HPLC and HRF switching method based on dual-mode channel difference perception

Abstract

The invention discloses an HPLC and HRF switching method based on dual-mode channel difference perception. The method comprises the following steps: acquiring a dual-mode channel data set of an HPLC mode and an HRF mode; on the basis of the data set, heterogeneous channel characteristics representing physical characteristics of the two modes respectively are extracted; calculating a performance difference degree on the basis of heterogeneous features in combination with a service spectrum weight; based on the dual-mode channel data set and the heterogeneous channel characteristics, determining the switching benefit of switching between the HPLC mode and the HRF mode; and generating a switching control signal according to the performance difference degree and the switching income. According to the invention, the accuracy and robustness of the switching decision are improved by introducing the depth perception of the channel spectrum shape and the causal income evaluation of the switching behavior.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of communication, and in particular to a HPLC and HRF switching method based on dual-mode channel difference perception. BACKGROUND

[0002] The dual-mode fusion technology of high-frequency power line carrier communication (HPLC) and high-speed radio frequency communication (HRF) provides a key technical path for building ubiquitous and high-resilience communication networks by complementing advantages and intelligently switching between two physically different channels. Therefore, researching an efficient, accurate and stable dual-mode switching method has important theoretical and engineering significance for ensuring the continuity of key services and improving the quality of communication services in complex environments.

[0003] Currently, the existing research on switching technology between heterogeneous networks mainly focuses on decision mechanisms based on macro physical layer parameters such as received signal strength (RSSI) and signal-to-noise ratio (SNR). These methods usually set fixed switching thresholds and supplement certain hysteresis (Hysteresis) time or switching counters to avoid excessive frequent ping-pong switching. In some schemes, simple considerations of business types or load states are also introduced, such as setting different switching thresholds for different businesses. In addition, some research also begins to use fuzzy logic or simple machine learning models to weight and fuse multiple input parameters in order to obtain a smoother and more robust decision boundary than single threshold decision, so as to adapt to the dynamic changes of the channel environment.

[0004] However, the existing technology still has deep technical problems when dealing with the dual-mode scene of HPLC and HRF with great differences in physical characteristics, mainly reflected in the superficiality of the decision basis and the relevance rather than causality of the decision logic. These problems together lead to suboptimal switching decisions, limiting the full potential of dual-mode systems. SUMMARY

[0005] The application aims to provide a HPLC and HRF switching method based on dual-mode channel difference perception to solve the above problems existing in the prior art.

[0006] Technical scheme, the HPLC and HRF switching method based on dual-mode channel difference perception, comprising:

[0007] Obtaining a dual-mode channel data set of high-frequency power line carrier communication (HPLC) mode and high-speed radio frequency communication (HRF) mode;

[0008] Based on the dual-mode channel data set, extracting a first channel feature for representing the channel characteristics of the HPLC mode and a second channel feature for representing the channel characteristics of the HRF mode to form heterogeneous channel features.

[0009] calculate a performance difference degree between the first channel feature and the second channel feature based on the heterogeneous channel feature;

[0010] determine a switching benefit of switching between the HPLC mode and the HRF mode based on the dual-mode channel dataset and the heterogeneous channel feature;

[0011] generate a switching control signal for switching between the HPLC mode and the HRF mode according to the performance difference degree and the switching benefit.

[0012] Beneficial effects, the present application improves the accuracy and robustness of switching decision by introducing deep perception of channel spectrum shape and causal benefit evaluation of switching behavior. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A step flowchart of a dual-mode channel difference perception-based HPLC and HRF switching method provided for an embodiment of the present application.

[0014] Figure 2 A step flowchart of extracting a second channel feature provided for an embodiment of the present application.

[0015] Figure 3 A step flowchart of determining a switching benefit provided for an embodiment of the present application.

[0016] Figure 4 A step flowchart of generating a switching control signal provided for an embodiment of the present application. DETAILED DESCRIPTION

[0017] In order to make the personnel in the art better understand the present application scheme, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0018] It should be noted that the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units need not be limited to those clearly listed steps or units, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0019] In the research, it is found that the existing method relies on scalar indicators such as RSSI or SNR, which are essentially coarse-grained, energy-dimension measurements of channel quality, ignoring the shape information of channel spectral structure, leading to the system's inability to distinguish between two performance degradation modes: one is the overall lifting of wideband background noise, and the other is the appearance of narrowband strong interference in the key business frequency band. Although these two modes may correspond to similar SNR values, their impact on the quality of service (QoS) of specific services (such as video streaming) is completely different. Based on this shape-blind difference perception, it is easy to make incorrect mode selection. In addition, the existing decision logic is usually based on direct comparison of the instantaneous performance indicators (such as estimated channel capacity) of the two modes, implicitly assuming that the performance difference is entirely caused by the superiority or inferiority of the channel state. However, in reality, performance is the result of the combined action of multiple factors such as channel, business load, and interference. This simple correlation comparison cannot isolate the net causal gain brought by the switching action itself, and thus cannot accurately answer the core question of how much benefit a single switch can bring, leading to the switching benefit being severely overestimated or underestimated.

[0020] As shown in Figure 1 , a dual-mode channel difference perception-based HPLC and HRF switching method is proposed, including the following steps:

[0021] A dual-mode channel data set of high-frequency power line carrier communication (HPLC) mode and high-speed radio frequency communication (HRF) mode is obtained.

[0022] In other words, a dual-mode channel data set of high-frequency power line carrier communication (HPLC) mode and high-speed radio frequency communication (HRF) mode is obtained, and the dual-mode channel data set is preprocessed to obtain a preprocessed dual-mode channel data set; wherein the preprocessing includes time synchronization, noise removal, and amplitude normalization.

[0023] In the embodiment, the dual-mode channel dataset refers to a time series dataset that is synchronized and pre-processed and can reflect the channel state of two communication modes. The process of obtaining the dataset, specifically, includes parallel collection of the baseband complex signal of the HPLC transceiver and the down-converted baseband signal of the HRF transceiver. Exemplarily, the sampling rate of the HPLC signal can be set to 2.4 MHz, and the sampling rate of the HRF signal can be set to 4 MHz. In order to ensure the accuracy of subsequent difference analysis, a key link is the synchronization and preprocessing of the signal. In a specific implementation, it can further include: using a GPS clock or a network time protocol (NTP) to accurately timestamp the dual-mode original signal, and aligning the signals of different sampling rates to a unified time reference through an interpolation algorithm to realize signal synchronization; and applying a Wiener filter to process the synchronized signal to effectively suppress channel noise and perform amplitude normalization. In this way, a high-quality pre-processed dual-mode channel dataset that can be used for subsequent analysis is obtained, which is a solid foundation for realizing channel difference perception.

[0024] Based on the dual-mode channel dataset, a first channel feature for characterizing the HPLC mode channel characteristics and a second channel feature for characterizing the HRF mode channel characteristics are extracted to constitute the heterogeneous channel features.

[0025] In the embodiment, the heterogeneous channel features refer to the extracted features that are essentially different in physical dimension and statistical characteristics because they are derived from two channels with different physical media and interference mechanisms. The core characteristics of the HPLC channel are affected by the periodic load change of the power grid, which is manifested as deterministic harmonic interference; while the core characteristics of the HRF channel are affected by the reflection and scattering of the physical environment, which is manifested as random multipath fading. Therefore, the original signal can be converted into a structured feature vector that can accurately depict the respective core physical characteristics. Specifically, the first channel feature can be a power grid harmonic feature vector representing power grid harmonic interference, and the second channel feature can be a multipath fading feature spectrum or a continuous time-delay-Doppler parameter pair representing multipath effect and mobility.

[0026] Based on the heterogeneous channel features, the performance difference degree between the first channel feature and the second channel feature is calculated.

