A deep learning-based RFID tag signal optimization method and system

Abstract

The application relates to the technical field of RFID tag signal optimization, and discloses an RFID tag signal optimization method and system based on deep learning.The application quantifies the core features of a multipath channel from the frequency domain and the time domain respectively by performing feature extraction on backscattering signals, and further obtains the numerical representation of the channel state, then takes the obtained data information as the input of a deep learning network, performs training and forward calculation, adjusts the original transmission parameters, and thus realizes the transformation of the transmission end from blind and extensive to precise regulation and control, so as to solve the problem that the parameter regulation and the channel state mismatch caused by the information split between the transmission end and the receiving end in the prior art, and then based on a discrete impulse response sequence and a noise interference feature set, a signal optimization method of joint adjustment of transmission and reception parameters is realized through an end-to-end deep learning architecture, and the optimization effect of the RFID tag signal in a complex environment is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of RFID tag signal optimization technology, and in particular to an RFID tag signal optimization method and system based on deep learning. Background Technology

[0002] RFID tags, short for Radio Frequency Identification tags, are electronic tags with built-in miniature antennas and chips. They can store the identity information and attribute data of target objects and communicate with readers via wireless radio frequency signals to achieve contactless data identification and transmission. They are widely used in scenarios such as logistics tracking, asset inventory, and access control. The signal transmission principle between RFID tags and readers is based on electromagnetic induction or electromagnetic backscatter coupling technology. The core is to complete data transmission through the electromagnetic wave interaction between the reader and the RFID tag. Specifically, the reader first actively emits electromagnetic waves of a specific frequency, forming a radio frequency energy field within a specific range. When the RFID tag... When a tag enters the energy field, its built-in antenna receives electromagnetic energy. If it is a passive tag, the received electromagnetic energy will be converted into electrical energy to power the tag chip. If it is an active tag, it will be activated directly. After the RFID tag chip is powered on, it will load the stored data onto its own antenna. By adjusting the antenna's load state, it will transmit the electromagnetic waves carrying the data back to the reader through reflection or backscattering. Finally, the reader receives the electromagnetic wave signal reflected by the RFID tag, and after demodulation and decoding, it extracts the data stored in the RFID tag and transmits it to the backend system for analysis and application, thus completing a complete identification process.

[0003] Existing RFID tag signal optimization mainly adopts two separate approaches: one is transmitter-side optimization, where the reader adjusts parameters such as transmission power and frequency based on simplified performance indicators such as read rate; the other is receiver-side optimization, where the reader uses independent signal processing technology to perform noise reduction and error correction on the backscattered signal. However, this separate signal optimization method has significant problems: the transmitter and receiver form an information barrier, and the transmission control module cannot obtain deep channel state information such as multipath interference and noise type contained in the received signal, and can only implement coarse control. Therefore, it cannot achieve signal optimization through transmission and reception coordination, resulting in poor signal optimization effect. Summary of the Invention

[0004] The main objective of this invention is to provide a method and system for optimizing RFID tag signals based on deep learning, aiming to solve the technical problems in the prior art.

[0005] This invention proposes a deep learning-based method for optimizing RFID tag signals, comprising:

[0006] Acquire the historical transmit and receive signals and original transmission parameters of the RFID reader, wherein the historical transmit and receive signals include backscatter signals and forward link signals;

[0007] The backscattered signal is analyzed to obtain the channel coherence bandwidth and discrete impulse response sequence;

[0008] The normalized path loss and noise interference feature set is obtained based on the historical transmitted and received signals, the original transmission parameters, and the discrete impulse response sequence.

[0009] The original transmission parameters are adjusted based on the normalized path loss and the channel coherence bandwidth to obtain optimized transmission parameters;

[0010] Based on the optimized transmission parameters, the RFID reader is driven to transmit electromagnetic waves to the RFID tag and receive the initial tag signal reflected by the RFID tag.

[0011] Based on the discrete impulse response sequence and the noise interference feature set, the initial tag signal is optimized using a signal optimization model to obtain an optimized tag signal;

[0012] The initial label signal is input into the signal optimization model for model evaluation to obtain the uncertainty index;

[0013] Determine whether the uncertainty index is greater than a preset threshold:

[0014] If the uncertainty index is greater than a preset threshold, the initial tag signal is included in the historical transmit and receive signals, and the process returns to the initial step to perform signal optimization.

[0015] Preferably, the step of analyzing the backscattered signal to obtain the channel coherence bandwidth and discrete impulse response sequence includes:

[0016] Acquire the guiding signal, and perform sliding cross-correlation calculation between the guiding signal and the backscattered signal to obtain a cross-correlation function sequence;

[0017] The preamble signal is extracted from the backscattered signal based on the cross-correlation function sequence, and the preamble signal is subjected to a fast Fourier transform to obtain the preamble frequency response.

[0018] Obtain the pilot signal frequency response, and obtain the channel frequency response sequence based on the pilot signal frequency response and the preamble frequency response;

[0019] The autocorrelation function output sequence is obtained based on the channel frequency response sequence, and the channel coherence bandwidth is obtained based on the autocorrelation function output sequence.

[0020] The cross-correlation function sequence is transformed using a channel impulse response estimation algorithm to obtain a discrete impulse response sequence.

[0021] Preferably, the step of obtaining the normalized path loss and noise interference feature set based on the historical transmitted and received signals, the original transmission parameters, and the discrete impulse response sequence includes:

[0022] The average received power value is obtained based on the backscattered signal;

[0023] The initial transmit power and initial antenna gain are obtained based on the original transmit parameters, and the normalized path loss is obtained based on the initial transmit power, initial antenna gain, and the average received power value.

[0024] The forward link signal is convolved based on the discrete impulse response sequence to generate a simulated reference signal sequence, and a residual signal sequence is obtained based on the simulated reference signal sequence and the backscattered signal.

[0025] A power spectral density curve is constructed based on the residual signal sequence, and the equivalent noise spectral density and narrowband interference spectrum set are obtained based on the power spectral density curve.

[0026] The residual signal sequence is analyzed to obtain a dynamic detection threshold, and the equivalent noise spectral density, narrowband interference spectrum set, and dynamic detection threshold are used as a noise interference feature set.

[0027] Preferably, the step of adjusting the original transmission parameters based on the normalized path loss and the channel coherence bandwidth to obtain optimized transmission parameters includes:

[0028] The normalized path loss and the channel coherence bandwidth are calculated forward using a deep learning network to obtain the real-time adaptive threshold and the real-time gradient shape factor.

[0029] Obtain the standard sub-channel bandwidth, and obtain the frequency aggregation determination threshold based on the standard sub-channel bandwidth and the real-time adaptive threshold;

[0030] The channel coherence bandwidth is compared with the frequency aggregation determination threshold to obtain frequency domain resource adjustment information;

[0031] The initial channel symbol rate and initial transmit power are obtained based on the original transmission parameters, and the initial channel symbol rate is adjusted based on the frequency domain resource adjustment information to obtain the optimized channel symbol rate.

[0032] Obtain the weight allocation strategy, and obtain the power weight vector based on the weight allocation strategy and the gradient shape factor;

[0033] The initial transmit power is adjusted based on the power weight vector and the frequency domain resource adjustment information to obtain the optimized transmit power, and the optimized channel symbol rate and the optimized transmit power are used as the optimized transmit parameters.

[0034] Preferably, the step of optimizing the initial tag signal based on the discrete impulse response sequence and the noise interference feature set using a signal optimization model to obtain the optimized tag signal includes:

[0035] The initial label signal and the discrete impulse response sequence are input into the encoder of the signal optimization model for encoding and feature extraction to obtain the intermediate layer feature tensor and conditional coding vector;

[0036] The conditional encoding vector and the intermediate layer feature tensor are fused by the conditional layer normalization operation to obtain the fused feature tensor. The fused feature tensor is then input into the signal optimization model for analysis to generate the correction feature tensor.

