Commodity management method and system based on electronic tag technology and internet of things architecture

Abstract

This application relates to the field of intelligent commodity management technology, and in particular to a commodity management method and system based on electronic tag technology and Internet of Things architecture. The method includes: establishing an intermodal correlation matrix for the tag data in the initial tag dataset of each commodity acquired through a reader, and obtaining the correlation coefficient between each tag data; the tag data is a received signal strength tag, a Doppler frequency shift tag, or a time difference of arrival tag; for each tag data, when the detected outlier value of the tag data is greater than a preset threshold, calculating a joint anomaly score for the tag data based on the outlier value of the tag, the correlation coefficient corresponding to the tag data, and the outlier values ​​of other tag data; when the joint anomaly score is greater than the preset threshold, deleting the tag data; and sequentially performing time-domain filtering, frequency-domain filtering, low-pass filtering, and standardization processing on the initial tag dataset after deletion to obtain a target dataset.

Description

Technical Field

[0001] This application relates to the field of intelligent commodity management technology, and in particular to a commodity management method and system based on electronic tag technology and Internet of Things architecture. Background Technology

[0002] Traditional retail systems have limitations in product positioning and environmental monitoring. They typically rely on signal strength and a triangulation algorithm for product location, neglecting the complex spatial relationship between the product and the reader, resulting in low positioning accuracy. Furthermore, environmental monitoring often involves periodically recording data such as temperature and humidity, making it difficult to predict environmental changes in real time, which can easily lead to product quality damage or improper storage conditions. Summary of the Invention

[0003] This invention provides a product management method based on electronic tag technology and an Internet of Things (IoT) architecture to address related problems in the prior art. The method includes: A modal correlation matrix is ​​established for the tag data in the initial tag dataset of each product obtained by the reader, and the correlation coefficient between each tag data is obtained; the tag data is a received signal strength tag, a Doppler frequency shift tag, or a time difference of arrival tag; For each tag data, when the detected outlier value of the tag data is greater than the preset threshold, the joint outlier score of the tag data is calculated based on the outlier value of the tag, the correlation coefficient corresponding to the tag data, and the outlier values ​​of other tag data. When the combined anomaly score is greater than the preset threshold, the label data is deleted; The initial label dataset after deletion is processed sequentially with time-domain filtering, frequency-domain filtering, low-pass filtering, and standardization to obtain the target dataset.

[0004] Specifically, obtaining the correlation coefficients between the various label data includes: When the tag data is the received signal strength tag, the correlation coefficient between the received signal strength tag and the Doppler frequency shift tag is obtained, as well as the correlation coefficient between the received signal strength tag and the time difference of arrival tag; When the tag data is the Doppler frequency shift tag, the correlation coefficient between the Doppler frequency shift tag and the received signal strength tag, and the correlation coefficient between the Doppler frequency shift tag and the time difference of arrival tag are obtained; When the tag data is the time difference of arrival tag, the correlation coefficient between the time difference of arrival tag and the received signal strength tag is obtained, as well as the correlation coefficient between the time difference of arrival tag and the Doppler frequency shift tag.

[0005] Specifically, the step of calculating the joint anomaly score of the tag data based on the outlier of the tag, the correlation coefficient corresponding to the tag data, and the outlier of other tag data includes: The joint anomaly score is calculated using the following formula: ; in, The abnormal value of the received signal strength tag; These are the outliers of the Doppler frequency shift tag; These are the outliers of the arrival time difference label; The correlation coefficient between the received signal strength tag and the Doppler frequency shift tag; The correlation coefficient between the received signal strength tag and the time difference of arrival tag; The joint anomaly score is the received signal strength label.

[0006] Specifically, after completing the standardization process but before obtaining the target dataset, the process further includes: The standardized data is used as an intermediate label dataset; Using an IoT architecture model, spatial relationship features between central data from intermediate tag datasets of different products are calculated; For each product, the tag data whose spatial relationship features meet the preset conditions are used as the target dataset for that product.

[0007] Specifically, the calculation of spatial relationship features between intermediate label datasets from different products includes: The spatial relationship features are calculated using the following formula. : ; in, and Used to describe the source of the label data; y is used to describe whether the label data comes from the same product, when y=1 indicates and Tag data from the same product, where y=0 indicates and Label data from different products; It is a sample and The spatial distance between them; m represents the minimum distance value between data from different labels; It is the regularization parameter.

[0008] Specifically, the calculation of spatial relationship features between intermediate label datasets from different products includes: The spatial relationship features are calculated using the following formula. :

[0009] Anchor point samples for label data; This indicates another label data source from the same product. This represents label data from different products. Anchor point sample and The distance between them. Anchor point sample and The distance between them. m represents the minimum distance difference between the anchor sample and the positive sample. These are preset weight parameters.

