Radio frequency channel door and object passage identification method based on multi-band collaborative control

Abstract

The present invention relates to the field of radio frequency channel gate technology, and specifically discloses a radio frequency channel gate and an article passage identification method based on multi-band collaborative control, wherein the radio frequency channel gate includes a multi-band reader / writer, a multi-band antenna array, a millimeter wave radar module, a metal shielding frame, a distributed signal processing module, an intelligent control module, a communication interface module, and an auxiliary indication device. The radio frequency channel gate and the article passage identification method of the embodiment of the present invention utilize millimeter wave radar and a metal attention module to generate a three-dimensional point cloud map and a spatial attention weight matrix in real time, effectively suppressing reflection interference and multipath effects in metal-dense scenes, and improving the reliability of tag signal recognition; with the help of a dynamic beamforming algorithm, the power ratio of the tag signal to the interference signal is maximized, and the LSTM model is combined to predict the tag trajectory and dynamically switch the NOMA / TDMA mode to achieve intelligent allocation of transmission power and time slots in high-traffic scenes.

Description

Technical Field

[0001] The present invention relates to the field of radio frequency channel door technology, and in particular to a radio frequency channel door based on multi-band collaborative control and an object passage identification method. Background Art

[0002] The traditional way to identify goods is to attach a barcode to the object and use the photoelectric effect generated by the barcode reader to read the barcode and thus identify the goods. However, the disadvantage of barcodes is that they have a short recognition distance and require a light source to illuminate the barcode.

[0003] In order to overcome the shortcomings of barcode readers, the current common practice is to use radio frequency identification technology (RFID) to attach electronic tags to objects. After the electronic tags enter the magnetic field, if they receive the ultra-high frequency signal emitted by the reader antenna, they can use the energy obtained from the induced current to send out the product information stored in the electronic chip, or actively send a signal of a certain frequency. After the reader reads and decodes the information, it sends it to the information system for data processing, thereby obtaining the item information.

[0004] In existing technologies, when radio frequency channel gates are placed in metal-dense scenarios (such as warehouses and security checkpoints), metal objects will cause strong reflection and multipath effects, causing tag signals to be drowned out and increasing the recognition miss rate. In addition, the existing TDMA protocol has a time slot collision rate of up to 25% in high-traffic scenarios (>80 tags / sec), and static resource allocation cannot adapt to burst traffic, which will cause a surge in recognition delays. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems in the related art to a certain extent. To this end, the present invention aims to propose a radio frequency channel gate and object passage identification method based on multi-band coordinated control to achieve improved identification accuracy and efficiency.

[0006] To achieve the above objectives, the first embodiment of the present invention proposes a radio frequency channel gate based on multi-band coordinated control, including:

[0007] Multi-band reader: The operating frequency band is 860-960MHz, divided into 16 sub-bands;

[0008] Multi-band antenna array: placed on both sides of the channel door frame, including UHF flat-panel directional antennas for transmitting and receiving RF signals;

[0009] Millimeter-wave radar module: operating frequency 77-81 GHz, outputs a 3D point cloud image of metal objects;

[0010] Metal shielding frame: embedded with flexible electromagnetic shielding layer, the spacing between adjacent antennas is λ / 2±10%, where λ is the center wavelength of the operating frequency band;

[0011] Distributed signal processing module: Each antenna node is equipped with a digital signal processor (DSP) and a built-in convolutional neural network (CNN) algorithm. The CNN includes a metal attention module (MAM), which is used to optimize the beamforming weight vector through a dynamic beamforming algorithm to maximize the power ratio of the tag signal to the interference signal.

[0012] Intelligent control module: uses reinforcement learning algorithm to dynamically optimize frequency band selection strategy;

[0013] Communication interface module: supports TCP / IP, RS485 and CAN bus protocols;

[0014] Auxiliary indicator device: includes LED status light and LCD display.

[0015] To achieve the above-mentioned object, a second embodiment of the present invention provides a method for identifying object passage, which is applied to the above-mentioned radio frequency channel door and includes the following steps:

[0016] S1. Initialize the system and load the pre-trained convolutional neural network and long short-term memory network models;

[0017] S2, millimeter wave radar scanning generates real-time 3D point cloud map;

[0018] S3. Execute a dynamic beamforming algorithm to suppress metal interference by optimizing the beamforming weight vector. The optimization goal of this algorithm is to maximize the power ratio of the tag signal to the interference signal. The updating process of the beamforming weight includes adaptively adjusting the step size factor.

[0019] S4, dynamically switching between non-orthogonal multiple access (NOMA) mode and time division multiple access (TDMA) mode according to the collision probability distribution graph;

[0020] S5, calculating the probability of item passage through a convolutional neural network enhanced by the metal attention module;

[0021] S6. When the probability of passing exceeds the preset threshold, passage is allowed; otherwise, an alarm of the corresponding level is triggered.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] The RF channel gate and object identification method based on multi-band coordinated control in the embodiments of the present invention uses millimeter-wave radar and a metal attention module to generate a three-dimensional point cloud map and a spatial attention weight matrix in real time, effectively suppressing reflection interference and multipath effects in dense metal scenes, and improving the reliability of tag signal recognition.

[0024] This object passage identification method uses a dynamic beamforming algorithm to maximize the power ratio of tag signals to interference signals, combines the LSTM model to predict tag trajectories and dynamically switches between NOMA / TDMA modes, realizes the intelligent allocation of transmission power and time slots in high-traffic scenarios, reduces the probability of signal conflicts, adapts to burst traffic demands, and solves the problems of high missed detection rate caused by metal interference and low efficiency of static resource allocation in existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0026] Figure 1 1 is a schematic diagram of the workflow of a radio frequency channel gate based on multi-band collaborative control in one embodiment of the present invention;

[0027] Figure 2 is a flow chart of a method for identifying passage of an object in another embodiment of the present invention;

[0028] Figure 3 Schematic diagram of the metal thermal properties of a metal shelf in one embodiment of the present invention;

[0029] Figure 4 3D spatial distribution diagram of power allocation of 50 tags in NOMA mode according to one embodiment of the present invention;

[0030] Figure 5 2 is a schematic diagram comparing collision rates among static TDMA, pure NOMA, and the dynamic switching mechanism of this solution in one embodiment of the present invention;

[0031] Figure 6 3. It is a schematic diagram comparing the identification delay between static TDMA and this solution in one embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0033] The following describes, with reference to the accompanying drawings, a radio frequency channel door and an object passage identification method based on multi-band collaborative control according to an embodiment of the present invention.

[0034] Radio frequency identification (RFID) technology has been widely used in scenarios such as warehousing and logistics, and security checkpoints. However, traditional RFID systems have significant performance bottlenecks in the following scenarios:

[0035] Traditional RFID channel gates: They operate in a single frequency band. Metal reflections in dense metal environments cause multipath effects, resulting in signal attenuation of up to 20-30dB. Furthermore, the static TDMA protocol has a collision rate of up to 25% in high-traffic scenarios, making it unable to adapt to burst traffic.

[0036] To this end, this solution proposes a radio frequency channel gate based on multi-band collaborative control, including:

[0037] Multi-band reader: The operating frequency band is 860-960 MHz, divided into 16 sub-bands, each with a bandwidth of 6.25 MHz. Frequency division multiplexing technology automatically switches from bands with strong metal reflections to sub-bands with less interference. For example, if the 915 MHz band is detected to be interfering with reflections from metal shelves, the intelligent control module uses a reinforcement learning algorithm to switch to the 875 MHz sub-band, improving the signal-to-noise ratio (SNR) by approximately 3-5 dB.

[0038] Typical available models include the Impinj SpeedwayR700 (supporting 860-960 MHz and featuring a built-in 16-channel digital receiver). The multi-band reader outputs frequency band selection instructions to the intelligent control module, which is dynamically optimized by a reinforcement learning algorithm and can also exchange tag signal strength data with the distributed signal processing module via the SPI bus.