[0037] The equivalent noise spectral density, narrowband interference spectrum set, and dynamic detection threshold are extracted from the noise interference feature set, and the correction feature tensor is filtered according to the narrowband interference spectrum set to obtain the interference suppression feature tensor.

[0038] Based on the dynamic detection threshold, the interference suppression feature tensor is smoothed and limited by a linear suppression function to obtain a cleaned feature tensor. The cleaned feature tensor and the equivalent noise spectral density are then input into the decoder of the signal optimization model for signal reconstruction to obtain the optimized label signal.

[0039] Preferably, the step of inputting the initial label signal into the signal optimization model for model evaluation to obtain the uncertainty index includes:

[0040] Obtain a preset neuron masking probability, and obtain the masking decision random number corresponding to each neuron in a specific hidden layer in the signal optimization model based on the neuron masking probability;

[0041] A binary random mask is obtained based on the neuron masking probability and the masking decision random number, and the initial label signal is repeatedly input into the encoder loaded with the binary random mask to obtain multiple deep feature tensors;

[0042] Calculate the cognitive consistency metric between any two deep feature tensors, and obtain an uncertainty index based on the multiple cognitive consistency metrics.

[0043] This application also provides a deep learning-based RFID tag signal optimization system, which includes multiple modules for implementing the steps of the deep learning-based RFID tag signal optimization method described above.

[0044] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described deep learning-based RFID tag signal optimization method.

[0045] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described deep learning-based RFID tag signal optimization method.

[0046] The beneficial effects of this invention are as follows: By extracting features from the backscattered signal, this invention outputs the channel coherence bandwidth, which characterizes the channel frequency selectivity, and the discrete impulse response sequence, which describes the multipath propagation characteristics. These two parameters quantify the core features of the multipath channel in the frequency and time domains, respectively, providing a key basis for understanding the distortion mechanism of signals in spatial propagation. Subsequently, by integrating historical transmitted and received signals, original transmission parameters, and the discrete impulse response sequence, a numerical representation of the channel state is further obtained. Then, the normalized path loss and channel coherence bandwidth are used as inputs to a fully connected neural network. By training the deep learning network and performing forward computation, the original transmission parameters are adjusted to obtain optimized transmission parameters. This achieves a transformation from blind and coarse control at the transmitter to precise and intelligent control, solving the problem of parameter control and channel state mismatch caused by the information separation between the transmitter and receiver in existing technologies. Next, the initial tag signal is input into the signal optimization model, which uses discrete impulse response sequences and noise interference feature sets as the basis for adjusting network weights. The model optimizes the signal through convolution operations and attention mechanisms. This signal optimization method, which establishes a direct correlation between the receiver signal optimization processing and the channel state information acquired by the transmitter, solves the problem in existing technologies where the lack of prior channel information in receiver signal processing leads to limited signal optimization effects. Finally, the signal optimization model is evaluated using the acquired uncertainty index, forming an adaptive optimization loop based on continuous learning. This closed-loop signal optimization method enables the signal optimization system to perform self-evaluation and continuous learning in unknown channel environments. Therefore, this invention achieves joint optimization of transmit and receive parameters through an end-to-end deep learning architecture, significantly improving the optimization effect of RFID tag signals in complex environments. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention.

[0049] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of this application.

[0050] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0051] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0052] like Figure 1 As shown, this application provides a deep learning-based RFID tag signal optimization method, including:

[0053] S1. Obtain the historical transmit and receive signals and original transmission parameters of the RFID reader, wherein the historical transmit and receive signals include backscatter signals and forward link signals;

[0054] S2. Analyze the backscattered signal to obtain the channel coherence bandwidth and discrete impulse response sequence;

[0055] S3. Obtain the normalized path loss and noise interference feature set based on the historical transmitted and received signals, the original transmission parameters and the discrete impulse response sequence;

[0056] S4. Adjust the original transmission parameters based on the normalized path loss and the channel coherence bandwidth to obtain optimized transmission parameters;

[0057] S5. Based on the optimized transmission parameters, drive the RFID reader to transmit electromagnetic waves to the RFID tag and receive the initial tag signal reflected by the RFID tag;

[0058] S6. Based on the discrete impulse response sequence and the noise interference feature set, the initial tag signal is optimized using a signal optimization model to obtain an optimized tag signal;

[0059] S7. Input the initial label signal into the signal optimization model for model evaluation to obtain the uncertainty index;

[0060] S8. Determine whether the uncertainty index is greater than a preset threshold:

[0061] If the uncertainty index is greater than a preset threshold, the initial tag signal is included in the historical transmit and receive signals, and the process returns to the initial step to perform signal optimization.

[0062] As described in steps S1-S8 above, this invention establishes a training dataset for channel state analysis by acquiring the historical transmit and receive signals and original transmission parameters of the RFID reader. The historical transmit and receive signals refer to all radio frequency signal samples recorded during bidirectional communication between the RFID reader and the RFID tag, including backscattered signals and forward link signals. The original transmission parameters refer to the set of radio frequency configuration parameters actually used by the reader during historical communication cycles, including transmission power and transmission frequency. Subsequently, by extracting features from the backscattered signals, the invention outputs the channel coherence bandwidth characterizing the channel frequency selectivity and the discrete impulse response sequence describing the multipath propagation characteristics. The channel coherence bandwidth refers to the maximum frequency interval when the channel frequency response maintains strong correlation, and the discrete impulse response sequence refers to the digitized sequence characterizing the time delay and attenuation coefficients of each propagation path in the multipath channel. These two quantify the core characteristics of the multipath channel from the frequency domain and time domain, respectively, providing a basis for understanding signal distortion in spatial propagation. The mechanism of change provides key evidence; subsequently, by integrating historical transmitted and received signals, original transmission parameters, and discrete impulse response sequences, normalized path loss and noise interference feature sets are further obtained. Here, normalized path loss refers to the difference in path loss relative to the ideal free space propagation model, and noise interference feature set refers to the quantitative description of interference signals extracted through spectrum analysis and statistical modeling, including rate spectral density distribution and time-domain statistical characteristics. These data information together constitute a numerical representation of the channel state. Then, normalized path loss and channel coherence bandwidth are used as inputs to a fully connected neural network. By training the deep learning network and performing forward computation, and adaptively selecting the operating frequency point based on the channel coherence bandwidth to avoid deep fading, the original transmission parameters are adjusted to obtain optimized transmission parameters. This realizes the transformation of the transmitter from blind and coarse to precise and intelligent control, so as to solve the problem of parameter control and channel state mismatch caused by the information separation between the transmitter and receiver in the existing technology.

[0063] Next, based on optimized transmission parameters, the RFID reader generates and transmits electromagnetic waves to the RFID tag, and receives the initial tag signal reflected by the RFID tag. The initial tag signal refers to the tag response signal received for the first time under the current optimized transmission parameter configuration, containing original noise and channel distortion. Then, the initial tag signal is input into a signal optimization model. This model uses discrete impulse response sequences and noise interference feature sets as the basis for adjusting network weights, and achieves enhanced reconstruction of signal features through convolution operations and attention mechanisms, outputting an optimized tag signal. The obtained normalized path loss and channel coherence bandwidth are used to guide the initial transmission... The adjustment of transmission parameters, along with the use of discrete impulse response sequences and noise interference feature sets to guide the optimized processing of the received signal, solves the optimization objective conflict problem caused by the use of the same coarse indicators at the transmitter and receiver in existing technologies. Furthermore, the signal optimization method of this invention establishes a direct correlation between the receiver's signal optimization processing and the channel state information acquired by the transmitter. By utilizing the precise channel characteristics obtained from transmitter analysis to guide the receiver's signal optimization process, it solves the problem of limited signal optimization effects due to the lack of prior channel information in receiver signal processing in existing technologies. Finally, through a confidence-based calculation... The evaluation module monitors the processing results by inputting the initial tag signal into the signal optimization model for model evaluation, obtaining an uncertainty index. This uncertainty index is a quantitative parameter characterizing the confidence level of the signal optimization model in the current input signal processing result. The uncertainty index is compared with a preset threshold. If the uncertainty index is not greater than the preset threshold, the current channel state is determined to be within the effective recognition range of the model. In this case, the current configuration of the optimized transmission parameters and the signal optimization model can be maintained, and the reader can continue to drive subsequent tag signal optimization and reading communication. If the uncertainty index is greater than the preset threshold, the current channel characteristics are determined to be beyond the coverage range of the model's training experience. In this case, an automatic data update mechanism needs to be triggered, incorporating the current initial tag signal into historical transmit and receive signals as a training dataset and initiating online fine-tuning of the model parameters. This forms an adaptive optimization loop based on continuous learning. By constructing a complete closed-loop optimization system, the signal optimization system achieves self-evaluation and continuous learning capabilities in unknown channel environments. This invention transforms the RF signal optimization problem into a neural network feature learning and parameter optimization process. Through an end-to-end deep learning architecture, it achieves joint optimization of transmit and receive parameters, significantly improving the optimization effect of RFID tag signals in complex environments.