[0010] Specifically, the method further includes: Based on the historical environmental information of the product obtained, the current environmental information of the product is predicted according to the following formula to obtain the current environmental prediction data. : ; Where t represents the time when data prediction is required. To predict environmental data, The data obtained at time t is the data acquired at the time preceding time t. The data is obtained one time after time t. For the preset time interval, is the base of the natural logarithm.

[0011] Specifically, all the obtained environmental data are normalized using the following formula: ; It is normalized environmental data. It's environmental data. It is the median of all environmental data. It is the interquartile range of all environmental data, that is, the difference between the third quartile (Q3) and the first quartile (Q1).

[0012] Specifically, the models required to implement the product management method include: Input layer, used to obtain the shape as The data, among which This indicates that the number of readers corresponding to product m is N, and 3 represents the number of readers. Get Products Tag data; The transformation layer is used to convert the input label data into a one-dimensional vector; The feature extraction layer is used to extract local features and non-linear characteristics between each reader and the product's tag; The feature fusion layer is used to integrate features from all readers.

[0013] This invention provides a product management system based on electronic tag technology and an Internet of Things (IoT) architecture. The system includes: The acquisition device is used to establish an intermodal correlation matrix for the tag data in the initial tag dataset of each product acquired by the reader, and to obtain the correlation coefficient between each tag data; the tag data is a received signal strength tag, a Doppler frequency shift tag, or a time difference of arrival tag; The detection device is used to calculate the joint anomaly score of each tag data when the detected anomaly value of the tag data is greater than a preset threshold, based on the anomaly value of the tag, the correlation coefficient corresponding to the tag data, and the anomaly values ​​of other tag data. A first processing device is used to delete the tag data when the joint anomaly score is greater than the preset threshold. The second processing unit is used to sequentially perform time-domain filtering, frequency-domain filtering, low-pass filtering, and standardization on the initial label dataset after the deletion operation to obtain the target dataset.

[0014] This method collects three types of data from goods: received signal strength tags, Doppler frequency shift tags, and time difference of arrival tags. Employing a multi-dimensional feature fusion approach, it accurately predicts the spatial location of goods, optimizing positioning accuracy and dynamic tracking capabilities. Simultaneously, the system uses real-time monitoring data from multiple sensors, including temperature, humidity, PM10, PM2.5, and light intensity, to accurately predict environmental changes within the next hour. This improves the efficiency of inventory management and energy scheduling, ensuring optimized product quality and operating costs. Attached Figure Description

[0015] Figure 1 A flowchart illustrating a product management method based on electronic tag technology and an Internet of Things architecture, provided for an embodiment of the present invention; Figure 2 This is an IoT architecture model provided in the embodiments of the present invention; Figure 3 A neural network structure diagram of the environmental prediction model provided in this embodiment of the invention; Figure 4 An architecture diagram of a commodity management system based on electronic tag technology and Internet of Things architecture provided for an embodiment of the present invention; Figure 5This is a structural diagram of a commodity management system based on electronic tag technology and Internet of Things architecture, provided for an embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0017] In the description of this application, it should be understood that if terms such as "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential" appear, these terms indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0018] Furthermore, where the terms "first" and "second" appear, these terms are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, where the term "multiple" appears, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0019] This invention provides a product management method and system based on electronic tag technology and an Internet of Things (IoT) architecture. The present invention provides the following embodiment 1, such as Figure 1 As shown, the specific steps of the method in Embodiment 1 are as follows: The first step is data collection. During this phase, the system uses multiple RFID readers to collect comprehensive information about the goods. The readers measure and record the following three key data points: Received Signal Strength Tag (RSSI): Reflects the distance between the tag on the product and the reader.

[0020] DopplerShift tag: Provides information on the speed and direction of movement of the product when there is relative motion between the tag on the product and the reader.

[0021] Time Difference of Arrival (TDOA) tagging: This method uses multiple readers to simultaneously measure the time difference between the arrival times of the signals from the tags on the product at each reader.

[0022] To simplify data description and subsequent calculations, the following symbols are defined: : Represents the RSSI data of the m-th item obtained by the n-th reader.

[0023] : Represents the Doppler Shift data of the m-th item obtained by the n-th reader.

[0024] : Represents the TDOA data of the m-th product obtained by the n-th reader.

[0025] in, Indicates the reader's serial number. This indicates the total number of readers. Indicates the product number. This indicates the total quantity of goods.

[0026] The initial label dataset includes all the acquired label data, such as received signal strength labels, Doppler shift labels, and time difference of arrival labels.

[0027] Step 11: Establish an intermodal correlation matrix for the tag data in the initial tag dataset of each product obtained by the reader, and obtain the correlation coefficient between each tag data; the tag data is a received signal strength tag, a Doppler frequency shift tag, or a time difference of arrival tag; the specific process of this step is as follows: First, we perform correlation modeling on the label data, using the obtained historical label data to build a correlation matrix between modalities for RSSI, Doppler Shift, and TDOA:

[0028] in, The correlation coefficient between the received signal strength tag and the Doppler frequency shift tag; The correlation coefficient between the received signal strength tag and the time difference of arrival tag; The correlation coefficient between the Doppler frequency shift tag and the received signal strength tag; The correlation coefficient between the Doppler frequency shift tag and the time difference of arrival tag; The correlation coefficient between the time difference of arrival tag and the received signal strength tag; The correlation coefficient is the relationship between the time difference of arrival (TDOA) tag and the Doppler shift tag.