[0027] Specifically, the performance difference degree is used to measure the performance gap between the two communication modes under the current channel state and service demand. Exemplarily, the calculation here is not a simple signal strength comparison, but a weighted, shape-sensitive comparison in the frequency domain combined with service demand. By calculating the difference degree, the system can perceive which mode's channel spectrum shape is more suitable for the current transmission needs of the service, thereby avoiding misjudgment due to fluctuations in a single indicator (such as RSSI).

[0028] Based on the dual-mode channel dataset and heterogeneous channel characteristics, determine the switching benefit of switching between the HPLC mode and the HRF mode.

[0029] Optionally, the switching benefit not only quantifies the performance gain that can be obtained after switching to the target mode, but also takes into account the cost of performing the switch itself. In a specific implementation, the performance gain can be determined by evaluating the channel capacity under the two modes or predicting the throughput improvement brought by the switch using a causal inference model; while the switching cost needs to be dynamically modeled, considering factors such as the basic switching delay, the cost of emptying the device cache of pending data, and the potential retransmission overhead. By balancing the benefits and costs, each switching decision can be profitable, thereby avoiding unnecessary switching overhead.

[0030] According to the performance difference degree and the switching benefit, a switching control signal is generated for switching between the HPLC mode and the HRF mode.

[0031] In this embodiment, the upstream analysis results - performance difference degree and switching benefit are taken as the core input, and a final binary switching instruction is generated through an intelligent decision mechanism. Optionally, the decision mechanism is not a simple static threshold decision, but can be an adaptive and statistically guaranteed decision process. For example, a decision score can be constructed that integrates multiple dimensions such as difference degree, benefit, cost, and channel complementarity, and historical data can be used to dynamically calibrate the decision boundary, thereby ensuring the accuracy of switching while effectively suppressing ping-pong effects and improving the overall stability and transmission efficiency of the system.

[0032] This embodiment realizes intelligent and efficient communication mode switching by deeply perceiving and differentiating the channel characteristics of high-frequency power line carrier communication (HPLC) and high-speed radio frequency communication (HRF).

[0033] According to one aspect of the present application, the performance difference degree is calculated, including:

[0034] The first power spectral density and the second power spectral density are derived from the first channel characteristics and the second channel characteristics, respectively.

[0035] Specifically, the first power spectral density is denoted as SHPLC(f), which can be directly generated from the power grid harmonic feature vector, where f represents the frequency. The second power spectral density is denoted as SHRF(f), which can be reconstructed from the HRF fading feature spectrum or multipath parameters.

[0036] Obtain the service spectrum weight associated with the current service type.

[0037] In the embodiment, the service spectrum weight is denoted as Wf(f), which is a function varying with frequency, and its value reflects the contribution or importance of different frequency bands to the user experience quality (QoS / QoE) under the current service type. In a possible implementation, the step of obtaining the service spectrum weight can be a rule matching method based on the service type. For example, the system first analyzes the service type of the current data flow. If it is a real-time control service, a spectrum weight function Wf(f) emphasizing the stability of low frequency is generated, such as an exponential decay function Wf(f) = exp(-f / f critical ), where f critical is the key frequency constant of the service. If it is a video service, a weight function with band-pass characteristics is generated, focusing on the frequency band where the main energy of the video stream is located. If it is a normal data service, a uniform weight Wf(f) = 1 can be used.

[0038] In a preferred implementation, in order to realize dynamic adaptation of the weight, a method of learning by reverse deduction from historical quality of service (QoE) logs can be used. Specifically, obtaining the service spectrum weight includes: accessing historical QoE logs, historical power spectrum logs, and service scenario labels associated with the two; analyzing the historical QoE logs and the historical power spectrum logs for a specific service scenario label; based on the analysis result, establishing a mapping relationship from the power spectrum feature to the QoE index, and based on the mapping relationship, reverse deducing the service spectrum weight representing the importance of different frequency bands. For example, a semi-supervised regression model with monotonic and sparse constraints can be constructed to associate QoE indicators such as stall time and throughput with power allocation of different frequency bands, so as to learn the service spectrum weight Wf(f).

[0039] The preset distribution metric distance is used in combination with the service spectrum weight to quantify the shape difference between the first power spectrum density and the second power spectrum density, and a performance difference degree is generated.

[0040] Specifically, the traditional Euclidean distance (or two-norm difference) mainly measures the energy difference, and is not sensitive to the structure and shape of the spectrum. For example, the interference of one channel is a single narrowband strong interference, and the interference of another channel is a wideband weak noise. The energy difference between the two can be very small, but the impact on the service is very different. Therefore, the embodiment preferably uses a distribution metric distance which is more sensitive to the shape. In a preferred implementation, the distribution metric distance can be a weighted Wasserstein-1 distance (also known as Earth Mover's Distance, EMD). At this time, the calculation process of the performance difference degree D(t) is as follows: the power spectrum is subjected to dimension alignment to better match the human perception characteristics: let X k = log( SHPLC(f k ) +ε), Y k= log( SHRF(f k ) +ε); wherein k is a discrete frequency index; f k is the frequency of the kth frequency point; ε is a very small positive number to avoid taking log of zero; SHPLC(f k ) is the power spectral density at frequency point f k under high frequency power line communication (HPLC) mode; SHRF(f k ) is the power spectral density at frequency point f k under high frequency radio frequency communication (HRF) mode; X k is the value after dimension alignment of the power spectrum under HPLC mode; Y k is the value after dimension alignment of the power spectrum under HRF mode. Discrete cumulative distribution functions (CDFs) of the logarithmic power spectrum are calculated respectively: CDF X (k) and CDF Y (k). A weighted EMD is calculated as the performance difference D(t), and the calculation formula is: D(t) = (Σ k Wf(f k ) * |CDF X (k) - CDF Y (k)| *Δf ) / (Σ k Wf(f k ) *Δf ); wherein D(t) is the performance difference at time t; Σ k indicates summation over all discrete frequency points k; Wf(f k ) is the service spectrum weight; |...| indicates taking absolute value; Δf is the frequency interval of the discrete frequency points; CDF X (k) is the discrete cumulative distribution function of the logarithmic power spectrum sequence {X k}; CDF Y (k) is the discrete cumulative distribution function of the logarithmic power spectrum sequence {Y k}. The calculation formula of the performance difference D(t) can be understood as the minimum weighted total cost of moving the distribution form of one power spectrum (regarded as a pile of sand) to the distribution form of another power spectrum, and thus is very sensitive to the shape information such as the position and width of the spectral peak.

[0041] In another optional implementation, the distribution metric distance can be a weighted Itakura-Saito divergence, and the calculation formula is: D(t) = (Σ k Wf(f k ) * ((SHPLC(f k )+ε) / (SHRF(f k )+ε) - log((SHPLC(f k) + ε) ) / (Σ k ) + ε) ) / (Σ k ) + ε) ) / (Σ k ) ). This divergence can effectively measure the shape difference between two power spectra as well.

[0042] Further, in order to enhance the foresight of decision-making, an autoregressive (AR) model can also be applied to the calculated performance difference D(t) time series for short-term prediction. The predicted difference value of one or more future time slots obtained in this way can be used as an important basis for subsequent determination of the optimal switching time, so that the switching decision is changed from passive response to active planning.

[0043] The present embodiment no longer relies on a single energy indicator such as RSSI, but introduces a deep perception of channel spectral shape. By using a shape-sensitive distribution metric such as weighted Wasserstein-1 distance, and combining the service spectral weight learned from user experience (QoE), the real impact of different degradation modes such as wideband noise and narrowband interference on the key service frequency band can be accurately distinguished, so that the difference perception is no longer blind.