[0064] In one embodiment, step S2, which involves analyzing the backscattered signal to obtain the channel coherence bandwidth and discrete impulse response sequence, includes:

[0065] S21. Obtain the guiding signal, and perform sliding cross-correlation calculation on the guiding signal and the backscattered signal to obtain a cross-correlation function sequence;

[0066] S22. Extract the preamble signal from the backscattered signal based on the cross-correlation function sequence, and perform a fast Fourier transform on the preamble signal to obtain the preamble frequency response;

[0067] S23. Obtain the pilot signal frequency response, and obtain the channel frequency response sequence based on the pilot signal frequency response and the preamble frequency response;

[0068] S24. Obtain the autocorrelation function output sequence based on the channel frequency response sequence, and obtain the channel coherence bandwidth based on the autocorrelation function output sequence;

[0069] S25. The cross-correlation function sequence is transformed using a channel impulse response estimation algorithm to obtain a discrete impulse response sequence.

[0070] As described in steps S21-S25 above, this invention uses a digital signal processor built into the RFID reader to identify and separate a preamble signal with a special predefined format from the backscattered signal, thereby providing the necessary prior signal reference for accurate estimation of channel state information. The specific implementation method is as follows: The digital signal processor pre-stores an ideal preamble signal that is completely consistent with the communication protocol standard as a local reference template. Due to the sharp autocorrelation characteristic of the preamble signal, this characteristic means that the cross-correlation result will only reach its maximum value when the signal sequence is perfectly time-aligned with the baseband waveform copy of the preamble signal. Subsequently, the digital signal processor uses the preamble signal as a... A reference signal is used, and a sliding cross-correlation calculation is performed between it and the backscattered signal to obtain a cross-correlation function sequence. Each point in this sequence corresponds to a specific relative time delay, and its value represents the similarity between the historical backscattered signal and the pilot signal at that time delay. When the pilot signal slides on the time axis to the precise moment when it is completely aligned with the preamble signal hidden in the backscattered signal, a global maximum point will be generated in the cross-correlation function sequence due to the sharp autocorrelation characteristics of the pilot signal. By detecting the time delay value corresponding to this maximum point, the digital signal processor can accurately locate the start time of the preamble signal and accurately extract the preamble signal from the backscattered signal according to the fixed length defined by the communication protocol standard.

[0071] In traditional methods, the preamble signal's primary function is frame synchronization. Once frame synchronization is achieved and a correct time base is established, the potential information inherent in its waveform is ignored. This invention, however, transforms the extracted preamble signal into the frequency domain using a Fast Fourier Transform (FFT), obtaining the preamble frequency response. The preamble frequency response is a discrete frequency domain sequence, where each index point corresponds to a specific frequency component. The complex values ​​of this sequence represent the amplitude and phase of the actually received signal at each frequency after channel interference and noise superposition. Subsequently, the locally stored preamble signal frequency response is acquired. The complex values ​​of this sequence represent the signal under ideal conditions without any distortion or interference. We expect the received signal to have the expected amplitude and phase at each frequency. Then, we perform complex division on the preamble frequency response and the pilot signal frequency response under the influence of the external channel at each corresponding index point to obtain the channel frequency response sequence. Each point in this sequence stores a complex value that precisely describes the complex gain experienced by that frequency component in the pilot signal after passing through the channel. Its magnitude represents the amplitude attenuation of that frequency component, while its argument represents the phase shift experienced by that frequency component. By analyzing the preamble signal, the mixed channel influences in the received signal can be separated, thereby accurately quantifying the amplitude attenuation and phase shift of each frequency component, providing direct frequency domain evidence for diagnosing channel defects.

[0072] Traditional methods for adjusting the transmission parameters of RFID readers often fail to adequately consider the channel coherence bandwidth, leading to overly arbitrary adjustment decisions. For example, when the received signal strength is low, the system reflexively increases the transmission power. However, if the channel coherence bandwidth is very small, increasing the power is ineffective for frequency components experiencing fading and may even exacerbate inter-channel interference. Therefore, this invention evaluates the correlation of the channel frequency response in the frequency domain by calculating the autocorrelation function of the channel frequency response sequence. The specific operation is as follows: First, a conjugate copy of the channel frequency response sequence is generated, i.e., the conjugate of each complex value in the channel frequency response sequence is taken to obtain the conjugate copy sequence. Then, the channel frequency response sequence and its conjugate copy sequence are processed using a frequency-domain sliding correlation algorithm: For each desired frequency interval value, the corresponding part of the channel frequency response sequence is compared with the corresponding part of the conjugate copy sequence shifted along the frequency axis by the frequency interval value. The autocorrelation function (ACF) is obtained by multiplying each element point by point and summing all the products. This summation is the value of the ACF at each frequency interval. By traversing all possible frequency interval values, a complete ACF output sequence is obtained. Finally, the absolute value of the frequency interval corresponding to the point where the ACF value drops from its maximum value to half of that maximum value is found in the ACF output sequence. This absolute value is the channel coherence bandwidth. The channel coherence bandwidth directly reflects the channel's ability to maintain frequency consistency and provides a quantitative basis for determining whether frequency-selective fading will occur. Then, the signal bandwidth is obtained. By comparing the signal bandwidth with the channel coherence bandwidth, a clear quantitative basis can be provided: if the signal bandwidth is much smaller than the channel coherence bandwidth, the channel is determined to be flat fading, and all frequency components experience similar distortion; if the signal bandwidth is greater than the channel coherence bandwidth, the channel is determined to be frequency-selective fading, and different spectral parts of the signal will experience uncorrelated fading, resulting in inter-symbol interference.

[0073] Then, the cross-correlation function sequence is converted into an equivalent discrete impulse response sequence using a standard channel impulse response estimation algorithm: ,in, The sequence length, i.e., the total number of elements contained in the sequence, is determined by the maximum observable delay that the system needs to analyze. Each unit of the sequence is called a "tap". This indicates the tap number, and its specific delay value is equal to the product of the tap number and the system sampling time period. This represents the complex value corresponding to each tap. This complex value is directly derived from the calculation results of the cross-correlation function sequence at the same time delay point. Specifically, the magnitude of the cross-correlation function at that time delay point is used for calculation. The modulus, which quantifies the degree of signal attenuation along the path, is used to calculate the phase information contained in the cross-correlation function at that time delay point. The argument, which quantifies the phase shift experienced by the signal, ultimately... It fully describes the complex changes experienced by all signal components that arrive at the reader after a specific time delay; the discrete impulse response sequence contains the multipath structure information of the channel, and its unique time delay and amplitude distribution can be used as features to identify specific propagation environments, clearly revealing the complete spatiotemporal characteristics such as the arrival time delay of each major multipath component in the channel and the corresponding complex channel gain.