[0039] Multi-band antenna array: 4×4 UHF flat-panel directional antenna arrays are deployed on each side frame, with an operating frequency of 920±40MHz, a horizontal coverage angle of ±60° (can cover a 3-meter wide channel), and a vertical coverage angle of 45° (to adapt to objects of different heights).

[0040] The multi-band antenna array is used to receive metal position data from the millimeter-wave radar module and dynamically adjust the beam direction. The distributed signal processing module controls the phase difference of each antenna through the digital beamforming (DBF) algorithm.

[0041] Millimeter-wave radar module: Operating frequency 77-81 GHz, transmitting frequency modulated continuous wave (FMCW) in the 77-81 GHz band, generating a three-dimensional point cloud by receiving reflected signals from metal objects.

[0042] For example, for metal shelves, point cloud data containing coordinates (x, y, z) and reflection intensity can be generated, with a positioning accuracy of ±2cm. Furthermore, the millimeter-wave radar module can create a metal fingerprint library and output a metal heat map matrix Q. Once the metal fingerprint library is established, the system automatically activates the metal interference suppression strategy when the same fingerprint is detected.

[0043] The metal shielding frame is embedded with a flexible electromagnetic shielding layer, and its outer layer can be a flexible polyurethane foam, which suppresses the metal frame's own reflections. The shielding layer shape is optimized through electromagnetic simulation to achieve phase cancellation of reflected waves. The internal spacing between adjacent antennas is λ / 2±10%, meeting the optimal coupling conditions of the Friis transmission formula, where λ is the center wavelength of the operating frequency band.

[0044] Distributed signal processing module: Each antenna node is equipped with a digital signal processor (DSP) and a built-in convolutional neural network (CNN) algorithm, which includes a metal attention module (MAM).

[0045] DSP can be selected from: Texas Instruments TMS320C6678 (8-core C66x DSP, floating-point computing capability of 40 GFLOPS) or Xilinx Zynq UltraScale+ MPSoC (integrated ARM Cortex-A53 and FPGA, supporting parallel computing).

[0046] Intelligent control module: uses reinforcement learning algorithm to dynamically optimize frequency band selection strategy and switch between NOMA / TDMA mode based on different situations.

[0047] Communication interface module: supports TCP / IP, RS485 and CAN bus protocols;

[0048] Auxiliary indicator device: includes LED status light and LCD display.

[0049] When the RF gate is working, it first initializes the system and loads the pre-trained convolutional neural network and long short-term memory network models. The millimeter-wave radar scans to generate a real-time three-dimensional point cloud image. The distributed signal processing module uses the metal attention module to convert the metal heat map matrix from the three-dimensional point cloud image. After 3×3 convolution and Sigmoid activation function, it generates a spatial attention weight matrix, weights the clean signal and shields the interference area. At the same time, it executes the dynamic beamforming algorithm and iteratively updates the weights using the minimum mean square error criterion to align the main lobe of the antenna array with the target tag and the null sink to the interference source to maximize the power ratio between the tag and the interference signal. Then, the intelligent control module uses the long short-term memory network model to predict the tag's future trajectory and movement speed, and generates a spatial conflict probability distribution map.

[0050] According to the probability of conflict, it dynamically switches between non-orthogonal multiple access (NOMA) mode and time division multiple access (TDMA) mode. In the non-orthogonal multiple access mode, the transmitter dynamically allocates power according to the signal-to-noise ratio. Tags with high signal-to-noise ratio are allocated less power, and the receiver executes a continuous interference cancellation algorithm. In the time division multiple access mode, the time slot length is adjusted according to the signal-to-noise ratio and the time slot is allocated with priority according to the distance from the tag to the center of the channel gate. The probability of the passage of the item is then calculated through the convolutional neural network enhanced by the metal attention module. When it exceeds the preset threshold, it is allowed to pass, otherwise a three-level alarm mechanism is triggered. At the same time, the passage records including timestamps, tag identity hash values, etc. are stored based on blockchain technology, and the intelligent control module uses a reinforcement learning algorithm to dynamically optimize the frequency band selection strategy to form a closed-loop system of perception, processing, decision-making, and feedback.

[0051] Among them, the core of the power allocation formula of the non-orthogonal multiple access mode is to dynamically adjust the transmission power according to the signal-to-noise ratio and collision probability of each tag. The metal attention module generates a weight matrix through convolution and activation functions to suppress interference.

[0052] In this solution, the RF gate uses millimeter-wave radar and metal attention module to generate three-dimensional point cloud maps and spatial attention weight matrices in real time, effectively suppressing reflection interference and multipath effects in metal-dense scenes, and improving the reliability of tag signal recognition; with the help of dynamic beamforming algorithm, the power ratio of tag signal to interference signal is maximized, combined with the long short-term memory network model to predict tag trajectory and dynamically switch non-orthogonal multiple access / time division multiple access mode, to achieve intelligent allocation of transmission power and time slots in high-traffic scenarios, reduce the probability of signal conflict, adapt to sudden traffic demands, and solve the problems of high missed detection rate and low efficiency of static resource allocation caused by metal interference in existing technologies. At the same time, the multi-level alarm mechanism and blockchain data storage improve security and data reliability.

[0053] In some embodiments of the present invention, the Metal Attention Module (MAM) is the core component of the distributed signal processing module, and its input and output form a closed-loop collaboration with various modules of the system.

[0054] Data input chain: millimeter wave radar module → 3D point cloud image → metal heat map matrix Q → MAM → spatial attention weight matrix A attn .

[0055] As an example, when a metal shelf enters an aisle, the 77GHz millimeter-wave radar generates a three-dimensional point cloud containing more than 1,000 points, which is then mapped to a 16×16 Q matrix after coordinate mapping and intensity normalization.

[0056] Processing output chain: A attn →weighted signal y out →Dynamic beamforming →NOMA / TDMA mode parameter adjustment →Pass probability calculation.

[0057] As an example, A attn The weight value of the metal area is close to 0, making y out The reflected signal from the middle shelf is suppressed by more than 20dB.

[0058] The Metal Attention Module (MAM) handles metal interference in the following way:

[0059] First, generate the spatial attention weight matrix A based on the 3D point cloud image attn :

[0060] A attn =σ(Conv 3×3 (Q)

[0061] Among them, Conv 3×3 This is a 3×3 convolutional layer that uses a zero-padded 3×3 convolution kernel to extract local spatial features of metal distribution, such as edges, corners, and other areas of strong reflection. The convolution kernel parameters are optimized by pre-training a CNN (such as ResNet-18) on a metal reflection dataset.

[0062] σ(·) is the Sigmoid activation function, which maps the convolution output to the [0,1] interval and generates normalized attention weights. For example, in the metal dense area A attn ≈0.1, tag signal area A attn ≈0.9;

[0063] Q is the metal heat map matrix, which is obtained by voxelizing the 3D point cloud. Each pixel value in the metal heat map represents the metal reflection intensity (unit: dBm) of the corresponding spatial area.

[0064] Then output the weighted signal:

[0065] y out =A attn ⊙y clean +(1-A attn )⊙y null

[0066] Among them, ⊙ is the Hadamard product, and spatial filtering is achieved by element-by-element multiplication;

[0067] y null is a shielding signal matrix composed of all zero elements, which is the same as (1-A attn ) multiplied to zero the metal area signal;

[0068] y clean It is a clean signal matrix, that is, the RF signal after interference suppression. The specific methods include denoising through wavelet transform, retaining the 920MHz±40MHz tag frequency band, and suppressing metal reflection noise.