[0044] In an exemplary embodiment, the switching gain can be determined based on a direct evaluation of the channel capacity under two modes, and subtracting the cost brought by the switching behavior itself. The channel capacity here refers to the maximum information rate of error-free transmission in theory under given channel conditions. Specifically, it includes: calculating the HPLC channel mutual information I HPLC (t) considering the power grid harmonic interference. In a specific calculation example, the channel is divided into multiple sub-channels, and the total capacity is the sum of the sub-channel capacities, which can be expressed as: C HPLC (t) = Σ i B i i log2(1 + SINR HPLC_i (t)) ; wherein C HPLC (t) is the total capacity of the HPLC channel at time t; i is the sub-channel index; B i is the bandwidth of the i-th sub-channel; log2 represents the logarithm with base 2; SINR HPLC_i (t) is the signal-to-interference-plus-noise ratio of the i-th sub-channel at time t. It should be particularly noted that the calculation of SINR HPLC_i (t) fully considers the unique properties of the HPLC channel, and its expression is: SINR HPLC_i (t) = P signal_i / (P noise_i + Σ k α k × Ak(t) 2 ) ; wherein Psignal_i P is the signal power; P noise_i Ak(t) is the amplitude of the kth harmonic obtained from the first channel characteristics; a k is the interference coefficient of the kth harmonic to the current sub-channel. The harmonic interference is precisely quantified and incorporated into the capacity evaluation. The HPLC channel mutual information I HPLC (t) is the HPLC channel total capacity. With the HRF fading characteristic spectrum SHRF(f, t), the water-filling algorithm is used to optimize the power allocation P(f) to maximize the channel capacity. The channel mutual information I HRF (t) is calculated as: I HRF (t) = ∫ log2(1 + P(f) x SHRF(f, t) / N0) df; where P(f) is the optimized power allocation function, satisfying the total power constraint ∫P(f)df = P total ; N0 is the channel noise power spectral density; df is the frequency integral variable; P total is the total power constraint. The instantaneous switching gain G(t) is defined as the difference between the expected information transmission gain in the remaining transmission time T remain (t) and the dynamic switching cost C switch (t), i.e.: G(t) = [I HRF (t) - I HPLC (t)] x T remain (t) - C switch (t).

[0045] In a preferred embodiment, to more accurately quantify the net causal effect of the mode switching intervention behavior on the communication performance, and to avoid the confounding bias that may be introduced by traditional correlation analysis (for example, the traffic burst may both cause the throughput to rise and trigger the switching decision), a causal inference-based method can be used to determine the switching gain. Specifically, as shown in Figure 3 , determining the switching gain includes:

[0046] Based on the pre-stored historical performance data, the dual-mode channel data set, and the heterogeneous channel characteristics, an anti-factual data set for causal analysis is constructed.

[0047] In this embodiment, the counterfactual dataset is a collection of data used for model training, which pairs each decision moment in history with a virtual sample assuming the opposite decision was made, so that the model can learn the pure impact of the decision behavior itself. For example, the construction process is as follows: records containing channel features (such as difference D(t), complementarity C(t)), link statistics (such as cache depth, retransmission rate), and corresponding handover decisions (handover / non-handover) and results (throughput, delay) are extracted from the historical database. For each sample that actually occurred handover (as the intervention group), one or more samples under similar channel and traffic conditions but without handover (as the control group) are found in the historical data by feature similarity matching (such as K-Nearest Neighbor algorithm), and vice versa. A large number of historical data points can be constructed to form a counterfactual dataset of semi-synthetic counterfactual paired samples.

[0048] According to the counterfactual dataset, a pre-configured causal inference model is trained, which is used to quantify the net causal effect of mode switching intervention on communication performance.

[0049] Optionally, the causal inference model can adopt a two-model structure (Two-Model Approach, or T-Learner), including: a first model component configured to learn and predict communication performance under handover operation based on intervention group data corresponding to performing mode switching in the counterfactual dataset. This model (such as a deep neural network or gradient boosting machine) learns the mapping relationship M treat from channel / traffic features to communication performance indicators (such as throughput Thr treat ). A second model component is configured to learn and predict communication performance under non-handover operation based on control group data corresponding to not performing mode switching in the counterfactual dataset. This model learns the performance mapping relationship M control under the same features, and predicts the throughput Thr control .

[0050] Using the trained causal inference model, combining the current dual-mode channel dataset and heterogeneous channel features, the expected performance gain is estimated, and the performance gain is used as the switching benefit.

[0051] In this embodiment, the performance gain is determined according to the prediction outputs of the first model component and the second model component. Specifically, when making real-time decisions, the channel and traffic feature vector X t of the current time is input into the two trained model components at the same time. The net causal effect (Causal Uplift) of switching, i.e. the switching benefit G(t), is determined by the difference between the prediction outputs of the two models: G(t) = M treat (X t) - M control (X t The G(t) value directly quantifies how much more throughput or latency improvement is expected to be achieved by performing a switch compared to not performing a switch under the current conditions. It is a more accurate benefit assessment than the traditional capacity difference method.

[0052] Optionally, the switching benefits also include the switching cost. In one possible implementation, the switching cost C switch (t) can be modeled as: C switch (t) = C base + β×B buffer (t) + γ×R retrans (t); where C base The fixed overhead based on the handover latency (e.g., 5ms); B buffer (t) represents the amount of data to be transmitted in the current device buffer; R retrans (t) represents the expected retransmission rate that may result from the handover; β and γ are weighting coefficients.

[0053] In a preferred implementation, a risk-based modeling approach is employed to make cost assessment more risk-aware. Specifically, determining the handover cost includes: obtaining the device buffer state reflecting the current amount of data to be transmitted; and assessing the transmission risk metric associated with the mode switch. For example, failure events are defined as performance degradation, connection interruption, or the need for immediate revert after handover. A conditional risk function (such as a survival analysis model) is built based on historical data to predict the failure probability Risk(t) of performing the handover under the current channel / service conditions. The handover cost is dynamically calculated based on the device buffer state and the transmission risk metric. The final handover cost C is then calculated. switch (t) is modeled as the sum of basic overhead and expected risk loss: C switch (t) = C base + E[Interruption time| Risk(t)] + E[Retransmission loss| Risk(t)], where E[ ] is the expectation function.

[0054] This embodiment, by constructing a counterfactual dataset and training a causal inference model, elevates the evaluation of switching benefits from simple performance metric comparisons to a level that quantifies the net causal gain brought about by the switching intervention itself. This allows the system to isolate the interference of confounding factors, answering the core question of how much benefit a single switch actually brings, and solving the problem of overestimating or underestimating benefits.

[0055] According to an aspect of the present application, the decision process for generating the handover control signal can be based on a fuzzy logic controller. The hard handover boundaries are fuzzified to handle the uncertainty when the channel quality fluctuates around the handover threshold, so as to suppress the ping-pong effect. Specifically, the process includes: fuzzifying the performance difference D(t) as an input, i.e. calculating its membership degree for a pre-defined fuzzy set. For example, three fuzzy sets can be defined based on the performance difference D(t): HPLC advantage, performance similar, and HRF advantage, and membership functions are designed for them, such as: HPLC (D) = 1 / (1+exp(5×(D-0.3))) ; μ equal (D) = exp(-10×(D-0.5) 2 ) ; μ HRF (D) = 1 / (1+exp(-5×( D-0.7))). Through a series of pre-defined fuzzy logic rules (IF-THEN rules), the multiple fuzzified inputs (such as the membership degree of D(t), the handover gain G(t), the channel complementarity C(t), etc.) are fused to perform fuzzy reasoning, and finally a quantized handover decision confidence Conf switch is generated. The calculated confidence Conf switch is compared with a pre-defined threshold (such as 0.8), and when Conf switch is greater than the threshold, the handover control signal Switch cmd = 1 is generated.

[0056] According to another aspect of the present application, in order to make the handover decision have strict statistical performance guarantee, instead of relying on empirical rules and thresholds, a calibration method based on conformal prediction can be used. Specifically, as shown in Figure 4 , generating the handover control signal includes:

[0057] Based on the performance difference and the handover gain, a non- consistency score for measuring the risk of the handover decision is constructed.