[0074] In one embodiment, step S3, which involves obtaining the normalized path loss and noise interference feature set based on the historical transmit and receive signals, the original transmit parameters, and the discrete impulse response sequence, includes:

[0075] S31. Obtain the average received power value based on the backscattered signal;

[0076] S32. Obtain the initial transmit power and initial antenna gain based on the original transmit parameters, and obtain the normalized path loss based on the initial transmit power, initial antenna gain and the average received power value;

[0077] S33. Perform convolution operation on the forward link signal based on the discrete impulse response sequence to generate a simulated reference signal sequence, and obtain the residual signal sequence based on the simulated reference signal sequence and the backscattered signal;

[0078] S34. Construct a power spectral density curve based on the residual signal sequence, and obtain the equivalent noise spectral density and narrowband interference spectrum set based on the power spectral density curve;

[0079] S35. Analyze the residual signal sequence to obtain the dynamic detection threshold, and use the equivalent noise spectral density, narrowband interference spectrum set and the dynamic detection threshold as the noise interference feature set.

[0080] As described in steps S31-S35 above, this invention converts electromagnetic wave signals into precise, calculable digital features by quantizing the signal power of the backscattered signal. The specific operation is as follows: Since the backscattered signal is a series of discrete instantaneous voltage values ​​sampled by an analog-to-digital converter, each sampling point corresponds to a specific instantaneous voltage value. In communication systems, it is usually assumed that the load impedance is normalized to "1" ohm. The instantaneous power value at that moment can be obtained by squaring the instantaneous voltage value at each sampling point. Next, the entire backscattered signal is defined as a complete sampling window, that is, the total duration of the time series is the length of a sampling window. Within this sampling window, the instantaneous power values ​​are calculated... The arithmetic mean is taken to obtain an average power value representing the average signal strength level over the time period. Since the dynamic range of signal power in wireless communication is extremely large, to facilitate subsequent processing and conform to engineering conventions, this calculated average power value is multiplied by 1000 to convert it to milliwatts. Then, it is substituted into the standard decibel conversion formula to obtain the average received power value. By converting linear-scale operations to logarithmic-scale operations, the dynamic range of the data is greatly compressed, and subsequent calculations are simplified. Next, the initial transmit power and initial antenna gain are obtained based on the original transmit parameters of the RFID reader. These parameters, along with the average received power value, are then substituted into the expression of the Fries free-space propagation model in the decibel system. in, This represents the average received power value. Indicates the initial transmit power. This represents the initial antenna gain. The actual path loss can be directly calculated by inversely solving for the actual path loss corresponding to the sampling window, which includes both inherent system loss and environmental propagation loss. Then, a reference path loss value is obtained. This reference path loss is a calibration value obtained before system deployment in an ideal, reflection-free microwave anechoic chamber at a fixed and known standard distance, using the same hardware equipment and configuration, and through the exact same measurement and calculation process. It accurately characterizes the systematic fixed deviation introduced by the inherent insertion loss and efficiency issues of specific readers, antennas, cables, and other hardware. Finally, the actual path loss corresponding to the sampling window is algebraically compared with the reference path loss value. Subtraction yields the normalized path loss, which effectively filters out fixed system biases that do not change with the environment. The resulting normalized path loss purely reflects the additional path loss caused by time-varying factors such as distance, obstacles, and multipath effects in the current actual deployment environment, thus quantifying the relative severity of the current propagation environment. Traditional methods rely solely on a single received signal strength indicator for rough judgment when adjusting transmission parameters. However, this invention achieves precise separation between environmentally-added losses and inherent hardware losses through rigorous power quantization and normalized path loss calculation, thereby solving the defect of traditional methods that misjudge environmental quality due to changes in hardware status.

[0081] Next, the forward link signal, time-aligned with the backscattered signal, is acquired and used as a known input reference. Subsequently, convolution is performed on the forward link signal based on the discrete impulse response sequence to generate a simulated reference signal sequence. The simulated reference signal sequence refers to an idealized received signal that reconstructs only multipath effects but does not contain noise or interference. Then, the backscattered signal and the simulated reference signal are subtracted at the sampling point level to obtain the residual signal sequence. This residual signal sequence mainly contains random components not captured by the preceding model, namely, the composite features of environmental noise, device thermal noise, and other unstructured interference. Traditional receivers treat multipath and noise as a whole, while this invention decomposes the original aliased signal sequence into two feature sets: deterministic multipath components and random noise interference components by actively reconstructing and eliminating known multipath components. This solves the core problem that traditional methods cannot accurately extract noise and interference features in strong multipath environments, which are often overwhelmed. It can provide structured and decoupled feature inputs for subsequent deep learning models, thereby significantly improving the model's recognition accuracy and learning efficiency of noise and interference features, and laying a data foundation for achieving accurate signal optimization.

[0082] Traditional methods for understanding noise are limited to its strength. This invention, however, uses a Fast Fourier Transform (FFT) on the residual signal sequence to obtain the noise interference spectrum. Then, the Welch method is employed to estimate the noise interference spectrum, yielding a discrete power spectral density (PSD) estimate sequence. This sequence, graphically represented as a PSD curve, clearly characterizes the power distribution of noise and interference with frequency. Based on this, within the system's effective bandwidth corresponding to the PSD curve, the PSD values ​​at discrete frequency points are summed to obtain the total noise power. This value represents the total energy of the noise within the system's effective bandwidth. Subsequently, in order to establish... A frequency domain benchmark for noise type identification is used. The ratio of total noise power to the system's effective bandwidth is calculated to obtain the equivalent noise spectral density. This value represents the power spectral density of ideal white noise that is equal to the currently measured total noise power and absolutely flat within the system's effective bandwidth. Subsequently, a spectral flatness test is performed: the entire power spectral density curve is scanned, iterating through the power spectral density values ​​corresponding to each frequency point. If the fluctuation range of the power spectral density values ​​corresponding to all frequency points relative to the equivalent noise spectral density is within ±1.5 dB, the current noise type is determined to be white noise. In this case, the equivalent noise spectral density is recorded to characterize the noise intensity; if... If the power spectral density value corresponding to any one or more frequency points fluctuates within a range exceeding ±1.5 dB relative to the equivalent noise spectral density, the current noise type is determined to be colored noise or narrowband interference. In this case, not only is the equivalent noise spectral density recorded to characterize the background noise floor, but also the power spectral density values ​​corresponding to all frequency points exceeding the fluctuation range are recorded to accurately characterize the spectral location and intensity of the interference. If the power spectral density value corresponding to a specific frequency on the power spectral density curve exceeds the equivalent noise spectral density by more than 10 dB, a single-tone interference is determined to exist, and the specific frequency corresponding to the narrowband spectral peak and its corresponding power spectral density value are recorded, thus forming a packet... A narrowband interference spectrum set containing specific frequencies and their corresponding power spectral density values ​​is generated. Through this multidimensional analysis of noise spectrum and statistics, a precise diagnosis of noise type is achieved, solving the problem that traditional methods cannot distinguish between broadband white noise and narrowband interference. At the same time, statistical analysis is performed on the residual signal sequence to calculate its fourth central moment and obtain the noise kurtosis index. In probability statistics, the theoretical benchmark value of Gaussian distribution kurtosis is three. If the calculated noise kurtosis index of the residual signal is greater than this benchmark value, it is determined that its statistical distribution has heavy-tailed characteristics, which means that impulse noise exists. The standard deviation of the residual signal sequence is calculated to quantify the amplitude dispersion of the background noise.Subsequently, combining the calculated noise kurtosis index, which characterizes the heavy-tailed distribution pattern, an adaptive method based on standard deviation multiples is used to determine the dynamic detection threshold: when the noise kurtosis index is significantly greater than three, it indicates that the distribution has heavy-tailed characteristics and there are impulse samples with amplitudes much higher than the normal level. In this case, the dynamic detection threshold is set to three to five times the standard deviation; if the noise kurtosis index is close to three, the dynamic detection threshold is set to two to three times the standard deviation. Finally, the equivalent noise spectral density, narrowband interference spectrum set, and dynamic detection threshold together constitute a noise interference feature set for directly guiding subsequent signal purification processing. This feature set not only quantifies the overall intensity of noise and interference but also accurately characterizes its spectral distribution and statistical characteristics, thereby improving the system's perception capability from macroscopic link quality assessment to fine-grained identification of interference types and components, providing a key decision-making basis for subsequent targeted signal optimization.

