buffer is updated to form a closed-loop feedback to achieve adaptive optimization and continuous improvement of the algorithm. In this way, the accuracy of the sensing recognition and positioning tracking system can be continuously improved to meet the application needs in different scenarios.

[0068] The present invention provides an intelligent shelf control system based on RFID, which mainly includes: a material feature library module, which is used to pre-establish a material feature library according to the differences in material and shape of the materials, obtain the material reflection coefficient and shape profile data of each material, and determine the material type through feature library matching; a multi-band RFID signal acquisition module, which is used to adopt multi-band RFID signal acquisition technology to obtain signal strength data of different frequency bands according to the placement position and environmental interference, and remove environmental noise through a signal filtering algorithm to obtain a pure RFID signal; a multi-tag recognition module, which is used to design a multi-tag recognition model based on deep learning according to a scene where multiple materials are mixed, input a pure RFID signal and material feature library data, judge the binding relationship between the material and the tag, and output the binding result; a triangulation positioning module, which is used to adopt a triangulation positioning algorithm based on signal strength to obtain the signal strength data of multiple RFID readers according to the positioning and tracking needs. Signal strength data is used to determine the precise location of materials by calculating the signal attenuation model; the dynamic shelf mapping module is used to design a dynamic shelf mapping algorithm based on shelf adjustments, obtain the spatial coordinate data after the shelf structure is adjusted, and update the material location information through coordinate transformation; the edge computing signal processing module is used to adopt real-time requirements, design a signal processing module based on edge computing, obtain real-time signal data from RFID readers, and realize rapid signal processing through a distributed computing framework; the software and hardware collaborative optimization module is used to build a software and hardware collaborative optimization framework based on algorithm design and hardware design, obtain algorithm processing results and hardware performance data, and optimize system operation efficiency through resource scheduling strategies; the adaptive optimization module is used to design an adaptive optimization model based on reinforcement learning for the accuracy requirements of sensing recognition and positioning tracking, obtain recognition and positioning error data, adjust algorithm parameters through feedback mechanisms, and improve system performance.

[0069] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. An RFID-based intelligent shelf control method and system, characterized in that: The following steps are involved: According to the differences in material and shape of materials, a material feature library is established in advance to obtain the material reflection coefficient and shape profile data of each material, and the material type is determined by matching the feature library; in view of the placement position and environmental interference, multi-band RFID signal acquisition technology is used to obtain signal strength data of different frequency bands, and the environmental noise is removed through the signal filtering algorithm to obtain a pure RFID signal; According to the scenario of mixed materials, a multi-label recognition model based on deep learning is designed. The pure RFID signal and material feature library data are input to determine the binding relationship between materials and labels, and the binding results are output. In response to the needs of positioning and tracking, a triangulation positioning algorithm based on signal strength is used to obtain the signal strength data of multiple RFID readers and writers, and the precise location of the materials is determined by calculating the signal attenuation model. According to the adjustment and changes of the shelves, a dynamic shelf mapping algorithm is designed to obtain the spatial coordinate data after the shelf structure is adjusted, and the material location information is updated through coordinate transformation. Based on the real-time requirement, a signal processing module based on edge computing is designed to obtain the real-time signal data of the RFID reader and writer, and realize rapid signal processing through a distributed computing framework. According to the algorithm design and hardware design, a software and hardware collaborative optimization framework is constructed to obtain the algorithm processing results and hardware performance data, and the system operation efficiency is optimized through resource scheduling strategies. According to the accuracy requirements of sensor recognition and positioning tracking, an adaptive optimization model based on reinforcement learning is designed to obtain the error data of recognition and positioning, and the algorithm parameters are adjusted through the feedback mechanism to improve the system performance.

2. The method according to claim 1, characterized in that: According to the differences in material and shape of materials, a material feature library is pre-established to obtain the material reflection coefficient and shape profile data of each material, and the material type is determined by matching the feature library, including: According to the differences in material and shape of materials, a feature database containing multiple material types is pre-established, and each material type corresponds to a set of material reflection coefficient and shape profile data; An optical sensor is used to obtain the reflectance spectrum data of the surface of the material to be identified, and the material reflectance coefficient of the material is obtained by analyzing the peaks and troughs of the spectrum data; Use machine vision technology to image the materials to be identified, extract the contour area of ​​the materials through image segmentation algorithms, and obtain contour data that characterizes the shape of the materials; The acquired material reflection coefficient and shape profile data are input into the material feature database, and the feature similarity matching algorithm is used to calculate the similarity between the material to be identified and each material type in the database; Set a similarity threshold. If the similarity between the material to be identified and a certain material type is greater than the threshold, the material is judged to belong to that type, otherwise it is judged to be an unknown type. If the material to be identified is judged as an unknown type, its material reflectance coefficient and shape profile data are added to the material feature database as a new material type to complete the automatic update of the database; According to the identification results of material types, the materials to be identified are classified and stored or subjected to subsequent processing operations to realize intelligent management of materials.