[0058] In the embodiment, the non-consistency score is a risk measure that fuses multiple decision factors, and measures the degree of deviation of the current state from the state suitable for keeping the current mode.

[0059] In a possible implementation manner, constructing the non-consistency score includes:

[0060] Based on the heterogeneous channel features, the channel time-frequency complementarity is determined.

[0061] Specifically, the channel time-frequency complementarity C(t) can quantify the potential of the two channels to cancel each other out in the frequency domain. Preferably, it is calculated as follows: calculate the cross-spectrum of the HPLC and HRF power spectrum; calculate the normalized cross-spectrum coherence function Coh(Δτ) within a preset finite time shift set Δτ, which is defined as: Coh(Δτ) = (Σ k SHPLC(f k ) * SHRF(f k ) * exp(-j2πf k Δτ)) / (sqrt(Σ k SHPLC(f k ) 2 ) * sqrt(Σ k SHRF(f k ) 2 )); take the negative of the peak value of the function as the complementarity indicator C(t), i.e. C(t) = max Δτ (-Re{Coh(Δτ)}) whose value range is [0, 1], and the greater the value, the stronger the complementarity.

[0062] According to a preset function, the non-uniformity score Score(t) is calculated by fusing the switching cost C switch (t), the switching gain G(t) and the channel time-frequency complementarity C(t).

[0063] For example, the calculation formula of the non-uniformity score can be: Score(t) = C switch (t) - G(t) - α *C(t); where α is an adjustable weight coefficient. The non-uniformity score can be understood as: when the switching cost is high, the gain is low, and the complementarity of the two channels is also low, the value of Score(t) is high, indicating that the risk of switching is high.

[0064] According to the pre-stored historical switching records, the non-uniformity score is subjected to conformal calibration to generate a decision threshold band with a preset confidence level.

[0065] In this embodiment, the conformal calibration can provide effective confidence guarantee for prediction. The specific process is as follows: use a calibration data set containing a large number of historical non-uniformity scores, which all correspond to known correct decisions (for example, it is correct to keep the current mode). Calculate the η quantile of the scores in this calibration data set, denoted as q η . For example, if η = 0.95, then q η is the 95th largest value in this set of historical scores. q ηThe current non-uniformity score is compared with the decision threshold band to determine whether to generate a switching control signal.

[0066] The current non-uniformity score is compared with the decision threshold band to determine whether to generate a switching control signal.

[0067] Specifically, when making a real-time decision, if the current calculated non-uniformity score Score(t) is lower than the decision threshold q η obtained through conformal calibration, it means that the current state is significantly different from the historical state that should be maintained, and thus a decision can be made with a confidence of no less than (1-η) that switching should be performed, and a switching control signal is generated. In order to avoid ping-pong switching caused by signal jitter near the decision boundary, the comparison here preferably introduces a hysteresis mechanism. Specifically, the decision threshold band is converted into a hysteresis interval with different upper and lower thresholds. Only when Score(t) crosses the lower threshold explicitly is switching triggered, and reverse switching is suppressed until Score(t) rises to the upper threshold explicitly, thereby enhancing the stability of the decision.

[0068] According to another aspect of the present application, the optimal execution timing can also be determined before the switching signal is generated. Specifically, generating the switching control signal can also include: predicting a performance difference degree sequence and a switching benefit sequence in a future time domain based on the current performance difference degree and the switching benefit; applying an optimization algorithm to analyze the performance difference degree sequence and the switching benefit sequence to determine an optimal switching timing; and generating the switching control signal based on the optimal switching timing.

[0069] In a preferred implementation, the optimization algorithm can be an approximate exponential strategy (derived from the Restless Bandit model), which can consider the expected future benefits, opportunity costs and switching costs within a rolling time window to solve the optimal switching time t opt that maximizes the total benefits in the future. The final switching control signal will be generated and issued based on the system state at t opt .

[0070] According to another aspect of the present application, generating the switching control signal can also include: determining channel time-frequency complementarity based on heterogeneous channel characteristics; and cooperatively processing the performance difference degree, the switching benefit and the channel time-frequency complementarity according to a preset fusion rule to generate the switching control signal.

[0071] In an embodiment of the present application, the traditional short-time Fourier transform (STFT) based method has a fixed time-frequency resolution contradiction, which is difficult to accurately capture the frequency and amplitude of harmonics at the same time, and therefore an adaptive harmonic tracking method is proposed to realize efficient and accurate extraction of power grid harmonic characteristics. Specifically, the first channel feature is extracted, including:

[0072] The HPLC baseband signal is obtained from the dual-mode channel data set, and the HPLC baseband signal is processed to track the power grid fundamental frequency in real time.

[0073] In the present embodiment, since the reference frequency (nominal 50Hz or 60Hz) of the power grid is not constant, it will produce a slight drift due to the real-time change of the power grid load, and if the fixed nominal frequency is still used in subsequent analysis, it will cause serious spectral leakage and amplitude estimation error. Therefore, real-time and accurate tracking of the power grid fundamental frequency is the premise of all subsequent harmonic analysis. Specifically, it includes: extracting the real part signal from the pre-processed HPLC baseband signal I HPLC (t)+jQ HPLC (t), where j is the imaginary unit, Q HPLC (t) is the imaginary part of the HPLC baseband signal at time t; through the zero-crossing detection algorithm, the number of signal zero-crossing points N measure is counted within the measurement time T cross , so as to calculate the instantaneous power grid frequency f grid (t)=N cross / (2×T measure ); in order to smooth the jitter of the instantaneous frequency, a Kalman filter is introduced to filter the frequency estimation value sequence, and a stable and accurate power grid fundamental frequency estimation value f grid_filtered (t) is output.

[0074] Based on the power grid fundamental frequency, the frequencies of a predetermined number of harmonics are adaptively predicted.

[0075] Specifically, due to the existence of nonlinear devices in the power system, the actual center frequency of the harmonics is also not a strict integer multiple of the fundamental frequency, and there is also a frequency shift. Optionally, based on the accurately tracked power grid fundamental frequency, the actual frequency of each harmonic is adaptively predicted. In a specific implementation, the frequencies of a predetermined number of harmonics are adaptively predicted, including: applying a pre-configured autoregressive model, learning and predicting the harmonic frequency shift based on pre-stored historical frequency data; combining the real-time tracked power grid fundamental frequency and the harmonic frequency shift to determine the predicted value of the predetermined number of harmonic frequencies. More specifically, for each harmonic order k (for example, k from 1 to 20), the predicted harmonic frequency f k (t) is superimposed by k times of the fundamental frequency and the predicted frequency shift Δf k (t), that is, f k (t)=k×fgrid_filtered (t)+Δf k (t); where the frequency offset Δf k (t) can be learned and predicted from the historical frequency offset data of this harmonic through a first-order autoregressive model. Finally, a prediction vector F containing all the harmonic frequencies to be analyzed is generated. harm (t)=[f1(t), f2(t), ..., f 20 (t)].

[0076] Construct a parallel Goertzel filter bank and dynamically adjust the filter parameters of the parallel Goertzel filter bank according to a predetermined number of harmonic frequencies.

[0077] Specifically, the Goertzel algorithm has a computational complexity of O(N), far lower than the O(NlogN) of the Fast Fourier Transform (FFT) for calculating the entire spectrum. This embodiment utilizes this to construct a dedicated Goertzel filter for each harmonic frequency to be analyzed, forming a parallel filter bank, thereby achieving targeted and efficient extraction of harmonic energy. Specifically, for the harmonic frequency prediction vector F... harm Each frequency component f in (t) k (t), which updates the core coefficients coeff of the corresponding Goertzel filter in real time. k (t)=2cos(2πf k (t) / f s ), where f s The signal sampling rate is used. In this way, a second-order IIR filter H consisting of 20 parallel, dynamically adaptive parameters is constructed. k (z)=1 / (1-coeff k (t)×z -1 +z -2 A filter bank consisting of z, where z represents the frequency domain representation of the discrete-time signal. This ensures that each filter is always precisely aligned to the real-time frequency of its target harmonic.