[0069] like Figure 3 The metal heat map of the metal shelf is shown, with the X-axis representing the horizontal direction and the Y-axis representing the depth direction. Figure 3 In the Q matrix, metal shelves are represented by pixel values ​​> 0.8 in the center and < 0.3 at the edges. The pixel values ​​in the metal shelf area are significantly higher, indicating stronger reflection intensity, while the reflection interference from the metal edges is weaker, but still higher than in the metal-free area.

[0070] For example, the pixel Q value in the center of the shelf reaches 0.9dBm, while the edge area is around 0.4dBm. This indicates that the metal shelf has a significant reflection interference on the RF signal, seriously affecting the accurate reading of the tag signal.

[0071] As an example, consider a large metal shelf in a smart warehouse environment. The shelf is 5 meters long, 2 meters wide, and 3 meters high, made of 3mm thick steel. Multiple items with radio frequency tags are placed on the shelf. As these items pass through the radio frequency access door, the millimeter-wave radar activates, generating a point cloud image at 10 frames per second. At a specific moment in the point cloud image, the coordinates of the metal shelf in 3D space range from (x1, y1, z1) to (x2, y2, z2). These coordinates are converted into a 16×16 metal heat map matrix Q through voxelization.

[0072] Then, the Metal Attention Module (MAM) comes into play. First, the Q matrix is ​​extracted through a 3×3 convolutional layer (zero padding). The parameters of this convolutional layer are pre-trained based on a large number of RF signal datasets containing scenes such as metal shelves and security gates. In the processing of the specific scene of metal shelves, the convolutional layer can accurately extract the features of strongly reflective areas such as metal edges and corners. After convolution processing, the characteristic response value of the central area of ​​the metal shelf reaches above 0.7, clearly identifying the strongly reflective area;

[0073] Then, the spatial attention weight moment A is generated by the Sigmoid activation function attn In this example, the A in the metal shelf area attn The value is calculated to be 0.15, and the A attn The value is as high as 0.9. This means that in the subsequent signal processing, the signal in the metal shelf area will be greatly suppressed, while the tag signal will be retained to the greatest extent;

[0074] Finally, the output signal is calculated by the Hadamard product, where y cleanThis is the signal after wavelet transform denoising, which preserves the effective signal components within the 920MHz±40MHz frequency band. After MAM processing, the signal strength in the metal shelf area is suppressed by 22dB, while the integrity of the tag signal is effectively protected, with signal strength loss within 3dB.

[0075] The above application examples clearly demonstrate that the Metal Attention Module (MAM) accurately identifies metal interference sources and appropriately weights the signal based on the interference level, effectively suppressing the impact of metal interference on tag signals and providing a high-quality signal foundation for subsequent tag recognition and passage determination. This highly effective metal interference suppression mechanism significantly improves system accuracy and reliability in numerous scenarios, including smart warehousing, airport security, and port container clearance, offering significant practical value.

[0076] In some embodiments of the present invention, the metal reflection fingerprint library established based on the millimeter wave radar module is parsed, and its functions include:

[0077] For example, when a metal forklift enters the passage, the fingerprint library matches its unique features, triggering the MAM to shield the signal of the area.

[0078] The millimeter-wave radar module establishes a metal reflection fingerprint library in the following ways:

[0079] First, a detection signal is emitted in a tag-free environment to collect the intensity, delay, and distance characteristics of the metal reflector, where:

[0080] Intensity characteristics: reflected signal power spectrum density, unit dBm / Hz, acquisition formula is:

[0081]

[0082] Where s(t) is the received signal, T is the sampling time, and f is the frequency.

[0083] Delay characteristics: Based on frequency modulated continuous wave (FMCW) time of flight (ToF) measurement, the formula is:

[0084]

[0085] Where d is the distance from the metal object to the radar, c is the speed of light, and the ranging error is <5cm.

[0086] Distance characteristics: Calculated by FMCW sweep bandwidth Δf, the distance resolution is:

[0087]

[0088] As an example, when Δf=4 GHz, Δd=3.75 cm.

[0089] Then, the phase change rate of the reflected signal is extracted as the metal fingerprint feature parameter;

[0090] The extraction of phase change rate includes:

[0091] Instantaneous phase calculation: Perform Hilbert transform on the reflected signal s(t) to extract the phase:

[0092]

[0093] Phase change rate (key characteristic parameter of metal fingerprint):

[0094]

[0095] Because the PCR characteristics of metal objects are unique, for example, the PCR of a metal plate is ≈ 0.2 rad / s (at rest), while the PCR of a metal cylinder is ≈ 0.5 rad / s (slightly shaking).

[0096] Finally, the statistical distribution characteristics of the Doppler frequency shift are calculated to establish a unique identification fingerprint library for metal objects.

[0097] Doppler shift formula:

[0098]

[0099] Where v is the object velocity and θ is the angle between the radar line of sight and the direction of motion.

[0100] As an example, consider a 10m x 8m warehouse with metal shelves (2m long x 3m high x 0.5m wide) and a radar scanning frequency of 10Hz. The fingerprint creation process involves scanning for 10 seconds in a tagless environment, collecting 100 frames of data. The calculated values ​​are an average intensity of -15dBm, a phase change rate of 0.15rad / s, and an average Doppler shift of 0.1Hz (at rest). A unique fingerprint hash value, 0x5f2a..., is generated and stored in the database.

[0101] Interference suppression effect: When the shelf passes through the channel, the system identifies and matches the fingerprint, and MAM suppresses its reflected signal by 30dB, increasing the tag recognition rate from 72% to 98%.

[0102] To achieve the above objectives, this solution also proposes a method for identifying object passage, which is applied to the above-mentioned radio frequency channel door and includes the following steps:

[0103] S1. Initialize the system and load the pre-trained convolutional neural network and long short-term memory network models;

[0104] Convolutional Neural Network (CNN): Based on the ResNet-18 architecture, it is pre-trained on massive amounts of RFID scene data containing metal interference and can quickly extract the spatial features of tag signals;

[0105] Long Short-Term Memory (LSTM) network: This network uses a two-layer bidirectional structure to learn historical movement trajectories of labels, accurately predict future positions within a short period of time (e.g., within 0.5 seconds), and assist in subsequent interference suppression and decision-making.

[0106] The pre-trained model eliminates the need for the system to learn environmental features from scratch. Initialization can be completed within 50 milliseconds after startup, which is many times faster than traditional online learning solutions.

[0107] S2. Utilizing a frequency-modulated continuous wave (FMCW) radar, a range-dimensional FFT (Fast Fourier Transform), a velocity-dimensional FFT, and an angle-dimensional MUSIC algorithm are sequentially used to generate a three-dimensional point cloud (point cloud image) containing the location and reflection intensity of metal objects, clearly demonstrating the distribution of metal interference within the channel. The original point cloud is then matched against a pre-stored metal fingerprint library (covering intensity, phase, and motion characteristics) to generate a "metal heat map." This heat map uses numerical values ​​to intuitively demonstrate the strength of metal interference, providing a basis for subsequent beamforming to accurately locate interference.

[0108] S3. Execute a dynamic beamforming algorithm to suppress metal interference by optimizing the beamforming weight vector. The optimization goal is to maximize the power ratio of the tag signal to the interference signal. With the goal of maximizing the tag signal to interference signal power ratio (SIR), the beam direction and gain are dynamically adjusted. The main lobe of the beam is aligned with the tag, and the null (signal suppression area) is directed toward the metal interference, significantly reducing the impact of metal reflection on tag recognition.

[0109] S4. Dynamically switch between non-orthogonal multiple access (NOMA) mode and time division multiple access (TDMA) mode according to the collision probability distribution diagram; calculate the tag collision probability (the possibility of identification failure caused by overlapping of multiple tag signals) in real time:

[0110] When the collision probability is low (<30%), the NOMA mode (power domain multiplexing) is used to improve traffic efficiency.

[0111] When the collision probability is moderate (30%-70%), a hybrid NOMA-TDMA mode is used to prioritize independent time slots for high-collision areas.