3. The method according to claim 1, characterized in that: The multi-band RFID signal acquisition technology is used to obtain signal strength data of different frequency bands in view of the placement position and environmental interference, and the environmental noise is removed by the signal filtering algorithm to obtain a pure RFID signal, including: According to the placement of the RFID tag, determine the RFID signal frequency range that needs to be collected, and set the corresponding signal receiving antenna and receiver for different frequency bands; Adopt multi-band RFID signal acquisition technology to simultaneously receive RFID signals from multiple frequency bands and obtain RFID tag signal strength data in different frequency bands; Preprocess the acquired multi-band RFID signal strength data to remove obvious outliers and invalid data, and obtain a preliminary filtered signal strength data set; According to the characteristics of environmental interference, an environmental noise model is established, and an adaptive signal filtering algorithm is used to automatically adjust the filter parameters to dynamically remove the impact of environmental noise on RFID signals; After being processed by the adaptive signal filtering algorithm, the pure RFID signal strength data after removing the environmental noise is obtained, which serves as the input data for subsequent tag positioning and identification; Adopting a multi-band RFID tag positioning algorithm based on signal strength, comprehensively utilizing the signal strength data of different frequency bands, and calculating the spatial position coordinates of the RFID tag through triangulation positioning or fingerprint matching methods; The RFID tag location coordinates obtained by positioning are matched with the pre-established environmental map to determine the specific location of the tag, including the shelf number and area name, to complete the RFID tag positioning and identification process.

4. The method according to claim 1, characterized in that: According to the scenario of multiple mixed materials, a multi-label recognition model based on deep learning is designed, pure RFID signals and material feature library data are input, the binding relationship between materials and labels is determined, and the binding results are output, including: Obtain RFID signal data and material feature library data in a mixed multi-material scenario as input to the deep learning multi-label recognition model; Preprocess the acquired RFID signal data to remove noise and interference and extract pure RFID signal features; Match the preprocessed RFID signal features with the material feature library data, and find the material features that best match each RFID tag through similarity calculation; According to the matching results, the support vector machine algorithm is used to perform multi-label classification on materials and RFID tags to obtain the preliminary binding relationship between materials and tags; The convolutional neural network algorithm is used to take material characteristics and RFID signal characteristics as input to further judge and optimize the binding relationship between materials and tags; By integrating the judgment results of support vector machines and convolutional neural networks, the matching degree between material characteristics and RFID signals is comprehensively considered to determine the final binding relationship between materials and tags. The binding relationship between materials and labels is output to complete label recognition and binding tasks in scenarios with mixed materials.

5. The method according to claim 1, characterized in that In view of the positioning and tracking needs, a triangulation positioning algorithm based on signal strength is used to obtain the signal strength data of multiple RFID readers and writers, and the precise location of the materials is determined by calculating the signal attenuation model, including: According to the pre-established RFID reader layout, obtain signal strength data of at least three readers; Based on the multiple signal strength data obtained, the triangulation positioning algorithm is used to calculate the approximate location coordinates of the target materials; According to the approximate location coordinates obtained by the triangulation positioning algorithm and the preset signal attenuation model, the distance between the target material and each reader is calculated; If the calculated distance value is less than the preset threshold, the location is determined to be the precise location of the target material; If it is greater than the threshold, return to step 1 and continue to obtain signal strength data of more readers; The historical positioning data is trained through machine learning algorithms to obtain optimized signal attenuation model parameters to improve the accuracy of subsequent positioning; Bind the precise location coordinates with the material ID, update the material location information in real time, and store it in the database; When you need to query the location of a specific material, get the latest location coordinates of the material from the database and visualize them on the map to achieve real-time positioning and tracking.

6. The method according to claim 1, characterized in that According to the shelf adjustment changes, a dynamic shelf mapping algorithm is designed to obtain the spatial coordinate data after the shelf structure is adjusted, and the material location information is updated through coordinate transformation, including: According to the shelf adjustment changes, the adjusted shelf structure data and space coordinate data are obtained; Adopt dynamic mapping algorithm to establish the mapping relationship between shelf structure and coordinates before and after adjustment; Calculate the new coordinate position of the materials in the adjusted shelf structure through coordinate transformation; Update the material location information database according to the new coordinates of the materials; If the location of materials changes, the automatic update mechanism is triggered to synchronize the updated location information to the relevant systems; Obtain updated material location information and generate shelf adjustment change reports to facilitate inventory management; Through visualization technology, the material distribution before and after shelf adjustment is displayed in real time, supporting dynamic query and positioning of materials.

7. The method according to claim 1, characterized in that The real-time requirement is adopted to design a signal processing module based on edge computing, obtain the real-time signal data of the RFID reader, and realize fast signal processing through a distributed computing framework, including: Obtain real-time signal data collected by RFID readers and transmit the data to edge computing nodes for preprocessing; The pre-processed signal data is parallelized through a distributed computing framework, and the fast Fourier transform algorithm is used to analyze the signal in the frequency domain to obtain the frequency characteristics of the signal; According to the preset frequency characteristic threshold, determine whether the signal contains valid RFID tag information. If it does, extract the tag ID; otherwise, discard the signal data; For the extracted RFID tag ID, the corresponding item information is obtained by querying the pre-established mapping table between the tag ID and the item information; Associate the acquired item information with the corresponding timestamp and reader ID to generate structured item tracking data; Use incremental learning algorithms to analyze item tracking data in real time, and determine the current location and status of items based on their movement trajectory and status changes; The analyzed item location and status information is pushed to the upper-level application in real time for the business system to make decisions and optimizations, thus achieving real-time tracking and management of items.