[0078] The HPLC baseband signal was processed using a dynamically adjusted parallel Goertzel filter bank to extract the grid harmonic feature vector, which was then used as the first channel feature.

[0079] In this embodiment, the preprocessed HPLC baseband signal is simultaneously fed into a parallel filter bank for filtering, with each filter H... k (z) will output the complex harmonic component X at its corresponding harmonic frequency. k (t). Calculate the amplitude A of each complex harmonic component. k_raw (t)=|X k(t) | / N, where N is the number of processed sample points. Since there can be spectral leakage between adjacent harmonics, in order to improve the accuracy of the amplitude estimation, preferably, a leakage compensation mechanism can be introduced. Specifically, a leakage compensation matrix L is constructed, where the element L {k，m} = |sinc(π(f k -f m )T) | represents the coefficient of the mth harmonic leaking to the kth harmonic, and T is the time window length used for spectral analysis. By solving the linear equation set A raw = L × A true , the compensated real harmonic amplitude vector A true can be obtained, where A raw is the original harmonic amplitude vector without leakage compensation. This compensated harmonic amplitude vector constitutes the final power grid harmonic feature vector H HPLC (k, t) = [A1(t), A2(t),..., A 20 (t)].

[0080] In another embodiment of the present application, since the traditional spectral analysis method does not utilize the priori information that the harmonic components in the HPLC signal have typical cyclostationary characteristics, i.e., their statistical characteristics change periodically with time, therefore, by introducing cyclo-spectral analysis, the periodic harmonic signal can be more robustly separated from the stationary noise, and combined with continuous parametric spectral line estimation, super-resolution feature extraction can be realized. Specifically, the first channel feature can also be:

[0081] Obtain the HPLC baseband signal from the dual-mode channel data set, and calculate the spectral correlation function of the HPLC baseband signal for a preset set of cyclic frequencies.

[0082] In this embodiment, the spectral correlation function (Spectral Correlation Function, SCF) is denoted as SCf(f, a), which is a two-dimensional function revealing whether there is a correlation with a period of a cyclic frequency a between the spectral components of the signal at different frequencies f. For the HPLC signal, its set of cyclic frequencies is the set of the fundamental frequency and its harmonic frequencies of the power grid. Only the periodic signal components (such as power harmonics) will show non-zero correlation at these specific cyclic frequencies, while the spectral correlation function of the stationary random noise is theoretically zero. Specifically, on the preset set of cyclic frequencies (for example, 50Hz, 100Hz, 150Hz...), the spectral correlation function of the preprocessed HPLC time domain signal is calculated, thereby upgrading the analysis of the signal from one-dimensional power spectrum to two-dimensional cyclic spectrum domain.

[0083] Based on the spectral correlation function, a cycle-enhanced spectrum is derived, which enhances the periodic harmonic components in the HPLC baseband signal.

[0084] In this embodiment, a one-dimensional power spectrum with significantly enhanced harmonic components is generated by using a two-dimensional spectral correlation function. In a specific implementation, the cycle-enhanced spectrum HPLC is obtained by integrating or summing the amplitude peaks of the spectral correlation function at each harmonic corresponding to the cycle frequency a, and then mapping the energy back to the one-dimensional power spectrum coordinates after normalization. _循环增强谱_SCHPLC (f) Since only the cycle-stationary harmonic signal contributes to this spectrum, the signal-to-noise ratio of the harmonic components is improved on this spectrum, while the background noise is effectively suppressed, thereby achieving enhancement of the periodic harmonic components.

[0085] Continuous parameter spectral line estimation is performed on the cycle-enhanced spectrum to resolve the continuous harmonic parameters with continuous frequency values and corresponding amplitudes.

[0086] In this embodiment, since the traditional discrete Fourier transform (DFT) has a fence effect, i.e., when the true frequency of the signal falls between two discrete frequency points, energy leakage and frequency estimation deviation will occur. Therefore, the continuous parameter spectral line estimation (also known as super-resolution spectral estimation) technique is adopted. Specifically, the Prony method or the ESPRIT (Estimation of Signal Parameters via Rotational Invariance Techniques) algorithm can be used to process the cycle-enhanced spectrum. These algorithms model the spectral line as a sum of multiple complex exponential signals, and can directly resolve the parameters of these exponential signals from the data, i.e., the harmonic center frequencies with continuous (rather than discrete) frequency values, and their corresponding complex amplitudes (including amplitude and phase information).

[0087] According to the continuous harmonic parameters, a power grid harmonic feature vector is generated, and the power grid harmonic feature vector is taken as the first channel feature.

[0088] In this embodiment, the super-resolution analysis result is converted into a standard format for use by other modules of the system. Specifically, the continuous amplitude values of the resolved harmonics are combined to form the final power grid harmonic feature vector. Alternatively, in order to be compatible with subsequent modules based on discrete frequency points (such as difference calculation), the estimated continuous frequencies can also be rasterized to discrete frequency points according to the principle of energy conservation to form a power spectrum or a harmonic feature vector. This embodiment can obtain more accurate and robust harmonic features than traditional methods, thereby improving the perception accuracy of the entire switching decision system.

[0089] In an exemplary embodiment, before extracting the second channel features, a delay-Doppler overcomplete dictionary can be constructed. This involves creating a set of basis functions, where each basis function (or atom) represents a potential multipath component with a specific delay and a specific Doppler shift. Through linear combinations of these atoms, arbitrarily complex channel impulse responses can be represented. Overcompleteness means that the number of atoms in the dictionary is much greater than the minimum dimension required to represent the channel, which provides the possibility for sparse representations of the channel. Specifically, the process includes: based on the channel bandwidth B and the maximum multipath delay τ... max Construct a discretized time-delay grid τ i (For example, the interval Δτ = 1 / B); simultaneously, based on the observation duration T obs Construct a discretized Doppler frequency shift grid ν j (For example, the interval Δν = 1 / T) obs Generate dictionary atom φ. {i，j} (τ,t)=sinc(B(τ-τ i ))×exp(j2πν j t). All these atoms φ {i，j} Together they form the overcomplete dictionary matrix Φ.

[0090] In a further embodiment, extracting the second channel features includes:

[0091] The HRF channel impulse response obtained from the dual-mode channel dataset is subjected to structured sparse decomposition to obtain a sparse coefficient vector characterizing the channel multipath components.

[0092] In this embodiment, the minimum number of values ​​that can most accurately reconstruct the currently measured HRF channel impulse response h is found from a large, overcomplete dictionary. HRF The atomic combination of (τ, t). The result is a sparse coefficient vector x, where the positions of the non-zero elements indicate which atoms are selected, and the values ​​of the non-zero elements represent the complex amplitudes of the corresponding multipath components. In a preferred implementation, structured sparsity prior information is introduced, since multipath components in the physical world often appear in clusters in the time-delay-Doppler domain, rather than being completely random and isolated. Therefore, the following optimization objective function is constructed: min ||h HRF - Φx|| 2 + λ||x||1 + γR struct (x); where h HRF The vectorized channel impulse response; Φ is the dictionary matrix; x is the sparse coefficient vector to be solved; ||...|| 2 Let L2 be the norm, representing the reconstruction error term; λ be the regularization coefficient; ||...||1 be the L1 norm, used to make the solution x sparse; γ be the structure regularization coefficient; R struct(x) is a structured regularization term, whose expression can be R. struct (x)=Σ i，j w ij |x i -x j |, weight w ij The design aims to encourage the coefficients of adjacent atoms (i.e., atoms with similar delay and Doppler characteristics) to converge, for example, w ij =exp(-|τ i -τ j | / τ c )×exp(-|ν i -ν j | / ν c ), where τ c and ν c τ is the correlation scaling factor; i and τ j ν represents the time delay positions of the i-th and j-th atoms; i and ν j Let x be the Doppler frequency shift positions of the i-th and j-th atoms. This optimization problem can be solved by algorithms such as the improved Orthogonal Matching Pursuit (OMP) algorithm or Basis Pursuit, thereby obtaining the sparse coefficient vector x.