[0029] Therefore, the correlation coefficients between the various label data are obtained, including: When the tag data is the received signal strength tag, the correlation coefficient between the received signal strength tag and the Doppler frequency shift tag is obtained, as well as the correlation coefficient between the received signal strength tag and the time difference of arrival tag; When the tag data is the Doppler frequency shift tag, the correlation coefficient between the Doppler frequency shift tag and the received signal strength tag, and the correlation coefficient between the Doppler frequency shift tag and the time difference of arrival tag are obtained; When the tag data is the time difference of arrival tag, the correlation coefficient between the time difference of arrival tag and the received signal strength tag is obtained, as well as the correlation coefficient between the time difference of arrival tag and the Doppler frequency shift tag.

[0030] Step 12: For each tag data, when the detected outlier of that tag data exceeds a preset threshold, calculate the joint outlier score of that tag data based on the outlier of that tag, the correlation coefficient corresponding to that tag data, and the outliers of other tag data; this step includes: The Z-Score method is used for outlier detection. This method determines whether a data point is an outlier (Z) by calculating the standard deviation distance between the data point and the mean. For example, if the outlier Z value of RSSI exceeds the threshold of 3 (ZRSSI>3), then the joint outlier score of RSSI is calculated according to the following formula (I): Formula (1) in, The abnormal value of the received signal strength tag; These are the outliers of the Doppler frequency shift tag; These are the outliers of the arrival time difference label; The correlation coefficient between the received signal strength tag and the Doppler frequency shift tag; The correlation coefficient between the received signal strength tag and the time difference of arrival tag; The joint anomaly score is the received signal strength label.

[0031] The joint anomaly scores of the Doppler frequency shift tags and the joint anomaly scores of the time difference of arrival tags are calculated sequentially as described above.

[0032] Step 13: When the joint anomaly score is greater than the preset threshold, delete the label data; If the joint anomaly score of the received signal strength tag If the RSSI value still exceeds the threshold of 3, it is considered an outlier; otherwise, it is a short-term mismatch between modalities. A short-term mismatch means that although the raw data for one modality may appear abnormal, considering the correlation with other modalities, this anomaly may be temporary or caused by the influence of other modalities. For example, if environmental interference causes a small fluctuation in RSSI, but the trends of Doppler Shift and TDOA do not change synchronously, this can be considered a short-term mismatch. Even if initial detection of RSSI data (ZRSSI > 3) shows a deviation from the normal value, joint score detection reveals that the outlier deviation can be partially explained by the trend changes of Doppler Shift and TDOA. In this case, the RSSI deviation is not independent but caused by short-term dynamic fluctuations between multimodal data, thus constituting a short-term mismatch rather than a true anomaly. If the ARSSI value is large, it indicates that the RSSI data deviation is independent of other modalities and may be caused by a true anomaly (such as environmental interference). Abnormal RSSI data are directly discarded, and the remaining usable RSSI data can be used for further processing.

[0033] For Doppler Shift and TDOA data, outliers are identified using the intermodal correlation matrix established above. Consistent with RSSI data processing, when their Z values ​​exceed their respective set thresholds, a joint outlier score is calculated based on other modal data for further assessment. Once identified as an outlier, it is directly removed, and subsequent processing utilizes the remaining valid data without requiring correction of the outlier, thus ensuring data reliability and the accuracy of the localization results.

[0034] During data acquisition, outliers may appear in RSSI, Doppler Shift, and TDOA data due to environmental factors, sensor errors, and other reasons. Outliers not only affect the training performance of the model but may also lead to biased prediction results. Therefore, it is essential to detect and remove outliers during the data processing stage.

[0035] Step 14: Perform time-domain filtering, frequency-domain filtering, low-pass filtering, and standardization on the initial label dataset after deletion to obtain the target dataset; specifically, this includes: To enhance signal quality, a unified three-stage filtering strategy is adopted for RSSI, Doppler Shift, and TDOA data.

[0036] The first stage is time-domain filtering, employing a median filtering method to remove sudden outliers and short-term noise. Specifically, a sliding window median filter is applied to the acquired signal, with the window size flexibly chosen based on the sampling frequency. This method is applicable to all three types of data, and is particularly effective in processing outliers caused by reflection interference in RSSI data. The sliding window median filter has significant advantages: it effectively handles sudden outliers and short-term noise, successfully removes spike interference from the signal, is unaffected by outliers, and simultaneously preserves the main characteristics and trends of the signal.