8. The method according to claim 1, characterized in that According to the algorithm design and hardware design, a software and hardware collaborative optimization framework is constructed to obtain the algorithm processing results and hardware performance data, and optimize the system operation efficiency through resource scheduling strategies, including: According to the design principles of the software-hardware collaborative optimization framework, a comprehensive optimization system is constructed by combining algorithm design and hardware design; Obtain the result data of algorithm processing and the performance data of hardware operation through the comprehensive optimization system, and use them as input for resource scheduling strategy optimization; Based on the obtained algorithm results and hardware performance data, a machine learning algorithm is used to optimize the resource scheduling strategy to obtain an optimized resource scheduling strategy; Apply the optimized resource scheduling strategy to the software-hardware collaborative optimization framework to improve system operation efficiency by dynamically adjusting the allocation of system resources; During system operation, the algorithm processing results and hardware performance data are continuously monitored. If the system operation efficiency is found to be reduced, the resource scheduling strategy is re-optimized. According to the optimized resource scheduling strategy, dynamically adjust the allocation of software and hardware resources to ensure that the system operation efficiency is maintained at a high level; The above process is continuously iterated and optimized, and the adaptive adjustment of the software and hardware collaborative optimization framework is used to achieve dynamic optimization of system operation efficiency and ensure the stability and improvement of system performance.

9. The method according to claim 1, characterized in that: In view of the accuracy requirements of sensing recognition and positioning tracking, an adaptive optimization model based on reinforcement learning is designed to obtain the error data of recognition and positioning, and the algorithm parameters are adjusted through the feedback mechanism to improve the system performance, including: Obtain the error data generated during the sensing, identification and positioning tracking process, use it as the input of the reinforcement learning model, and evaluate the performance of the model through the reward function; According to the output results of the reinforcement learning model, the key parameters in the sensing recognition and positioning tracking algorithms are dynamically adjusted, including the threshold of the feature extraction algorithm and the similarity threshold of the matching algorithm, to meet the accuracy requirements in different scenarios; In the process of adjusting the algorithm parameters, an adaptive optimization strategy is adopted to automatically select gradient descent or Newton method according to the distribution characteristics and change trends of the error data to speed up the parameter convergence; By setting a reasonable reward function, the reinforcement learning model is guided to optimize in the direction of reducing recognition and positioning errors, while avoiding overfitting and underfitting problems and improving the generalization ability of the model; During the training process of the reinforcement learning model, an experience replay mechanism is used to store historical error data in the replay buffer and randomly extract samples for training to break the correlation between data and improve the stability of the model. Regularly evaluate the performance of the reinforcement learning model, determine whether the model has converged to the optimal state by calculating the mean and variance of the recognition and positioning errors, and adjust the model structure and hyperparameters if necessary; Apply the optimized algorithm parameters to actual sensing recognition and positioning tracking systems, continuously monitor system performance, collect new error data, form a closed-loop feedback mechanism, and achieve adaptive optimization and continuous improvement of the algorithm.

10. An intelligent shelf control system based on RFID, characterized in that: The system includes: a material feature library module, which is used to pre-establish a material feature library according to the differences in material and shape of the materials, obtain the material reflection coefficient and shape profile data of each material, and determine the material type through feature library matching; a multi-band RFID signal acquisition module, which is used to adopt multi-band RFID signal acquisition technology to obtain signal strength data of different frequency bands according to the placement position and environmental interference, and remove environmental noise through a signal filtering algorithm to obtain a pure RFID signal; a multi-tag recognition module, which is used to design a multi-tag recognition model based on deep learning according to a scene with multiple materials mixed, input a pure RFID signal and material feature library data, judge the binding relationship between the material and the tag, and output the binding result; a triangulation positioning module, which is used to adopt a triangulation positioning algorithm based on signal strength to obtain the signal strength data of multiple RFID readers according to the positioning and tracking needs, and calculate the signal strength data of multiple RFID readers through signal filtering. The signal attenuation model determines the precise location of materials; the dynamic shelf mapping module is used to design a dynamic shelf mapping algorithm according to shelf adjustments, obtain the spatial coordinate data after the shelf structure is adjusted, and update the material location information through coordinate transformation; the edge computing signal processing module is used to adopt real-time requirements, design a signal processing module based on edge computing, obtain real-time signal data from the RFID reader, and realize rapid signal processing through a distributed computing framework; the software and hardware collaborative optimization module is used to build a software and hardware collaborative optimization framework based on algorithm design and hardware design, obtain algorithm processing results and hardware performance data, and optimize system operation efficiency through resource scheduling strategies; the adaptive optimization module is used to design an adaptive optimization model based on reinforcement learning for the accuracy requirements of sensing recognition and positioning tracking, obtain recognition and positioning error data, adjust algorithm parameters through feedback mechanism, and improve system performance.

Images