[0093] Based on the preset stability criterion, the sparse coefficient vector is analyzed, and the multipath components are divided into a stable multipath set and a dynamic multipath set.

[0094] Specifically, after obtaining the sparse coefficient vector x, each non-zero element x k Each corresponds to an identified multipath component. By analyzing the behavior of these components over a period of time, their stability can be classified. For example, this process includes: calculating each non-zero sparse coefficient x within a preset time window. k The mean amplitude |x k | mean and amplitude variance var(x) k The classification is based on preset criteria. For example, one criterion could be: if |x k | mean > TH mean And var(x) k ) < TH var If so, the multipath component is marked as a stable multipath and included in the stable multipath set S. stable These components typically correspond to line-of-sight paths or fixed, strong-reflection paths and are reliable components of the channel. If |x k | mean > TH mean And var(x) k) >= TH var , the multipath component is labeled as dynamic multipath and is included in the dynamic multipath set S dynamic Such components usually correspond to moving reflectors or fast changing scattering paths, and are the main cause of fast fading. Here TH mean and TH var are preset mean and variance thresholds of the amplitude.

[0095] According to the sparse coefficient vector and the classification results of the multipath components, the HRF fading feature spectrum is generated, and the coherence bandwidth and the coherence time are determined to jointly constitute the second channel feature.

[0096] In this embodiment, based on the refined analysis results, the final channel feature is generated. Specifically, the channel parameters are treated differently for different types of multipath components. For example, the coherence bandwidth B c (t) for evaluating the channel frequency selective fading is mainly determined by the stable multipaths, and thus its calculation relies on the root mean square delay spread τ rms_stable , which is calculated only from the components in the stable multipath set S stable : B c (t) ≈ 1 / (5xτ rms_stable ). While the coherence time T c (t) for evaluating the channel time selective fading is determined by the fastest changing components in the channel, and thus its calculation relies on the maximum Doppler shift v max , which needs to be extracted from all the identified multipath components (including stable and dynamic): T c (t) ≈ 0.423 / v max . By taking the Fourier transform of the channel impulse response reconstructed from the sparse coefficient vector x and the dictionary Φ and taking the modulus, the HRF fading feature spectrum S HRF (f, t) can be generated. The system can obtain a more in-depth structured understanding of the HRF channel, and provide more abundant and physically meaningful channel features for subsequent handover decisions.

[0097] In another exemplary embodiment, as shown in Figure 2 , the second channel feature can also be extracted as:

[0098] For the HRF channel impulse response obtained from the dual-mode channel data set, a channel measurement model is constructed in the continuous delay domain and the continuous Doppler domain.

[0099] In this embodiment, the channel estimation problem is mathematically generalized from discrete domain to continuous domain. Instead of constructing a discrete dictionary matrix, an atom set A is defined, which contains all possible multipath components, each atom a(τ, v) is parameterized by a pair of continuous time delay τ and continuous Doppler shift v. The channel impulse response h is modeled as a weighted integral (or finite sum) of these continuous atoms, and the measurement model describes the relationship between the measurement value y and the true channel h, for example, y = M(h) + n, where M is the measurement operator and n is the noise.

[0100] Based on the channel measurement model, the continuous time-delay-Doppler parameter pairs representing the multipath components of the channel are recovered by solving an atom norm minimization problem.

[0101] Specifically, the atom norm is denoted as ||h|| A , which is a norm associated with the atom set A and can be regarded as a generalization of the L1 norm in continuous domain. The atom norm minimization problem is solved, i.e., solving the following convex optimization problem: min ||h|| A , with the constraint ||y - M(h)||2≤ε, where ε is the noise level. The solution h of the optimization problem can be represented as a linear combination of a sparse number of atoms in the atom set A. By solving this convex optimization problem (for example, by methods such as semi-definite programming (SDP)), the continuous parameters of the sparse multipath components, i.e., a set of continuous time-delay-Doppler parameter pairs (τ k , v k ) and their corresponding complex amplitudes c k , can be directly recovered without any pre-set discrete grid.

[0102] According to the recovered continuous time-delay-Doppler parameter pairs, the HRF fading feature spectrum is generated, and the coherence bandwidth and the coherence time are determined to jointly constitute the second channel feature.

[0103] Specifically, the coherence bandwidth B c (t) is calculated according to the root mean square delay spread calculated from the set of stable multipaths; the coherence time T c (t) is calculated according to the maximum Doppler shift found from all recovered multipath components. The HRF fading feature spectrum S HRF (f, t) is obtained by Fourier transforming the continuous channel response reconstructed from all recovered continuous multipath components. This embodiment can obtain a fine characterization of the HRF channel with super-resolution and no grid mismatch error, which improves the accuracy of the channel feature and provides a solid foundation for realizing higher performance handover decision.

[0104] Optionally, extracting the second channel feature further includes:

[0105] analyzing amplitude statistical properties of each multipath component characterized by a continuous delay-Doppler parameter pair; and, dividing the multipath components into a stable multipath set and a dynamic multipath set according to a preset classification criterion; wherein the second channel feature is generated based on the stable multipath set and the dynamic multipath set.

[0106] The operating object of the embodiment is the continuous parameter multipath components recovered by the meshless method. Specifically, within a preset time window, for each recovered multipath component k, the change of its complex amplitude c k is tracked, and the mean and variance of its amplitude are calculated. According to a preset classification criterion (for example, a threshold based on the mean and variance of the amplitude), the multipath components characterized by the continuous parameters are divided into a stable multipath set and a dynamic multipath set.

[0107] In an embodiment of the present application, it further comprises:

[0108] Performing mode switching according to the switching control signal.

[0109] In the embodiment, when the upstream decision module generates a switching control signal (for example, a binary flag with a value of 1), the signal is sent to the physical layer together with the related resource scheduling information for execution. In a specific implementation, if the system decision switches from HPLC to HRF, in addition to sending the switching instruction itself, the system will also generate a dual-mode resource pre-allocation scheme in advance according to the optimal switching time t opt . The scheme can include radio resource blocks (RBs), transmission power levels, and initial modulation and coding schemes reserved for HRF mode. It is sent to the transceiver device through the control channel, and at the same time, the system accurately records the time when the switching instruction is sent as the switching start time t switch_start for subsequent performance evaluation. This resource pre-allocation mechanism is the key to achieving low-latency, quasi-seamless switching.

[0110] Monitoring the post-switching performance metrics after the mode switching is completed.

[0111] In the embodiment, the actual effect of a switching operation can be objectively and quantitatively evaluated. Specifically, after the switching is completed, the system will collect a series of key performance indicators, exemplarily including: the actual switching delay Δt switch : by recording the time t switch_end when the switching is completed and calculating the difference between it and the switching start time t switch_start , i.e., Δt switch = t switch_end - t switch_start , which reflects the efficiency of the switching process, where t switch_start is the switching start time; the data interruption duration T interrupt: The impact of handover on traffic continuity is measured by estimating the number or duration of un-acknowledged (ACK) packets during handover; pre-and post-handover BER variation: The average BER before handover for a period of time before and the average BER after handover for a period of time after are recorded respectively to evaluate the actual improvement (or deterioration) of transmission quality due to handover.