[0112] When the collision probability is high (≥70%), switch to pure TDMA mode to avoid signal collision.

[0113] S5. Calculate the probability of item passage through a convolutional neural network enhanced by the metal attention module.

[0114] Fusion of CNN, MAM, and LSTM: CNN extracts tag signal features, MAM generates "attention weights" to shield the features of metal interference areas and highlight tag signals; LSTM combines the tag's historical trajectory to predict future locations and comprehensively calculate the item's "passing probability" (the probability that the tag is legal and free of interference and obstruction).

[0115] S6. When the probability of passing exceeds the preset threshold, passage is allowed; otherwise, an alarm of the corresponding level is triggered.

[0116] This method constructs an item access identification system with "precise perception, intelligent suppression, and reliable decision-making." It fundamentally solves the RFID identification problem in metal-intensive scenarios, significantly improves system performance and adaptability, and provides a highly reliable access gate solution for scenarios such as smart logistics and intelligent manufacturing. The convolutional neural network (CNN) and long short-term memory network (LSTM) loaded by S1 provide tag signal feature extraction capabilities for beamforming weight optimization in S3 and a spatiotemporal trajectory prediction model for the calculation of access probability in S5. The three-dimensional point cloud generated by S2, after being processed by a metal reflection fingerprint library (including intensity, phase, and Doppler characteristics), outputs a spatial distribution matrix of metal interference, which serves as input for beam nulling positioning in S3 and provides tag overlap information caused by metal occlusion for the calculation of collision probability in S4. The tag signal-to-noise ratio (SNR) after metal interference suppression in S3 also directly affects the calculation of collision probability in S4.

[0117] In some embodiments of the present invention, to address the issue of "beamforming weights not matching interference / signal changes in dynamic environments," this solution proposes an adaptive beamforming weight update scheme. Through real-time monitoring, minimum mean square error (LMS) iteration, and step size self-adjustment, this scheme achieves precise mainlobe tracking of tags and dynamic nulling to suppress interference. The beamforming weights in step S3 are updated using an adaptive algorithm:

[0118] 1. Real-time monitoring of environmental interference changes and tag signal strength. The monitoring data directly calls the millimeter-wave radar output and reader signal processing results of step S2 to provide real-time input for updating beamforming weights.

[0119] Among them, the monitoring objects and parameters include:

[0120] Environmental interference changes: The millimeter-wave radar module (step S2) collects the interference power (I(t), unit: dBm) and arrival angle (θI(t), unit: degrees) of the metal reflection signal in real time, reflecting the intensity and spatial position of the interference source;

[0121] Tag signal strength: The received power (S(t), unit: dBm) and signal-to-noise ratio (SNR) (SNR(t), unit: dB) of the tag signal are measured at the reader receiving end, reflecting the signal quality of the target tag.

[0122] 2. Use the minimum mean square error (LMS) criterion to iteratively update the beamforming weights so that the main lobe of the antenna array points to the target tag and the null points to the interference source;

[0123] LMS iteration formula: Assume that the weight vector of the antenna array is w(n) (n is the iteration step), the desired signal is d(n), and the received signal is x(n), then the weight update rule is:

[0124] w(n+1)=w(n)+μ(n)·e * (n)·x(n)

[0125] Where, e(n)=d(n)-w H (n) x(n): Error signal ((·) H is the conjugate transpose), reflecting the deviation between the current weight and the ideal weight;

[0126] μ(n): step size factor (dynamically adjusted, unit: dimensionless quantity), controls the update speed of weights;

[0127] d(n): expected signal (reference signal of the target tag, generated by tag ID precoding).

[0128] Iteration goal: Through LMS iteration, minimize the error signal e(n) so that the main lobe (maximum gain direction) of the antenna array points to the target tag (θ S (t) is the tag arrival angle), the null (minimum gain direction) points to the interference source (θ I (t) is the interference arrival angle). The arrival angle of the target tag θ S (t) Through the millimeter wave radar point cloud solution in step S2, the interference source arrival angle θ I (t) Obtained through metal thermal map and reflection signal analysis to ensure the accuracy of main lobe / null pointing.

[0129] 3. The step size factor during the update process is adaptively adjusted according to the signal change speed.

[0130] Step size adjustment formula:

[0131]

[0132] Where, μ0: basic step size (e.g. 0.01, the initial value of experimental optimization);

[0133] ΔSNR(n) = SNR(n) - SNR(n-1): Signal-to-noise ratio change rate, reflecting the speed of tag signal fluctuation;

[0134] ΔI(n)=I(n)-I(n-1): Interference power change rate, reflecting the fluctuation speed of environmental interference;

[0135] SNR ref , I ref : Reference value (such as SNR ref =10dB,I ref =10dBm), and normalize the denominator to avoid too large a step size.

[0136] The adaptive adjustment principle is as follows: When the signal / interference changes rapidly (e.g., a metal forklift passes quickly, |ΔI(n)| > 5dBm), the step size μ(n) is increased (e.g., μ(n) = 0.03) to accelerate weight updates. When the changes are slow (e.g., a static metal shelf, |ΔI(n)| < 1dBm), the step size is reduced (e.g., μ(n) = 0.005) to stabilize the weights and prevent oscillation. Both ΔSNR(n) and ΔI(n) are derived from real-time monitoring data in step S2, ensuring that step size adjustments are synchronized with environmental changes.

[0137] For the first time, the interference change rate + signal change rate are incorporated into the beamforming weight update, building a complete closed loop of "real-time monitoring-iterative optimization-step size self-adjustment", breaking through the limitations of the traditional LMS algorithm of "static step size and response lag".

[0138] In some embodiments of the present invention, the dynamic switching logic between the NOMA mode and the TDMA mode in step S4 is analyzed.

[0139] In RF channel gate applications, when objects (carrying tags) pass through, the signal environment becomes complex due to the dynamic change in the number of tags and metal interference. Traditional fixed multiple access modes (such as single NOMA or TDMA) have obvious drawbacks:

[0140] Fixed NOMA mode: This mode improves transmission efficiency in low-contention scenarios. However, in high-contention scenarios, signal collisions are severe, causing a sharp drop in tag recognition rates. It cannot cope with the increased interference caused by dense tag aggregation and metal obstructions.

[0141] Fixed TDMA mode: It can avoid collisions in high-conflict scenarios, but it causes redundant resource allocation in low-conflict scenarios, reducing channel efficiency and making it difficult to adapt to dynamic traffic changes such as logistics peaks and rapid personnel passage.

[0142] To address the issue of "the mismatch between the multiple access mode and the actual signal conflict status in dynamic scenarios, resulting in low recognition efficiency and poor reliability", this solution proposes dynamic switching logic between NOMA and TDMA to achieve intelligent mode adaptation according to the signal conflict status, balancing recognition efficiency and reliability.

[0143] In the dynamic switching logic, we first predict the tag trajectory and speed based on the long short-term memory network (LSTM). The following definitions are important:

[0144] The Long Short-Term Memory (LSTM) model is a variant of the Recurrent Neural Network (RNN) that has the ability to learn long-term dependencies in time series data and is used to process time series information of tag motion trajectories.

[0145] Tag future trajectory position: Based on historical trajectory data, predict the 2D / 3D coordinates (x t+Δt ,y t+Δt ,z t+Δt ), corresponding to the channel's horizontal, vertical, and height positions);

[0146] Moving speed: The movement rate of the tag during the prediction period, which is calculated by the rate of change of the trajectory coordinates, that is, Reflects the dynamic characteristics of tag movement.

[0147] The input of the LSTM model is a sequence of label history trajectories:

[0148] {(x1,y1,z1,t1),(x2,y2,z2,t2),…,(x t ,y t ,z t ,t t )} (including coordinates and timestamps). After cyclic calculations through the forget gate, input gate, and output gate, it learns the trajectory change pattern and outputs the predicted future trajectory position and speed. This prediction provides a basis for dynamic label distribution for subsequent collision probability analysis. It works in conjunction with the "millimeter-wave radar scanning to generate real-time 3D point cloud images" in this solution. The former provides trajectory trend prediction, while the latter provides real-time perception of the current environment, jointly supporting intelligent decision-making.