[0083] In one embodiment, step S4, which adjusts the original transmission parameters based on the normalized path loss and the channel coherence bandwidth to obtain optimized transmission parameters, includes:

[0084] S41. The normalized path loss and the channel coherence bandwidth are calculated forward through a deep learning network to obtain the real-time adaptive threshold and the real-time gradient shape factor.

[0085] S42. Obtain the standard sub-channel bandwidth, and obtain the frequency aggregation determination threshold based on the standard sub-channel bandwidth and the real-time adaptive threshold;

[0086] S43. Compare the channel coherence bandwidth with the frequency aggregation determination threshold to obtain frequency domain resource adjustment information;

[0087] S44. Obtain the initial channel symbol rate and initial transmit power based on the original transmission parameters, and adjust the initial channel symbol rate based on the frequency domain resource adjustment information to obtain the optimized channel symbol rate;

[0088] S45. Obtain the weight allocation strategy, and obtain the power weight vector according to the weight allocation strategy and the gradient shape factor;

[0089] S46. Adjust the initial transmit power based on the power weight vector and the frequency domain resource adjustment information to obtain the optimized transmit power, and use the optimized channel symbol rate and the optimized transmit power as optimized transmit parameters.

[0090] As described in steps S41-S46 above, this invention provides a complete data foundation with clear optimization objectives for subsequent deep learning model training by acquiring a historical experience database. The historical experience database stores a large number of data samples collected at different times and under different environments. Each sample contains a four-tuple of information: a historical channel feature vector as input conditions, i.e., a vector composed of historical normalized path loss and historical channel coherence bandwidth; and optimal system configuration parameters obtained through post-analysis optimization as supervision labels. The optimal system configuration parameters include a dynamic adaptive threshold and a gradient shape factor. The method for obtaining the optimal system configuration parameters is as follows: During the learning phase after system deployment, for each channel feature vector, the system will conduct a small-scale exploration within its parameter space, i.e., try multiple different configurations, and record the comprehensive system performance index achieved after executing a complete communication process under each configuration. This index is the read success rate, read latency, and reader energy consumption. The weighted harmonic mean of the parameters is calculated. Then, through comparative analysis, the configuration parameters that bring the highest comprehensive performance index to the channel feature vector are recorded as the optimal system configuration parameters corresponding to the channel feature vector and stored in the historical experience database. By using a large number of such quadruplet samples to supervise the training of the deep learning network, the network becomes a regression prediction model that can map any given channel feature vector to its corresponding, approximately optimal, optimal system configuration parameters. Through this data-driven approach, the system automatically learns and internalizes the complex nonlinear mapping relationship between the channel state and the globally optimal system configuration, thereby providing the transmitter with unprecedented intelligent decision-making capabilities. In actual operation, a real-time channel feature vector is first constructed based on the normalized path loss and channel coherence bandwidth. Then, the real-time channel feature vector is input into the trained deep learning network for forward computation, and the predicted real-time adaptive threshold and real-time gradient shape factor are output.

[0091] Subsequently, the system coordinates the adjustment of the RFID reader's original transmission parameters based on these two predicted parameters. This adjustment process includes two core operations: dynamic frequency aggregation and intelligent gradient power allocation. In the dynamic frequency aggregation operation, firstly, the system obtains the preset standard sub-channel bandwidth and calculates the product of the standard sub-channel bandwidth and the real-time adaptive threshold to obtain the frequency aggregation judgment threshold. Then, the channel coherence bandwidth is compared with the frequency aggregation judgment threshold to obtain frequency domain resource adjustment information, which includes frequency domain resource aggregation information and frequency domain resource allocation information. If the channel coherence bandwidth is greater than the frequency aggregation judgment threshold, the frequency domain resource configuration in the original transmission parameters is aggregated. Aggregation adjustment: The spectrum resources within the RFID reader's baseband processor, originally mapped to multiple independent low-speed data streams with adjacent frequencies and each having a bandwidth equal to the standard sub-channel bandwidth, are aggregated into a single wideband logical channel with a total bandwidth equal to the product of the standard sub-channel bandwidth and the number of sub-channels. This channel carries a single high-speed data stream. This aggregation operation improves the system's initial channel symbol rate from a single value to a new optimized value, i.e., the optimized channel symbol rate. The optimized channel symbol rate is equal to the product of the initial channel symbol rate, the number of aggregated sub-channels, and the number of bits carried by each symbol under the high-order modulation scheme used by the system, thereby achieving a leap in throughput. Conversely, if the channel is coherent... If the bandwidth is not greater than the frequency aggregation threshold, the frequency domain resource configuration in the original transmission parameters is adjusted by partitioning: the original single-carrier broadband signal is divided into multiple independent narrowband transmission units within the baseband. The number of units is equal to the ratio of the total system bandwidth to the real-time channel coherence bandwidth, ensuring that the bandwidth of each unit is not greater than the real-time channel coherence bandwidth. The single high-speed data stream to be transmitted is demultiplexed into multiple parallel low-speed streams. This partitioning operation adjusts the initial channel symbol rate of the system from a single value to the sum of the symbol rates of multiple parallel channels, thus obtaining the optimized channel symbol rate. The optimized channel symbol rate of each independent narrowband transmission unit is set to the initial channel symbol rate and... The ratio of the number of narrowband transmission units is used to reduce the symbol rate of a single channel in exchange for transmission reliability. The above parameter adjustment method can solve the problems of the traditional method, which is one-sided and has the opposite effect: when the channel coherence bandwidth is diagnosed as insufficient, the present invention does not blindly increase the transmission power based solely on the received signal strength as in the traditional method (this operation will aggravate the interference between multipath components in the frequency selective channel, leading to an increase in the bit error rate), but fundamentally reconstructs the spectral structure of the transmitted signal based on the core feature of the channel coherence bandwidth, ensuring that the spectral characteristics of each component match the frequency consistency capability of the channel, thereby avoiding the generation of inter-symbol interference from a physical mechanism.

[0092] After completing the aforementioned dynamic frequency aggregation or partitioning operations to determine the signal's spectral structure and channel symbol rate, an intelligent gradient power allocation operation is then performed on the initial transmit power extracted from the original transmit parameters to ultimately determine the power spectrum shape of the transmitted signal. When aggregation adjustments are required, intelligent gradient power allocation is performed based on the normalized path loss and the gradient shape factor output by the deep learning network. This gradient shape factor is a specific numerical value that directly determines the steepness of the power distribution profile. Its specific implementation involves: first, obtaining the weight allocation strategy, which is pre-set before system deployment, and its core mathematical relationship is designed as follows: each The weight of each subcarrier is determined by the absolute value of its normalized offset from the center frequency of the logical channel, and the gradient shape factor. It strictly follows a decreasing relationship: the larger the frequency offset, the smaller the weight. Simultaneously, the gradient shape factor is used as an exponent or gain coefficient of this decreasing relationship. This ensures that a larger gradient shape factor amplifies the difference between the center and edge weights, resulting in a steeper power distribution profile. Based on this gradient shape factor, a power weight vector is calculated and generated using a weight allocation strategy. The dimension of the power weight vector is consistent with the total number of subcarriers in the aggregated broadband channel, with the subcarriers located near the center frequency band of the logical channel being weighted. Subcarriers are assigned larger weights, while those in the edge bands are assigned smaller weights. The larger the gradient shape factor, the greater the difference between the center and edge weights, resulting in a steeper power distribution profile. The sum of all weights in the entire vector equals the total number of subcarriers to ensure the total transmit power remains constant before and after allocation. Subsequently, the product of the power weight for each subcarrier and its initial transmit power is calculated to obtain the optimized transmit power. This power allocation method elevates power control from traditional total control to the dimension of spectrum shape management, intelligently utilizing the gradient shape factor—a quantization command derived from deep learning—to achieve this. By precisely focusing power energy on the center frequency band where channel conditions are relatively better, it is beneficial to maximize energy delivery efficiency under complex channels. Conversely, when division and adjustment are required, the power allocation strategy aims to ensure the basic communication quality of each independent narrowband channel. At this time, the system distributes the total transmit power equally to each narrowband transmit unit formed after division, that is, calculates the ratio of the initial transmit power to the total number of narrowband transmit units to obtain the optimized transmit power. Through this collaborative adjustment operation, the present invention completely breaks the limitation of isolated adjustment of transmit parameters (such as power and rate) in traditional methods, and realizes joint optimization of transmit parameters based on deep diagnosis of channel characteristics.