[0037] The second stage is frequency domain filtering, which aims to convert the signal to a frequency domain representation using FFT, thereby removing high-frequency noise or periodic interference. Specifically, this involves performing a Fast Fourier Transform on the segmented signal, followed by designing appropriate filters based on the spectral characteristics, such as bandpass or notch filters. This stage primarily targets RSSI and Doppler Shift data with significant periodic interference.

[0038] The third stage implements low-pass filtering, the goal of which is to smooth signal changes and extract the main trends. By using a low-pass filter and setting a reasonable cutoff frequency, such as fc = 0.1Hz, this filtering method is applicable to all data types and effectively ensures the smoothness and stability of the signal.

[0039] This unified three-stage filtering strategy is applicable to all three signal types. This not only greatly simplifies system complexity but also effectively ensures coordinated processing between different signals, significantly improving overall signal quality. The unified filtering strategy ensures that RSSI, TDOA, and Doppler Shift data remain highly consistent during processing to obtain the final target data, effectively avoiding inconsistencies that might arise from independent processing. Furthermore, by dynamically adjusting filter parameters, this strategy maintains good performance under various environmental conditions. This adaptability allows the system to better cope with complex and changing real-world environments, significantly improving its robustness.

[0040] Since different input data have different numerical ranges and distributions, data standardization is crucial. First, three types of data are collected from product m by reader n. , and The mean and standard deviation were calculated independently, and then the Z-score standardization method was applied to make the mean of each input data point 0 and the standard deviation 1, as shown in Formula (II): Formula (II) in, It is standardized label data. It is the raw tag data, namely the received signal strength tag, Doppler frequency shift tag, or time difference of arrival tag; It is the mean of the label data. This is the standard deviation of the labeled data. This standardization method transforms all data to the same scale, accelerating the convergence of neural networks and reducing the risk of gradient explosion or vanishing gradients.

[0041] During the training phase of a neural network, the dataset must first be appropriately divided into a training set, a validation set, and a test set, with the training set accounting for 80%, and the validation and test sets each accounting for 10%, to ensure the model's generalization ability. Since multiple products need to be input into the model separately for independent training and prediction, a separate dataset is constructed for each product, ensuring the independence and consistency of the data partitioning.

[0042] During training, mean squared error (MSE) is used as the loss function to quantify the deviation between the model's predicted values ​​and the true values. The model is trained separately for each product's data, with specific steps including forward propagation and backpropagation. During forward propagation, the network performs calculations based on the input data to obtain predicted values; during backpropagation, the model adjusts the network weights using the mean squared error based on the difference between the predicted and true values, gradually optimizing the network parameters. After multiple iterations, the neural network gradually learns and captures the complex relationship between the input data and the spatial location of the product, achieving accurate prediction of the product's location. To prevent overfitting during training and ensure the stability of the validation set error, L2 regularization is employed to effectively control model complexity and avoid excessive fluctuations in the model's performance on the validation set.

[0043] The Adam optimizer was used in the optimization phase. The Adam optimizer can automatically and adaptively adjust the learning rate based on the mean and variance of each parameter, thus effectively handling tricky problems such as vanishing or exploding gradients that may occur during training. The learning rate was set to 0.001, the batch size to 32, and the number of training epochs to 50. Through these optimization strategies and reasonable parameter configurations, the neural network can converge more efficiently and stably during training, continuously optimizing network parameters. Finally, the model's prediction accuracy is accurately evaluated using a test set, thereby determining the model's quality and generalization ability.

[0044] In addition, to achieve real-time monitoring of air quality and the environment within the supermarket, a merchandise management system based on electronic tag technology and an IoT architecture can deploy various types of sensors in multiple key areas (such as the fruit and vegetable section, fresh produce section, and refrigerated section), including temperature, humidity, PM10, PM2.5, and light intensity sensors. Each sensor collects data at a frequency of once per minute and transmits the data to a central server via a wireless network. This ensures the real-time nature and accuracy of the data, adapting to rapid changes in the supermarket environment. The acquired environmental data can also be used for the following operations: When processing supermarket temperature, humidity, PM2.5, PM10, and light intensity monitoring data, the first step is to address missing values ​​caused by sensor malfunctions or unstable data transmission. To resolve this issue, the current environmental information of the products can be predicted based on historical environmental information obtained from the products, using the following formula. The current environmental prediction data is obtained according to formula (III). : Formula (3) Where t represents the time when data prediction is required. To predict environmental data, The data obtained at time t is the data acquired at the time preceding time t. The data is obtained one time after time t. For the preset time interval, is the base of the natural logarithm.

[0045] This method takes into account the rate of change between data points and attempts to smooth this change through exponential decay, making the results more stable and natural.