[0149] Then, a spatial conflict probability distribution map is generated based on the predicted trajectory, where:

[0150] Spatial conflict probability distribution map: The spatial range is the channel gate area (e.g., divided into M×N grids, where M and N are the number of horizontal and vertical grids). Each grid cell corresponds to a conflict probability value, reflecting the possibility of signal conflict at that location due to tag aggregation, signal superposition, etc. The probability value range is [0, 1], and the larger the value, the higher the conflict risk.

[0151] Signal conflict refers to the phenomenon in which multiple tag signals overlap in the same space-time region, making it difficult for the receiving end to accurately decode and identify them. It is affected by factors such as the number of tags, distribution density, and signal power.

[0152] During implementation, based on the future trajectory of the labels predicted by LSTM, the number of labels n in each grid cell at the future moment is counted. r,l(r, l are grid cell coordinates), combined with the tag signal coverage and power attenuation model (for example, signal strength decays exponentially with increasing distance), calculate the collision probability:

[0153] P r,l =f(n r,l ,d r,l ,P tag )

[0154] Among them, d r,l is the distance from the grid unit to the reader, P tag is the tag transmission power.

[0155] For example, the greater the number of tags and the closer they are to the reader, the higher the collision probability. This distribution map is linked to the metal interference heatmap generated by the Metal Attention Module (MAM) to jointly characterize the signal interference state within the channel. Metal interference causes signal distortion, and the collision probability analysis considers the superposition of signals between tags. The two complement each other to improve interference assessment.

[0156] It's important to note that non-orthogonal multiple access (NOMA) is a multiple access technology that allows multiple tags to transmit signals in the same time slot and frequency band. The receiver uses power and coding fields to distinguish between different tag signals, enabling signal multiplexing and improving transmission efficiency. However, this technology has a low tolerance for signal collisions. Time division multiple access (TDMA) divides time into several independent time slots, allocating a dedicated time slot for each tag to transmit. Different tag signals are isolated in the time domain, preventing signal collisions. However, resource allocation is relatively redundant, and efficiency is limited by the number of time slots.

[0157] Therefore, before switching, according to the channel door application scenario (such as logistics channel, access control channel), label density, recognition accuracy requirements, etc., the conflict probability critical value (such as low threshold P th1 =0.3, high threshold P th2 =0.7), used to determine the signal conflict status and trigger mode switching.

[0158] Switching principle: Real-time calculation of the overall conflict probability of the channel gate area (such as weighted average of the probability values ​​of all grid cells in the spatial conflict probability distribution map). <P th1 When the NOMA mode is activated, the power reuse feature is used to increase the number of tags recognized per unit time in scenarios where the tags are sparsely distributed and the conflict risk is low, so as to meet the efficient passage requirements such as logistics peak. When the conflict probability P>P th2When the time is up, it switches to TDMA mode, assigning each tag an independent time slot and forcing signal isolation. This ensures accurate tag recognition in high-contention scenarios and addresses recognition failures caused by metal obstruction and dense tag aggregation. This switching logic works in conjunction with the "dynamic beamforming algorithm to suppress metal interference." The former optimizes resource allocation at the multiple access level, while the latter enhances anti-interference capabilities at the signal transmission level, jointly ensuring channel gate recognition performance.

[0159] Under this switching logic, LSTM predicts tag motion trends, enabling early detection of signal conflict risks and intelligent switching between NOMA and TDMA modes. This ensures both traffic efficiency in low-conflict scenarios (NOMA improves throughput) and recognition reliability in high-conflict scenarios (TDMA avoids signal collisions), making it suitable for diverse dynamic scenarios such as logistics and access control. Furthermore, this solution integrates trajectory prediction and conflict probability analysis to compensate for the traditional fixed mode's limited adaptability to dynamic interference (such as metal occlusion and sudden changes in tag speed). Even in complex environments, it maintains high recognition rates and efficiency, addressing the shortcomings of existing technologies, such as single scenarios and poor interference resistance.

[0160] In some embodiments of the present invention, in order to solve the "adaptability problem of NOMA mode power allocation and interference elimination in dynamic scenarios", this solution proposes a NOMA optimization solution based on the collaboration of conflict probability and signal-to-noise ratio to achieve intelligent power allocation at the transmitting end and precise interference elimination at the receiving end, thereby improving the efficiency and reliability of multi-tag recognition.

[0161] The NOMA model is implemented as follows:

[0162] 1. At the transmitter, transmit power is dynamically allocated based on the signal-to-noise ratio of each tag. Tags with high signal-to-noise ratios are allocated lower power, while tags with low signal-to-noise ratios are allocated higher power.

[0163] The formula for transmitting power allocation at the transmitting end is expressed as:

[0164] β=δ×P coll (x i ,y i )

[0165] Among them, P i is the transmit power allocated to the i-th tag (unit: dBm), dynamic allocation result;

[0166] j represents the index of any tag among all currently activated tags;

[0167] P max The maximum allowable transmit power of the reader to ensure signal coverage and compliance;

[0168] SNR jis the signal-to-noise ratio of the jth tag, reflecting the tag signal quality;

[0169] β is the power allocation weight factor, which is a collaborative adjustment parameter integrating conflict probability and signal-to-noise ratio;

[0170] δ is a proportional constant, ranging from 1 to 10. The experimental optimization value is 5, which controls the impact of the conflict probability on β.

[0171] P coll (x i ,y i ) is (x i ,y i ), which ranges from 0 to 1 and is output by the conflict probability distribution graph, reflecting the signal collision risk at that location; M is the total number of currently activated tags.

[0172] Traditional NOMA power allocation only depends on SNR. This scheme uses β to allocate P coll (x i ,y i ) into decision-making - high conflict areas (P coll >0.5), β increases, and higher power is allocated to low SNR tags first, forcing them to increase their signal strength; in low conflict areas (P coll <0.3), β is reduced to allow high SNR tags to moderately "give up" power to avoid resource waste. coll (x i ,y i ) The "spatial conflict probability distribution map" from step S4 is coordinated with LSTM trajectory prediction and millimeter-wave radar point cloud to achieve closed-loop control of "environmental perception → conflict quantification → power adaptation".

[0173] like Figure 4 The three-dimensional spatial distribution diagram of the power allocation of 50 tags in NOMA mode is shown. The spatial position of each tag in the figure is determined by its horizontal coordinate (x-axis) and vertical coordinate (y-axis), and the allocated power (z-axis) is doubly characterized by the color gradient (blue low power to red high power) and the height of the column. Although the tag in the metal interference area marked by the black dotted box (x∈[-1,1]m,y∈[2,4]m) has a low signal-to-noise ratio (5-8dB), it obtains a high power compensation of 25-28dBm due to the high collision probability (0.75-0.85); the high collision area marked by the red circle (around the center of the channel y=3m) has a significantly higher power than the edge area (>20dBmvs<3dBm), and the edge of the channel is only allocated 0.5-3dBm of power due to the low collision probability (0.1-0.3) and high signal-to-noise ratio (18-22dB). Figure 4 The technical effect of increasing the power in the metal interference area achieved by the above-mentioned transmission power allocation formula is intuitively verified.

[0174] 2. Execute the continuous interference cancellation algorithm at the receiving end, decoding each tag signal in order from high to low power. After decoding each tag, its interference is eliminated from the total received signal.

[0175] The continuous interference cancellation algorithm refers to a technology that sorts the tag signals from high to low according to their power at the receiving end, decodes the tag signals in sequence, and subtracts the interference components of the decoded tags from the total received signal, gradually eliminating the interference of multiple tags.