[0093] In one embodiment, step S6, which optimizes the initial tag signal based on the discrete impulse response sequence and the noise interference feature set using a signal optimization model to obtain the optimized tag signal, includes:

[0094] S61. Input the initial label signal and the discrete impulse response sequence into the encoder of the signal optimization model for encoding and feature extraction to obtain the intermediate layer feature tensor and conditional coding vector.

[0095] S62. The conditional coding vector and the intermediate layer feature tensor are fused through the conditional layer normalization operation to obtain the fused feature tensor, and the fused feature tensor is input into the signal optimization model for analysis to generate the correction feature tensor.

[0096] S63. Extract the equivalent noise spectral density, narrowband interference spectrum set, and dynamic detection threshold from the noise interference feature set, and filter the correction feature tensor according to the narrowband interference spectrum set to obtain the interference suppression feature tensor.

[0097] S64. Based on the dynamic detection threshold, the interference suppression feature tensor is smoothed and limited by a linear suppression function to obtain a cleaned feature tensor. The cleaned feature tensor and the equivalent noise spectral density are then input into the decoder of the signal optimization model for signal reconstruction to obtain the optimized label signal.

[0098] As described in steps S61-S64 above, this invention performs final signal optimization by applying a pre-trained signal optimization model. This model has been trained offline using a large amount of historical data and has reached convergence before deployment, thus its functionality is already fixed. In the actual optimization process, the initial label signal received in real-time is first input to the input layer of the signal optimization model. Simultaneously, the discrete impulse response sequence representing the multipath propagation path of the signal is input to a dedicated conditional vector fully connected layer. This layer maps the discrete impulse response sequence into a conditional encoding vector that matches the dimension of the encoder's intermediate feature layer. Subsequently, the encoder processes the input initial label signal... After convolutional and nonlinear transformation processing, when the signal is processed to a preset intermediate level, the encoder outputs an intermediate-level feature tensor that carries the abstract representation of the signal at this level. Simultaneously, the signal optimization model passes the conditional coding vector through a linear projection layer to generate a set of scaling parameters and a set of translation parameters corresponding to the channel dimensions of the intermediate-level feature tensor. Next, the signal optimization model initiates its feature fusion mechanism. This mechanism first performs standard layer normalization on the intermediate-level feature tensor to stabilize its data distribution. Then, it performs a conditional layer normalization transformation, that is, multiplying the normalized intermediate-level feature tensor channel by channel using the scaling parameters, and then performing a channel-by-channel transformation. By adding translation parameters, a new fused feature tensor is obtained, whose data distribution has been dynamically modulated by the multipath channel information represented by the conditional coding vector. This method of dynamically readjusting the statistical distribution of deep features using the multipath channel information carried by the conditional coding vector can amplify feature channels related to multipath interference and suppress irrelevant channels. This allows the signal optimization model to activate the corresponding multipath interference suppression mode in its feature space based on the specific multipath characteristics of the current channel. The fused feature tensor serves as the input to the encoder at subsequent deeper levels, ensuring that the model's further abstraction and refinement of signal features is always influenced by the current multipath channel characteristics. The traditional method uses blind equalization or adaptive equalization based on training sequences. The update of the filter coefficients lacks explicit and deep coupling with the physical channel impulse response. However, this invention uses a specific network structure of conditional layer normalization to take the channel state information represented by the discrete impulse response sequence as a priori condition and directly modulate the distribution of the signal in the feature space. This method introduces channel distortion compensation in the feature extraction stage. In essence, it constructs a nonlinear filter conditioned on the channel impulse response. Compared with traditional linear or finite nonlinear equalizers, this method can more accurately model and cancel the nonlinear inter-symbol interference caused by multipath propagation.Based on this, the fused feature tensor is fed into a multi-head scaled dot product self-attention layer for computation to generate a dynamic temporal attention weight map. The specific generation process includes: first, linearly projecting the fused feature tensor into a query matrix, a key matrix, and a value matrix using three independent trainable weight matrices; then calculating the product of the query matrix and the transpose of the key matrix; and scaling and normalizing the result using an exponential function to obtain an attention score matrix, where each element represents the correlation strength between features at different time points in the signal sequence. During model training, by using an ideal, distortion-free, clean signal as the supervision target, the model is forced to learn an inverse attention allocation strategy: that is, for time point regions where symbol energy overlaps due to multipath delays, resulting in high correlation, the attention weight map will assign significantly lower weight values. After multiplying the domain attention weight map with the value matrix, the signal component weights in these overlapping symbol regions are precisely weakened at the feature level, resulting in a corrected feature tensor that has undergone multipath distortion correction. Traditional time-domain equalizers, such as linear transverse filters or decision feedback equalizers, are limited by the number of taps and linear processing capabilities, making it difficult to handle frequency-selective channels with deep fading. The multi-head self-attention mechanism adopted in this invention calculates the global dependencies of the sequence, and the dynamically generated time-domain attention weight map can accurately identify the energy dispersion regions caused by multipath delay spread and suppress the features of these regions. This mechanism is equivalent to an adaptive equalizer with long memory depth and strong nonlinear processing capabilities. Its tap coefficients (attention weights) are dynamically optimized according to the joint features of signal content and channel state, thereby achieving intelligent cancellation of multipath distortion in the deep feature space.

[0099] Next, the noise interference feature set is used as a series of control parameters input to the signal optimization model to guide it in performing targeted noise filtering. The specific implementation process is as follows: The signal optimization model first performs a Fast Fourier Transform on the correction feature tensor output by the encoder, transforming it from a time-domain feature representation to a frequency-domain feature representation, obtaining a complex-form frequency-domain feature tensor. Each specific dimension of this tensor corresponds to a frequency component of the signal. Subsequently, a narrowband interference spectrum set is extracted from the noise interference feature set. This set pre-records the specific frequencies of all identified narrowband interferences and their corresponding power spectral density values. For each specific frequency that needs to be suppressed, the signal optimization model simulates the notch filtering effect in its internal frequency-domain feature tensor using a differentiable, distance-based Gaussian weighted function. This function generates a continuous mask value between 0 and 1 on several frequency channels near the target specific frequency. The closer the channel is to the specific frequency, the closer its mask value is to 0. As the distance between the frequency and the specific frequency increases, the mask value smoothly transitions to 1, thus obtaining an adaptive Gaussian notch filter mask. Then, The generated adaptive Gaussian notch mask is multiplied pointwise by a pointwise complex multiplication operation with the frequency domain feature tensor. This operation significantly attenuates the amplitude of eigenvalues ​​near specific frequencies of narrowband interference, while preserving normal signal features outside the transition band almost without loss, thus obtaining a complex frequency domain feature tensor. The model then performs an inverse fast Fourier transform on the complex frequency domain feature tensor to restore it to its time domain feature representation, resulting in the interference suppression feature tensor. The processing of the deep learning model is a phased, layer-by-layer abstraction process. The encoder is responsible for analyzing and extracting features, while the decoder is responsible for synthesizing and reconstructing based on these features. The interference suppression feature tensor ensures that all information obtained through the encoder and frequency domain filtering module can be completely and efficiently passed to the decoder for signal reconstruction. The differentiability of the entire process ensures that the adaptive notch behavior can be optimized end-to-end with other parts of the model, enabling the model to not only learn how to apply a mask according to a given interference frequency during training, but also to optimize the width and steepness of the mask, thereby achieving the optimal balance between suppressing interference and preserving useful signals.