[0046] During data acquisition, outliers may appear due to environmental factors, sensor errors, and other reasons. Outliers not only affect the training performance of the model but may also lead to biased prediction results. Therefore, it is essential to detect and remove outliers during the data processing stage. A commonly used outlier detection method is the Z-Score method, which determines whether a data point is an outlier by calculating the standard deviation distance between the data point and the mean. If the absolute value of the Z-score of a data point is greater than a set threshold (usually 3), the data point is considered an outlier. After identifying outliers, they can be replaced with the average value within the same time window using mean substitution, or interpolation can be used to repair them, in order to maintain the smoothness and continuity of the data and ensure that the accuracy of model training and prediction is not affected.

[0047] Finally, to improve model training efficiency and prediction accuracy, all obtained environmental data are normalized using the following formula (iv): Formula (IV) It is normalized environmental data. It's environmental data. It is the median of all environmental data. It is the interquartile range of all environmental data, that is, the difference between the third quartile (Q3) and the first quartile (Q1).

[0048] The neural network structure diagram of the environmental prediction model in a commodity management system based on electronic tag technology and IoT architecture is shown below. Figure 3 As shown, the model's input consists of five independent time-series data points from the past 24 hours: temperature, humidity, PM10, PM2.5, and light intensity. The input data has a shape of (1440, 1), where 1440 represents the number of sampling points in the past 24 hours. Each time-series data point is used as a separate input channel for the model, which receives a total of five channels of time-series data, corresponding to the dynamic changes in temperature, humidity, PM10, PM2.5, and light intensity. The feature extraction layer includes two bidirectional LSTM layers responsible for extracting temporal features from the input time-series data. The first layer is a bidirectional LSTM layer with 128 neurons. This layer independently captures the temporal dependencies of each input channel's time-series data, extracting the temporal dynamic features of each data point. Subsequently, a second bidirectional LSTM layer with 64 neurons further extracts higher-level temporal features, providing a deeper understanding of the evolution of each channel's data in the temporal dimension and offering richer temporal features for subsequent feature fusion. The feature fusion layer uses two fully connected layers to deeply fuse temporal features extracted from five independent time series, uncovering potential nonlinear coupling relationships between different environmental data. The first fully connected layer contains 32 neurons, initially integrating the various temporal features and using a ReLU nonlinear activation function to achieve nonlinear transformations between features, enhancing their interactions. The second fully connected layer contains 16 neurons, also incorporating a ReLU nonlinear activation function, further performing nonlinear transformations and compression on the fused features, extracting key high-dimensional feature representations, and deeply exploring the complex interrelationships between the five data points: temperature, humidity, PM10, PM2.5, and light intensity. For example, changes in humidity may cause fluctuations in PM2.5 and PM10 concentrations, while changes in light intensity may affect temperature. Through the efficient modeling of the feature fusion layer, the system can capture the nonlinear coupling patterns between these data, thereby significantly improving the ability to predict future environmental changes. Ultimately, the output layer contains 5 neurons, each corresponding to a predicted value of temperature, humidity, PM10, PM2.5, and light intensity for the next 60 minutes. The output result has a shape of (60, 5), achieving accurate prediction of the changing trends of the five independent data.

[0049] In the training process of the neural network model for environmental prediction, data from each key area (such as the vegetable and fruit area, fresh produce area, and cold storage area) is collected separately, and the corresponding neural network is trained separately. First, the datasets for each area are divided into training, validation, and test sets, with the training set accounting for 80%, and the validation and test sets each accounting for 10%. To capture the trend and periodic characteristics of multivariate environmental data (temperature, humidity, PM10, PM2.5, and light intensity) in each area, a data window size of 24 hours is set, containing data from the past 24 hours for these five environmental variables as input; the target window size is 1 hour, used to predict the changing trends of these five variables in the next hour. A sliding time window technique is used to split the training set for each area, moving the window forward by one hour each time to generate a new training sample. In this way, the data for each area is separately divided into multiple samples, each used to train the corresponding model for that area.

[0050] To measure the predictive performance of the model for each region, the root mean square error (RMSE) is used as the loss function, directly reflecting the error between the predicted and true values. The model training process is set to 50 iterations. In each iteration, the model for each region calculates the predicted value through forward propagation and updates the network weights based on the error between the predicted and true values ​​using backpropagation. During training, the RMSE value on the validation set is continuously monitored. If the validation set RMSE no longer improves after several iterations, training is terminated early using an early stopping mechanism to prevent overfitting.

[0051] To further suppress overfitting, L2 regularization was introduced during model training for each region. By imposing constraints on the model weights, the magnitude of the weights was reasonably controlled, reducing the model's sensitivity to noisy data. In the optimization phase, the Adam optimizer was used for parameter updates. The Adam optimizer has an adaptive learning rate adjustment mechanism, effectively avoiding gradient explosion and gradient vanishing problems. The initial learning rate was set to 0.001, and the batch size was set to 32 to ensure that the model converged quickly for each region with reasonable computational resources.