[0176] Power order from large to small: Based on the power allocation result of the transmitter (P i The larger the value, the stronger the signal power), determine the decoding priority (such as P1>P2>…>P M , then decode the tag 1 first).

[0177] As an example, in the scenario of smart warehouse access door, assume the following conditions:

[0178] Reader maximum transmission power: P max =30dBm; Current number of active tags: M = 3 (tags A, B, C); Proportional constant: δ = 5 (take the middle value);

[0179] Label position conflict probability (generated by LSTM prediction):

[0180] Tag A:P coll (x A ,y A ) = 0.2 (edge ​​area, low conflict);

[0181] Tag B:P coll (x B ,y B ) = 0.5 (transition region, medium conflict);

[0182] Tag C:P coll (x C ,y C ) = 0.8 (central area, high conflict);

[0183] Tag signal-to-noise ratio (after beamforming optimization):

[0184] Label A: SNR A =20dB (high quality signal);

[0185] Label B: SNR B =15dB (medium quality signal);

[0186] Label C: SNR C =10dB (low quality signal, interfered by metal).

[0187] Then, record the power allocation calculation process:

[0188] Step 1: Calculate the power allocation weight factor β i :

[0189] β i =δ×P coll (x i ,y i )

[0190] Tag A: β A =5×0.2=1.0;

[0191] Label B: Beta B =5×0.5=2.5;

[0192] Tag C: Beta C =5×0.8=4.0;

[0193] Step 2: Calculate the exponential term numerator exp(β·SNR i ):

[0194] Label A: exp(1.0×20)=exp(20)≈4.8517×10 8 ;

[0195] Label B: exp(2.5×15)=exp(37.5)≈1.6218×1016;

[0196] Label C: exp(4.0×10)=exp(40)≈2.3530×1017;

[0197] Step 3: Calculate the denominator

[0198] ∑=exp(20)+exp(37.5)+exp(40)

[0199] =4.8517×10 8 +1.6218×10 16 +2.3530×10 17

[0200] =2.5152×10 17

[0201] Step 4: Calculate the power P allocated to each tag i :

[0202]

[0203] Tag A (high SNR + low collision):

[0204]

[0205] Label B (medium SNR + medium collision):

[0206]

[0207] Label C (low SNR + high collision):

[0208]

[0209] Based on the above data, the following table can be obtained:

[0210]

[0211] As can be seen from the table, tag C is located near the metal shelf (collision probability 0.8), and is allocated nearly 94% of the system power (28.06 / 30), significantly improving the recognition reliability of weak signal tags; tag A is in a low-interference area and is only allocated 0.02% of the power, avoiding resource waste; and the weight factor β reduces the collision probability P coll Multiplied by SNR, dual regulation is achieved: high-conflict areas (such as metal-dense areas) automatically trigger power compensation; low-conflict areas (such as channel edges) suppress power overflow.

[0212] The SIC decoding order at the receiving end is: decoding from large to small according to the allocated power: Label C → Label B → Label A.

[0213] After decoding tag C, its interference is eliminated, and then tag B is decoded (power difference 26.13dB), and finally tag A is decoded (power difference 28.05dB), ensuring that the high-power tag does not mask the low-power tag.

[0214] Compared with the traditional TDMA mode, if static TDMA is used to allocate equal power (10dBm / tag): tag C cannot be identified due to metal interference (signal strength < detection threshold), and the recognition success rate is only 66.7% (tags A and B are successful).

[0215] As can be seen, under this solution, tag C's power increased by 18.06dB, the signal-to-noise ratio increased from 10dB to 28.06dB, the recognition success rate increased to 100% (all three tags were successfully decoded), and power utilization increased by 41% (total power 30dBm vs. TDMA requires 30dBm). By coupling spatial collision probability (LSTM prediction) with channel quality (SNR), intelligent dynamic power allocation is achieved, effectively solving the problem of missed tag detection in dense metal scenes.

[0216] Traditional SIC decodes tags based on fixed rules (e.g., SNR from high to low). This solution prioritizes decoding by transmit power and works in conjunction with the transmitter's power allocation strategy. High-power tags (mostly those with low SNR and high-collision areas) are decoded first, ensuring that weak-signal tags are "rescued" first and avoiding decoding failures caused by strong residual interference. Combining the "dynamic beamforming" in step S3 with the "metal attention module (MAM)" in step S5, beamforming suppresses metal reflection interference, the MAM shields metal area features, and SIC focuses on inter-tag interference elimination, forming a dual-dimensional anti-interference system in the "spatial domain + power domain."

[0217] In some embodiments of the present invention, to address the issue of "TDMA mode time slot allocation not matching the dynamic environment in high-conflict scenarios," this solution proposes a TDMA optimization solution based on the coordination of signal-to-noise ratio and spatial interference, enabling dynamic adjustment of time slot lengths and intelligent avoidance of interfering time slots, thereby improving recognition reliability and resource utilization in high-conflict scenarios. The TDMA mode is implemented as follows:

[0218] 1. Dynamically adjust the time slot length based on the tag signal-to-noise ratio. Tags with low signal-to-noise ratio are assigned longer time slots to improve transmission reliability.

[0219] Slot length adjustment rule: Assume that the basic slot length is T0 (e.g. 2ms, the minimum decodable slot in the system), and define the slot adjustment coefficient α:

[0220] α=1+k·max(0,SNR th -SNR i )

[0221] Then the time slot length of the i-th tag is: T i =α·T0

[0222] Among them, SNR th : Signal-to-noise ratio threshold (e.g. 5dB), when the signal-to-noise ratio is lower than this value, the time slot extension is started;

[0223] k: adjustment coefficient (value ranges from 1 to 3, such as 2), which controls the time slot extension amplitude (SNR i Each below SNR th 1dB, time slot extension k·T0);

[0224] T i : The final time slot length of the i-th tag (unit: ms), dynamically adapting to the signal quality.

[0225] Low SNR tags (such as tags blocked by metal, SNR i <3dB) allocate longer time slots (such as T i =6ms), improve decoding reliability by increasing signal transmission time; high SNR tag (such as no interference tag in the center of the channel, SNRi >10dB) to allocate short time slots (such as T i =2ms), releasing resources for other tags.

[0226] 2. Establish a time slot allocation priority strategy to prioritize low-interference time slots for high-priority tags. The priority is determined based on the distance between the tag and the geometric center of the channel gate.

[0227] First, the priority calculation rules:

[0228]

[0229] Among them, d i : The distance from tag i to the geometric center of the channel gate (unit: m). The closer the distance, the higher the priority (the channel center is the key area for identification);

[0230] ∈: minimum value (such as 0.1m), avoid d i =0, the priority is infinite;

[0231] Definition of low-interference time slots: Combine the "millimeter-wave radar 3D point cloud" of step S2 and the "metal heat map" of step S3, and mark the time slots where the interference intensity in the channel is lower than the threshold (such as interference power <-20dBm) as "low-interference time slots" and give priority to high-priority tags.

[0232] Allocation principle: high priority label (channel center label, d i <1m) prioritize low-interference time slots to avoid metal reflection interference (such as strong interference time slots near shelves) and ensure its recognition reliability; low-priority tags (channel edge tags, d i >2m) allocate the remaining time slots, even if affected by interference, the impact on the key area (center) of the overall recognition is small.

[0233] In this scheme, d i Through real-time calculation of millimeter-wave radar point clouds, metal heat maps mark interference areas and, in conjunction with the "conflict probability distribution map" in step S4, implement dual-dimensional time slot scheduling based on "spatial position + interference intensity." Furthermore, for the first time, the tag's spatial position (channel center distance) is combined with dynamic signal-to-noise ratio (SNR) to construct a dual-dimensional "quality-spatial" time slot scheduling strategy. This breaks through the traditional TDMA model of "allocating only by time or SNR" and provides an innovative solution for high-conflict scenarios.