[0112] The performance metrics after handover are used to update the historical performance database.

[0113] In this embodiment, an experience knowledge base is built to associate decision scenarios, actions taken, and final outcomes. Specifically, after each handover event is completed, a new record is generated and stored in the historical performance database. The record is a data tuple containing at least: {the performance metrics of this handover (such as At switch , T interrupt , BER after / BER before ), the channel and traffic characteristics at the moment of decision making (such as D(t switch ), G(t switch ), C(t switch ), and the decision outcome (e.g., switching from HPLC to HRF)}. This database is not only the data foundation for adaptive adjustment, but also an important data source for building counterfactual data sets.

[0114] Based on the updated historical performance database, the decision parameters for generating handover control signals are adjusted adaptively.

[0115] In this embodiment, accumulated historical experience is used to continuously optimize future decision-making behavior. In a specific implementation, this adaptive adjustment is reflected in multiple levels: slow adaptive fine-tuning of decision thresholds: the system can periodically analyze the historical performance database to calculate a comprehensive historical satisfaction index S perf , for example, S perf = exp(-T int_avg / T target ) × exp(-f switch / f target ), where T int_avg is the average interruption duration, f switch is the handover frequency, T target and f targetis the desired target value. Based on this satisfaction index, the key parameters in the decision model are fine-tuned with a small learning rate, e.g. adjusting the center point of the fuzzy logic membership function or the confidence level η of the conformal calibration threshold band. Fast response and handling of abnormal switching: In addition to the long-term slow optimization, the system also has a fast response mechanism. For example, if it is monitored that a certain switching is an abnormal switching (such as Δt switch > 20ms or BER after > 10x BER before ), the system will immediately trigger a temporary fallback or suppression mechanism. Specifically, the threshold of the switching decision can be temporarily and significantly increased (for example, the confidence level of conformal calibration is temporarily increased from 95% to 99.9%), so that when the channel environment is extremely unstable, the conservative strategy of not cutting or cutting randomly can avoid the system performance avalanche caused by continuous wrong decisions.

[0116] In further embodiments, in order to achieve more refined pre-allocation management, risk threshold controlled quota pre-warming can also be introduced, so that the pre-allocated resource quota is dynamically associated with the risk or confidence of the decision. Specifically, it includes: based on the switching risk C switch or the decision confidence Conf switch , a confidence-resource quota curve is established; pre-allocation (i.e. pre-warming) is triggered only when the decision confidence exceeds a certain starting threshold; and the amount of pre-allocated resources (such as the number of RBs, the size of the cache) is proportional to the confidence: when the confidence is low, only a small amount of core resources is reserved for rapid startup; when the confidence is very high, all resources to meet the expected traffic volume are fully reserved. The pre-allocation is upgraded from a binary decision of yes / no to a smooth, risk-matched analog decision.

[0117] Optionally, a combined option (Option) strategy can also be introduced, modeling the resource pre-allocation behavior as the purchase and exercise of an option. Specifically, it includes: when the system predicts that switching may be needed in the future, it does not immediately occupy resources, but first buys an option to use resources on the target mode, which corresponds to a light-weight resource reservation signaling. When the confidence Conf switch of the switching decision finally rises above the upper threshold, the option is automatically exercised and the resources are formally allocated and activated. If before the exercise, the channel conditions change and Conf switch drops below the lower threshold, the option is voided, the reservation signaling is canceled, and the actual allocation of resources is avoided. By introducing the exercise logic, the dynamic changes of the decision can be more flexibly handled, and the resource allocation ping-pong effect caused by decision oscillation can be effectively suppressed.

[0118] In one specific embodiment, assume that at a decision time t, the system needs to evaluate whether it should switch from the current working HPLC mode to the HRF mode. At this time, through the processing of the upstream module, the following initial parameters have been obtained: discrete power spectral density: assume that at 5 discrete frequency points k = {1, 2, 3, 4, 5}, the measured log power spectrum (unit: dBm) is respectively: log(SHPLC(f k )) = [-20, -25, -50, -55, -60] (reflecting the harmonic characteristics of HPLC energy concentrated in low frequencies); log(SHRF(f k )) = [-40, -42, -35, -45, -48] (reflecting the characteristics of HRF performing better in the middle frequency band). The current service is a video stream, and its service spectrum weight Wf(f k ) is determined as: [0.1, 0.2, 0.5, 0.2, 0.0] (highly concerned about the transmission quality in the middle frequency band). Through causal inference and risk modeling, the net switching benefit G(t) = 5.0 (unit: equivalent Mbps) has been calculated; the switching cost C switch (t) = 2.0 (unit: equivalent Mbps). The channel complementarity C(t) = 0.8. In the decision model parameters, the weight coefficient a in the non-consistency score is a = 1.0. Through historical data conformal calibration, the decision threshold q 0.95 at the confidence level η = 95% is obtained as -2.5. Calculate the performance difference D(t) using the weighted Wasserstein-1 distance method: log power spectrum vector (for simplicity, directly use the dBm value): X k = [-20, -25, -50, -55, -60]; Y k = [-40, -42, -35, -45, -48]. Calculate the cumulative distribution function (CDF) (for simplicity, no strict probability normalization is performed, only the calculation idea is shown): CDF X (k) = [-20, -45, -95, -150, -210]; CDF Y (k) = [-40, -82, -117, -162, -210]. Calculate the absolute value of the CDF difference: |CDF X - CDF Y | = [20, 37, 22, 12, 0]. Perform weighted summation (numerator part, assume Δf = 1): numerator = Σ k Wf(f k ) * |CDF X (k) - CDF Y(k) = (0.1 * 20) + (0.2 * 37) + (0.5 * 22) + (0.2 * 12) + (0.0 * 0) = 2.0 + 7.4 + 11.0 + 2.4 + 0 = 22.8. Calculate the sum of weights (denominator part): Denominator = ∑ k Wf(f k ) = 0.1 + 0.2 + 0.5 + 0.2 + 0.0 = 1.0; get the final performance difference: D(t) = 22.8 / 1.0 = 22.8. This value itself is a relative value, its trend of change and relationship with threshold are more meaningful. Calculate the inconsistency score Score(t): Score(t) = C switch (t) - G(t) - a * C(t). Substitute the initial parameters: Score(t) = 2.0 - 5.0 - 1.0 * 0.8 Score(t) = -3.8. This score is a risk measure, negative and large in absolute value, which intuitively indicates that the benefit of switching is much greater than the risk. Compare the currently calculated inconsistency score with the decision threshold obtained by conformal calibration. Current score Score(t) = -3.8; decision threshold q 0.95 = -2.5 at 95% confidence level; decision rule: if Score(t) < q 0.95 , then execute the switch with no less than 95% confidence. Comparison result: -3.8 < -2.5, condition is met. Since the currently calculated inconsistency score -3.8 is lower than the decision threshold -2.5 with 95% confidence guarantee, the system determines that the current state is different from the historical state that should maintain the HPLC mode. Therefore, the system will generate a switching control signal instructing the device to switch from the HPLC mode to the HRF mode.

[0119] In an optional embodiment, the power allocation of the HRF channel can also be based on the idea of Distributionally Robust Optimization (DRO). Specifically, distributional robustness means that the optimization goal of power allocation is no longer to maximize the channel capacity under a single predicted interference scenario, but to maximize the channel capacity in the worst case within an uncertainty set containing all possible interference scenarios. Specifically, it can include: based on the predicted value given by the interference prediction model, and the statistical distribution of its historical prediction error, construct an interference power spectrum distribution family (i.e. uncertainty set) with the predicted value as the center and a certain probability (for example, defined by the conformal quantile interval) as the radius. The water-filling algorithm is extended to the distributionally robust water-filling algorithm, and the objective function of the algorithm is changed to find the worst interference distribution in the uncertainty set, and in this worst case, maximize the channel capacity by adjusting the power allocation P(f). This problem can be solved by convex optimization theory to find an approximate closed-form solution or by a differentiable programming layer to efficiently solve. The power allocation scheme obtained in this way naturally contains a certain safety margin, which can effectively resist the interference fluctuations beyond the expectation, and improve the transmission reliability of the system in the real complex electromagnetic environment.