[0052] Through the above optimization strategies, data from each key region is trained and validated using independent neural networks. This ensures that the model can more accurately learn the environmental change characteristics of that region, improving prediction performance and generalization ability. Finally, the model for each region is evaluated separately on the test set to verify its accuracy in predicting five variables—temperature, humidity, PM10, PM2.5, and light intensity—over the next hour. This ensures the effectiveness and reliability of the system in real-world application scenarios across different regions.

[0053] The architecture of the commodity management system based on electronic tag technology and Internet of Things architecture involved in this invention is as follows: Figure 4 As shown, it mainly consists of three parts: the perception layer, the network layer, and the application layer. The perception layer collects product location information and supermarket environmental data through RFID readers and environmental sensors (such as temperature and humidity sensors, PM2.5, PM10, and light intensity sensors). RFID readers are used to acquire the RSSI, Doppler Shift, and TDOA data of products, tracking their location in real time. Environmental sensors monitor environmental variables such as temperature, humidity, PM2.5, and PM10 to ensure that the product storage environment meets standards. The perception layer also performs preliminary data processing tasks, such as noise filtering, outlier removal, data missing filling, and standardization, providing accurate data for subsequent analysis.

[0054] The network layer is responsible for securely and reliably transmitting data collected by the perception layer to the central server or other processing nodes. It is compatible with and supports various communication technologies, such as Wi-Fi, Zigbee, and LoRa, ensuring seamless integration of data from different sensors. The network layer not only manages data transmission but also employs encryption technology to protect data security, preventing leakage and tampering. Through these technologies, the network layer can efficiently manage large amounts of data, providing a solid data foundation for the application layer.

[0055] As the highest layer of the system architecture, the application layer is responsible for integrating and processing data from the perception and network layers. Through a user interface, data analysis tools, and intelligent decision-making modules, it transforms raw data into meaningful information and decision support. It utilizes neural networks with fully connected layers for product positioning prediction, optimizing shelf layout to improve sales efficiency. Furthermore, the application layer uses deep learning models to monitor and predict environmental data in real time, providing early warnings of future environmental changes and adjusting parameters such as temperature and humidity to ensure optimal storage conditions for goods. The system also provides a visual user interface to help managers monitor and adjust operational strategies in real time, improving the supermarket's operational efficiency and competitiveness.

[0056] The solution provided in this application can track the location of goods in a supermarket in real time. By analyzing RSSI, DopplerShift, and TDOA data through a neural network model, it can accurately predict the spatial coordinates of the goods. This function not only helps optimize inventory management and reduce product loss, but also improves logistics efficiency. Through precise product positioning, supermarkets can more rationally plan shelf layouts and implement dynamic pricing strategies, thereby improving operational efficiency. In addition, the system also provides automatic replenishment suggestions, further enhancing the efficiency and accuracy of supply chain management.

[0057] By deploying environmental sensors in key areas of the supermarket, the system monitors changes in data such as temperature, humidity, PM10, and PM2.5 in real time, and uses deep learning models to predict environmental trends for the next hour. When abnormal environmental changes are detected, the system automatically adjusts temperature, humidity, or other environmental parameters to maintain optimal storage conditions for goods. Through real-time monitoring and prediction, the system significantly reduces the risk of product quality damage, reduces energy consumption, and improves operational efficiency.

[0058] Equipped with advanced anti-theft early warning functions, it can detect abnormal movement and unauthorized removal of goods in real time. By analyzing the RSSI, Doppler shift, and TDOA data of RFID tags, the system can identify abnormal movement trajectories of goods and immediately trigger an alarm when potential theft is detected, notifying security personnel to take action.

[0059] It not only provides real-time monitoring and forecasting capabilities but also offers intuitive data support to management through intelligent analytics and visualization tools. By deeply mining historical data, the system can identify factors such as sales patterns, inventory turnover rates, and environmental impacts, helping supermarkets optimize operational strategies and improve profitability. Furthermore, the system supports the automatic generation of reports and analytical charts, enabling management to easily grasp key indicators and make more scientific and rational decisions.

[0060] To more clearly distinguish product labels of different categories and improve product positioning accuracy, this invention provides Embodiment 2, which includes the following steps after completing the standardization process and before obtaining the target dataset: Use such as Figure 2 The IoT architecture model shown calculates the spatial relationship features between central data from intermediate tag datasets of different products; The specific IoT architecture model is as follows: First, input the shape through the input layer. The data, among which This indicates that the number of readers corresponding to product m is N, and 3 represents the number of readers. Get Products The three data ( , and Next, the data enters the flatten layer, which flattens the input data, converting the multidimensional input into a one-dimensional vector. This allows subsequent fully connected layers to better process the data, improving the efficiency and accuracy of feature extraction. Following this, the data enters the feature extraction layer. The three data points from each reader are processed through independent fully connected networks. The first layer contains 16 neurons, and the second layer contains 8 neurons, both using the ReLU activation function. This extracts local features between each reader and the label and captures their non-linear characteristics, thus learning the local spatial relationship between the reader and the label. Next, in the feature fusion layer, the local features from all readers are concatenated into a single global feature vector and processed through two fully connected networks. The first layer contains 128 neurons, and the second layer contains 64 neurons, both using the ReLU activation function. This stage integrates features from multiple readers, further learning the complex global relationship between the reader and the label, including their relative positions and mutual influences in three-dimensional space. Finally, after passing through an output layer containing 3 neurons and using a linear activation function, the model predicts the product. Specific location coordinates in three-dimensional space This enables accurate estimation of the location of goods.