[0234] like Figure 5 The collision rate comparison chart between static TDMA, pure NOMA and this scheme is shown. Figure 6 A comparison chart of identification delay between static TDMA and this scheme is shown.

[0235] Figure 5In the figure, the red curve (static TDMA) maintains a low collision rate (<5%) when the tag flow rate is ≤60 tags / sec. However, when the flow rate is >80 tags / sec, the collision rate rises sharply to 25% (80 tags / sec) and even 32% (150 tags / sec). This is consistent with the defect of the existing TDMA protocol pointed out in the background technology of this application that the time slot collision rate is as high as 25% in high traffic scenarios (>80 tags / sec).

[0236] The blue curve (pure NOMA) performs best at low traffic (<50 tags / sec) (collision rate ≈ 10%). However, as traffic increases, metal interference and multipath effects cause the collision rate to continue to rise, reaching 28% at 120 tags / sec. This causes tag signals to be drowned out in dense metal environments.

[0237] Green curve (this application's solution): The collision rate is only 4.2% in the high-traffic scenario of 80 tags / sec, which is significantly lower than the 25% of the static TDMA solution and the 22% of the pure NOMA solution. As the traffic increases to 150 tags / sec, it remains stable below 7.5%, directly verifying the effectiveness of the dynamic switching mechanism. Figure 5 The specially marked 80 tags / sec benchmark point (green solution 4.2% vs red TDMA 25%) also proves that metal interference suppression and intelligent time slot allocation work together to reduce the probability of signal collision.

[0238] Figure 6 Middle, red dashed line (static TDMA): In a high-traffic scenario (100 tags / sec), the recognition delay reaches 800ms. It can be seen that static resource allocation causes a surge in recognition delay.

[0239] The green curve (the solution of this application): the delay is only 120ms at a flow rate of 100 tags / sec, which is 85% lower than the 800ms of static TDMA. The green curve's delay performance of less than 200ms throughout the entire process confirms the efficiency improvement effect of the adaptive beamforming algorithm and dynamic mode switching. The delay mark (120ms vs 800ms) at 100 tags / sec in the figure quantifies the technical effect of this application in solving the problem of surge in recognition delays, and the delay improvement trend is mutually confirmed by the system-level optimization of the 41% increase in power utilization in the above-mentioned NOMA implementation.

[0240] In some embodiments of the present invention, in order to solve the problem of "the alarm mechanism in dynamic scenarios does not match the actual security status, resulting in false alarms and missed alarms, and insufficient security risk prevention and control", the present invention proposes a multi-level alarm mechanism, which achieves accurate alarms and active prevention and control through layered responses to instantaneous interference, continuous anomalies, and comprehensive security risks.

[0241] The multi-level alarm mechanism includes:

[0242] Level 1 alarm: When the confidence level of a single detection is lower than 0.7, a yellow indicator light will be triggered to warn;

[0243] Clarify the concept: Detection confidence refers to the model's confidence probability for the tag recognition result, denoted as C(t) (range [0,1]), which is output by the convolutional neural network (pre-trained model in step S1) and reflects the reliability of the current recognition result. The triggering condition for the first-level alarm is that when C(t) < 0.7, the yellow indicator light will be triggered to warn that "there is a certain degree of uncertainty in the current recognition result, and attention is required." Low single-time confidence may be caused by instantaneous metal interference (such as a metal object passing quickly) or temporary attenuation of the tag signal. Therefore, the yellow warning serves as a "lightweight prompt" that responds to anomalies while avoiding misjudgments due to instantaneous interference (such as the traditional solution that directly locks the channel, which may affect normal traffic).

[0244] Level 2 Alarm: When the confidence level of three consecutive tests is less than 0.5, a red indicator light and an audible and visual alarm are triggered. To reach the level 2 alarm, three consecutive tests are required: using the time series {t1, t2, t3} as a window (e.g., 100ms interval, 300ms total), the confidence sequence {C(t1), C(t2), C(t3)} is obtained. The triggering conditions for the level 2 alarm are: when C(t1) < 0.5, C(t2) < 0.5, and C(t3) < 0.5, a red indicator light and an audible and visual alarm are triggered, indicating "Continued recognition anomaly, manual intervention required." Due to three consecutive low confidence levels, transient interference is ruled out, indicating that the channel door has a persistent recognition failure (such as a damaged tag or an unremoved metal interference source). The red alarm forces manual intervention to avoid the risk of missed detection.

[0245] Level 3 alarm: When the comprehensive safety score S security When the value falls below the critical value, the passage door will be automatically locked. The safety score takes into account the probability of passage, detection confidence and historical passage records.

[0246] Comprehensive safety score S security , fusion pass probability (P pass , output of step S5), detection confidence (C(t)), historical pass records (H, the number of historical anomalies stored in the blockchain), the formula is:

[0247] S security =ω1·P pass +ω2·mean(C(t))-ω3·H

[0248] Where: ω1, ω2, ω3 represent weights (e.g., ω1 = 0.5, ω2 = 0.3, ω3 = 0.2), which are determined based on experimental optimization and reflect the degree of impact of each factor on safety;

[0249] mean(C(t)): The mean of the recent detection confidence (such as the average confidence of the past five recognitions).

[0250] The triggering condition of the third level alarm is: security th (critical value, such as 0.6), the passage door is automatically locked to prevent abnormal passage.

[0251] Traditional solutions only focus on the current recognition results. This solution incorporates historical anomalies (such as repeated missed detections and false alarms) to more comprehensively reflect the security status of the channel door. For example, for a channel with many historical anomalies (large H), even if the current probability of passing is high, S security Lockout may also be triggered.

[0252] In summary, this solution proposes for the first time a multi-level mechanism of "layered alarm + historical data fusion". The first-level warning (yellow) filters instantaneous interference and reduces false alarms; the second-level alarm (red) identifies persistent anomalies and avoids missed detections; the third-level lockdown assesses security in multiple dimensions and proactively prevents and controls risks.

[0253] In some embodiments of the present invention, in order to solve the "deficiencies in security, traceability, and mutual trust of pass data storage", this solution proposes a pass data storage system based on blockchain technology. By leveraging the "decentralized, tamper-proof, and traceable" characteristics of blockchain, data security is guaranteed, providing reliable data support for multi-level alarms, security scoring and other functions.

[0254] Establishing a data storage system based on blockchain technology includes:

[0255] (1) Design of traffic record structure

[0256] Each historical transaction record H, as a "transaction" in the blockchain, contains the following key information:

[0257] Timestamp: Accurately record the time when the traffic data is generated (such as "2025-06-3014:23:56"), marking the moment when the data is generated for easy subsequent tracing;

[0258] Tag identity hash value: Hash the tag's unique code (such as the electronic tag ID) (for example, using the SHA-256 algorithm) to hide the tag's true ID, protect privacy, and retain identity recognition;

[0259] Pass status code: marks the pass result, which is divided into three categories: normal pass (corresponding to step S6 green light release); level one alarm (corresponding to step S6 yellow warning); level two and above alarm (corresponding to step S6 red alarm, channel locked);

[0260] ​Previous record hash value: Associated with the hash value of the previous pass record, stringing the records together like a chain. Once a record is changed, subsequent records will fail hash verification and the tampering can be detected.

[0261] This structure associates each record with historical records to form a "block chain", ensuring that data is traceable and cannot be tampered with from generation to storage. It is deeply bound to the multi-level alarm results of step S6 and the identification process of step S5 to ensure consistency between stored data and identification logic.

[0262] (2) Data immutability protection

[0263] Hash operation: The complete contents of each record (timestamp, tag hash value, status code, and previous record hash value) are concatenated and processed using a hash algorithm such as SHA-256 to generate the current record hash value. The hash value changes if any content in the record (for example, the status code changes from "Level 1 Alarm" to "Normal Passage").