[0100] Simultaneously, a dynamic detection threshold is extracted from the noise interference feature set. Based on this, the internal feature flow of the model is monitored and processed in real time. This processing is not a simple "if-else" logical judgment, but is implemented through a differentiable, parameterized nonlinear suppression function. This function uses the dynamic detection threshold as the core control parameter. Specifically, the nonlinear suppression function is mathematically represented as a continuous and differentiable composite function. Its function curve is designed such that when the input activation value is much lower than the dynamic detection threshold, its output approximates the input, thus preserving normal signal characteristics without loss. When the input activation value is close to the dynamic detection threshold, the function curve provides a smooth transition range. When the input activation value significantly exceeds the dynamic detection threshold, the output becomes approximately equal to the input, thus preserving normal signal characteristics without loss. When the threshold is measured, the output gain of the function drops sharply and gradually approaches saturation, stably limiting the output value to a preset upper limit. At the signal processing level, the input-output relationship described by the nonlinear suppression function is equivalent to an adaptive soft limiter whose limiting threshold is dynamically controlled by a dynamic detection threshold. The fully differentiable nature of this nonlinear suppression function ensures that the entire impulse noise suppression process can be seamlessly integrated into the model's end-to-end training framework. Through the backpropagation algorithm, the model not only learns how to identify impulse noise but also jointly optimizes the specific parameters of the nonlinear suppression function to achieve accurate and smooth suppression of abnormal activation values ​​while minimizing damage to normal signal features. Deep neural networks... Networks risk overfitting the training target in signal reconstruction tasks, potentially leading to excessive smoothing of the model's output and damaging weak, useful signal components with power levels close to the background noise. Therefore, this invention extracts the equivalent noise spectral density from the noise interference feature set and uses it as a reference benchmark for the background noise level, incorporating it as a regularization constraint into the decoder's signal reconstruction process. Specifically, during the training phase, an additional spectral density constraint term is added to the model's reconstruction loss function. This term first calculates the difference between the power spectral density of the final output signal and the equivalent noise spectral density in non-signal frequency bands, and uses the norm of this difference as an additional regularization loss term. By minimizing this loss, the model is forced to... The requirement that the background noise component in the output optimized signal must not be significantly lower than the provided equivalent noise spectral density benchmark in terms of power spectrum shape effectively prevents the decoder from falling into overfitting in pursuit of an excessively high output signal-to-noise ratio, thereby suppressing weak useful signals with amplitudes similar to the background noise, and ultimately achieving an optimal balance between noise suppression and signal preservation. At this point, the deep feature sequence, which has undergone the above-mentioned multipath compensation, narrowband interference suppression, impulse noise suppression, and background noise constraint processing, is fed into the decoder part of the model for signal reconstruction. The decoder outputs an optimized label signal with a clear time-domain waveform and significantly suppressed interference components through a series of preset upsampling and nonlinear transformation layers.

[0101] In one embodiment, step S7, which involves inputting the initial label signal into a signal optimization model for model evaluation to obtain an uncertainty index, includes:

[0102] S71. Obtain the preset neuron masking probability, and obtain the masking decision random number corresponding to each neuron in a specific hidden layer in the signal optimization model according to the neuron masking probability.

[0103] S72. Obtain a binary random mask based on the neuron masking probability and the masking decision random number, and repeatedly input the initial label signal into the encoder loaded with the binary random mask to obtain multiple deep feature tensors;

[0104] S73. Calculate the cognitive consistency metric between any two deep feature tensors respectively, and obtain the uncertainty index based on the multiple cognitive consistency metrics.

[0105] As described in steps S71-S73 above, this invention obtains a preset neuron shielding probability for controlling the shielding strength. This probability defines the likelihood that the output of each neuron in a specified hidden layer of the encoder in the signal optimization model is set to 0 during forward propagation. This probability serves as the core parameter for controlling the random perturbation strength of the signal optimization model. Then, based on the neuron shielding probability, a shielding decision random number is independently generated for each neuron in a specific hidden layer of the encoder. The specific implementation method is as follows: First, the pseudo-random number generator iteratively updates its internal current state. This update process is defined by a linear recursive function, such as using the Mason rotation algorithm, which generates a new state value through its built-in shift, AND, OR, and NOT bit operations. Subsequently, based on the updated internal state, a floating-point number satisfying a uniform distribution characteristic over an interval is extracted and mapped from the state value through a specific output conversion function. For each neuron to be processed in a specific hidden layer of the encoder, the complete calculation process from state update to output conversion is independently repeated once, thereby assigning a statistically independent random number to each neuron, thus obtaining the shielding result. A random number is generated for the decision, and then the random number for the mask decision is compared with the masking probability of the neuron to generate a binary random mask that conforms to the Bernoulli distribution. If the random number for the mask decision is greater than the masking probability of the neuron, the mask value is set to 1, indicating that the neuron is activated and its output value will be retained in this forward propagation and participate in subsequent calculations. If the random number for the mask decision is not greater than the masking probability of the neuron, the mask value is set to 0, indicating that the neuron is temporarily masked and its output value is forcibly set to zero in this forward propagation, simulating the interruption of the connection path. Traditional methods for post-event evaluation of signal quality by receiving signal strength indication or bit error rate measure the results after channel degradation and are susceptible to specific types of interference. This invention abandons this external dependence and creatively transforms the Monte Carlo method in deep learning into a self-diagnostic mechanism embedded in the model itself and placed before signal optimization. By generating and applying a binary random mask, the system actively introduces controlled structural perturbations into the model. This is not for training, but to actively explore the stability of the model's cognition during the inference stage, laying the foundation for subsequent uncertainty quantification.

[0106] Subsequently, the generated binary random mask is multiplied element-wise with the output tensor of the neuron in its hidden layer after being transformed by a nonlinear activation function. This operation constructs slightly different network connectivity paths in each forward propagation. Based on this, the initial label signal and the neuron masking probability are repeatedly input into the encoder of the signal optimization model. Through the above controlled random perturbation mechanism, based on the same input signal, i.e., the initial label signal, multiple deep feature tensors with subtle differences are finally output, and a feature sampling set is constructed based on multiple deep feature tensors. Traditional signal quality assessment methods cannot effectively identify the following two situations: one is that the input signal itself is of extremely high quality, and its clear features make it difficult for any... The model can easily make accurate interpretations; secondly, the input signal quality is poor, but the optimized model, with its complex nonlinear mapping ability, accidentally outputs a result that seems correct on a specific index. These two situations are treated as equal in traditional, single-output deterministic models. However, this invention transforms the single, deterministic feature extraction process into a probabilistic, statistical sampling process by constructing a feature sampling set. Each deep feature tensor in this set represents the model's understanding of the signal in a network state under controlled random perturbation. The whole set constitutes an approximate sample of the model's cognitive posterior. By analyzing the statistical characteristics of this sampling set, the two essentially different situations mentioned above can be distinguished.

[0107] Then, by calculating the cosine distance between each pair of deep feature tensors in the feature sampling set and using this value as a cognitive consistency metric, the arithmetic mean of these cognitive consistency metrics is obtained to yield an uncertainty index. This value quantifies the model's cognitive uncertainty regarding the current input signal. A high uncertainty index indicates poor consistency in the feature representation generated by the model under random perturbations, reflecting that the input signal has been severely distorted due to poor channel conditions and deviates from the model's learned distribution, thus triggering system-level optimization decisions. A low uncertainty index indicates good current channel propagation conditions, less distortion introduced by multipath effects and noise interference in the initial label signal, and high confidence in the signal optimization model's feature extraction and semantic understanding. The currently adopted optimization method... The emission parameters and signal cleanup strategies remain effective, eliminating the need for system reconfiguration. This invention transforms traditional signal quality assessment, which relies on external physical layer indicators, into uncertainty quantification based on the model's own cognitive state by actively injecting controlled random perturbations during the model inference phase and analyzing their output consistency. This is a higher-dimensional, introspective evaluation paradigm based on the model's internal state. Furthermore, the uncertainty index is endogenously obtained by the signal optimization model during forward propagation through controlled random perturbations and analysis of the consistency of its output in the feature space. Therefore, the uncertainty index is deeply coupled with the architecture and forward computation process of the signal optimization model, breaking through the paradigm where the design and implementation of signal evaluation and signal processing functional modules in traditional communication systems are independent.