[0061] The spatial relationship features between data points in the intermediate label datasets from different products are calculated as follows: During the training process after the model is built, given the significant differences in the numerical range and distribution of RSSI, Doppler Shift, and TDOA, in order to ensure that data from different modalities can be processed in the same feature space and to improve the model's accuracy in distinguishing goods, for label data from the same label, the contrastive loss function minimizes the distance between them, thus making the label data from the same label closer in the feature space. For label data from different labels, the loss function maximizes the distance between them, and a threshold m (set to 0.3) is introduced to ensure the minimum distance between negative sample pairs. This helps to increase the discriminative power of data from different labels. Spatial relationship features are calculated according to the following formula (V). : Formula (5); in, and Used to describe the source of the label data; y is used to describe whether the label data comes from the same product, when y=1 indicates and Tag data from the same product, where y=0 indicates and Label data from different products; It is a sample and The spatial distance between them; m represents the minimum distance value between data from different labels; This is the regularization parameter, used to control the effects of the non-linear and regularization terms in the loss function. For labeled data from the same label, this function minimizes the distance between them; for labeled data from different labels, it maximizes the distance between them and ensures that the minimum distance is m.

[0062] Further improvements enhance the consistency of label data within the same label, making label data from different labels more spatially dispersed. Spatial relationship features can also be calculated using the following formula (VI). : Formula (VI); Anchor point samples for label data; This indicates another label data source from the same product. This represents label data from different products. Anchor point sample and The distance between them. Anchor point sample and The distance between them. m represents the minimum distance difference between the anchor sample and the positive sample. These are preset weight parameters.

[0063] By introducing contrastive loss functions and triplet loss functions, the model's performance in the feature space can be significantly improved. First, these loss functions enhance the aggregation ability of data with the same label, ensuring that RSSI, Doppler Shift, and TDOA data form tighter clusters in the feature space. This clustering not only reduces the distance between data points with the same label but also makes the multimodal data representation more consistent, thereby improving the model's ability to understand and process data of the same category. Second, by increasing the distance between different labeled data in the feature space, the model can more clearly distinguish product labels of different categories, improving the accuracy of product location.

[0064] For each product, the tag data whose spatial relationship features meet the preset conditions are used as the target dataset for that product.

[0065] like Figure 5 As shown, this embodiment of the invention provides a product management system based on electronic tag technology and an Internet of Things (IoT) architecture. The system includes: The acquisition device 51 is used to establish an intermodal correlation matrix for the tag data in the initial tag dataset of each product acquired by the reader, and to obtain the correlation coefficient between each tag data; the tag data is a received signal strength tag, a Doppler frequency shift tag, or a time difference of arrival tag; The detection device 52 is used to calculate the joint anomaly score of each tag data when the detected anomaly value of the tag data is greater than a preset threshold, based on the anomaly value of the tag, the correlation coefficient corresponding to the tag data, and the anomaly values ​​of other tag data. The first processing device 53 is used to delete the tag data when the joint anomaly score is greater than the preset threshold. The second processing unit 54 is used to sequentially perform time-domain filtering, frequency-domain filtering, low-pass filtering, and standardization on the initial label dataset after the deletion operation to obtain the target dataset.

[0066] This invention provides an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. The program or instructions are executed by the processor to perform the steps of the above-described commodity management method based on electronic tag technology and Internet of Things architecture.

[0067] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0068] In this application, unless otherwise expressly specified and limited, the use of descriptions such as "above" or "below" the second feature indicates that the first and second features are in direct contact or indirect contact via an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. Similarly, "below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0069] It should be noted that if an element is referred to as being "fixed to" or "set on" another element, it can be directly on the other element or there may be an intermediate element. If an element is considered to be "connected to" another element, it can be directly connected to the other element or there may be an intermediate element present. Where applicable, the terms "vertical," "horizontal," "upper," "lower," "left," "right," and similar expressions used in this application are for illustrative purposes only and do not represent the only possible implementations. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; however, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0070] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A commodity management method based on electronic tag technology and Internet of Things architecture, characterized in that, The method includes: A modal correlation matrix is ​​established for the tag data in the initial tag dataset of each product obtained by the reader, and the correlation coefficient between each tag data is obtained; the tag data is a received signal strength tag, a Doppler frequency shift tag, or a time difference of arrival tag; For each tag data, when the detected outlier value of the tag data is greater than the preset threshold, the joint outlier score of the tag data is calculated based on the outlier value of the tag, the correlation coefficient corresponding to the tag data, and the outlier values ​​of other tag data. When the combined anomaly score is greater than the preset threshold, the label data is deleted; The initial label dataset after deletion is processed sequentially with time-domain filtering, frequency-domain filtering, low-pass filtering, and standardization to obtain the target dataset.