[0264] Tampering detection: Because records are linked by hashes, if one record is changed, the verification of the "previous record hash value" of subsequent records will fail. Nodes in the blockchain (local channel gates, cloud servers, regulatory nodes, etc.) will reject such tampering to ensure data integrity.

[0265] Furthermore, historical access records H are decentralized and distributed across multiple nodes, without a single control center, eliminating the risk of single-point tampering. This unalterable data provides a true and reliable historical record of the number of historical anomalies in step S6, the "Comprehensive Safety Score," ensuring the accuracy of the safety score.

[0266] (3) Distributed Node Verification and Traceability Query

[0267] Node Verification: Using consensus algorithms like Practical Byzantine Fault Tolerance (PBFT), gatekeepers, servers, and regulatory agencies vote together to verify newly generated access records. Only when more than two-thirds of the nodes approve (with consistent hashes and valid status codes) will the record be officially stored on the blockchain, ensuring the legitimacy and authenticity of the data on the chain.

[0268] Traceability query: Provides an API interface. By entering the tag hash value or timestamp, you can query all historical travel records of the tag, including the time, status, and associated hash of the previous record of each travel.

[0269] This traceability capability provides historical data (counting historical anomalies) for the "Comprehensive Security Score" in step S6, enabling a multi-dimensional security assessment combining "current identification results + historical anomalies." Furthermore, regulatory authorities can verify compliance with door identification through traceability queries, resolving data trust issues in multi-system collaboration. This eliminates the need for third-party guarantees when sharing data between logistics, security, and other systems, reducing trust costs.

[0270] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0271] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0272] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A radio frequency channel gate based on multi-band coordinated control, characterized in that: include: Multi-band reader: The operating frequency band is 860-960MHz, divided into 16 sub-bands; Multi-band antenna array: placed on both sides of the channel door frame, including UHF flat-panel directional antennas for transmitting and receiving RF signals; Millimeter-wave radar module: operating frequency 77-81 GHz, outputs a 3D point cloud image of metal objects; Metal shielding frame: embedded with flexible electromagnetic shielding layer, the spacing between adjacent antennas is λ / 2±10%, where λ is the center wavelength of the operating frequency band; Distributed signal processing module: Each antenna node is equipped with a digital signal processor (DSP) and a built-in convolutional neural network (CNN) algorithm. The CNN includes a metal attention module (MAM), which optimizes the beamforming weight vector using a dynamic beamforming algorithm to maximize the power ratio between the tag signal and the interference signal. The MAM handles metal interference in the following ways: Generate spatial attention weight matrix A based on the 3D point cloud attn : From attn =σ(Conv 3×3 (Q)) Among them, σ(·) is the Sigmoid activation function, Conv 3×3 is a 3×3 convolutional layer, Q is the metal heat map matrix, which is converted from the 3D point cloud image; Output weighted signal: and out =A attn ⊙and clean +(1-A attn )⊙and null Among them, ⊙ is the Hadamard product, y null is a shielding signal matrix composed of all zero elements, y clean is the clean signal matrix, i.e. the RF signal after interference suppression; Intelligent control module: uses reinforcement learning algorithm to dynamically optimize frequency band selection strategy; Communication interface module: supports TCP / IP, RS485 and CAN bus protocols; Auxiliary indicator device: includes LED status light and LCD display.

2. The radio frequency channel door according to claim 1, characterized in that: The millimeter wave radar module establishes a metal reflection fingerprint library in the following way: In a tag-free environment, a detection signal is emitted to collect the intensity, delay, and distance characteristics of metal reflectors; Extract the phase change rate of the reflected signal as the metal fingerprint feature parameter; The statistical distribution characteristics of Doppler frequency shift are calculated to establish a unique identification fingerprint library for metal objects.

3. A method for identifying passage of an object, applied to the radio frequency channel door according to any one of claims 1-2, characterized in that: The following steps are involved: S1. Initialize the system and load the pre-trained convolutional neural network and long short-term memory network models; S2, millimeter wave radar scanning generates real-time 3D point cloud map; S3. Execute a dynamic beamforming algorithm to suppress metal interference by optimizing the beamforming weight vector. The optimization goal of this algorithm is to maximize the power ratio of the tag signal to the interference signal. The updating process of the beamforming weight includes adaptively adjusting the step size factor. S4, dynamically switching between a non-orthogonal multiple access (NOMA) mode and a time division multiple access (TDMA) mode according to a collision probability distribution graph; S5, calculating the probability of item passage through a convolutional neural network enhanced by the metal attention module; S6. When the probability of passing exceeds the preset threshold, passage is allowed; otherwise, an alarm of the corresponding level is triggered.

4. The method according to claim 3, characterized in that The dynamic switching logic between NOMA mode and TDMA mode in step S4 is: Predict the tag's future trajectory position and movement speed through the long short-term memory network model; generating a spatial conflict probability distribution map based on the predicted trajectory, wherein the distribution map reflects the probability of signal conflicts occurring at different locations within the channel gate area; When the collision probability is lower than the set threshold, the non-orthogonal multiple access mode is activated, allowing multiple tags to transmit simultaneously; When the collision probability is higher than the set threshold, it switches to time division multiple access mode and allocates independent time slots to different tags.

5. The method according to claim 4, characterized in that The NOMA mode is implemented as follows: At the transmitter, transmit power is dynamically allocated based on the signal-to-noise ratio of each tag. Tags with high signal-to-noise ratios are allocated lower power, while tags with low signal-to-noise ratios are allocated higher power. The formula for transmitting power allocation at the transmitting end is expressed as: Among them, P i is the transmission power allocated to the i-th tag, j represents the index of any tag among all currently activated tags, P max is the maximum allowed transmission power of the reader, SNR j is the signal-to-noise ratio of the jth tag, β is the power allocation weight factor, δ is the proportional constant, ranging from 1 to 10, P coll (x i ,y i ) is (x i ,y i ), the collision probability value at the location ranges from 0 to 1, and M is the total number of currently activated tags; At the receiving end, a continuous interference cancellation algorithm is executed to decode the tag signals in order from high to low power. After each tag is decoded, its interference is eliminated from the total received signal.

6. The method according to claim 4, characterized in that The TDMA mode is implemented as follows: Dynamically adjust the time slot length based on the tag signal-to-noise ratio. Tags with low signal-to-noise ratio are assigned longer time slots to improve transmission reliability. A time slot allocation priority strategy is established to preferentially allocate low-interference time slots to high-priority tags. The priority is determined based on the distance between the tag and the geometric center of the channel gate.

7. The method according to claim 3, characterized in that The updating of the beamforming weights in step S3 adopts an adaptive algorithm: Real-time monitoring of environmental interference changes and tag signal strength; The minimum mean square error criterion is used to iteratively update the beamforming weights so that the main lobe of the antenna array points to the target tag and the null points to the interference source; The specific method of adaptively adjusting the step factor is: dynamically adjusting the size of the step factor according to the signal change speed.

8. The method according to claim 3, characterized in that Establish a multi-level alarm mechanism: Level 1 alarm: When the confidence level of a single detection is lower than 0.7, a yellow indicator light will be triggered to warn; Second level alarm: When the confidence level of three consecutive tests is lower than 0.5, a red indicator light and an audible and visual alarm will be triggered; Level 3 alarm: When the comprehensive safety score S security When the threshold is lower than the critical value, the passage door is automatically locked. The safety score comprehensively considers the probability of passage, detection confidence and historical passage records.

9. The method according to claim 3, characterized in that Establish a data storage system based on blockchain technology: Each pass record contains a timestamp, a hash value of the tag identifier, a pass status code, and a hash value of the previous record; The hash chain structure is used to ensure that data cannot be tampered with. The status codes include normal passage, level one alarm, and level two and above alarms. Supports distributed node verification and data traceability query functions.

Images