[0108] This application also provides a deep learning-based RFID tag signal optimization system, comprising:

[0109] The signal acquisition module is used to acquire the historical transmit and receive signals and original transmission parameters of the RFID reader, wherein the historical transmit and receive signals include backscatter signals and forward link signals;

[0110] The signal analysis module is used to analyze the backscattered signal to obtain the channel coherence bandwidth and discrete impulse response sequence;

[0111] The feature acquisition module is used to acquire a normalized path loss and noise interference feature set based on the historical transmit and receive signals, the original transmission parameters and the discrete impulse response sequence;

[0112] The parameter optimization module is used to adjust the original transmission parameters based on the normalized path loss and the channel coherence bandwidth to obtain optimized transmission parameters;

[0113] The signal transceiver module is used to drive the RFID reader to transmit electromagnetic waves to the RFID tag based on the optimized transmission parameters, and to receive the initial tag signal reflected by the RFID tag.

[0114] The signal optimization module is used to optimize the initial tag signal based on the discrete impulse response sequence and the noise interference feature set through a signal optimization model to obtain an optimized tag signal;

[0115] The model evaluation module is used to evaluate the input signal optimization model with the initial label signal to obtain the uncertainty index.

[0116] The decision control module is used to determine whether the uncertainty index is greater than a preset threshold.

[0117] If the uncertainty index is greater than a preset threshold, the initial tag signal is included in the historical transmit and receive signals, and the process returns to the initial step to perform signal optimization.

[0118] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described deep learning-based RFID tag signal optimization method.

[0119] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described deep learning-based RFID tag signal optimization method.

[0120] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0121] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0122] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A deep learning-based RFID tag signal optimization method, characterized by, The method comprises the following steps: acquiring historical transceiving signals and original transmission parameters of an RFID reader, wherein the historical transceiving signals comprise backscatter signals and forward link signals; analyzing the backscatter signals to obtain channel coherence bandwidth and discrete impulse response sequence; acquiring normalized path loss and noise interference feature set according to the historical transceiving signals, original transmission parameters and the discrete impulse response sequence; adjusting the original transmission parameters based on the normalized path loss and the channel coherence bandwidth to obtain optimized transmission parameters; driving the RFID reader to transmit electromagnetic waves to an RFID tag based on the optimized transmission parameters, and receiving initial tag signals reflected by the RFID tag; optimizing the initial tag signals based on the discrete impulse response sequence and the noise interference feature set through a signal optimization model to obtain optimized tag signals; inputting the initial tag signals into the signal optimization model for model evaluation to obtain an uncertainty index; judging whether the uncertainty index is greater than a preset threshold: if the uncertainty index is greater than the preset threshold, the initial tag signals are included in the historical transceiving signals, and the signal optimization is performed again from the initial step. 2.The deep learning-based RFID tag signal optimization method of claim 1, wherein, The step of analyzing the backscatter signals to obtain channel coherence bandwidth and discrete impulse response sequence comprises the following steps: acquiring a pilot signal, and performing a sliding cross-correlation calculation on the pilot signal and the backscatter signal to obtain a cross-correlation function sequence; extracting a preamble signal from the backscatter signal based on the cross-correlation function sequence, and performing a fast Fourier transform on the preamble signal to obtain a preamble frequency response; acquiring a pilot signal frequency response, and acquiring a channel frequency response sequence based on the pilot signal frequency response and the preamble frequency response; acquiring an autocorrelation function output sequence based on the channel frequency response sequence, and acquiring channel coherence bandwidth based on the autocorrelation function output sequence; converting the cross-correlation function sequence by using a channel impulse response estimation algorithm to obtain a discrete impulse response sequence. 3.The deep learning-based RFID tag signal optimization method of claim 1, wherein, The step of acquiring normalized path loss and noise interference feature set according to the historical transceiving signals, original transmission parameters and the discrete impulse response sequence comprises the following steps: acquiring an average received power value based on the backscatter signals; acquiring an initial transmission power and an initial antenna gain based on the original transmission parameters, and acquiring a normalized path loss based on the initial transmission power, the initial antenna gain and the average received power value; performing a convolution operation on the forward link signals based on the discrete impulse response sequence to generate an analog reference signal sequence, and acquiring a residual signal sequence based on the analog reference signal sequence and the backscatter signals; constructing a power spectral density curve based on the residual signal sequence, and acquiring an equivalent noise spectral density and a narrowband interference spectrum set based on the power spectral density curve; analyzing the residual signal sequence to obtain a dynamic detection threshold, and taking the equivalent noise spectral density, the narrowband interference spectrum set and the dynamic detection threshold as a noise interference feature set. 4.The deep learning-based RFID tag signal optimization method of claim 1, wherein, The step of adjusting the original transmission parameter based on the normalized path loss and the channel coherence bandwidth to obtain an optimized transmission parameter comprises: The normalized path loss and the channel coherence bandwidth are forward calculated through a deep learning network to obtain a real-time adaptive threshold and a real-time gradient shape factor; A standard sub-channel bandwidth is obtained, and a frequency aggregation judgment threshold is obtained according to the standard sub-channel bandwidth and the real-time adaptive threshold; The channel coherence bandwidth is compared with the frequency aggregation judgment threshold to obtain frequency domain resource adjustment information; An initial channel symbol rate and an initial transmission power are obtained according to the original transmission parameter, and the initial channel symbol rate is adjusted based on the frequency domain resource adjustment information to obtain an optimized channel symbol rate; A weight allocation strategy is obtained, and a power weight vector is obtained according to the weight allocation strategy and the gradient shape factor; The initial transmission power is adjusted based on the power weight vector and the frequency domain resource adjustment information to obtain an optimized transmission power, and the optimized channel symbol rate and the optimized transmission power are taken as the optimized transmission parameter. 5.The deep learning-based RFID tag signal optimization method of claim 1, wherein, The step of optimizing the initial tag signal based on the discrete impulse response sequence and the noise interference feature set through a signal optimization model to obtain an optimized tag signal comprises: The initial tag signal and the discrete impulse response sequence are input into an encoder of the signal optimization model for encoding and feature extraction to obtain an intermediate layer feature tensor and a conditional encoding vector; The conditional encoding vector and the intermediate layer feature tensor are fused through a conditional layer normalization operation to obtain a fusion feature tensor, and the fusion feature tensor is input into the signal optimization model for analysis to generate a correction feature tensor; An equivalent noise spectral density, a narrowband interference spectrum set and a dynamic detection threshold are extracted from the noise interference feature set, and the correction feature tensor is filtered according to the narrowband interference spectrum set to obtain an interference suppression feature tensor; The interference suppression feature tensor is smoothed by a linear clipping function based on the dynamic detection threshold to obtain a purified feature tensor, and the purified feature tensor and the equivalent noise spectral density are input into a decoder of the signal optimization model for signal reconstruction to obtain an optimized tag signal. 6.The deep learning-based RFID tag signal optimization method of claim 1, wherein, The step of inputting the initial tag signal into the signal optimization model for model evaluation to obtain an uncertainty index comprises: A preset neuron masking probability is obtained, and a masking decision random number corresponding to each neuron of a specific hidden layer in the signal optimization model is obtained according to the neuron masking probability; A binary random mask is obtained according to the neuron masking probability and the masking decision random number, and the initial tag signal is repeatedly input into the encoder loaded with the binary random mask to obtain a plurality of deep layer feature tensors; Cognitive consistency metric values between any two deep layer feature tensors are calculated respectively, and an uncertainty index is obtained according to a plurality of the cognitive consistency metric values. 7.A deep learning based RFID tag signal optimization system, characterized by, The system comprises a plurality of modules for implementing the steps of the method of any one of claims 1 to 6.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

Images