2. The method as described in claim 1, characterized in that, The process of obtaining the correlation coefficients between the various tag data includes: When the tag data is the received signal strength tag, the correlation coefficient between the received signal strength tag and the Doppler frequency shift tag is obtained, as well as the correlation coefficient between the received signal strength tag and the time difference of arrival tag; When the tag data is the Doppler frequency shift tag, the correlation coefficient between the Doppler frequency shift tag and the received signal strength tag, and the correlation coefficient between the Doppler frequency shift tag and the time difference of arrival tag are obtained; When the tag data is the time difference of arrival tag, the correlation coefficient between the time difference of arrival tag and the received signal strength tag is obtained, as well as the correlation coefficient between the time difference of arrival tag and the Doppler frequency shift tag.

3. The method as described in claim 1, characterized in that, The step of calculating the joint anomaly score of the tag data based on the outlier of the tag, the correlation coefficient corresponding to the tag data, and the outlier of other tag data includes: The joint anomaly score is calculated using the following formula: ； in, The abnormal value of the received signal strength tag; These are the outliers of the Doppler frequency shift tag; These are the outliers of the arrival time difference label; The correlation coefficient between the received signal strength tag and the Doppler frequency shift tag; The correlation coefficient between the received signal strength tag and the time difference of arrival tag; The joint anomaly score is the received signal strength label.

4. The method as described in claim 1, characterized in that, After completing the standardization process and before obtaining the target dataset, the following further steps are included: The standardized data is used as an intermediate label dataset; Using an IoT architecture model, spatial relationship features between central data from intermediate tag datasets of different products are calculated; For each product, the tag data whose spatial relationship features meet the preset conditions are used as the target dataset for that product.

5. The method as described in claim 4, characterized in that, The calculation of spatial relationship features between intermediate label datasets from different products includes: The spatial relationship features are calculated using the following formula. ; ； in, and Used to describe the source of the label data; y is used to describe whether the label data comes from the same product, when y=1 indicates and Tag data from the same product, where y=0 indicates and Label data from different products; It is a sample and The spatial distance between them; m represents the minimum distance value between data from different labels; It is the regularization parameter.

6. The method as described in claim 4, characterized in that, The calculation of spatial relationship features between intermediate label datasets from different products includes: The spatial relationship features are calculated using the following formula. ; ; Anchor point samples for label data; This indicates another label data from the same product. This represents label data from different products. Anchor point sample and The distance between them. Anchor point sample and The distance between them, m represents the minimum distance difference between the anchor sample and the positive sample. These are preset weight parameters.

7. The method as described in claim 1, characterized in that, The method further includes: Based on the historical environmental information of the product obtained, the current environmental information of the product is predicted according to the following formula to obtain the current environmental prediction data. ; ; Where t represents the time when data prediction is required. To predict environmental data, The data obtained at time t is the data acquired at the time preceding time t. The data is obtained one time after time t. For the preset time interval, is the base of the natural logarithm.

8. The method as described in claim 7, characterized in that, All obtained environmental data were normalized using the following formula: ； It is normalized environmental data. It's environmental data. It is the median of all environmental data. It is the interquartile range in all environmental data, that is, the difference between the third quartile (Q3) and the first quartile (Q1).

9. The method as described in claim 1, characterized in that, The models required to implement the merchandise management method include: Input layer, used to obtain the shape as The data, among which This indicates that the number of readers corresponding to product m is N, and 3 represents the number of readers. Get Products Tag data; The transformation layer is used to convert the input label data into a one-dimensional vector; The feature extraction layer is used to extract local features and non-linear characteristics between each reader and the product's tag; The feature fusion layer is used to integrate features from all readers.

10. A commodity management system based on electronic tag technology and Internet of Things architecture, characterized in that, The system includes: The acquisition device is used to establish an intermodal correlation matrix for the tag data in the initial tag dataset of each product acquired by the reader, and to obtain the correlation coefficient between each tag data; the tag data is a received signal strength tag, a Doppler frequency shift tag, or a time difference of arrival tag; The detection device is used to calculate the joint anomaly score of each tag data when the detected anomaly value of the tag data is greater than a preset threshold, based on the anomaly value of the tag, the correlation coefficient corresponding to the tag data, and the anomaly values ​​of other tag data. A first processing device is used to delete the tag data when the joint anomaly score is greater than the preset threshold. The second processing unit is used to sequentially perform time-domain filtering, frequency-domain filtering, low-pass filtering, and standardization on the initial label dataset after the deletion operation to obtain the target dataset.

Images