[0120] In another optional embodiment, the idea of end-to-end integrated optimization can also be introduced. Specifically, the differentiable programming layer technology in deep learning can be used to embed traditional, non-differentiable optimization algorithms (such as the water-filling algorithm) into a neural network, so that it becomes a differentiable module. A preferred implementation is a differentiable layer based on the KKT (Karush-Kuhn-Tucker) condition, including: constructing a prediction-optimization integrated network, the input of which is the current channel state information, and the output is directly the optimized power allocation vector; embed a KKT layer in the last one or several layers of the network. The role of this layer is to ensure that any output of the network strictly satisfies all the constraint conditions of the power allocation problem (such as total power constraint, spectrum mask constraint, etc.). In other words, this layer maps an unconstrained neural network output to a feasible solution space that satisfies all physical constraints. Since the entire network (including the KKT layer) is end-to-end differentiable, it can directly use the final performance indicator of the system (such as maximizing the channel capacity) as the loss function, and train it through the back propagation algorithm. The prediction module can perceive the impact of its prediction results on the final power allocation decision during the training process, and thus learn the optimal feature representation for the final goal. A feasible and optimal fast decision pipeline is realized, which improves the real-time performance and global optimality of the decision.

[0121] According to an aspect of the present application, a dual-mode communication switching device comprises:

[0122] The acquisition unit is configured to acquire a dual-mode channel data set;

[0123] The feature extraction unit is configured to extract, based on the dual-mode channel data set, a first channel feature representing channel characteristics of the first communication mode and a second channel feature representing channel characteristics of the second communication mode.

[0124] The difference calculation unit is configured to calculate, based on the first channel feature and the second channel feature, a performance difference.

[0125] The benefit determination unit is configured to determine, based on the dual-mode channel data set, the first channel feature and the second channel feature, a switching benefit.

[0126] The decision generation unit is configured to generate, according to the performance difference and the switching benefit, a switching control signal.

[0127] According to another aspect of the present application, a communication device comprises a dual-mode communication switching apparatus.

[0128] According to another aspect of the present application, a computer readable storage medium has a computer program stored thereon, and the computer program, when executed by a processor, implements the method according to any one of the above embodiments.

[0129] In an optional embodiment, in the dual-mode channel difference-aware based HPLC and HRF switching method, the method further comprises: calculating a weighted performance difference between the first channel feature and the second channel feature based on the heterogeneous channel features, wherein the weighted performance difference quantifies shape difference rather than absolute energy difference in the frequency domain by introducing a traffic spectrum weight function; and further calculating a channel time-frequency complementarity index based on the heterogeneous channel features, wherein the complementarity index determines the complementarity degree by analyzing the negative correlation characteristics of the two modes.

[0130] The above detailed the preferred embodiments of the present application, but the present application is not limited to the specific details in the above embodiments, and various equivalent transformations can be made to the technical solutions of the present application within the technical concept of the present application, and these equivalent transformations all belong to the protection scope of the present application.

Claims

1. A method for HPLC and HRF handover based on dual-mode channel diversity awareness, characterized in that, The method comprises the following steps: obtaining a dual-mode channel data set of a high-frequency power line carrier communication (HPLC) mode and a high-speed radio frequency (HRF) mode; based on the dual-mode channel data set, extracting a first channel feature for characterizing the HPLC mode channel and a second channel feature for characterizing the HRF mode channel to form a heterogeneous channel feature; based on the heterogeneous channel feature, calculating a performance difference between the first channel feature and the second channel feature; based on the dual-mode channel data set and the heterogeneous channel feature, determining a switching benefit of switching between the HPLC mode and the HRF mode; generating a switching control signal for switching between the HPLC mode and the HRF mode according to the performance difference and the switching benefit.

2. The method of claim 1, wherein, The extraction of the second channel feature comprises: constructing a channel measurement model on a continuous time delay domain and a continuous Doppler domain for the HRF channel impulse response obtained from the dual-mode channel data set; based on the channel measurement model, solving an atomic norm minimization problem to recover a continuous time delay-Doppler parameter pair representing a channel multipath component; generating an HRF fading feature spectrum and determining a coherence bandwidth and a coherence time according to the recovered continuous time delay-Doppler parameter pair to jointly form the second channel feature.

3. The method of claim 1, wherein, The determination of the switching benefit comprises: based on pre-stored historical performance data, the dual-mode channel data set and the heterogeneous channel feature, constructing a counterfactual data set for causal analysis; training a pre-configured causal inference model according to the counterfactual data set to quantify the net causal effect of mode switching intervention on communication performance; using the trained causal inference model, combining the current dual-mode channel data set and the heterogeneous channel feature to estimate the expected performance gain, which is taken as the switching benefit.

4. The method of claim 1, wherein, The generation of the switching control signal comprises: based on the performance difference and the switching benefit, constructing a non-consistency score for measuring the risk of switching decision; performing conformal calibration on the non-consistency score according to pre-stored historical switching records to generate a decision threshold band with a pre-set confidence level; comparing the current non-consistency score with the decision threshold band to determine whether to generate the switching control signal.

5. The method of claim 1, wherein, The calculation of the performance difference comprises: respectively deriving a first power spectral density and a second power spectral density from the first channel feature and the second channel feature; obtaining a service spectrum weight associated with a current service type; using a pre-set distribution metric distance and combining the service spectrum weight to quantify the shape difference between the first power spectral density and the second power spectral density to generate the performance difference.

6. The method of claim 1, wherein, The extraction of the first channel feature comprises: obtaining an HPLC baseband signal from the dual-mode channel data set and processing it to track the power grid fundamental frequency in real time; based on the power grid fundamental frequency, adaptively predicting a predetermined number of harmonic frequencies; constructing a parallel Goertzel filter bank and dynamically adjusting the filter parameters of the parallel Goertzel filter bank according to the predetermined number of harmonic frequencies; processing the HPLC baseband signal using the dynamically adjusted parallel Goertzel filter bank to analyze an electric grid harmonic feature vector, which is taken as the first channel feature.

7. The method of claim 5, wherein, The obtaining of the service spectrum weight comprises: Accessing historical service quality logs, historical power spectrum logs, and service scenario labels associated with both; Analyzing the historical service quality logs and the historical power spectrum logs for specific service scenario labels; Based on the analysis results, establishing a mapping relationship from power spectrum features to service quality indicators, and accordingly deducing service spectrum weights representing the importance of different frequency bands.

8. The method of claim 1, wherein, Further comprising: Performing mode switching according to the switching control signal; Monitoring post-switching performance metrics after the mode switching is completed; Updating the historical performance database using the post-switching performance metrics; Adaptively adjusting decision parameters for generating the switching control signal based on the updated historical performance database.

9. The method of claim 3, wherein, The causal inference model includes: A first model component configured to learn and predict communication performance under switching operation based on intervention group data corresponding to performing mode switching in counterfactual data set; and A second model component configured to learn and predict communication performance under non-switching operation based on control group data corresponding to not performing mode switching in the counterfactual data set; Wherein, the performance gain is determined according to the predicted outputs of the first model component and the second model component. Adaptively predicting a predetermined number of harmonic frequencies, including:

10. The method of claim 6, wherein, Applying a pre-configured autoregressive model to learn and predict harmonic frequency shifts based on pre-stored historical frequency data; Combining the real-time tracked power grid fundamental frequency and the harmonic frequency shift to determine the predicted value of the predetermined number of harmonic frequencies. ​

Images